ACS Publications. Most Trusted. Most Cited. Most Read
Molecular Dynamics (MD)-Derived Features for Canonical and Noncanonical Amino Acids
More

OR SEARCH CITATIONS

My Activity
Publications
  • Open Access
Machine Learning and Deep Learning

Molecular Dynamics (MD)-Derived Features for Canonical and Noncanonical Amino Acids
Click to copy article linkArticle link copied!

Open PDFSupporting Information (1)

Journal of Chemical Information and Modeling

Cite this: J. Chem. Inf. Model. 2025, 65, 4, 1837–1849
Click to copy citationCitation copied!
https://doi.org/10.1021/acs.jcim.4c02102
Published February 2, 2025

Copyright © 2025 The Authors. Published by American Chemical Society. This publication is licensed under

CC-BY-NC-ND 4.0 .

Abstract

Click to copy section linkSection link copied!
High Resolution Image
Download MS PowerPoint Slide

Machine learning (ML) models have become increasingly popular for predicting and designing structures and properties of peptides and proteins. These ML models typically use peptides and proteins containing only canonical amino acids as the training data. Consequently, these models struggle to make accurate predictions for peptides and proteins containing new amino acids that are absent in the training data set (e.g., noncanonical amino acids). One approach to improve the accuracy of the models is to collect more training data with the desired amino acids. However, this strategy is suboptimal as new data may not be easily attainable, and additional time is required to retrain the ML models. Alternatively, the extendibility of the ML models can be improved if the amino acid features used are representative and generalizable to the unseen amino acids. Herein, we develop amino acid features using molecular dynamics (MD) simulation results. Specifically, for a given amino acid, we perform MD simulation of its dipeptide to create features based on its backbone (ϕ, ψ) distributions and its electrostatic potentials. We demonstrate that these new features enable our ML models to more accurately predict the structural ensembles of cyclic peptides containing amino acids not present in the original training data set. For example, we build ML models to predict cyclic pentapeptide structures, with the training data set containing a library of 15 amino acids and the test data set containing the same 15-amino-acid library or an extended 50-amino-acid library. When using popular features such as Morgan fingerprints and MACCS keys to represent amino acids, the ML models achieve R2 = 0.963 for structural predictions of test cyclic pentapeptides containing the same 15-amino-acid library. However, these models’ performances decrease significantly to R2 = 0.430 and R2 = 0.508, respectively, when tasked to predict the structures of cyclic pentapeptides containing a library of 50 amino acids. On the other hand, the model using our backbone (ϕ, ψ) features outperforms those using Morgan fingerprints and MACCS keys, with R2 = 0.700. Overall, instead of having to collect more training data, our new features enable predictions of peptide sequences containing amino acids not originally present in the training data set at the mere cost of performing new dipeptide simulations for the new amino acids.

This publication is licensed under

CC-BY-NC-ND 4.0 .
  • cc licence
  • by licence
  • nc licence
  • nd licence
Copyright © 2025 The Authors. Published by American Chemical Society

Subjects

what are subjects

Introduction

Click to copy section linkSection link copied!
Peptides and proteins are crucial in maintaining biological processes. These molecules can also be designed to bind to specific disease-relevant targets, making them attractive drug candidates. (1,2) The ability to efficiently and accurately predict peptide or protein structures and properties would greatly improve our ability to rationally design peptide and protein therapeutics. Accordingly, developing machine learning (ML) models for predicting the structures and properties of peptides and proteins has become increasingly popular. (3−10)
While there is much literature on the successes of peptide- and protein-related ML models, questions about their applicability domain remain. For example, many of these models are trained using peptides with a limited set of amino acids (often canonical amino acids), and their performance is rarely demonstrated for peptides containing new amino acids that are absent in the training data set. Since it is a common strategy in the drug design process to incorporate new amino acids (e.g., noncanonical amino acids) to enhance properties like proteolytic resistance and membrane permeability, it would be highly beneficial if the ML models could be applied to new amino acids with good performance. (11−14)
To accurately make predictions for peptides and proteins containing new amino acids, one could collect more training data containing the new amino acids and retrain the ML models. However, collecting more training data could be time-consuming and difficult, making this approach undesirable. Alternatively, the features chosen to represent peptide and protein sequences for the ML models can offer a strategy to improve the models’ extendibility. If these features capture generalizable information across amino acids, then they have the potential to help the models make accurate predictions for peptides and proteins with new amino acids.
For example, generalization is very difficult for one-hot encoding, as the digits representing new amino acids do not exist in the original model, and there would be no data related to these new digits in the original model. However, outputs from SMILES, (15) CHUCKLES, (16) CHORTLES, (17) PLN, (18) BigSMILES, (19) and HELM (20) can be tokenized to create features that may be more generalizable. (21−24) For example, Morgan fingerprints (FPs) and molecular access system (MACCS) keys can be used to encode amino acids, where an amino acid is represented by a bit vector indicating the presence or absence of “sub-fragments” or chemical moieties. (25,26) Using these encoding schemes to extrapolate to new amino acids is possible because, in principle, if the new amino acids are similar to and share fragments that were seen in the training data set, then the ML model has some context about the new amino acids. However, new amino acids may contain chemical moieties that are simply absent in the original set of amino acids, and thus, more consideration may be necessary. For example, Schissel and co-workers used Morgan FPs to represent amino acids and developed an ML model to predict if a miniprotein could deliver an antisense oligomer to the nucleus; however, when models were trained using data sets that exclude sequences with a specific amino acid, the models performed poorly when tasked with making predictions for sequences containing the excluded amino acid. (27)
We hypothesize that more informative features for amino acids could improve the generalizability of ML models. While features like Morgan FPs and MACCS keys can embed useful information about molecular “fragments” or chemical moieties, if there are multiple instances of the same chemical moiety at different positions in the molecule, the encodings may not differentiate them. Furthermore, for amino acids, information about the relative locations of chemical moieties on the side chain and the backbone atoms may be useful in prediction tasks. To create a richer FP feature that includes this information, we develop position-aware side chain (PASC) features (Figure 1A).

Figure 1

Figure 1. Overview of custom features for amino acids. (A) Position-aware side chain (PASC) fingerprints are based on a “heavy-atom walk” along the amino acid side chain. Starting at the Cα atom, a Morgan fingerprint with a radius of 1 is generated (red circle). Morgan fingerprints with a radius of 1 centered at the Cβ atom (orange circle), Cγ atom (yellow circle), etc., are similarly generated to provide the PASC features. (B) MD simulations of amino acid dipeptides are used to generate MD-derived features. For the backbone (BB) features, the (ϕ, ψ) distribution is calculated from the dipeptide simulation and binned in a 2D grid. Then, the resulting 2D probability density is flattened into a 1D vector. For the voxel (VOX) features, the simulation frames are aligned to reference coordinates for C, Cα, and N, where the Cα atom is at the origin, the N atom is at (1.449, 0.000, 0.000), and the C atom is at (−0.523, 1.429, 0.000) (in Å). Then, frame-averaged molecular electrostatic potential is calculated on a 3D voxel and flattened into a 1D vector. See the Methods section for more details.

High Resolution Image
Download MS PowerPoint Slide
In addition to the efforts toward building upon the FP features, we also aim to develop novel molecular dynamics (MD)-derived amino acid features. We create features for amino acids based on the MD simulations of the dipeptides of the amino acids. We develop backbone (BB) features to capture the backbone conformational preferences of an amino acid by analyzing the dipeptides’ backbone (ϕ, ψ) distributions, and we develop voxel (VOX) features to capture the electrostatic environment created by an amino acid by calculating their frame-averaged electrostatic potentials on a voxel (Figure 1B). When extending the model to a data set that includes new amino acids, all one needs to do is run dipeptide simulations of the new amino acids (ranging from 100 to 300 ns of two parallel bias-exchange metadynamics (BE-META) MD simulations using two initial starting structures), calculate their (ϕ, ψ) distribution and electrostatic potentials to create encodings for the new amino acids, and then make the predictions without needing to retrain the model.
Herein, we demonstrate the applicability and value of our new features using our previously developed ML models for cyclic peptide structure prediction, namely, the Structural Ensembles Achieved by Molecular Dynamics and Machine Learning (StrEAMM) models. (5,6) Cyclic peptides have emerged as a promising drug modality, and efficiently characterizing their structures could aid in their design. (28−30) Given a cyclic peptide sequence, the StrEAMM models predict a population vector, where each value represents the population of a specific cyclic peptide structure in the structural ensemble. (5,6) The StrEAMM models are trained to reproduce the structural ensembles observed in MD simulations, where the backbone dihedrals of the cyclic peptides are measured and discretized, each frame in the simulations is given a structural digit string, and the numbers of frames with the same structural digit strings are counted to obtain the population vectors. Our original training and test data sets included cyclic pentapeptides and cyclic hexapeptides containing a limited 15 amino-acid (15-aa) library (Figure 2, black box). In this current work, we evaluate our new amino acid features on how well the models trained on the data set with a limited amino acid library predict test sequences containing an expanded 37-aa library (Figure 2, blue box) and an expanded 50-aa library that includes noncanonical amino acids (Figure 2, purple box). Compared to Morgan FPs and MACCS keys, our new MD-derived features perform better on test sets containing amino acids that were absent in the training data set. For example, models using Morgan FPs, MACCS keys, or our BB features predict the structures of cyclic pentapeptides containing an expanded library of 50 amino acids with R2 = 0.430, R2 = 0.508, and R2 = 0.700, respectively. Our MD-derived features offer a convenient way to improve a model’s generalizability without spending more time collecting new training data.

Figure 2

Figure 2. Amino acids included in the training, validation, and test data sets for the StrEAMM models. The training and validation data sets include cyclic peptide sequences from a 15-amino-acid (15-aa) library (black box). The test data sets include sequences containing amino acids in the same 15-aa library, the 37-aa library (blue box), or the 50-aa library (purple box). *For brevity, only the L-forms of chiral amino acids are depicted, but their mirror images are also included in the library.

High Resolution Image
Download MS PowerPoint Slide

Methods

Click to copy section linkSection link copied!

Charge Derivation for Noncanonical Amino Acids (ncAAs)

To derive the backbone (BB) and voxel (VOX) features for the noncanonical amino acids (ncAAs), namely, L/D-PCF, L/D-F4C, L/D-BFA, L/D-NAL, L/D-HSE, L/D-GML, and AIB (Figure 2), we first needed to derive the charges for these ncAAs so that we could perform MD simulations of the dipeptides of L-PCF, L-F4C, L-BFA, L-NAL, L-HSE, L-GML, and AIB. The BB and VOX features for the D-amino acids can be easily derived from those for the L-amino acids by symmetry.

Initial Charge Derivation

We first converted the SMILES string for each of the ncAAs into a 3D structure with RDKit (version 2021.03.05). (31) Then, the structures of each dipeptide were capped by an acetyl N-terminal cap (ACE) and an N-methyl C-terminal cap (NME). Two conformations were built for each dipeptide (except for AIB) with (ϕ, ψ) in the α-helix region (−60, −40°) and the β-sheet region (−150, 150°) using PyMOL, respectively. (32) For the AIB dipeptide, only the α-helix conformation was used. Then, the software Antechamber (33) (AmberTools version 22) was used to prepare the ncAA topology files. The initial partial charges of the ncAA were derived in Antechamber using AM1-BCC. (34) However, these charges were only used for the initial simulated annealing simulations, which were used to sample side chain conformations for deriving the final partial charges using restrained electrostatic potential (RESP) fitting. (35) The parmchk2 module within AmberTools was used to assign AMBER ff99SB force field (36) parameters, and any missing parameters were supplemented with the parameters from the generalized AMBER force field (GAFF). (37) Additionally, modifications from residue-specific force field 2 (RSFF2) were also applied to the ncAAs. (38) Specifically, the RSFF2 modifications for L-LEU, L-PHE, L-PHE, L-TYR, L-TYR, and L-ALA, were applied to L-GML, L-BFA, L-NAL, L-F4C, L-PCF, and L-HSE, respectively. The RSFF2 modifications from both L-ALA and D-ALA were applied to AIB.

Simulated Annealing

To sample the side chain conformations and generate structures for RESP charge derivation, simulated annealing simulations starting from each backbone conformation of the ncAA dipeptides were carried out in GROMACS (version 2018.6). (39) Dihedral restraints with a force constant of 1,000 kJ·mol–1·rad–2 were applied to the (ϕ, ψ) dihedral angles of each ncAA dipeptide to maintain its corresponding backbone conformation. Each initial structure of the ncAA dipeptide was first energy-minimized in vacuum. The energy-minimized structure was used to generate 25 replicas with random initial velocities at 300 K. Next, each replica was heated from 300 to 800 K in vacuum in 100 ps and held at 800 K for another 100 ps in an NVT ensemble. For simulation steps in vacuum, neighbor search and nonbonded interactions were truncated at 999.0 nm, and the neighbor list was only built once and never updated. The resulting structure was solvated using TIP3P water. (40) The minimum distance between each dipeptide and the box walls was set to 1.0 nm. The system was neutralized with minimal counterions (Na+ or Cl–) when applicable. Next, the steepest descent algorithm was employed to perform energy minimization for each solvated and neutralized system. Then, the system was relaxed at 300 K for 500 ps in an NVT ensemble. After NVT equilibration, an annealing process was carried out in an NPT ensemble. Specifically, the system was heated from 300 to 500 K in 100 ps, held at 500 K for 100 ps, cooled to 300 K in 500 ps, kept at 300 K for 100 ps, and then cooled again to 5 K in 200 ps. The final dipeptide structures from the 50 replicas (25 replicas for AIB) were used for subsequent charge derivations.
For simulation steps in explicit solvent, neighbor search and nonbonded interactions were truncated at 1.0 nm. Long-range electrostatics interactions beyond the cutoff were treated with particle mesh Ewald method (41) and a Fourier spacing of 0.12 nm and cubic interpolation. A long-range dispersion correction for energy and pressure was used to account for the 1.0 nm cutoff of the Lennard–Jones interactions. The Berendsen barostat with a time coupling constant of 2.0 ps and isothermal compressibility of 4.5 × 10–5 bar–1 was applied to maintain a pressure of 1 bar. The v-rescale thermostat with a time coupling constant of 0.1 ps was employed for temperature control. The LINCS constraint algorithm was applied to bonds connecting to hydrogen atoms. The leapfrog integrator was used with a time step of 2 fs. Extra improper terms related to C, O, N, and H atoms were applied to the two peptide bonds of each dipeptide to maintain planarity and keep a trans-amide configuration.

Final Charge Derivation

First, the final structures during simulated annealing simulations of each ncAA dipeptide were geometry-optimized using Gaussian16 software with tight convergence criteria at HF/6-31G(d) level of theory and the (ϕ, ψ) dihedral angles were frozen for each structure. (42) For any structures that did not reach convergence after 100 optimization steps, they were rerun with additional frozen χ1 dihedral angles where applicable. Similarly, reruns with additional frozen χ2 dihedral angles (where applicable) were performed when geometry optimization with additional frozen χ1 dihedral angles failed to converge. Structures that did not converge after additional optimizations with frozen χ2 dihedral angles were not used in subsequent charge derivation. Out of the 50 structures for each dipeptide (except 25 for AIB), one β-sheet conformation structure of GML and one β-sheet conformation structure of NAL did not converge when having additional frozen χ1 dihedral angles. However, this β-sheet conformation structure of GML converged after freezing the additional χ2 dihedral angle, while the β-sheet conformation structure of NAL did not.
The optimized structures of each dipeptide were submitted to the R.E.D. web server for charge derivation. (43) Gaussian16 was used to compute the molecular electrostatic potential with IOp(6/33 = 2,6/41 = 4,6/42 = 6) at HF/6-31G(d) level of theory to obtain reliable RESP charges following a two-stage RESP charge fitting method (35) to calculate partial charges. We assigned the charges for the backbone carbonyl C and O atoms in the ncAA, as well as those for the backbone amide N and H atoms, to be consistent with the charges for all the neutral canonical amino acids in the AMBER ff99SB force field. (36) Since the caps were used to help mimic the peptide environment, we set the charges for the N and H atoms in the NME cap, and the C and O atoms in the ACE cap, to also be consistent with those of the backbone atoms of the ncAA. Additionally, the total net charge of ACE and NME groups were constrained to zero. The derived partial charges of ncAA fragment were used for subsequent MD simulations of ncAA-dipeptides and cyclic peptides with ncAAs, while the partial charges of the caps from these RESP fittings were discarded.

MD Simulations of Amino Acid Dipeptides

Bias-exchange metadynamics (44) (BE-META) simulations for the two initial structures (α-helix and β-sheet conformations) for each dipeptide were run in parallel using GROMACS (version 2018.6) (39) patched by the PLUMED 2.5.1 plugin. (45) The initial structures were solvated in a cubic water box similar to the simulated annealing simulations, with a minimum distance between the dipeptide and the walls of the box set to 1.0 nm. The steepest descent algorithm was used to minimize the solvated structure. Upon minimization, the system was equilibrated in two stages: first, all heavy atoms of the dipeptide were position-restrained and 50 ps of NVT simulation at 300 K and 50 ps of NPT simulation at 300 K and 1 bar were performed. In the second stage, the position restraints were removed, and the same sequence of NVT and NPT simulations were run again for 100 ps each. BE-META production runs of 100 ns were then performed in the NPT ensemble, at 300 K and 1 bar, with a 2 fs time step. Each BE-META production run had one bias replica with a 2D bias, (ϕ, ψ) dihedral angles of each dipeptide, and also five neutral replicas with no bias.
All simulations were run using the leapfrog algorithm, with water geometry maintained using SETTLE and hydrogen-containing bonds constrained to equilibrium lengths using LINCS. The nonbonded interaction cutoff was set to 1.0 nm, with Coulombic interactions beyond the cutoff computed using particle mesh Ewald summation, with a Fourier spacing of 0.12 nm and cubic interpolation. Dispersion corrections for both energy and pressure were applied to the long-range van der Waals interactions. Temperature was controlled by velocity rescaling, with a coupling time constant of 0.1 ps. The Parrinello–Rahman barostat was used for pressure control, with a coupling time constant of 2.0 ps and isothermal compressibility of 4.5 × 10–5 bar–1.
To monitor simulation convergence, we first computed the 2D density distributions of the (ϕ, ψ) angles for the two parallel simulations of each dipeptide system using the last 50 ns of the neutral replicas. Then, normalized integrated products (NIPs) between the density distributions of two parallel simulations were calculated, where an NIP value of 1.0 represents perfect similarity. (46) Most dipeptides converged with NIP ≥ 0.99. For PHE, THR, TYR, and VAL dipeptides, the BE-META simulations were extended to 200 ns, and NIP ≥ 0.99 was achieved when using the 100–200 ns trajectories of the neutral replicas. For the F4C dipeptide, the BE-META simulations were extended to 300 ns, and NIP ≥ 0.99 was achieved when using 200–300 ns trajectories of the neutral replicas.

Cyclic Pentapeptide and Cyclic Hexapeptide Data Sets

The training data set of 705 cyclic pentapeptide sequences generated from the 15-aa library and the test data set of 50 cyclic pentapeptide sequences from the same 15-aa library are from our previously published work developing the StrEAMM models. (5) The training data set of 705 cyclic hexapeptide sequences generated from the 15-aa library comes from 555 cyclic hexapeptide sequences used in the previous StrEAMM work (6) and an additional 150 cyclic hexapeptides simulated in this work. The 49 test cyclic hexapeptides from the same 15-aa library, as well as 25 test cyclic pentapeptides and 25 test cyclic hexapeptides from the expanded 37-aa library, were also from the previous StrEAMM work. (6) In this work, we aimed to follow the same MD simulation protocol as our previously published work to curate approximately 50 test cyclic pentapeptides and 50 test cyclic hexapeptides for the expanded 37-aa library and also for the expanded 50-aa library. (5,6) We performed two sets of BE-META simulations starting from two initial structures for each cyclic peptide and monitored the simulation convergence for each cyclic peptide by performing dihedral principal component analysis using the backbone dihedrals, calculating the density distributions in the 3D principal component space, and calculating the NIP between the 3D density distributions from the two parallel simulations. If the NIP between the two parallel simulations was ≥ 0.9, we assumed the simulations had converged. For most cyclic peptides, this was achieved when using the last 50 ns of the neutral replicas of 100 ns BE-META simulations, which were used for subsequent structural analysis. For the 25 additional cyclic pentapeptides containing amino acids from the 37-aa library, four sequences were extended to 200 ns, and convergence was achieved using the last 100 ns of the neutral replicas. For the 50 cyclic pentapeptides containing amino acids from the 50-aa library, one sequence was extended to 200 ns, and convergence was achieved using the last 100 ns of the neutral replicas. For the 25 additional cyclic hexapeptides containing amino acids from the 37-aa library, six sequences were extended to 200 ns, and convergence was achieved using the last 100 ns of the neutral replicas; one sequence was extended to 300 ns, and convergence was achived using the last 200 ns of the neutral replicas. For the 50 cyclic hexapeptides containing amino acids from the 50-aa library, fourteen sequences were extended to 200 ns, and convergence was achieved using the last 100 ns of the neutral replicas; four sequences were extended to 300 ns, and convergence was achieved using the last 200 ns of the neutral replicas; three sequences did not converge in 300 ns and were discarded to ensure that we only included high quality data.

Position-Aware Side Chain (PASC) Fingerprints

To embed information about the positions of amino acid side chain chemical moieties relative to the backbone, we generated position-aware side chain (PASC) features that encode amino acids based on the concatenated fingerprints generated from a “heavy-atom walk” of the side chain. To generate the PASC fingerprints, RDKit (version 2021.03.05) was used to construct a 1024-bit Morgan fingerprint (25) centered on each heavy atom along the length of an amino acid’s side chain using a radius of 1. The length of the side chain was determined to start from Cα and continue along the side chain until reaching the second to last terminal atom. The last heavy atom in the side chain is not included in the encoding because the previous adjacent atom’s fingerprint should capture neighboring atoms within 1 bond away, given the set radius of 1. For an amino acid with a side chain length N (defined by heavy atoms), each side chain position’s fingerprint was concatenated together to obtain an (N – 1) × 1024 bit vector, such that the bits toward the end of the concatenated feature vector reflect chemical moieties furthest away from the backbone. To deal with the fact that different amino acids have varying side chain lengths, and thus encoding vectors for different amino acids could have different lengths, we used a fixed length from the longest side chain. The longest side chain in the 50-aa library was L-BFA with a length of 10, and the total encoding for each amino acid had 9 × 1024 bits = 9216 total bits. For amino acids with shorter side chains than L-BFA, the fingerprint was generated for all the heavy atoms in the side chain, and then the encoding was padded with zeros to make the final encoding length up to 9216 bits. To deal with multiple heavy atoms that are in the same position relative to the backbone, a fingerprint was made for each heavy atom, and the average of the fingerprints was used in the final concatenated bit vector.

Backbone (BB) Features

To generate the backbone (BB) features, we first binned the (ϕ, ψ) space for each amino acid dipeptide into an N × N grid. The N × N grids for the D-amino acids were derived from those of the L-amino acids by centro-symmetry. The N × N grids were then flattened into 1D feature vectors (Figure 1B). While the time required to train ML models would increase with the size of this feature vector, having too small of a feature vector would result in a large loss of resolution of the (ϕ, ψ) distribution. Therefore, we aimed to balance efficiency and accuracy and considered various N values between 2 and 100. The 40 × 40 grid size was ultimately chosen for our BB features as it achieved similar resolutions as those from grids made with larger N values (Figures S1–S4). After choosing the grid size, we normalized the probability densities using min-max normalization and obtained feature values within the range of 0 and 1. This normalization was performed in two ways; in one method, we performed a “per-plot” normalization where each amino acid’s 40 × 40 grid was normalized based on its own minimum and maximum grid points. We also performed a “per-grid” normalization that considers minimum and maximum values for each grid point across all 15 amino acids in the 15-aa library. We tested the different normalization schemes by training StrEAMM models on either non-normalized BB features (BB0), the “per-plot” normalization (BB1), and the “per-grid” normalization (BB2), and observed that the BB2 models perform the best on the validation data set for cyclic pentapeptides and cyclic hexapeptides (Figure S5). Therefore, the BB features referred to in the main text are “per-grid” normalized.

Electrostatic Potential Voxel (VOX) Features

Frame-averaged molecular electrostatic potentials were created for each amino acid from their dipeptide simulation trajectories. To calculate electrostatic potentials for each dipeptide, first, the backbone N, Cα, and C atoms of all frames of the dipeptide were aligned to a three-atom amino acid backbone template structure, where the Cα atom was at the origin, the N atom was at (1.449, 0.000, 0.000), and the C atom was at (−0.523, 1.429, 0.000) (in Å, Figure 1B). The partial charges on each atom were taken from the simulation force field and were represented as Gaussian charge distributions with the standard deviation equal to the van der Waals radius for each atom type. The electrostatic potential at a point r and frame (or time) t can be calculated using the following equation written in atomic units
ϕt(r)=∑inQi|ri,t−r|erf⁡⁡(|ri,t−r|RvdW,i2)
where ri,t is the position of atom i of the dipeptide at time t; Qi and RvdW,i are the partial charge and the van der Waals radius of atom i of the dipeptide; and n is the number of atoms of the dipeptide. Furthermore, the term in the sum can be simplified as the distance between a grid point and an atom approaches zero,
lim|ri,t−r|→0⁡Qi|ri,t−r|erf⁡⁡(|ri,t−r|RvdW,i2)=QiRvdW,i2π
This form is useful for the Cα of the dipeptide, which is always aligned with the grid point at the origin. The electrostatic potential was calculated for each frame of the dipeptide simulation and then the time-average molecular electrostatic potential Φ(r) was obtained by averaging ϕt(r) over all frames.
The frame-averaged molecular electrostatic potential was calculated on a grid. The grid points were evenly spaced in the x, y, and z directions using 1 Å intervals. To reduce the number of grid points, a grid was also generated using 2.5 Å intervals, which we used to generate our voxel features. The size of the grid was determined by the extreme position of atoms across all frames and across all dipeptides to accommodate all amino acids on a single grid size. A 5 Å buffer was included between the edges of the box and the extrema positions to calculate the electrostatic potential extending outward from the extreme atoms. For the 1 Å-grid, the edges of the grid were rounded to the nearest whole number (Table S1). For the 2.5 Å-grid, the edges of the grid were rounded so that the total distance along the grid would be a multiple of 2.5 (Table S1). Molecular electrostatic potential representations for D-amino acids were calculated by flipping the L-amino acid representations across the xy-plane.
Similar to the BB features, we tested different normalization schemes by training StrEAMM models on non-normalized VOX features (VOX0), the “per-plot” normalization (VOX1), and the “per-grid” normalization (VOX2). We observed that the VOX0 models perform best on the validation data set for cyclic pentapeptides and hexapeptides (Figure S6). Therefore, the VOX features referred to in the main text are not normalized.

StrEAMM Neural Network Models

The ML models used for the StrEAMM method are convolutional neural networks (CNNs), as described in previous work. (6) We chose the CNN architectures that performed the best for the cyclic pentapeptide data set (“CNN (1,2)” models) and for the cyclic hexapeptide data set (“CNN (1,2)+(1,3)+(1,4)” model) as a starting point. (6) In brief, the StrEAMM CNN models represented a cyclic peptide sequence with C amino acids as a matrix with C columns and R rows, where R is the dimension of the feature (for example, R = 1600 for the BB features made from the 40 × 40 grid). For a “CNN (1,2)” model, this R × C matrix was concatenated with another matrix that encodes the cyclic peptide sequence permutated by one position (i.e., if the original sequence is cyclic-(ABCDE), the permutated sequence is cyclic-(BCDEA)). The resulting concatenated matrix was (2 × R) × C so that the features representing the amino acids within a (1, 2) pair (i.e., directly adjacent amino acids in the sequence) were also in the same column. Then, this feature matrix was input into a 1D convolutional layer, followed by a fully connected hidden layer of a multilayer perceptron (MLP) with a rectified linear unit (ReLU) activation. (47) The output layer (with a dimension based on the number of structures in the ensemble) used a SoftMax activation function to make the predicted populations sum to 1. The loss function used was the summation of the squared error loss.
For each feature type, we performed hyperparameter tuning with 3-fold cross-validation for up to 2000 epochs and using an initial grid search where the learning rates tested were 1 × 10–3, 1 × 10–4, and 1 × 10–5. The grid search also included the number of filters, including 128, 256, 512, and 1024, and the number of nodes, including 256, 512, and 1024. The best-performing models were determined based on the best average mean squared error (across three folds) for the validation data sets, which include sequences containing the same 15-aa library used for the training data sets. We also extended the initial grid when the best model had its hyperparameter values at any of the intial grid’s extrema. The optimal epoch value was selected using generalized error and percent change calculations that are used as overfitting criterion (Figure S7). (48) All model training was performed using the python package PyTorch (version 1.9.062). (49)

Results and Discussion

Click to copy section linkSection link copied!

Application of FPs, MACCS Keys, and Our New Features for Cyclic Peptide Structural Ensemble Prediction: 15-aa Training Data Set, 15-aa Test Data Set

To evaluate the generalizability of these new features, we apply them to predict cyclic peptide structural ensembles. We had previously developed the StrEAMM method, which enabled the prediction of cyclic peptide structural ensembles by using ML models trained on MD simulation results. (5) In the previous work, we simulated 705 cyclic pentapeptides containing a 15-aa library (Figure 2, black box) to generate a training data set. (5) Then, we trained linear regression and neural networks to predict the structural ensembles for a test data set containing 50 cyclic pentapeptides using the same 15-aa library as the training data set. (5,6) The StrEAMM models achieved good performance on this test set containing the 15-aa library. For example, the StrEAMM convolutional neural network (CNN) models, which used Morgan fingerprint (FP) features, reported an average R2 of 0.963 (Figure 3A, “15 AAs,” FP; training and validation results shown in Figure S8). The average performance on the test data set is calculated by performing 3-fold cross-validation and using each of the resulting three models to make predictions for the test data set.

Figure 3

Figure 3. Performance of different amino acid features on different cyclic pentapeptide test data sets. (A) The models are trained using 3-fold cross-validation. The table reports the average R2 (coefficient of determination) and standard deviation across the 3 folds. (B) The performance for one out of the three models from the 3-fold cross-validation is plotted. The predicted population of each structure in the cyclic peptides’ structural ensemble is compared to its populations observed in MD simulations. For clarity, only structures in the ensembles for all cyclic peptides in the test data sets with either a predicted or observed (in MD) percent population of >1% are plotted.

High Resolution Image
Download MS PowerPoint Slide
Herein, we train StrEAMM CNN models using features other than Morgan FPs, like MACCS keys and our new features. When we compare the models’ performances on the test set from the 15-aa library, we observe similar performances across all the features evaluated (Figure 3A, “15 AAs”). The average R2 for the models trained using MACCS keys, position-aware side chain fingerprints (PASC), backbone (BB), and voxel (VOX) features are 0.963, 0.963, 0.961, and 0.950, respectively. Parity plots show that the predicted populations of the test cyclic peptides are close to the populations observed in our MD simulation results (Figure 3B, top row). Statistical testing shows that there is no significant difference between the performance of all five features (Figure S9A, “15 AAs”).
In addition to developing models to predict cyclic pentapeptide structural ensembles, we also apply our StrEAMM method to cyclic hexapeptides. We have a training data set of 705 cyclic hexapeptides containing the 15-aa library and a test data set of 49 cyclic hexapeptides containing the same 15-aa library. Models trained using Morgan FPs and MACCS keys report average R2 of 0.898 and 0.877 on the test data set, respectively (Figure 4A, “15 AAs,” FP and MACCS; training and validation results shown in Figure S10). Consistent with what is observed from the cyclic pentapeptide model performances, our new PASC, BB, and VOX features perform similarly well, which have average R2 of 0.877, 0.892, and 0.847, respectively (Figure 4A, “15 AAs,” PASC, BB, and VOX; Figure 4B, top row; see Figure S9B “15 AAs” for statistical testing results).

Figure 4

Figure 4. Performance of different amino acid features on different cyclic hexapeptide test data sets. (A) The models are trained using 3-fold cross-validation. The table reports the average R2 (coefficient of determination) and standard deviation across the 3 folds. (B) The performance for one out of the three models from the 3-fold cross-validation is plotted. The predicted population of each structure in the cyclic peptides’ structural ensemble is compared to its populations observed in MD simulations. For clarity, only structures in the ensembles for all cyclic peptides in the test data sets with either a predicted or observed (in MD) percent population of >1% are plotted.

High Resolution Image
Download MS PowerPoint Slide
We further evaluate the models’ performance using weighted errors (WEs) as a metric. The general conclusions remain the same, i.e., all five features perform similarly for cyclic pentapeptides and cyclic hexapeptides (Figures S11 and S12, “15 AAs”).

Application of FPs, MACCS Keys, and Our New Features on Extended Test Data Sets: 15-aa Training Data Set, 37-aa and 50-aa Test Data Sets

In this present work, we expand the test data sets to include cyclic pentapeptides and cyclic hexapeptides containing a 37-aa library (Figure 2, blue box). While the 15-aa library is a representative subset of the canonical amino acids and their mirror images, the 37-aa library contains all the canonical amino acids (except for proline) and their mirror images. To assess the generalizability of the amino acid features, StrEAMM cyclic pentapeptide models trained on the data set generated from the 15-aa library are used to predict a test data set containing the 37-aa library. The models using the Morgan FPs and MACCS keys report poor performances of R2 of 0.364 and 0.531, respectively (Figure 3A, “37 AAs,” FP and MACCS). The models using the new PASC and VOX features also report poor performances, with R2 values of 0.352 and 0.406, respectively (Figure 3A, “37 AAs,” PASC and VOX). However, when using the BB features, the model performance increases to an average R2 of 0.735 (Figure 3A, “37 AAs,” BB). The BB features also outperform the other features in the StrEAMM cyclic hexapeptide models, reporting an average R2 of 0.683 (Figure 4A, “37 AAs,” BB). Statistical testing shows that the performance of the BB feature is statistically different from (better than) all the other features for the 37-aa tests (Figure S9, “37 AAs”).
We also test the generalizability of our new features to include noncanonical amino acids (ncAAs) by constructing a 50-aa library (Figure 2, purple box). The ncAAs selected in this study are the L- and D-form of p-chloro-phenylalanine (PCF), 4-carbamoyl-phenylalanine (F4C), biphenylalanine (BFA), β-(2-naphthyl)-alanine (NAL), homoserine (HSE), γ-methyl-leucine (GML), as well as 2-aminoisobutyric acid (AIB). The first six ncAAs are chosen because they have been included in peptide design schemes; when screening for peptides that have the potential to target some proteins of interest, these ncAAs were present in hits or optimized versions of hits. (50−52) AIB is chosen because of its achirality and ability to stabilize α-helices. (53,54) Using the same StrEAMM cyclic pentapeptide models trained on the data set generated from the 15-aa library, we assess the performance of the different amino acid features on a test data set generated from the 50-aa library. Similar to the models’ results on the test data set generated from the 37-aa library, we observe that the models using Morgan FPs and MACCS keys report poor performances on the test data set generated from the 50-aa library, with average R2 of 0.430 and 0.508, respectively (Figure 3A, “50 AAs,” FP and MACCS). The decreases in model performance on the extended test data sets using the FP-type encodings could be because there are bits in the bit-vectors of the new amino acids that are absent in the bits from the amino acids in the training data set. For example, for the Morgan FPs, while there are 97 unique bits present from all the amino acids in the 15-aa library (considering a 2048-bit vector with radius = 3), there are 247 unique bits present in the 37-aa library and 305 unique bits present in the 50-aa library. Therefore, there are 247 – 97 = 150 bits in the 37-aa library and 305 – 97 = 208 bits in the 50-aa library that are not present in the original training data set containing the 15-aa library.
On the other hand, our BB features outperform all the other features on the test data set generated from the 50-aa library, with an average R2 of 0.700 (Figure 3A, “50 AAs,” BB). The BB features also outperform the other amino acid features on the 50-aa cyclic hexapeptide data set, reporting R2 = 0.592 (Figure 4A, “50 AAs”). Statistical testing shows that the performance of the BB feature is statistically different from (better than) all the other features for the 50-aa tests (Figure S9, “50 AAs”). We further evaluated the models’ performance using weighted errors (WEs) as a metric. The BB features outperform the other features for cyclic pentapeptides (Figure S11, “37 AAs” and “50 AAs”). For cyclic hexapeptides, both the BB and VOX features perform similarly well (Figure S12, “37 AAs” and “50 AAs”).
It is noted that the improved performance on the 37-aa and 50-aa test data sets when using the BB features is not as pronounced for the cyclic hexapeptide models compared to the cyclic pentapeptide models. We hypothesize that for cyclic hexapeptides, the information related to the side chain, such as the voxel (VOX) features, may be almost as useful as backbone-based features, compared to cyclic pentapeptides, for which BB features are very informative for the models’ structural ensemble predictions.

Evaluation of the StrEAMM Model Performances on the Extended Test Data Sets Using Combinations of Features

The amino acid features included in our comparisons represent different aspects of amino acids; for example, the BB features aim to capture the (ϕ, ψ) preferences for each amino acid, while the FP-type encodings capture the presence and absence of specific chemical moieties. Both types of information may be useful in the prediction tasks, and we evaluate the performance of the StrEAMM models using different pairs of features.
First, for the 15-aa test data set, we observe that the cyclic pentapeptide models using different feature-pair combinations perform similarly to those using only a single feature. For example, while the cyclic pentapeptide model using only BB features has an average R2 of 0.961, the models using BB+VOX, BB+PASC, BB+MACCS keys, and BB+FP have R2 of 0.961, 0.961, 0.962, and 0.963 respectively (Figure 5, top left; training and validation results shown in Figure S13). Similarly, the cyclic hexapeptide models using feature-pair combinations perform as well as those using only a single feature. For example, the model using only BB features has an average R2 of 0.892, and the models using BB+VOX, BB+PASC, BB+MACCS keys, and BB+FP have an average R2 of 0.898, 0.894, 0.900, and 0.899, respectively (Figure 5, bottom left).

Figure 5

Figure 5. Performance of different combinations of amino acid features on cyclic pentapeptide (top) and hexapeptide (bottom) test data sets containing 15 AAs (left), 37 AAs (middle), and 50 AAs (right). The models are trained using 3-fold cross-validation, and the average R2 and standard deviation are reported. The models using a single type of feature (e.g., “BB only”) are represented on the diagonals. The best-performing models for the 37 AA and 50 AA test data sets, based on the average R2, are boxed with bold black outlines.

High Resolution Image
Download MS PowerPoint Slide
We then use the feature-pairs to predict the more challenging test data sets generated from the 37-aa and 50-aa libraries. For the cyclic pentapeptide models predicting the 37-aa test data set, the best-performing model based on the average R2 uses BB+VOX keys, reporting an average R2 of 0.754 (Figure 5, top middle). This result is an improvement from the best single-feature model, BB-only, with an average R2 of 0.735 (Figure 5, top middle). For the cyclic pentapeptide models predicting the 50-aa test data set, the best-performing model is also a combination of features. The best-performing model uses BB+MACCS keys, reporting an average R2 of 0.742, which is better than the best single-feature model, BB-only, with an average R2 of 0.700 (Figure 5, top right).
However, for the cyclic hexapeptide models, the feature-pair combination models do not necessarily perform better than the single-feature models. For the cyclic hexapeptide models predicting the 37-aa test data set, the overall best-performing model is the BB-only model with an average R2 of 0.683, while the best-performing feature-pair model is BB+MACCS with an average R2 of 0.666 (Figure 5, bottom middle). For the cyclic hexapeptide models predicting the 50-aa test, the overall best-performing model is the BB-only model with an average R2 of 0.592, while the best-performing feature-pair model is BB+VOX with an average R2 of 0.586 (Figure 5, bottom right). However, the performance of the BB-only model is within the standard deviation of the BB+VOX model. One possible reason why the models using feature-pair combinations are not better than those using the single features when predicting the 37-aa and 50-aa test data sets could be that we select the best hyperparameters based on the validation data sets (which are composed of amino acids from the 15-aa library), and the models may be biased/overfit to the 15-aa library containing only a subset of canonical amino acids. Hence, there is a challenge to the assumption that optimizing the hyperparameters based on the 15-aa validation data set will be optimal for the 37-aa and 50-aa libraries, and we are interested in exploring the regularization of these models in the future. We also plan to perform, for example, SHAP analysis (55) to understand which specific bits in the single and combined features contribute the most to the model’s predictions.
In principle, the StrEAMM platform can be applied to predict the structural ensembles of linear peptides and stapled peptides as well. The new BB features are readily applicable to those cases and to any ML models for peptides and proteins in general.

Conclusions

Click to copy section linkSection link copied!
We develop new amino acid features with the intention of improving the generalizability of ML models without needing to pay the additional cost of collecting more training data and retraining models. Toward this end, we create a position-aware side chain (PASC) fingerprint feature and features based on MD simulation results from amino acid dipeptides. We demonstrate that models implementing the commonly used features, namely, Morgan fingerprints (FPs) and MACCS keys, can accurately predict the structural ensembles of cyclic pentapeptides and cyclic hexapeptides when the test data set is generated from the same 15-aa library as the training data set. However, we show that the performance of these Morgan FPs and MACCS keys-based models decreases significantly when tasked to predict the structural ensembles of cyclic pentapeptides and cyclic hexapeptides containing amino acids that were not originally present in the training data set (i.e., cyclic peptides generated from the 37-aa and 50-aa libraries).
The new amino acid backbone (BB) features consistently outperform Morgan FPs and MACCS keys when applied to the test data sets from the expanded 37-aa and 50-aa libraries, highlighting the ability of the BB features to improve the extendibility of the models. For example, when compared to the models using Morgan FPs and MACCS keys to predict the structural ensembles of the test cyclic pentapeptides from a 37-aa library, our BB features outperform the other features with an average R2 of 0.735, compared to the average R2 of 0.364 and 0.531 for Morgan FPs and MACCS keys, respectively. This trend is also observed when predicting the structural ensembles of the test cyclic pentapeptides from a 50-aa library containing noncanonical amino acids, with our BB features reporting an average R2 of 0.700, compared to the average R2 of 0.430 and 0.508 for Morgan FPs and MACCS keys, respectively. The improved performances from the models using our BB features demonstrate both the utility of the features to help the extendibility of the models, as well as the advantage of leveraging information from MD simulation data. Specifically, the (ϕ, ψ) preferences of the amino acids are important for the cyclic peptide structural ensemble prediction task, and they can now be gauged using the dipeptide simulations and represented as features for the amino acids.
Overall, the promising results shown herein using expanded test data sets and new features highlight that our BB features improve the generalizability of the StrEAMM models for cyclic peptide structure predictions. We envision that these features could also be applied to and improve the generalizability of other ML models for peptides and proteins. The BB features only require dipeptide simulations of the amino acids to generate, and such MD simulation data provide valuable information about the amino acids’ backbone structural preferences. ML models trained using BB features are thus more transferable to new amino acids. If one is interested in making predictions for new amino acids not seen in the training data set, the cost is a mere dipeptide simulation, which is presumably minimal compared to the cost of collecting new training data that include the new amino acid and retraining the model. Therefore, the new BB features offer an attractive, cost-effective way to improve model extendibility.

Data Availability

Click to copy section linkSection link copied!

The new features that were used to train the models in this study can be found in the public GitHub repository, https://github.com/thui16/MD_derived_features_for_aas/. The StrEAMM models and MD simulation data used to train these models are under a patent application (please see Competing Interests section) and are not publicly available.

Supporting Information

Click to copy section linkSection link copied!

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.jcim.4c02102.

  • Example of a (ϕ, ψ) distribution binned using various grid sizes; comparison of backbone (BB) features of amino acids in the different amino acid libraries; comparison of the different normalization schemes applied to the BB and voxel (VOX) features; examples of learning curves from hyperparameter tuning; model performances (reporting R2) on the training and validation data sets for the cyclic pentapeptides; p-values from t-tests comparing different features; model performances (reporting R2) on the training and validation data sets for the cyclic hexapeptides; model performances (reporting weighted error, WE) on the various test data sets for the cyclic pentapeptides and cyclic hexapeptides; model performances (reporting R2) using combinations of features on the training and validation data sets for the cyclic pentapeptides and cyclic hexapeptides (PDF)

Terms & Conditions

Most electronic Supporting Information files are available without a subscription to ACS Web Editions. Such files may be downloaded by article for research use (if there is a public use license linked to the relevant article, that license may permit other uses). Permission may be obtained from ACS for other uses through requests via the RightsLink permission system: http://pubs.acs.org/page/copyright/permissions.html.

Author Information

Click to copy section linkSection link copied!
  • Corresponding Author
  • Authors
    • Tiffani Hui - Department of Chemistry, Tufts University, Medford, Massachusetts 02155, United States;  Orcidhttps://orcid.org/0000-0002-1355-389X
    • Maxim Secor - Department of Chemistry, Tufts University, Medford, Massachusetts 02155, United States
    • Minh Ngoc Ho - Department of Chemistry, Tufts University, Medford, Massachusetts 02155, United States;  Orcidhttps://orcid.org/0009-0000-8054-0362
    • Nomindari Bayaraa - Department of Chemistry, Tufts University, Medford, Massachusetts 02155, United States
  • Author Contributions

    T.H., M.S., and M.N.H. contributed equally to this work. T.H. contributed to methodology, model training, analysis, and writing of the manuscript. M.S. contributed to methodology, analysis, and writing the manuscript. M.N.H., contributed to data collection, analysis, and writing the manuscript. N.B. contributed to data collection and editing the manuscript.

  • Notes
    The authors declare the following competing financial interest(s): A patent application PCT/US2022/072941, Cyclic peptide structure prediction via structural ensembles achieved by molecular dynamics and machine learning was filed on 2022/6/14. Y.-S.L. is on the scientific advisory board for Zonsen PepLib Biotech.

Acknowledgments

Click to copy section linkSection link copied!

This work was supported by the National Institute of General Medical Sciences of the National Institutes of Health under award number R01GM124160 (PI: Y.-S.L.). We are grateful for the support from the Tufts Technology Services and for the computing resources at the Tufts Research Cluster. Initial structures for the simulations were built using UCSF Chimera, developed by the Resource for Biocomputing, Visualization, and Informatics at the University of California, San Francisco, with support from NIH Grant P41-GM103311.

References

Click to copy section linkSection link copied!

This article references 55 other publications.

  1. 1
    Wang, L.; Wang, N.; Zhang, W.; Cheng, X.; Yan, Z.; Shao, G.; Wang, X.; Wang, R.; Fu, C. Therapeutic peptides: current applications and future directions. Signal Transduction Targeted Ther. 2022, 7, 48  DOI: 10.1038/s41392-022-00904-4
    Google Scholar
    1
    Therapeutic peptides: current applications and future directions
    Wang, Lei; Wang, Nanxi; Zhang, Wenping; Cheng, Xurui; Yan, Zhibin; Shao, Gang; Wang, Xi; Wang, Rui; Fu, Caiyun
    Signal Transduction and Targeted Therapy (2022), 7 (1), 48CODEN: STTTCB; ISSN:2059-3635. (Nature Portfolio)
    A review. Peptide drug development has made great progress in the last decade thanks to new prodn., modification, and analytic technologies. Peptides have been produced and modified using both chem. and biol. methods, together with novel design and delivery strategies, which have helped to overcome the inherent drawbacks of peptides and have allowed the continued advancement of this field. A wide variety of natural and modified peptides have been obtained and studied, covering multiple therapeutic areas. This review summarizes the efforts and achievements in peptide drug discovery, prodn., and modification, and their current applications. We also discuss the value and challenges assocd. with future developments in therapeutic peptides.
  2. 2
    Liu, K.; Li, M.; Li, Y.; Li, Y.; Chen, Z.; Tang, Y.; Yang, M.; Deng, G.; Liu, H. A review of the clinical efficacy of FDA-approved antibody–drug conjugates in human cancers. Mol. Cancer 2024, 23, 62  DOI: 10.1186/s12943-024-01963-7
    Google Scholar
    There is no corresponding record for this reference.
  3. 3
    Jumper, J.; Evans, R.; Pritzel, A.; Green, T.; Figurnov, M.; Ronneberger, O.; Tunyasuvunakool, K.; Bates, R.; Žídek, A.; Potapenko, A.; Bridgland, A.; Meyer, C.; Kohl, S. A. A.; Ballard, A. J.; Cowie, A.; Romera-Paredes, B.; Nikolov, S.; Jain, R.; Adler, J.; Back, T.; Petersen, S.; Reiman, D.; Clancy, E.; Zielinski, M.; Steinegger, M.; Pacholska, M.; Berghammer, T.; Bodenstein, S.; Silver, D.; Vinyals, O.; Senior, A. W.; Kavukcuoglu, K.; Kohli, P.; Hassabis, D. Highly accurate protein structure prediction with AlphaFold. Nature 2021, 596, 583– 589,  DOI: 10.1038/s41586-021-03819-2
    Google Scholar
    3
    Highly accurate protein structure prediction with AlphaFold
    Jumper, John; Evans, Richard; Pritzel, Alexander; Green, Tim; Figurnov, Michael; Ronneberger, Olaf; Tunyasuvunakool, Kathryn; Bates, Russ; Zidek, Augustin; Potapenko, Anna; Bridgland, Alex; Meyer, Clemens; Kohl, Simon A. A.; Ballard, Andrew J.; Cowie, Andrew; Romera-Paredes, Bernardino; Nikolov, Stanislav; Jain, Rishub; Adler, Jonas; Back, Trevor; Petersen, Stig; Reiman, David; Clancy, Ellen; Zielinski, Michal; Steinegger, Martin; Pacholska, Michalina; Berghammer, Tamas; Bodenstein, Sebastian; Silver, David; Vinyals, Oriol; Senior, Andrew W.; Kavukcuoglu, Koray; Kohli, Pushmeet; Hassabis, Demis
    Nature (London, United Kingdom) (2021), 596 (7873), 583-589CODEN: NATUAS; ISSN:0028-0836. (Nature Portfolio)
    Proteins are essential to life, and understanding their structure can facilitate a mechanistic understanding of their function. Through an enormous exptl. effort, the structures of around 100,000 unique proteins have been detd., but this represents a small fraction of the billions of known protein sequences. Structural coverage is bottlenecked by the months to years of painstaking effort required to det. a single protein structure. Accurate computational approaches are needed to address this gap and to enable large-scale structural bioinformatics. Predicting the three-dimensional structure that a protein will adopt based solely on its amino acid sequence-the structure prediction component of the 'protein folding problem'-has been an important open research problem for more than 50 years. Despite recent progress, existing methods fall far short of at. accuracy, esp. when no homologous structure is available. Here we provide the first computational method that can regularly predict protein structures with at. accuracy even in cases in which no similar structure is known. We validated an entirely redesigned version of our neural network-based model, AlphaFold, in the challenging 14th Crit. Assessment of protein Structure Prediction (CASP14), demonstrating accuracy competitive with exptl. structures in a majority of cases and greatly outperforming other methods. Underpinning the latest version of AlphaFold is a novel machine learning approach that incorporates phys. and biol. knowledge about protein structure, leveraging multi-sequence alignments, into the design of the deep learning algorithm.
  4. 4
    Baek, M.; DiMaio, F.; Anishchenko, I.; Dauparas, J.; Ovchinnikov, S.; Lee, G. R.; Wang, J.; Cong, Q.; Kinch, L. N.; Schaeffer, R. D.; Millán, C.; Park, H.; Adams, C.; Glassman, C. R.; DeGiovanni, A.; Pereira, J. H.; Rodrigues, A. V.; van Dijk, A. A.; Ebrecht, A. C.; Opperman, D. J.; Sagmeister, T.; Buhlheller, C.; Pavkov-Keller, T.; Rathinaswamy, M. K.; Dalwadi, U.; Yip, C. K.; Burke, J. E.; Garcia, K. C.; Grishin, N. V.; Adams, P. D.; Read, R. J.; Baker, D. Accurate prediction of protein structures and interactions using a three-track neural network. Science 2021, 373, 871– 876,  DOI: 10.1126/science.abj8754
    Google Scholar
    4
    Accurate prediction of protein structures and interactions using a three-track neural network
    Baek, Minkyung; DiMaio, Frank; Anishchenko, Ivan; Dauparas, Justas; Ovchinnikov, Sergey; Lee, Gyu Rie; Wang, Jue; Cong, Qian; Kinch, Lisa N.; Schaeffer, R. Dustin; Millan, Claudia; Park, Hahnbeom; Adams, Carson; Glassman, Caleb R.; DeGiovanni, Andy; Pereira, Jose H.; Rodrigues, Andria V.; van Dijk, Alberdina A.; Ebrecht, Ana C.; Opperman, Diederik J.; Sagmeister, Theo; Buhlheller, Christoph; Pavkov-Keller, Tea; Rathinaswamy, Manoj K.; Dalwadi, Udit; Yip, Calvin K.; Burke, John E.; Garcia, K. Christopher; Grishin, Nick V.; Adams, Paul D.; Read, Randy J.; Baker, David
    Science (Washington, DC, United States) (2021), 373 (6557), 871-876CODEN: SCIEAS; ISSN:1095-9203. (American Association for the Advancement of Science)
    DeepMind presented notably accurate predictions at the recent 14th Crit. Assessment of Structure Prediction (CASP14) conference. We explored network architectures that incorporate related ideas and obtained the best performance with a three-track network in which information at the one-dimensional (1D) sequence level, the 2D distance map level, and the 3D coordinate level is successively transformed and integrated. The three-track network produces structure predictions with accuracies approaching those of DeepMind in CASP14, enables the rapid soln. of challenging x-ray crystallog. and cryo-electron microscopy structure modeling problems, and provides insights into the functions of proteins of currently unknown structure. The network also enables rapid generation of accurate protein-protein complex models from sequence information alone, short-circuiting traditional approaches that require modeling of individual subunits followed by docking. We make the method available to the scientific community to speed biol. research.
  5. 5
    Miao, J.; Descoteaux, M. L.; Lin, Y.-S. Structure prediction of cyclic peptides by molecular dynamics + machine learning. Chem. Sci. 2021, 12, 14927– 14936,  DOI: 10.1039/D1SC05562C
    Google Scholar
    5
    Structure prediction of cyclic peptides by molecular dynamics + machine learning
    Miao, Jiayuan; Descoteaux, Marc L.; Lin, Yu-Shan
    Chemical Science (2021), 12 (44), 14927-14936CODEN: CSHCCN; ISSN:2041-6520. (Royal Society of Chemistry)
    Recent computational methods have made strides in discovering well-structured cyclic peptides that preferentially populate a single conformation. However, many successful cyclic-peptide therapeutics adopt multiple conformations in soln. In fact, the chameleonic properties of some cyclic peptides are likely responsible for their high cell membrane permeability. Thus, we require the ability to predict complete structural ensembles for cyclic peptides, including the majority of cyclic peptides that have broad structural ensembles, to significantly improve our ability to rationally design cyclic-peptide therapeutics. Here, we introduce the idea of using mol. dynamics simulation results to train machine learning models to enable efficient structure prediction for cyclic peptides. Using mol. dynamics simulation results for several hundred cyclic pentapeptides as the training datasets, we developed machine-learning models that can provide mol. dynamics simulation-quality predictions of structural ensembles for all the hundreds of thousands of sequences in the entire sequence space. The prediction for each individual cyclic peptide can be made using less than 1 s of computation time. Even for the most challenging classes of poorly structured cyclic peptides with broad conformational ensembles, our predictions were similar to those one would normally obtain only after running multiple days of explicit-solvent mol. dynamics simulations. The resulting method, termed StrEAMM (Structural Ensembles Achieved by Mol. Dynamics and Machine Learning), is the first technique capable of efficiently predicting complete structural ensembles of cyclic peptides without relying on addnl. mol. dynamics simulations, constituting a seven-order-of-magnitude improvement in speed while retaining the same accuracy as explicit-solvent simulations.
  6. 6
    Hui, T.; Descoteaux, M. L.; Miao, J.; Lin, Y.-S. Training neural network models using molecular dynamics simulation results to efficiently predict cyclic hexapeptide structural ensembles. J. Chem. Theory Comput. 2023, 19, 4757– 4769,  DOI: 10.1021/acs.jctc.3c00154
    Google Scholar
    There is no corresponding record for this reference.
  7. 7
    Wan, F.; Kontogiorgos-Heintz, D.; de la Fuente-Nunez, C. Deep generative models for peptide design. Digital Discovery 2022, 1, 195– 208,  DOI: 10.1039/D1DD00024A
    Google Scholar
    There is no corresponding record for this reference.
  8. 8
    Ferguson, A. L.; Ranganathan, R. 100th anniversary of macromolecular science viewpoint: data-driven protein design. ACS Macro Lett. 2021, 10, 327– 340,  DOI: 10.1021/acsmacrolett.0c00885
    Google Scholar
    8
    100Th Anniversary of Macromolecular Science Viewpoint: Data-Driven Protein Design
    Ferguson, Andrew L.; Ranganathan, Rama
    ACS Macro Letters (2021), 10 (3), 327-340CODEN: AMLCCD; ISSN:2161-1653. (American Chemical Society)
    A review. The design of synthetic proteins with the desired function is a long-standing goal in biomol. science, with broad applications in biochem. engineering, agriculture, medicine, and public health. Rational de novo design and exptl. directed evolution have achieved remarkable successes but are challenged by the requirement to find functional "needles" in the vast "haystack" of protein sequence space. Data-driven models for fitness landscapes provide a predictive map between protein sequence and function and can prospectively identify functional candidates for exptl. testing to greatly improve the efficiency of this search. This Viewpoint reviews the applications of machine learning and, in particular, deep learning as part of data-driven protein engineering platforms. We highlight recent successes, review promising computational methodologies, and provide an outlook on future challenges and opportunities. The article is written for a broad audience comprising both polymer and protein scientists and computer and data scientists interested in an up-to-date review of recent innovations and opportunities in this rapidly evolving field.
  9. 9
    Strokach, A.; Kim, P. M. Deep generative modeling for protein design. Curr. Opin. Struct. Biol. 2022, 72, 226– 236,  DOI: 10.1016/j.sbi.2021.11.008
    Google Scholar
    9
    Deep generative modeling for protein design
    Strokach, Alexey; Kim, Philip M.
    Current Opinion in Structural Biology (2022), 72 (), 226-236CODEN: COSBEF; ISSN:0959-440X. (Elsevier Ltd.)
    A review. Deep learning approaches have produced substantial breakthroughs in fields such as image classification and natural language processing and are making rapid inroads in the area of protein design. Many generative models of proteins have been developed that encompass all known protein sequences, model specific protein families, or extrapolate the dynamics of individual proteins. Those generative models can learn protein representations that are often more informative of protein structure and function than hand-engineered features. Furthermore, they can be used to quickly propose millions of novel proteins that resemble the native counterparts in terms of expression level, stability, or other attributes. The protein design process can further be guided by discriminative oracles to select candidates with the highest probability of having the desired properties. In this review, we discuss five classes of generative models that have been most successful at modeling proteins and provide a framework for model guided protein design.
  10. 10
    Chandra, A.; Tünnermann, L.; Löfstedt, T.; Gratz, R. Transformer-based deep learning for predicting protein properties in the life sciences. eLife 2023, 12, e82819  DOI: 10.7554/eLife.82819
    Google Scholar
    There is no corresponding record for this reference.
  11. 11
    Oliva, R.; Chino, M.; Pane, K.; Pistorio, V.; De Santis, A.; Pizzo, E.; D’Errico, G.; Pavone, V.; Lombardi, A.; Del Vecchio, P.; Notomista, E.; Nastri, F.; Petraccone, L. Exploring the role of unnatural amino acids in antimicrobial peptides. Sci. Rep. 2018, 8, 8888  DOI: 10.1038/s41598-018-27231-5
    Google Scholar
    11
    Exploring the role of unnatural amino acids in antimicrobial peptides
    Oliva Rosario; Chino Marco; De Santis Augusta; D'Errico Gerardino; Pavone Vincenzo; Lombardi Angela; Del Vecchio Pompea; Nastri Flavia; Petraccone Luigi; Pane Katia; Pizzo Elio; Notomista Eugenio; Pistorio Valeria
    Scientific reports (2018), 8 (1), 8888 ISSN:.
    Cationic antimicrobial peptides (CAMPs) are a promising alternative to treat multidrug-resistant bacteria, which have developed resistance to all the commonly used antimicrobial, and therefore represent a serious threat to human health. One of the major drawbacks of CAMPs is their sensitivity to proteases, which drastically limits their half-life. Here we describe the design and synthesis of three nine-residue CAMPs, which showed high stability in serum and broad spectrum antimicrobial activity. As for all peptides a very low selectivity between bacterial and eukaryotic cells was observed, we performed a detailed biophysical characterization of the interaction of one of these peptides with liposomes mimicking bacterial and eukaryotic membranes. Our results show a surface binding on the DPPC/DPPG vesicles, coupled with lipid domain formation, and, above a threshold concentration, a deep insertion into the bilayer hydrophobic core. On the contrary, mainly surface binding of the peptide on the DPPC bilayer was observed. These observed differences in the peptide interaction with the two model membranes suggest a divergence in the mechanisms responsible for the antimicrobial activity and for the observed high toxicity toward mammalian cell lines. These results could represent an important contribution to unravel some open and unresolved issues in the development of synthetic CAMPs.
  12. 12
    Lu, J.; Xu, H.; Xia, J.; Ma, J.; Xu, J.; Li, Y.; Feng, J. D- and unnatural amino acid substituted antimicrobial peptides with improved proteolytic resistance and their proteolytic degradation characteristics. Front. Microbiol. 2020, 11, 563030  DOI: 10.3389/fmicb.2020.563030
    Google Scholar
    12
    D- and Unnatural Amino Acid Substituted Antimicrobial Peptides With Improved Proteolytic Resistance and Their Proteolytic Degradation Characteristics
    Lu Jianguang; Xu Hongjiang; Xia Jianghua; Li Yanan; Feng Jun; Lu Jianguang; Ma Jie; Xu Jun; Feng Jun; Xu Hongjiang; Li Yanan
    Frontiers in microbiology (2020), 11 (), 563030 ISSN:1664-302X.
    The transition of antimicrobial peptides (AMPs) from the laboratory to market has been severely hindered by their instability toward proteases in biological systems. In the present study, we synthesized derivatives of the cationic AMP Pep05 (KRLFKKLLKYLRKF) by substituting L-amino acid residues with D- and unnatural amino acids, such as D-lysine, D-arginine, L-2,4-diaminobutanoic acid (Dab), L-2,3-diaminopropionic acid (Dap), L-homoarginine, 4-aminobutanoic acid (Aib), and L-thienylalanine, and evaluated their antimicrobial activities, toxicities, and stabilities toward trypsin, plasma proteases, and secreted bacterial proteases. In addition to measuring changes in the concentration of the intact peptides, LC-MS was used to identify the degradation products of the modified AMPs in the presence of trypsin and plasma proteases to determine degradation pathways and examine whether the amino acid substitutions afforded improved proteolytic resistance. The results revealed that both D- and unnatural amino acids enhanced the stabilities of the peptides toward proteases. The derivative DP06, in which all of the L-lysine and L-arginine residues were replaced by D-amino acids, displayed remarkable stability and mild toxicity in vitro but only slight activity and severe toxicity in vivo, indicating a significant difference between the in vivo and in vitro results. Unexpectedly, we found that the incorporation of a single Aib residue at the N-terminus of compound UP09 afforded remarkably enhanced plasma stability and improved activity in vivo. Hence, this derivative may represent a candidate AMP for further optimization, providing a new strategy for the design of novel AMPs with improved bioavailability.
  13. 13
    Taechalertpaisarn, J.; Ono, S.; Okada, O.; Johnstone, T. C.; Lokey, R. S. A new amino acid for improving permeability and solubility in macrocyclic peptides through side chain-to-backbone hydrogen bonding. J. Med. Chem. 2022, 65, 5072– 5084,  DOI: 10.1021/acs.jmedchem.2c00010
    Google Scholar
    13
    A new amino acid for improving permeability and solubility in macrocyclic peptides through side chain-to-backbone hydrogen bonding
    Taechalertpaisarn, Jaru; Ono, Satoshi; Okada, Okimasa; Johnstone, Timothy C.; Lokey, R. Scott
    Journal of Medicinal Chemistry (2022), 65 (6), 5072-5084CODEN: JMCMAR; ISSN:0022-2623. (American Chemical Society)
    Despite the notoriously poor membrane permeability of peptides, many cyclic peptide natural products show high passive membrane permeability and potently inhibit a variety of "undruggable" intracellular targets. A major impediment to the design of cyclic peptides with good permeability is the high desolvation energy assocd. with the peptide backbone amide NH groups. While several strategies have been proposed to mitigate this deleterious effect, only few studies have used polar side chains to sequester backbone NH groups. We investigated the ability of N,N-pyrrolidinylglutamine (Pye), whose side chain contains a powerful hydrogen-bond-accepting C:O amide group but no hydrogen-bond donors, to sequester exposed backbone NH groups in a series of cyclic hexapeptide diastereomers. Analyses revealed that specific Leu-to-Pye substitutions conferred dramatic improvements in aq. soly. and permeability in a scaffold- and position-dependent manner. Therefore, this approach offers a complementary tool for improving membrane permeability and soly. in cyclic peptides.
  14. 14
    Geurink, P. P.; van der Linden, W. A.; Mirabella, A. C.; Gallastegui, N.; de Bruin, G.; Blom, A. E.; Voges, M. J.; Mock, E. D.; Florea, B. I.; van der Marel, G. A.; Driessen, C.; van der Stelt, M.; Groll, M.; Overkleeft, H. S.; Kisselev, A. F. Incorporation of non-natural amino acids improves cell permeability and potency of specific inhibitors of proteasome trypsin-like sites. J. Med. Chem. 2013, 56, 1262– 1275,  DOI: 10.1021/jm3016987
    Google Scholar
    14
    Incorporation of non-natural amino acids improves cell permeability and potency of specific inhibitors of proteasome trypsin-like sites
    Geurink, Paul P.; van der Linden, Wouter A.; Mirabella, Anne C.; Gallastegui, Nerea; de Bruin, Gerjan; Blom, Annet E. M.; Voges, Mathias J.; Mock, Elliot D.; Florea, Bogdan I.; van der Marel, Gijs A.; Driessen, Christoph; van der Stelt, Mario; Groll, Michael; Overkleeft, Herman S.; Kisselev, Alexei F.
    Journal of Medicinal Chemistry (2013), 56 (3), 1262-1275CODEN: JMCMAR; ISSN:0022-2623. (American Chemical Society)
    Proteasomes degrade the majority of proteins in mammalian cells by a concerted action of three distinct pairs of active sites. The chymotrypsin-like sites are targets of antimyeloma agents bortezomib and carfilzomib. Inhibitors of the trypsin-like site sensitize multiple myeloma cells to these agents. Here we describe systematic effort to develop inhibitors with improved potency and cell permeability, yielding azido-Phe-Leu-Leu-4-aminomethyl-Phe-Me vinyl sulfone (I), LU-102, and a fluorescent activity-based probe for this site. X-ray structures of I and related inhibitors complexed with yeast proteasomes revealed the structural basis for specificity. Nontoxic to myeloma cells when used as a single agent, I sensitized them to bortezomib and carfilzomib. This sensitizing effect was much stronger than the synergistic effects of histone acetylase inhibitors or additive effects of doxorubicin and dexamethasone, raising the possibility that combinations of inhibitors of the trypsin-like site with bortezomib or carfilzomib would have stronger antineoplastic activity than combinations currently used clin.
  15. 15
    Weininger, D. SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules. J. Chem. Inf. Comput. Sci. 1988, 28, 31– 36,  DOI: 10.1021/ci00057a005
    Google Scholar
    15
    SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules
    Weininger, David
    Journal of Chemical Information and Computer Sciences (1988), 28 (1), 31-6CODEN: JCISD8; ISSN:0095-2338.
    The SMILES (simplified mol. input line entry system) chem. notation system is described for information processing. The system is based on principles of mol. graph theory and it allows structure specification by use of a very small and natural grammar well suited for high-speed machine processing. The system is easy to use, has high machine compatibility, and allows many computer applications, including notation generation, const. speed database retrieval, substructure searching, and property prediction models.
  16. 16
    Siani, M. A.; Weininger, D.; Blaney, J. M. CHUCKLES: A method for representing and searching peptide and peptoid sequences on both monomer and atomic levels. J. Chem. Inf. Comput. Sci. 1994, 34, 588– 593,  DOI: 10.1021/ci00019a017
    Google Scholar
    16
    CHUCKLES: A method for representing and searching peptide and peptoid sequences on both monomer and atomic levels
    Siani, Michael A.; Weininger, David; Blaney, Jeffrey M.
    Journal of Chemical Information and Computer Sciences (1994), 34 (3), 588-93CODEN: JCISD8; ISSN:0095-2338.
    Dual representation of peptide and non-peptide structures in a chem. database as at.-level mol. graphs and sequence strings permits chem. substructure and similarity searches as well as sequence-based substring and regular expression searches. CHUCKLES interconverts monomer-based sequences with SMILES, which represent at.-level mol. graphs. Forward-translation maps peptide or other sequences into SMILES. Back-translation exts. monomer sequences from SMILES. This approach permits a generalized representation of monomers allowing user specification of any monomer. CHUCKLES allows mixing of atoms with user-defined monomer names; i.e., monomer representation is consistent with SMILES notation. In addn., oligomer branching and cyclization are handled.
  17. 17
    Siani, M. A.; Weininger, D.; James, C. A.; Blaney, J. M. CHORTLES: A method for representing oligomeric and template-based mixtures. J. Chem. Inf. Comput. Sci. 1995, 35, 1026– 1033,  DOI: 10.1021/ci00028a012
    Google Scholar
    17
    CHORTLES: A Method for Representing Oligomeric and Template-Based Mixtures
    Siani, Michael A.; Weininger, David; James, Craig A.; Blaney, Jeffrey M.
    Journal of Chemical Information and Computer Sciences (1995), 35 (6), 1026-33CODEN: JCISD8; ISSN:0095-2338. (American Chemical Society)
    Screening mixts. of synthetic oligomers or fixed templates (e.g., rings) with varying substituents is increasingly the focus of drug discovery programs. CHORTLES is designed and implemented to facilitate representation, storage, and searching of oligomeric and template-based mixts. of any size. Building upon the CHUCKLES method of representing oligomers as both monomer-based sequences and all-atom structures, CHORTLES compactly represents a mixt. without explicitly enumerating individual mols. This method lends itself to a hierarchy relating mixts. to submixts. and individual compds., as one finds when deconvoluting mixts. in drug lead discovery programs. In addn., we describe two methods of searching mixts. at the monomer level. We also present a simple pictorial representation for describing all components in a mixt., which becomes essential as the list of monomer names is expanded beyond common names (e.g., amino acids).
  18. 18
    Jensen, J. H.; Hoeg-Jensen, T.; Padkjær, S. B. Building a biochemformatics database. J. Chem. Inf. Model. 2008, 48, 2404– 2413,  DOI: 10.1021/ci800128b
    Google Scholar
    18
    Building a BioChemformatics Database
    Jensen, Jan H.; Hoeg-Jensen, Thomas; Padkjaer, Soeren B.
    Journal of Chemical Information and Modeling (2008), 48 (12), 2404-2413CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    The structural registration of chem. modified macromols. is vital for the development of biopharmaceuticals. However, registration and search of such complex mols. has so far posed formidable challenges performance-wise, since today's chem.-oriented databases do not scale well to macromols. As a practical consequence, macromols. tend to be stored in protein databases with a focus on protein sequence only, and salient chem. details are therefore lost. This article describes protein format extensions and the use of pseudoatoms for representing natural amino acids in chem. structures to allow high-performance registration and retrieval of large macromols. The representations include exact chem. modifications and enable lossless conversion between chem. and sequence formats. Registration is done in parallel in both sequence and chem. formats, and users can register and retrieve mols. in either format as they choose, resulting in what we call a BioChemformatics database. Having both sequence and chem. formats available on-demand allows for the construction of protein SAR tables with mixed sequence and chem. information. Likewise, searching may combine sequence and chem. terms and be performed in std. vendor applications like MDL's ISIS/Base or inhouse applications using std. SQL queries.
  19. 19
    Lin, T.-S.; Coley, C. W.; Mochigase, H.; Beech, H. K.; Wang, W.; Wang, Z.; Woods, E.; Craig, S. L.; Johnson, J. A.; Kalow, J. A.; Jensen, K. F.; Olsen, B. D. BigSMILES: A structurally-based line notation for describing macromolecules. ACS Cent. Sci. 2019, 5, 1523– 1531,  DOI: 10.1021/acscentsci.9b00476
    Google Scholar
    19
    BigSMILES: A Structurally-Based Line Notation for Describing Macromolecules
    Lin, Tzyy-Shyang; Coley, Connor W.; Mochigase, Hidenobu; Beech, Haley K.; Wang, Wencong; Wang, Zi; Woods, Eliot; Craig, Stephen L.; Johnson, Jeremiah A.; Kalow, Julia A.; Jensen, Klavs F.; Olsen, Bradley D.
    ACS Central Science (2019), 5 (9), 1523-1531CODEN: ACSCII; ISSN:2374-7951. (American Chemical Society)
    Having a compact yet robust structurally based identifier or representation system is a key enabling factor for efficient sharing and dissemination of research results within the chem. community, and such systems lay down the essential foundations for future informatics and data-driven research. While substantial advances have been made for small mols., the polymer community has struggled in coming up with an efficient representation system. This is because, unlike other disciplines in chem., the basic premise that each distinct chem. species corresponds to a well-defined chem. structure does not hold for polymers. Polymers are intrinsically stochastic mols. that are often ensembles with a distribution of chem. structures. This difficulty limits the applicability of all deterministic representations developed for small mols. In this work, a new representation system that is capable of handling the stochastic nature of polymers is proposed. The new system is based on the popular "simplified mol.-input line-entry system" (SMILES), and it aims to provide representations that can be used as indexing identifiers for entries in polymer databases. As a pilot test, the entries of the std. data set of the glass transition temp. of linear polymers (Bicerano, 2002) were converted into the new BigSMILES language. Furthermore, it is hoped that the proposed system will provide a more effective language for communication within the polymer community and increase cohesion between the researchers within the community.
  20. 20
    Zhang, T.; Li, H.; Xi, H.; Stanton, R. V.; Rotstein, S. H. HELM: A hierarchical notation language for complex biomolecule structure representation. J. Chem. Inf. Model. 2012, 52, 2796– 2806,  DOI: 10.1021/ci3001925
    Google Scholar
    20
    HELM: A Hierarchical Notation Language for Complex Biomolecule Structure Representation
    Zhang, Tianhong; Li, Hongli; Xi, Hualin; Stanton, Robert V.; Rotstein, Sergio H.
    Journal of Chemical Information and Modeling (2012), 52 (10), 2796-2806CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    When biol. macromols. are used as therapeutic agents, it is often necessary to introduce non-natural chem. modifications to improve their pharmaceutical properties. The final products are complex structures where entities such as proteins, peptides, oligonucleotides, and small mol. drugs may be covalently linked to each other, or may include chem. modified biol. moieties. An accurate in silico representation of these complex structures is essential, as it forms the basis for their electronic registration, storage, anal., and visualization. The size of these mols. (henceforth referred to as "biomols.") often makes them too unwieldy and impractical to represent at the at. level, while the presence of non-natural chem. modifications makes it impossible to represent them by sequence alone. Here we describe the Hierarchical Editing Language for Macromols. ("HELM") and demonstrate its utility in the representation of structures such as antisense oligonucleotides, short interference RNAs, peptides, proteins, and antibody drug conjugates.
  21. 21
    David, L.; Thakkar, A.; Mercado, R.; Engkvist, O. Molecular representations in AI-driven drug discovery: a review and practical guide. J. Cheminf. 2020, 12, 56  DOI: 10.1186/s13321-020-00460-5
    Google Scholar
    21
    Molecular representations in AI-driven drug discovery: a review and practical guide
    David, Laurianne; Thakkar, Amol; Mercado, Rocio; Engkvist, Ola
    Journal of Cheminformatics (2020), 12 (1), 56CODEN: JCOHB3; ISSN:1758-2946. (SpringerOpen)
    A review. Abstr.: The technol. advances of the past century, marked by the computer revolution and the advent of high-throughput screening technologies in drug discovery, opened the path to the computational anal. and visualization of bioactive mols. For this purpose, it became necessary to represent mols. in a syntax that would be readable by computers and understandable by scientists of various fields. A large no. of chem. representations have been developed over the years, their numerosity being due to the fast development of computers and the complexity of producing a representation that encompasses all structural and chem. characteristics. We present here some of the most popular electronic mol. and macromol. representations used in drug discovery, many of which are based on graph representations. Furthermore, we describe applications of these representations in AI-driven drug discovery. Our aim is to provide a brief guide on structural representations that are essential to the practise of AI in drug discovery. This review serves as a guide for researchers who have little experience with the handling of chem. representations and plan to work on applications at the interface of these fields.
  22. 22
    Nguyen-Vo, T.-H.; Teesdale-Spittle, P.; Harvey, J. E.; Nguyen, B. P. Molecular representations in bio-cheminformatics. Memetic Comput. 2024, 16, 519– 536,  DOI: 10.1007/s12293-024-00414-6
    Google Scholar
    There is no corresponding record for this reference.
  23. 23
    Wigh, D. S.; Goodman, J. M.; Lapkin, A. A. A review of molecular representation in the age of machine learning. WIREs Comput. Mol. Sci. 2022, 12, e1603  DOI: 10.1002/wcms.1603
    Google Scholar
    There is no corresponding record for this reference.
  24. 24
    Sousa, T.; Correia, J.; Pereira, V.; Rocha, M. Generative deep learning for targeted compound design. J. Chem. Inf. Model. 2021, 61, 5343– 5361,  DOI: 10.1021/acs.jcim.0c01496
    Google Scholar
    24
    Generative Deep Learning for Targeted Compound Design
    Sousa, Tiago; Correia, Joao; Pereira, Vitor; Rocha, Miguel
    Journal of Chemical Information and Modeling (2021), 61 (11), 5343-5361CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    A review. In the past few years, de novo mol. design has increasingly been using generative models from the emergent field of Deep Learning, proposing novel compds. that are likely to possess desired properties or activities. De novo mol. design finds applications in different fields ranging from drug discovery and materials sciences to biotechnol. A panoply of deep generative models, including architectures as Recurrent Neural Networks, Autoencoders, and Generative Adversarial Networks, can be trained on existing data sets and provide for the generation of novel compds. Typically, the new compds. follow the same underlying statistical distributions of properties exhibited on the training data set. Addnl., different optimization strategies, including transfer learning, Bayesian optimization, reinforcement learning, and conditional generation, can direct the generation process toward desired aims, regarding their biol. activities, synthesis processes or chem. features. Given the recent emergence of these technologies and their relevance, this work presents a systematic and crit. review on deep generative models and related optimization methods for targeted compd. design, and their applications.
  25. 25
    Morgan, H. L. The generation of a unique machine description for chemical structures-a technique developed at chemical abstracts service. J. Chem. Doc. 1965, 5, 107– 113,  DOI: 10.1021/c160017a018
    Google Scholar
    25
    Generation of a unique machine description for chemical structures--a technique developed at Chemical Abstracts Service
    Morgan, H. L.
    Journal of Chemical Documentation (1965), 5 (2), 107-13CODEN: JCHDAN; ISSN:0021-9576.
    The description employed is a uniquely ordered list of the node symbols of the structure (or graph) in which the value (at. symbol) of each node and its attachment (bonding) to the other nodes of the total structure. When the entire structure has been numbered according to a given set of rules, the connection table is formed by recording the structural relation by a process of successive partial orderings.
  26. 26
    Durant, J. L.; Leland, B. A.; Henry, D. R.; Nourse, J. G. Reoptimization of MDL Keys for use in drug discovery. J. Chem. Inf. Comput. Sci. 2002, 42, 1273– 1280,  DOI: 10.1021/ci010132r
    Google Scholar
    26
    Reoptimization of MDL Keys for Use in Drug Discovery
    Durant, Joseph L.; Leland, Burton A.; Henry, Douglas R.; Nourse, James G.
    Journal of Chemical Information and Computer Sciences (2002), 42 (6), 1273-1280CODEN: JCISD8; ISSN:0095-2338. (American Chemical Society)
    For a no. of years MDL products have exposed both 166 bit and 960 bit keysets based on 2D descriptors. These keysets were originally constructed and optimized for substructure searching. We report on improvements in the performance of MDL keysets which are reoptimized for use in mol. similarity. Classification performance for a test data set of 957 compds. was increased from 0.65 for the 166 bit keyset and 0.67 for the 960 bit keyset to 0.71 for a surprisal S/N pruned keyset contg. 208 bits and 0.71 for a genetic algorithm optimized keyset contg. 548 bits. We present an overview of the underlying technol. supporting the definition of descriptors and the encoding of these descriptors into keysets. This technol. allows definition of descriptors as combinations of atom properties, bond properties, and at. neighborhoods at various topol. sepns. as well as supporting a no. of custom descriptors. These descriptors can then be used to set one or more bits in a keyset. We constructed various keysets and optimized their performance in clustering bioactive substances. Performance was measured using methodol. developed by Briem and Lessel. "Directed pruning" was carried out by eliminating bits from the keysets on the basis of random selection, values of the surprisal of the bit, or values of the surprisal S/N ratio of the bit. The random pruning expt. highlighted the insensitivity of keyset performance for keyset lengths of more than 1000 bits. Contrary to initial expectations, pruning on the basis of the surprisal values of the various bits resulted in keysets which underperformed those resulting from random pruning. In contrast, pruning on the basis of the surprisal S/N ratio was found to yield keysets which performed better than those resulting from random pruning. We also explored the use of genetic algorithms in the selection of optimal keysets. Once more the performance was only a weak function of keyset size, and the optimizations failed to identify a single globally optimal keyset. Instead multiple, equally optimal keysets could be produced which had relatively low overlap of the descriptors they encoded.
  27. 27
    Schissel, C. K.; Mohapatra, S.; Wolfe, J. M.; Fadzen, C. M.; Bellovoda, K.; Wu, C.-L.; Wood, J. A.; Malmberg, A. B.; Loas, A.; Gómez-Bombarelli, R.; Pentelute, B. L. Deep learning to design nuclear-targeting abiotic miniproteins. Nat. Chem. 2021, 13, 992– 1000,  DOI: 10.1038/s41557-021-00766-3
    Google Scholar
    27
    Deep learning to design nuclear-targeting abiotic miniproteins
    Schissel, Carly K.; Mohapatra, Somesh; Wolfe, Justin M.; Fadzen, Colin M.; Bellovoda, Kamela; Wu, Chia-Ling; Wood, Jenna A.; Malmberg, Annika B.; Loas, Andrei; Gomez-Bombarelli, Rafael; Pentelute, Bradley L.
    Nature Chemistry (2021), 13 (10), 992-1000CODEN: NCAHBB; ISSN:1755-4330. (Nature Portfolio)
    There are more amino acid permutations within a 40-residue sequence than atoms on Earth. This vast chem. search space hinders the use of human learning to design functional polymers. Here we show how machine learning enables the de novo design of abiotic nuclear-targeting miniproteins to traffic antisense oligomers to the nucleus of cells. We combined high-throughput experimentation with a directed evolution-inspired deep-learning approach in which the mol. structures of natural and unnatural residues are represented as topol. fingerprints. The model is able to predict activities beyond the training dataset, and simultaneously deciphers and visualizes sequence-activity predictions. The predicted miniproteins, termed Mach reach an av. mass of 10 kDa, are more effective than any previously known variant in cells and can also deliver proteins into the cytosol. The Mach miniproteins are non-toxic and efficiently deliver antisense cargo in mice. These results demonstrate that deep learning can decipher design principles to generate highly active biomols. that are unlikely to be discovered by empirical approaches.
  28. 28
    Ji, X.; Nielsen, A. L.; Heinis, C. Cyclic peptides for drug development. Angew. Chem., Int. Ed. 2024, 63, e202308251  DOI: 10.1002/anie.202308251
    Google Scholar
    There is no corresponding record for this reference.
  29. 29
    Costa, L.; Sousa, E.; Fernandes, C. Cyclic peptides in pipeline: what future for these great molecules?. Pharmaceuticals 2023, 16, 996  DOI: 10.3390/ph16070996
    Google Scholar
    29
    Antibiotics, antifungals, anticancer, and immunosuppressants use of cyclic peptides in therapeutics for different diseases
    Costa, Lia; Sousa, Emilia; Fernandes, Carla
    Pharmaceuticals (2023), 16 (7), 996CODEN: PHARH2; ISSN:1424-8247. (MDPI AG)
    A review. Cyclic peptides are mols. that are already used as drugs in therapies approved for various pharmacol. activities, for example, as antibiotics, antifungals, anticancer, and immunosuppressants. Interest in these mols. has been growing due to the improved pharmacokinetic and pharmacodynamic properties of the cyclic structure over linear peptides and by the evolution of chem. synthesis, computational, and in vitro methods. To date, 53 cyclic peptides have been approved by different regulatory authorities, and many others are in clin. trials for a wide diversity of conditions. In this review, the potential of cyclic peptides is presented, and general aspects of their synthesis and development are discussed. Furthermore, an overview of already approved cyclic peptides is also given, and the cyclic peptides in clin. trials are summarized.
  30. 30
    Zhang, H.; Chen, S. Cyclic peptide drugs approved in the last two decades (2001–2021). RSC Chem. Biol. 2022, 3, 18– 31,  DOI: 10.1039/D1CB00154J
    Google Scholar
    30
    Cyclic peptide drugs approved in the last two decades (2001-2021)
    Zhang Huiya; Chen Shiyu
    RSC chemical biology (2022), 3 (1), 18-31 ISSN:.
    In contrast to the major families of small molecules and antibodies, cyclic peptides, as a family of synthesizable macromolecules, have distinct biochemical and therapeutic properties for pharmaceutical applications. Cyclic peptide-based drugs have increasingly been developed in the past two decades, confirming the common perception that cyclic peptides have high binding affinities and low metabolic toxicity as antibodies, good stability and ease of manufacture as small molecules. Natural peptides were the major source of cyclic peptide drugs in the last century, and cyclic peptides derived from novel screening and cyclization strategies are the new source. In this review, we will discuss and summarize 18 cyclic peptides approved for clinical use in the past two decades to provide a better understanding of cyclic peptide development and to inspire new perspectives. The purpose of the present review is to promote efforts to resolve the challenges in the development of cyclic peptide drugs that are more effective.
  31. 31
    Landrum, G. RDKit: Open-Source Cheminformatics Software. https://www.rdkit.org/.
    Google Scholar
    There is no corresponding record for this reference.
  32. 32
    The PyMOL Molecular Graphics System, version 3.0; Schrödinger, LLC.
    Google Scholar
    There is no corresponding record for this reference.
  33. 33
    Case, D. A.; Aktulga, H. M. A.; Belfon, K.; Ben-Shalom, I. Y.; Berryman, J. T.; Brozell, S. R.; Cerutti, D. S.; Cheatham, T. E., III; Cisneros, G. A.; Cruzeiro, V. W. D.; Darden, T. A.; Duke, R. E.; Giambasu, G.; Gilson, M. K.; Gohlke, H.; Goetz, A. W.; Harris, R.; Izadi, S.; Izmailov, S. A.; Kasavajhala, K.; Kaymak, M. C.; King, E.; Kovalenko, A.; Kurtzman, T.; Lee, T. S.; LeGrand, S.; Li, P.; Lin, C.; Liu, J.; Luchko, T.; Luo, R.; Machado, M.; Man, V.; Manathunga, M.; Merz, K. M.; Miao, Y.; Mikhailovskii, O.; Monard, G.; Nguyen, H.; O’Hearn, K. A.; Onufriev, A.; Pan, F.; Pantano, S.; Qi, R.; Rahnamoun, A.; Roe, D. R.; Roitberg, A.; Sagui, C.; Schott-Verdugo, S.; Shajan, A.; Shen, J.; Simmerling, C. L.; Skrynnikov, N. R.; Smith, J.; Swails, J.; Walker, R. C.; Wang, J.; Wang, J.; Wei, H.; Wolf, R. M.; Wu, X.; Xiong, Y.; Xue, Y.; York, D. M.; Zhao, S.; Kollman, P. A. Amber; University of California: San Francisco, 2022.
    Google Scholar
    There is no corresponding record for this reference.
  34. 34
    Jakalian, A.; Jack, D. B.; Bayly, C. I. Fast, efficient generation of high-quality atomic charges. AM1-BCC model: II. Parameterization and validation. J. Comput. Chem. 2002, 23, 1623– 4161,  DOI: 10.1002/jcc.10128
    Google Scholar
    34
    Fast, efficient generation of high-quality atomic charges. AM1-BCC model: II. parameterization and validation
    Jakalian, Araz; Jack, David B.; Bayly, Christopher I.
    Journal of Computational Chemistry (2002), 23 (16), 1623-1641CODEN: JCCHDD; ISSN:0192-8651. (John Wiley & Sons, Inc.)
    We present the first global parameterization and validation of a novel charge model, called AM1-BCC, which quickly and efficiently generates high-quality at. charges for computer simulations of org. mols. in polar media. The goal of the charge model is to produce at. charges that emulate the HF/6-31G* electrostatic potential (ESP) of a mol. Underlying electronic structure features, including formal charge and electron delocalization, are first captured by AM1 population charges; simple additive bond charge corrections (BCCs) are then applied to these AM1 at. charges to produce the AM1-BCC charges. The parameterization of BCCs was carried out by fitting to the HF/6-31G* ESP of a training set of >2700 mols. Most org. functional groups and their combinations were sampled, as well as an extensive variety of cyclic and fused bicyclic heteroaryl systems. The resulting BCC parameters allow the AM1-BCC charging scheme to handle virtually all types of org. compds. listed in The Merck Index and the NCI Database. Validation of the model was done through comparisons of hydrogen-bonded dimer energies and relative free energies of solvation using AM1-BCC charges in conjunction with the 1994 Cornell et al. forcefield for AMBER. Homo-dimer and hetero-dimer hydrogen-bond energies of a diverse set of org. mols. were reproduced to within 0.95 kcal/mol RMS deviation from the ab initio values, and for DNA dimers the energies were within 0.9 kcal/mol RMS deviation from ab initio values. The calcd. relative free energies of solvation for a diverse set of monofunctional isosteres were reproduced to within 0.69 kcal/mol of expt. In all these validation tests, AMBER with the AM1-BCC charge model maintained a correlation coeff. above 0.96. Thus, the parameters presented here for use with the AM1-BCC method present a fast, accurate, and robust alternative to HF/6-31G* ESP-fit charges for general use with the AMBER force field in computer simulations involving org. small mols.
  35. 35
    Bayly, C. I.; Cieplak, P.; Cornell, W.; Kollman, P. A. A well-behaved electrostatic potential based method using charge restraints for deriving atomic charges: the RESP model. J. Phys. Chem. A 1993, 97, 10269– 10280,  DOI: 10.1021/j100142a004
    Google Scholar
    35
    A well-behaved electrostatic potential based method using charge restraints for deriving atomic charges: the RESP model
    Bayly, Christopher I.; Cieplak, Piotr; Cornell, Wendy; Kollman, Peter A.
    Journal of Physical Chemistry (1993), 97 (40), 10269-80CODEN: JPCHAX; ISSN:0022-3654.
    The authors present a new approach to generating electrostatic potential (ESP) derived charges for mols. The major strength of electrostatic potential derived charges is that they optimally reproduce the intermol. interaction properties of mols. with a simple two-body additive potential, provided, of course, that a suitably accurate level of quantum mech. calcn. is used to derive the ESP around the mol. Previously, the major weaknesses of these charges have been that they were not easily transferably between common functional groups in related mols., they have often been conformationally dependent, and the large charges that frequently occur can be problematic for simulating intramol. interactions. Introducing restraints in the form of a penalty function into the fitting process considerably reduces the above problems, with only a minor decrease in the quality of the fit to the quantum mech. ESP. Several other refinements in addn. to the restrained electrostatic potential (RESP) fit yield a general and algorithmic charge fitting procedure for generating atom-centered point charges. This approach can thus be recommended for general use in mol. mechanics, mol. dynamics, and free energy calcns. for any org. or bioorg. system.
  36. 36
    Hornak, V.; Abel, R.; Okur, A.; Strockbine, B.; Roitberg, A.; Simmerling, C. Comparison of multiple amber force fields and development of improved protein backbone parameters. Proteins 2006, 65, 712– 725,  DOI: 10.1002/prot.21123
    Google Scholar
    36
    Comparison of multiple Amber force fields and development of improved protein backbone parameters
    Hornak, Viktor; Abel, Robert; Okur, Asim; Strockbine, Bentley; Roitberg, Adrian; Simmerling, Carlos
    Proteins: Structure, Function, and Bioinformatics (2006), 65 (3), 712-725CODEN: PSFBAF ISSN:. (Wiley-Liss, Inc.)
    The ff94 force field that is commonly assocd. with the Amber simulation package is one of the most widely used parameter sets for biomol. simulation. After a decade of extensive use and testing, limitations in this force field, such as over-stabilization of α-helixes, were reported by the authors and other researchers. This led to a no. of attempts to improve these parameters, resulting in a variety of "Amber" force fields and significant difficulty in detg. which should be used for a particular application. The authors show that several of these continue to suffer from inadequate balance between different secondary structure elements. In addn., the approach used in most of these studies neglected to account for the existence in Amber of two sets of backbone .vphi./ψ dihedral terms. This led to parameter sets that provide unreasonable conformational preferences for glycine. The authors report here an effort to improve the .vphi./ψ dihedral terms in the ff99 energy function. Dihedral term parameters are based on fitting the energies of multiple conformations of glycine and alanine tetrapeptides from high level ab initio quantum mech. calcns. The new parameters for backbone dihedrals replace those in the existing ff99 force field. This parameter set, which the authors denote ff99SB, achieves a better balance of secondary structure elements as judged by improved distribution of backbone dihedrals for glycine and alanine with respect to PDB survey data. It also accomplishes improved agreement with published exptl. data for conformational preferences of short alanine peptides and better accord with exptl. NMR relaxation data of test protein systems.
  37. 37
    Wang, J.; Wolf, R. M.; Caldwell, J. W.; Kollman, P. A.; Case, D. A. Development and testing of a general amber force field. J. Comput. Chem. 2004, 25, 1157– 1174,  DOI: 10.1002/jcc.20035
    Google Scholar
    37
    Development and testing of a general Amber force field
    Wang, Junmei; Wolf, Romain M.; Caldwell, James W.; Kollman, Peter A.; Case, David A.
    Journal of Computational Chemistry (2004), 25 (9), 1157-1174CODEN: JCCHDD; ISSN:0192-8651. (John Wiley & Sons, Inc.)
    We describe here a general Amber force field (GAFF) for org. mols. GAFF is designed to be compatible with existing Amber force fields for proteins and nucleic acids, and has parameters for most org. and pharmaceutical mols. that are composed of H, C, N, O, S, P, and halogens. It uses a simple functional form and a limited no. of atom types, but incorporates both empirical and heuristic models to est. force consts. and partial at. charges. The performance of GAFF in test cases is encouraging. In test I, 74 crystallog. structures were compared to GAFF minimized structures, with a root-mean-square displacement of 0.26 Å, which is comparable to that of the Tripos 5.2 force field (0.25 Å) and better than those of MMFF 94 and CHARMm (0.47 and 0.44 Å, resp.). In test II, gas phase minimizations were performed on 22 nucleic acid base pairs, and the minimized structures and intermol. energies were compared to MP2/6-31G* results. The RMS of displacements and relative energies were 0.25 Å and 1.2 kcal/mol, resp. These data are comparable to results from Parm99/RESP (0.16 Å and 1.18 kcal/mol, resp.), which were parameterized to these base pairs. Test III looked at the relative energies of 71 conformational pairs that were used in development of the Parm99 force field. The RMS error in relative energies (compared to expt.) is about 0.5 kcal/mol. GAFF can be applied to wide range of mols. in an automatic fashion, making it suitable for rational drug design and database searching.
  38. 38
    Zhou, C.-Y.; Jiang, F.; Wu, Y.-D. Residue-specific force field based on protein coil library. RSFF2: modification of AMBER ff99SB. J. Phys. Chem. B 2015, 119, 1035– 1047,  DOI: 10.1021/jp5064676
    Google Scholar
    38
    Residue-Specific Force Field Based on Protein Coil Library. RSFF2: Modification of AMBER ff99SB
    Zhou, Chen-Yang; Jiang, Fan; Wu, Yun-Dong
    Journal of Physical Chemistry B (2015), 119 (3), 1035-1047CODEN: JPCBFK; ISSN:1520-5207. (American Chemical Society)
    Recently, we developed a residue-specific force field (RSFF1) based on conformational free-energy distributions of the 20 amino acid residues from a protein coil library. Most parameters in RSFF1 were adopted from the OPLS-AA/L force field, but some van der Waals and torsional parameters that effectively affect local conformational preferences were introduced specifically for individual residues to fit the coil library distributions. Here a similar strategy has been applied to modify the Amber ff99SB force field, and a new force field named RSFF2 is developed. It can successfully fold α-helical structures such as polyalanine peptides, Trp-cage miniprotein, and villin headpiece subdomain and β-sheet structures such as Trpzip-2, GB1 β-hairpins, and the WW domain, simultaneously. The properties of various popular force fields in balancing between α-helix and β-sheet are analyzed based on their descriptions of local conformational features of various residues, and the anal. reveals the importance of accurate local free-energy distributions. Unlike the RSFF1, which overestimates the stability of both α-helix and β-sheet, RSFF2 gives melting curves of α-helical peptides and Trp-cage in good agreement with exptl. data. Fitting to the two-state model, RSFF2 gives folding enthalpies and entropies in reasonably good agreement with available exptl. results.
  39. 39
    Abraham, M. J.; Murtola, T.; Schulz, R.; Páll, S.; Smith, J. C.; Hess, B.; Lindahl, E. GROMACS: High performance molecular simulations through multi-level parallelism from laptops to supercomputers. SoftwareX 2015, 1–2, 19– 25,  DOI: 10.1016/j.softx.2015.06.001
    Google Scholar
    There is no corresponding record for this reference.
  40. 40
    Jorgensen, W. L.; Chandrasekhar, J.; Madura, J. D.; Impey, R. W.; Klein, M. L. Comparison of simple potential functions for simulating liquid water. J. Chem. Phys. 1983, 79, 926– 935,  DOI: 10.1063/1.445869
    Google Scholar
    40
    Comparison of simple potential functions for simulating liquid water
    Jorgensen, William L.; Chandrasekhar, Jayaraman; Madura, Jeffry D.; Impey, Roger W.; Klein, Michael L.
    Journal of Chemical Physics (1983), 79 (2), 926-35CODEN: JCPSA6; ISSN:0021-9606.
    Classical Monte Carlo simulations were carried out for liq. H2O in the NPT ensemble at 25° and 1 atm using 6 of the simpler intermol. potential functions for the dimer. Comparisons were made with exptl. thermodn. and structural data including the neutron diffraction results of Thiessen and Narten (1982). The computed densities and potential energies agree with expt. except for the original Bernal-Fowler model, which yields an 18% overest. of the d. and poor structural results. The discrepancy may be due to the correction terms needed in processing the neutron data or to an effect uniformly neglected in the computations. Comparisons were made for the self-diffusion coeffs. obtained from mol. dynamics simulations.
  41. 41
    Essmann, U.; Perera, L.; Berkowitz, M. L.; Darden, T.; Lee, H.; Pedersen, L. G. A smooth particle mesh Ewald method. J. Chem. Phys. 1995, 103, 8577– 8593,  DOI: 10.1063/1.470117
    Google Scholar
    41
    A smooth particle mesh Ewald method
    Essmann, Ulrich; Perera, Lalith; Berkowitz, Max L.; Darden, Tom; Lee, Hsing; Pedersen, Lee G.
    Journal of Chemical Physics (1995), 103 (19), 8577-93CODEN: JCPSA6; ISSN:0021-9606. (American Institute of Physics)
    The previously developed particle mesh Ewald method is reformulated in terms of efficient B-spline interpolation of the structure factors. This reformulation allows a natural extension of the method to potentials of the form 1/rp with p ≥ 1. Furthermore, efficient calcn. of the virial tensor follows. Use of B-splines in the place of Lagrange interpolation leads to analytic gradients as well as a significant improvement in the accuracy. The authors demonstrate that arbitrary accuracy can be achieved, independent of system size N, at a cost that scales as N log(N). For biomol. systems with many thousands of atoms and this method permits the use of Ewald summation at a computational cost comparable to that of a simple truncation method of 10 Å or less.
  42. 42
    Frisch, M. J.; Trucks, G. W.; Schlegel, H. B.; Scuseria, G. E.; Robb, M. A.; Cheeseman, J. R.; Scalmani, G.; Barone, V.; Petersson, G. A.; Nakatsuji, H.; Li, X.; Caricato, M.; Marenich, A. V.; Bloino, J.; Janesko, B. G.; Gomperts, R.; Mennucci, B.; Hratchian, H. P.; Ortiz, J. V.; Izmaylov, A. F.; Sonnenberg, J. L.; Williams; ; Ding, F.; Lipparini, F.; Egidi, F.; Goings, J.; Peng, B.; Petrone, A.; Henderson, T.; Ranasinghe, D.; Zakrzewski, V. G.; Gao, J.; Rega, N.; Zheng, G.; Liang, W.; Hada, M.; Ehara, M.; Toyota, K.; Fukuda, R.; Hasegawa, J.; Ishida, M.; Nakajima, T.; Honda, Y.; Kitao, O.; Nakai, H.; Vreven, T.; Throssell, K.; Montgomery, J. A., Jr.; Peralta, J. E.; Ogliaro, F.; Bearpark, M. J.; Heyd, J. J.; Brothers, E. N.; Kudin, K. N.; Staroverov, V. N.; Keith, T. A.; Kobayashi, R.; Normand, J.; Raghavachari, K.; Rendell, A. P.; Burant, J. C.; Iyengar, S. S.; Tomasi, J.; Cossi, M.; Millam, J. M.; Klene, M.; Adamo, C.; Cammi, R.; Ochterski, J. W.; Martin, R. L.; Morokuma, K.; Farkas, O.; Foresman, J. B.; Fox, D. J. Gaussian 16, rev. C.01; Gaussian Inc.: Wallingford, CT, 2016.
    Google Scholar
    There is no corresponding record for this reference.
  43. 43
    Vanquelef, E.; Simon, S.; Marquant, G.; Garcia, E.; Klimerak, G.; Delepine, J. C.; Cieplak, P.; Dupradeau, F. Y. R.E.D. Server: a web service for deriving RESP and ESP charges and building force field libraries for new molecules and molecular fragments. Nucleic Acids Res. 2011, 39, W511– 517,  DOI: 10.1093/nar/gkr288
    Google Scholar
    43
    R.E.D. Server: a web service for deriving RESP and ESP charges and building force field libraries for new molecules and molecular fragments
    Vanquelef, Enguerran; Simon, Sabrina; Marquant, Gaelle; Garcia, Elodie; Klimerak, Geoffroy; Delepine, Jean Charles; Cieplak, Piotr; Dupradeau, Francois-Yves
    Nucleic Acids Research (2011), 39 (Web Server), W511-W517CODEN: NARHAD; ISSN:0305-1048. (Oxford University Press)
    R.E.D. Server is a unique, open web service, designed to derive non-polarizable RESP and ESP charges and to build force field libraries for new mols./mol. fragments. It provides to computational biologists the means to derive rigorously mol. electrostatic potential-based charges embedded in force field libraries that are ready to be used in force field development, charge validation and mol. dynamics simulations. R.E.D. Server interfaces quantum mechanics programs, the RESP program and the latest version of the R.E.D. tools. A two step approach has been developed. The first one consists of prepg. P2N file(s) to rigorously define key elements such as atom names, topol. and chem. equivalencing needed when building a force field library. Then, P2N files are used to derive RESP or ESP charges embedded in force field libraries in the Tripos mol2 format. In complex cases an entire set of force field libraries or force field topol. database is generated. Other features developed in R.E.D. Server include help services, a demonstration, tutorials, frequently asked questions, Jmol-based tools useful to construct PDB input files and parse R.E.D. Server outputs as well as a graphical queuing system allowing any user to check the status of R.E.D. Server jobs.
  44. 44
    Piana, S.; Laio, A. A bias-exchange approach to protein folding. J. Phys. Chem. B 2007, 111, 4553– 4559,  DOI: 10.1021/jp067873l
    Google Scholar
    44
    A Bias-Exchange Approach to Protein Folding
    Piana, Stefano; Laio, Alessandro
    Journal of Physical Chemistry B (2007), 111 (17), 4553-4559CODEN: JPCBFK; ISSN:1520-6106. (American Chemical Society)
    By suitably extending a recent approach [Bussi, G., et al., 2006] the authors introduce a powerful methodol. that allows the parallel reconstruction of the free energy of a system in a virtually unlimited no. of variables. Multiple metadynamics simulations of the same system at the same temp. are performed, biasing each replica with a time-dependent potential constructed in a different set of collective variables. Exchanges between the bias potentials in the different variables are periodically allowed according to a replica exchange scheme. Due to the efficaciously multidimensional nature of the bias the method allows exploring complex free energy landscapes with high efficiency. The usefulness of the method is demonstrated by performing an atomistic simulation in explicit solvent of the folding of a Triptophane cage miniprotein. It is shown that the folding free energy landscape can be fully characterized starting from an extended conformation with use of only 40 ns of simulation on 8 replicas.
  45. 45
    Tribello, G. A.; Bonomi, M.; Branduardi, D.; Camilloni, C.; Bussi, G. PLUMED 2: New feathers for an old bird. Comput. Phys. Commun. 2014, 185, 604– 613,  DOI: 10.1016/j.cpc.2013.09.018
    Google Scholar
    45
    PLUMED 2: New feathers for an old bird
    Tribello, Gareth A.; Bonomi, Massimiliano; Branduardi, Davide; Camilloni, Carlo; Bussi, Giovanni
    Computer Physics Communications (2014), 185 (2), 604-613CODEN: CPHCBZ; ISSN:0010-4655. (Elsevier B.V.)
    Enhancing sampling and analyzing simulations are central issues in mol. simulation. Recently, we introduced PLUMED, an open-source plug-in that provides some of the most popular mol. dynamics (MD) codes with implementations of a variety of different enhanced sampling algorithms and collective variables (CVs). The rapid changes in this field, in particular new directions in enhanced sampling and dimensionality redn. together with new hardware, require a code that is more flexible and more efficient. We therefore present PLUMED 2 here-a complete rewrite of the code in an object-oriented programming language (C++). This new version introduces greater flexibility and greater modularity, which both extends its core capabilities and makes it far easier to add new methods and CVs. It also has a simpler interface with the MD engines and provides a single software library contg. both tools and core facilities. Ultimately, the new code better serves the ever-growing community of users and contributors in coping with the new challenges arising in the field.
  46. 46
    Damas, J. M.; Filipe, L. C.; Campos, S. R.; Lousa, D.; Victor, B. L.; Baptista, A. M.; Soares, C. M. Predicting the thermodynamics and kinetics of helix formation in a cyclic peptide model. J. Chem. Theory Comput. 2013, 9, 5148– 5157,  DOI: 10.1021/ct400529k
    Google Scholar
    46
    Predicting the thermodynamics and kinetics of helix formation in a cyclic peptide model
    Damas, Joao M.; Filipe, Luis C. S.; Campos, Sara R. R.; Lousa, Diana; Victor, Bruno L.; Baptista, Antonio M.; Soares, Claudio M.
    Journal of Chemical Theory and Computation (2013), 9 (11), 5148-5157CODEN: JCTCCE; ISSN:1549-9618. (American Chemical Society)
    The peptide, Ac-(cyclo-2,6)-R-[KAAAD]-NH2 (cyc-RKAAAD), is a short cyclic peptide known to adopt a remarkably stable single-turn α-helix in water. Due to its simplicity and the availability of thermodn. and kinetic exptl. data, cyc-RKAAAD poses as an ideal model for evaluating the aptness of current mol. dynamics (MD) simulation methodologies to accurately sample conformations that reproduce exptl. obsd. properties. Here, the authors extensively sampled the conformational space of cyc-RKAAAD using microsecond-timescale MD simulations. The authors characterized the peptide conformational preferences in terms of secondary structure propensities and, using Cartesian-coordinate principal component anal. (cPCA), constructed its free energy landscape, thus obtaining a detailed weighted discrimination between the helical and nonhelical subensembles. The cPCA state discrimination, together with a Markov model built from it, allowed the authors to est. the free energy of unfolding (-0.57 kJ/mol) and the relaxation time (∼0.435 μs) at 298.15 K, which were in excellent agreement with the exptl. reported values. Addnl., the authors presented simulations conducted using 2 enhanced sampling methods: replica-exchange mol. dynamics (REMD) and bias-exchange metadynamics (BE-MetaD). The authors compared the free energy landscape obtained by these 2 methods with the results from MD simulations and discussed the sampling and computational gains achieved. Overall, the results obtained attested to the suitability of modern simulation methods to explore the conformational behavior of peptide systems with a high level of realism.
  47. 47
    Nair, V.; Hinton, G. E. Rectified Linear Units Improve Restricted Boltzmann Machines, Proceedings of the 27th International Conference on Machine Learning (ICML’10), Haifa, Israel, June 21–24, Fürnkranz, J.; Joachims, T., Eds.; Omnipress: Madison, WI, 2010; pp 807– 814.
    Google Scholar
    There is no corresponding record for this reference.
  48. 48
    Prechelt, L. Automatic early stopping using cross validation: quantifying the criteria. Neural Networks 1998, 11, 761– 767,  DOI: 10.1016/S0893-6080(98)00010-0
    Google Scholar
    48
    Automatic early stopping using cross validation: quantifying the criteria
    Prechelt Lutz
    Neural networks : the official journal of the International Neural Network Society (1998), 11 (4), 761-767 ISSN:.
    Cross validation can be used to detect when overfitting starts during supervised training of a neural network; training is then stopped before convergence to avoid the overfitting ('early stopping'). The exact criterion used for cross validation based early stopping, however, is chosen in an ad-hoc fashion by most researchers or training is stopped interactively. To aid a more well-founded selection of the stopping criterion, 14 different automatic stopping criteria from three classes were evaluated empirically for their efficiency and effectiveness in 12 different classification and approximation tasks using multi-layer perceptrons with RPROP training. The experiments show that, on average, slower stopping criteria allow for small improvements in generalization (in the order of 4%), but cost about a factor of 4 longer in training time.
  49. 49
    Paszke, A. G. S.; Massa, F.; Lerer, A.; Bradbury, J.; Chanan, G.; Killeen, T. L.; Gimelshein, N.; Antiga, L.; Desmaison, A.; Kopf, A.; Yang, E. D. Z.; Raison, M.; Tejani, A.; Chilamkurthy, S.; Steiner, B. F.; Bai, J.; Chintala, S. An Imperative Style, High-Performance Deep Learning Library, Advances in Neural Information Processing Systems 32 (NeurIPS 2019); Wallach, H.; Larochelle, H.; Beygelzimer, A.; d’Alché-Buc, F.; Fox, E.; Garnett, R., Eds.; Curran Associates: Vancouver, Canada, 2019; pp 8024– 8035.
    Google Scholar
    There is no corresponding record for this reference.
  50. 50
    Hayashi, K.; Uehara, S.; Yamamoto, S.; Cary, D. R.; Nishikawa, J.; Ueda, T.; Ozasa, H.; Mihara, K.; Yoshimura, N.; Kawai, T.; Ono, T.; Yamamoto, S.; Fumoto, M.; Mikamiyama, H. Macrocyclic peptides as a novel class of NNMT inhibitors: A SAR study aimed at inhibitory activity in the cell. ACS Med. Chem. Lett. 2021, 12, 1093– 1101,  DOI: 10.1021/acsmedchemlett.1c00134
    Google Scholar
    50
    Macrocyclic Peptides as a Novel Class of NNMT Inhibitors: A SAR Study Aimed at Inhibitory Activity in the Cell
    Hayashi, Kyohei; Uehara, Shota; Yamamoto, Shiho; Cary, Douglas R.; Nishikawa, Junichi; Ueda, Taichi; Ozasa, Hiroki; Mihara, Kousuke; Yoshimura, Norito; Kawai, Taeko; Ono, Takashi; Yamamoto, Saki; Fumoto, Masataka; Mikamiyama, Hidenori
    ACS Medicinal Chemistry Letters (2021), 12 (7), 1093-1101CODEN: AMCLCT; ISSN:1948-5875. (American Chemical Society)
    Nicotinamide N-methyltransferase (NNMT), which catalyzes the methylation of nicotinamide, is a cytosolic enzyme that has attracted much attention as a therapeutic target for a variety of diseases. However, despite the considerable interest in this target, reports of NNMT inhibitors have still been limited to date. In this work, utilizing in vitro translated macrocyclic peptide libraries, we identified peptide 1 as a novel class of NNMT inhibitors. Further exploration based on the X-ray cocrystal structures of the peptides with NNMT provided a dramatic improvement in inhibitory activity (peptide 23: IC50 = 0.15 nM). Furthermore, by balance of the peptides' lipophilicity and biol. activity, inhibitory activity against NNMT in cell-based assay was successfully achieved (peptide 26: cell-based IC50 = 770 nM). These findings illuminate the potential of cyclic peptides as a relatively new drug discovery modality even for intracellular targets.
  51. 51
    Brousseau, M. E.; Clairmont, K. B.; Spraggon, G.; Flyer, A. N.; Golosov, A. A.; Grosche, P.; Amin, J.; Andre, J.; Burdick, D.; Caplan, S.; Chen, G.; Chopra, R.; Ames, L.; Dubiel, D.; Fan, L.; Gattlen, R.; Kelly-Sullivan, D.; Koch, A. W.; Lewis, I.; Li, J.; Liu, E.; Lubicka, D.; Marzinzik, A.; Nakajima, K.; Nettleton, D.; Ottl, J.; Pan, M.; Patel, T.; Perry, L.; Pickett, S.; Poirier, J.; Reid, P. C.; Pelle, X.; Seepersaud, M.; Subramanian, V.; Vera, V.; Xu, M.; Yang, L.; Yang, Q.; Yu, J.; Zhu, G.; Monovich, L. G. Identification of a PCSK9-LDLR disruptor peptide with in vivo function. Cell Chem. Biol. 2022, 29, 249– 258.e5,  DOI: 10.1016/j.chembiol.2021.08.012
    Google Scholar
    51
    Identification of a PCSK9-LDLR disruptor peptide with in vivo function
    Brousseau, Margaret E.; Clairmont, Kevin B.; Spraggon, Glen; Flyer, Alec N.; Golosov, Andrei A.; Grosche, Philipp; Amin, Jakal; Andre, Jerome; Burdick, Debra; Caplan, Shari; Chen, Guanjing; Chopra, Raj; Ames, Lisa; Dubiel, Diana; Fan, Li; Gattlen, Raphael; Kelly-Sullivan, Dawn; Koch, Alexander W.; Lewis, Ian; Li, Jingzhou; Liu, Eugene; Lubicka, Danuta; Marzinzik, Andreas; Nakajima, Katsumasa; Nettleton, David; Ottl, Johannes; Pan, Meihui; Patel, Tajesh; Perry, Lauren; Pickett, Stephanie; Poirier, Jennifer; Reid, Patrick C.; Pelle, Xavier; Seepersaud, Mohindra; Subramanian, Vanitha; Vera, Victoria; Xu, Mei; Yang, Lihua; Yang, Qing; Yu, Jinghua; Zhu, Guoming; Monovich, Lauren G.
    Cell Chemical Biology (2022), 29 (2), 249-258.e5CODEN: CCBEBM; ISSN:2451-9448. (Cell Press)
    Proprotein convertase subtilisin/kexin type 9 (PCSK9) regulates plasma low-d. lipoprotein cholesterol (LDL-C) levels by promoting hepatic LDL receptor (LDLR) degrdn. Therapeutic antibodies that disrupt PCSK9-LDLR binding reduce LDL-C concns. and cardiovascular disease risk. The epidermal growth factor precursor homol. domain A (EGF-A) of the LDLR serves as a primary contact with PCSK9 via a flat interface, presenting a challenge for identifying small mol. PCSK9-LDLR disruptors. We employ an affinity-based screen of 1013in vitro-translated macrocyclic peptides to identify high-affinity PCSK9 ligands that utilize a unique, induced-fit pocket and partially disrupt the PCSK9-LDLR interaction. Structure-based design led to mols. with enhanced function and pharmacokinetic properties (e.g., 13PCSK9i). In mice, 13PCSK9i reduces plasma cholesterol levels and increases hepatic LDLR d. in a dose-dependent manner. 13PCSK9i functions by a unique, allosteric mechanism and is the smallest mol. identified to date with in vivo PCSK9-LDLR disruptor function.
  52. 52
    Yoshida, S.; Uehara, S.; Kondo, N.; Takahashi, Y.; Yamamoto, S.; Kameda, A.; Kawagoe, S.; Inoue, N.; Yamada, M.; Yoshimura, N.; Tachibana, Y. Peptide-to-small molecule: a pharmacophore-guided small molecule lead generation strategy from high-affinity macrocyclic peptides. J. Med. Chem. 2022, 65, 10655– 10673,  DOI: 10.1021/acs.jmedchem.2c00919
    Google Scholar
    There is no corresponding record for this reference.
  53. 53
    Banerjee, R.; Basu, G.; Chène, P.; Roy, S. Aib-based peptide backbone as scaffolds for helical peptide mimics. J. Pept. Res. 2002, 60, 88– 94,  DOI: 10.1034/j.1399-3011.2002.201005.x
    Google Scholar
    53
    Aib-based peptide backbone as scaffolds for helical peptide mimics
    Banerjee, R.; Basu, G.; Chene, P.; Roy, S.
    Journal of Peptide Research (2002), 60 (2), 88-94CODEN: JPERFA; ISSN:1397-002X. (Blackwell Munksgaard)
    Helical peptides that can intervene and disrupt therapeutically important protein-protein interactions are attractive drug targets. In order to develop a general strategy for developing such helical peptide mimics, the authors have studied the effect of incorporating α-amino isobutyric acid (Aib), an amino acid with strong preference for helical backbone, as the sole helix promoter in designed peptides. Specifically, the focus is on the hdm2-p53 interaction, which is central to development of many types of cancer. The peptide corresponding to the hdm2 interacting part of p53, helical in bound state but devoid of structure in soln., served as the starting point for peptide design that involved replacement of noninteracting residues by Aib. Incorporation of Aib, while preserving the interacting residues, led to significant increase in helical structure, particularly at the C-terminal region as judged by NMR and CD. The interaction with hdm2 was also found to be enhanced. Most interestingly, trypsin cleavage was found to be retarded by several orders of magnitude. It is concluded that incorporation of Aib is a feasible strategy to create peptide helical mimics with enhanced receptor binding and lower protease cleavage rate.
  54. 54
    Karle, I. L. Controls exerted by the Aib residue: helix formation and helix reversal. Pept. Sci. 2001, 60, 351– 365,  DOI: 10.1002/1097-0282(2001)60:5<351::AID-BIP10174>3.0.CO;2-U
    Google Scholar
    There is no corresponding record for this reference.
  55. 55
    Lundberg, S. M.; Lee, S.-I. A Unified Approach to Interpreting Model Predictions; Proceedings of the 31st International Conference on Neural Information Processing Systems (NIPS’17), Long Beach, CA, USA, December 4–9, 2017; Von Luxburg, U.; Guyon, I.; Bengio, S.; Wallach, H.; Ferugs, R., Eds.; Curan Associates: Redhook, NY, 2017; pp 4768– 4777.
    Google Scholar
    There is no corresponding record for this reference.

Cited By

Click to copy section linkSection link copied!
Citation Statements
Explore this article's citation statements on scite.ai
powered by  Scite

This article is cited by 1 publications.

  1. Yanpeng Fang, Duoyang Fan, Bin Feng, Yingli Zhu, Ruyan Xie, Xiaorong Tan, Qianhui Liu, Jie Dong, Wenbin Zeng. Harnessing advanced computational approaches to design novel antimicrobial peptides against intracellular bacterial infections. Bioactive Materials 2025, 50 , 510-524. https://doi.org/10.1016/j.bioactmat.2025.04.016
Go to
Get e-Alerts
Get e-Alerts

Journal of Chemical Information and Modeling

Cite this: J. Chem. Inf. Model. 2025, 65, 4, 1837–1849
Click to copy citationCitation copied!
https://doi.org/10.1021/acs.jcim.4c02102
Published February 2, 2025

Copyright © 2025 The Authors. Published by American Chemical Society. This publication is licensed under

CC-BY-NC-ND 4.0 .

Article Views

1568

Altmetric

-

Citations

1
Learn about these metrics

Article Views are the COUNTER-compliant sum of full text article downloads since November 2008 (both PDF and HTML) across all institutions and individuals. These metrics are regularly updated to reflect usage leading up to the last few days.

Citations are the number of other articles citing this article, calculated by Crossref and updated daily. Find more information about Crossref citation counts.

The Altmetric Attention Score is a quantitative measure of the attention that a research article has received online. Clicking on the donut icon will load a page at altmetric.com with additional details about the score and the social media presence for the given article. Find more information on the Altmetric Attention Score and how the score is calculated.

  • Abstract

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 1

    Figure 1. Overview of custom features for amino acids. (A) Position-aware side chain (PASC) fingerprints are based on a “heavy-atom walk” along the amino acid side chain. Starting at the Cα atom, a Morgan fingerprint with a radius of 1 is generated (red circle). Morgan fingerprints with a radius of 1 centered at the Cβ atom (orange circle), Cγ atom (yellow circle), etc., are similarly generated to provide the PASC features. (B) MD simulations of amino acid dipeptides are used to generate MD-derived features. For the backbone (BB) features, the (ϕ, ψ) distribution is calculated from the dipeptide simulation and binned in a 2D grid. Then, the resulting 2D probability density is flattened into a 1D vector. For the voxel (VOX) features, the simulation frames are aligned to reference coordinates for C, Cα, and N, where the Cα atom is at the origin, the N atom is at (1.449, 0.000, 0.000), and the C atom is at (−0.523, 1.429, 0.000) (in Å). Then, frame-averaged molecular electrostatic potential is calculated on a 3D voxel and flattened into a 1D vector. See the Methods section for more details.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 2

    Figure 2. Amino acids included in the training, validation, and test data sets for the StrEAMM models. The training and validation data sets include cyclic peptide sequences from a 15-amino-acid (15-aa) library (black box). The test data sets include sequences containing amino acids in the same 15-aa library, the 37-aa library (blue box), or the 50-aa library (purple box). *For brevity, only the L-forms of chiral amino acids are depicted, but their mirror images are also included in the library.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 3

    Figure 3. Performance of different amino acid features on different cyclic pentapeptide test data sets. (A) The models are trained using 3-fold cross-validation. The table reports the average R2 (coefficient of determination) and standard deviation across the 3 folds. (B) The performance for one out of the three models from the 3-fold cross-validation is plotted. The predicted population of each structure in the cyclic peptides’ structural ensemble is compared to its populations observed in MD simulations. For clarity, only structures in the ensembles for all cyclic peptides in the test data sets with either a predicted or observed (in MD) percent population of >1% are plotted.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 4

    Figure 4. Performance of different amino acid features on different cyclic hexapeptide test data sets. (A) The models are trained using 3-fold cross-validation. The table reports the average R2 (coefficient of determination) and standard deviation across the 3 folds. (B) The performance for one out of the three models from the 3-fold cross-validation is plotted. The predicted population of each structure in the cyclic peptides’ structural ensemble is compared to its populations observed in MD simulations. For clarity, only structures in the ensembles for all cyclic peptides in the test data sets with either a predicted or observed (in MD) percent population of >1% are plotted.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 5

    Figure 5. Performance of different combinations of amino acid features on cyclic pentapeptide (top) and hexapeptide (bottom) test data sets containing 15 AAs (left), 37 AAs (middle), and 50 AAs (right). The models are trained using 3-fold cross-validation, and the average R2 and standard deviation are reported. The models using a single type of feature (e.g., “BB only”) are represented on the diagonals. The best-performing models for the 37 AA and 50 AA test data sets, based on the average R2, are boxed with bold black outlines.

    High Resolution Image
    Download MS PowerPoint Slide

  • References

    This article references 55 other publications.

    1. 1
      Wang, L.; Wang, N.; Zhang, W.; Cheng, X.; Yan, Z.; Shao, G.; Wang, X.; Wang, R.; Fu, C. Therapeutic peptides: current applications and future directions. Signal Transduction Targeted Ther. 2022, 7, 48  DOI: 10.1038/s41392-022-00904-4
      1
      Therapeutic peptides: current applications and future directions
      Wang, Lei; Wang, Nanxi; Zhang, Wenping; Cheng, Xurui; Yan, Zhibin; Shao, Gang; Wang, Xi; Wang, Rui; Fu, Caiyun
      Signal Transduction and Targeted Therapy (2022), 7 (1), 48CODEN: STTTCB; ISSN:2059-3635. (Nature Portfolio)
      A review. Peptide drug development has made great progress in the last decade thanks to new prodn., modification, and analytic technologies. Peptides have been produced and modified using both chem. and biol. methods, together with novel design and delivery strategies, which have helped to overcome the inherent drawbacks of peptides and have allowed the continued advancement of this field. A wide variety of natural and modified peptides have been obtained and studied, covering multiple therapeutic areas. This review summarizes the efforts and achievements in peptide drug discovery, prodn., and modification, and their current applications. We also discuss the value and challenges assocd. with future developments in therapeutic peptides.
    2. 2
      Liu, K.; Li, M.; Li, Y.; Li, Y.; Chen, Z.; Tang, Y.; Yang, M.; Deng, G.; Liu, H. A review of the clinical efficacy of FDA-approved antibody–drug conjugates in human cancers. Mol. Cancer 2024, 23, 62  DOI: 10.1186/s12943-024-01963-7
      There is no corresponding record for this reference.
    3. 3
      Jumper, J.; Evans, R.; Pritzel, A.; Green, T.; Figurnov, M.; Ronneberger, O.; Tunyasuvunakool, K.; Bates, R.; Žídek, A.; Potapenko, A.; Bridgland, A.; Meyer, C.; Kohl, S. A. A.; Ballard, A. J.; Cowie, A.; Romera-Paredes, B.; Nikolov, S.; Jain, R.; Adler, J.; Back, T.; Petersen, S.; Reiman, D.; Clancy, E.; Zielinski, M.; Steinegger, M.; Pacholska, M.; Berghammer, T.; Bodenstein, S.; Silver, D.; Vinyals, O.; Senior, A. W.; Kavukcuoglu, K.; Kohli, P.; Hassabis, D. Highly accurate protein structure prediction with AlphaFold. Nature 2021, 596, 583– 589,  DOI: 10.1038/s41586-021-03819-2
      3
      Highly accurate protein structure prediction with AlphaFold
      Jumper, John; Evans, Richard; Pritzel, Alexander; Green, Tim; Figurnov, Michael; Ronneberger, Olaf; Tunyasuvunakool, Kathryn; Bates, Russ; Zidek, Augustin; Potapenko, Anna; Bridgland, Alex; Meyer, Clemens; Kohl, Simon A. A.; Ballard, Andrew J.; Cowie, Andrew; Romera-Paredes, Bernardino; Nikolov, Stanislav; Jain, Rishub; Adler, Jonas; Back, Trevor; Petersen, Stig; Reiman, David; Clancy, Ellen; Zielinski, Michal; Steinegger, Martin; Pacholska, Michalina; Berghammer, Tamas; Bodenstein, Sebastian; Silver, David; Vinyals, Oriol; Senior, Andrew W.; Kavukcuoglu, Koray; Kohli, Pushmeet; Hassabis, Demis
      Nature (London, United Kingdom) (2021), 596 (7873), 583-589CODEN: NATUAS; ISSN:0028-0836. (Nature Portfolio)
      Proteins are essential to life, and understanding their structure can facilitate a mechanistic understanding of their function. Through an enormous exptl. effort, the structures of around 100,000 unique proteins have been detd., but this represents a small fraction of the billions of known protein sequences. Structural coverage is bottlenecked by the months to years of painstaking effort required to det. a single protein structure. Accurate computational approaches are needed to address this gap and to enable large-scale structural bioinformatics. Predicting the three-dimensional structure that a protein will adopt based solely on its amino acid sequence-the structure prediction component of the 'protein folding problem'-has been an important open research problem for more than 50 years. Despite recent progress, existing methods fall far short of at. accuracy, esp. when no homologous structure is available. Here we provide the first computational method that can regularly predict protein structures with at. accuracy even in cases in which no similar structure is known. We validated an entirely redesigned version of our neural network-based model, AlphaFold, in the challenging 14th Crit. Assessment of protein Structure Prediction (CASP14), demonstrating accuracy competitive with exptl. structures in a majority of cases and greatly outperforming other methods. Underpinning the latest version of AlphaFold is a novel machine learning approach that incorporates phys. and biol. knowledge about protein structure, leveraging multi-sequence alignments, into the design of the deep learning algorithm.
    4. 4
      Baek, M.; DiMaio, F.; Anishchenko, I.; Dauparas, J.; Ovchinnikov, S.; Lee, G. R.; Wang, J.; Cong, Q.; Kinch, L. N.; Schaeffer, R. D.; Millán, C.; Park, H.; Adams, C.; Glassman, C. R.; DeGiovanni, A.; Pereira, J. H.; Rodrigues, A. V.; van Dijk, A. A.; Ebrecht, A. C.; Opperman, D. J.; Sagmeister, T.; Buhlheller, C.; Pavkov-Keller, T.; Rathinaswamy, M. K.; Dalwadi, U.; Yip, C. K.; Burke, J. E.; Garcia, K. C.; Grishin, N. V.; Adams, P. D.; Read, R. J.; Baker, D. Accurate prediction of protein structures and interactions using a three-track neural network. Science 2021, 373, 871– 876,  DOI: 10.1126/science.abj8754
      4
      Accurate prediction of protein structures and interactions using a three-track neural network
      Baek, Minkyung; DiMaio, Frank; Anishchenko, Ivan; Dauparas, Justas; Ovchinnikov, Sergey; Lee, Gyu Rie; Wang, Jue; Cong, Qian; Kinch, Lisa N.; Schaeffer, R. Dustin; Millan, Claudia; Park, Hahnbeom; Adams, Carson; Glassman, Caleb R.; DeGiovanni, Andy; Pereira, Jose H.; Rodrigues, Andria V.; van Dijk, Alberdina A.; Ebrecht, Ana C.; Opperman, Diederik J.; Sagmeister, Theo; Buhlheller, Christoph; Pavkov-Keller, Tea; Rathinaswamy, Manoj K.; Dalwadi, Udit; Yip, Calvin K.; Burke, John E.; Garcia, K. Christopher; Grishin, Nick V.; Adams, Paul D.; Read, Randy J.; Baker, David
      Science (Washington, DC, United States) (2021), 373 (6557), 871-876CODEN: SCIEAS; ISSN:1095-9203. (American Association for the Advancement of Science)
      DeepMind presented notably accurate predictions at the recent 14th Crit. Assessment of Structure Prediction (CASP14) conference. We explored network architectures that incorporate related ideas and obtained the best performance with a three-track network in which information at the one-dimensional (1D) sequence level, the 2D distance map level, and the 3D coordinate level is successively transformed and integrated. The three-track network produces structure predictions with accuracies approaching those of DeepMind in CASP14, enables the rapid soln. of challenging x-ray crystallog. and cryo-electron microscopy structure modeling problems, and provides insights into the functions of proteins of currently unknown structure. The network also enables rapid generation of accurate protein-protein complex models from sequence information alone, short-circuiting traditional approaches that require modeling of individual subunits followed by docking. We make the method available to the scientific community to speed biol. research.
    5. 5
      Miao, J.; Descoteaux, M. L.; Lin, Y.-S. Structure prediction of cyclic peptides by molecular dynamics + machine learning. Chem. Sci. 2021, 12, 14927– 14936,  DOI: 10.1039/D1SC05562C
      5
      Structure prediction of cyclic peptides by molecular dynamics + machine learning
      Miao, Jiayuan; Descoteaux, Marc L.; Lin, Yu-Shan
      Chemical Science (2021), 12 (44), 14927-14936CODEN: CSHCCN; ISSN:2041-6520. (Royal Society of Chemistry)
      Recent computational methods have made strides in discovering well-structured cyclic peptides that preferentially populate a single conformation. However, many successful cyclic-peptide therapeutics adopt multiple conformations in soln. In fact, the chameleonic properties of some cyclic peptides are likely responsible for their high cell membrane permeability. Thus, we require the ability to predict complete structural ensembles for cyclic peptides, including the majority of cyclic peptides that have broad structural ensembles, to significantly improve our ability to rationally design cyclic-peptide therapeutics. Here, we introduce the idea of using mol. dynamics simulation results to train machine learning models to enable efficient structure prediction for cyclic peptides. Using mol. dynamics simulation results for several hundred cyclic pentapeptides as the training datasets, we developed machine-learning models that can provide mol. dynamics simulation-quality predictions of structural ensembles for all the hundreds of thousands of sequences in the entire sequence space. The prediction for each individual cyclic peptide can be made using less than 1 s of computation time. Even for the most challenging classes of poorly structured cyclic peptides with broad conformational ensembles, our predictions were similar to those one would normally obtain only after running multiple days of explicit-solvent mol. dynamics simulations. The resulting method, termed StrEAMM (Structural Ensembles Achieved by Mol. Dynamics and Machine Learning), is the first technique capable of efficiently predicting complete structural ensembles of cyclic peptides without relying on addnl. mol. dynamics simulations, constituting a seven-order-of-magnitude improvement in speed while retaining the same accuracy as explicit-solvent simulations.
    6. 6
      Hui, T.; Descoteaux, M. L.; Miao, J.; Lin, Y.-S. Training neural network models using molecular dynamics simulation results to efficiently predict cyclic hexapeptide structural ensembles. J. Chem. Theory Comput. 2023, 19, 4757– 4769,  DOI: 10.1021/acs.jctc.3c00154
      There is no corresponding record for this reference.
    7. 7
      Wan, F.; Kontogiorgos-Heintz, D.; de la Fuente-Nunez, C. Deep generative models for peptide design. Digital Discovery 2022, 1, 195– 208,  DOI: 10.1039/D1DD00024A
      There is no corresponding record for this reference.
    8. 8
      Ferguson, A. L.; Ranganathan, R. 100th anniversary of macromolecular science viewpoint: data-driven protein design. ACS Macro Lett. 2021, 10, 327– 340,  DOI: 10.1021/acsmacrolett.0c00885
      8
      100Th Anniversary of Macromolecular Science Viewpoint: Data-Driven Protein Design
      Ferguson, Andrew L.; Ranganathan, Rama
      ACS Macro Letters (2021), 10 (3), 327-340CODEN: AMLCCD; ISSN:2161-1653. (American Chemical Society)
      A review. The design of synthetic proteins with the desired function is a long-standing goal in biomol. science, with broad applications in biochem. engineering, agriculture, medicine, and public health. Rational de novo design and exptl. directed evolution have achieved remarkable successes but are challenged by the requirement to find functional "needles" in the vast "haystack" of protein sequence space. Data-driven models for fitness landscapes provide a predictive map between protein sequence and function and can prospectively identify functional candidates for exptl. testing to greatly improve the efficiency of this search. This Viewpoint reviews the applications of machine learning and, in particular, deep learning as part of data-driven protein engineering platforms. We highlight recent successes, review promising computational methodologies, and provide an outlook on future challenges and opportunities. The article is written for a broad audience comprising both polymer and protein scientists and computer and data scientists interested in an up-to-date review of recent innovations and opportunities in this rapidly evolving field.
    9. 9
      Strokach, A.; Kim, P. M. Deep generative modeling for protein design. Curr. Opin. Struct. Biol. 2022, 72, 226– 236,  DOI: 10.1016/j.sbi.2021.11.008
      9
      Deep generative modeling for protein design
      Strokach, Alexey; Kim, Philip M.
      Current Opinion in Structural Biology (2022), 72 (), 226-236CODEN: COSBEF; ISSN:0959-440X. (Elsevier Ltd.)
      A review. Deep learning approaches have produced substantial breakthroughs in fields such as image classification and natural language processing and are making rapid inroads in the area of protein design. Many generative models of proteins have been developed that encompass all known protein sequences, model specific protein families, or extrapolate the dynamics of individual proteins. Those generative models can learn protein representations that are often more informative of protein structure and function than hand-engineered features. Furthermore, they can be used to quickly propose millions of novel proteins that resemble the native counterparts in terms of expression level, stability, or other attributes. The protein design process can further be guided by discriminative oracles to select candidates with the highest probability of having the desired properties. In this review, we discuss five classes of generative models that have been most successful at modeling proteins and provide a framework for model guided protein design.
    10. 10
      Chandra, A.; Tünnermann, L.; Löfstedt, T.; Gratz, R. Transformer-based deep learning for predicting protein properties in the life sciences. eLife 2023, 12, e82819  DOI: 10.7554/eLife.82819
      There is no corresponding record for this reference.
    11. 11
      Oliva, R.; Chino, M.; Pane, K.; Pistorio, V.; De Santis, A.; Pizzo, E.; D’Errico, G.; Pavone, V.; Lombardi, A.; Del Vecchio, P.; Notomista, E.; Nastri, F.; Petraccone, L. Exploring the role of unnatural amino acids in antimicrobial peptides. Sci. Rep. 2018, 8, 8888  DOI: 10.1038/s41598-018-27231-5
      11
      Exploring the role of unnatural amino acids in antimicrobial peptides
      Oliva Rosario; Chino Marco; De Santis Augusta; D'Errico Gerardino; Pavone Vincenzo; Lombardi Angela; Del Vecchio Pompea; Nastri Flavia; Petraccone Luigi; Pane Katia; Pizzo Elio; Notomista Eugenio; Pistorio Valeria
      Scientific reports (2018), 8 (1), 8888 ISSN:.
      Cationic antimicrobial peptides (CAMPs) are a promising alternative to treat multidrug-resistant bacteria, which have developed resistance to all the commonly used antimicrobial, and therefore represent a serious threat to human health. One of the major drawbacks of CAMPs is their sensitivity to proteases, which drastically limits their half-life. Here we describe the design and synthesis of three nine-residue CAMPs, which showed high stability in serum and broad spectrum antimicrobial activity. As for all peptides a very low selectivity between bacterial and eukaryotic cells was observed, we performed a detailed biophysical characterization of the interaction of one of these peptides with liposomes mimicking bacterial and eukaryotic membranes. Our results show a surface binding on the DPPC/DPPG vesicles, coupled with lipid domain formation, and, above a threshold concentration, a deep insertion into the bilayer hydrophobic core. On the contrary, mainly surface binding of the peptide on the DPPC bilayer was observed. These observed differences in the peptide interaction with the two model membranes suggest a divergence in the mechanisms responsible for the antimicrobial activity and for the observed high toxicity toward mammalian cell lines. These results could represent an important contribution to unravel some open and unresolved issues in the development of synthetic CAMPs.
    12. 12
      Lu, J.; Xu, H.; Xia, J.; Ma, J.; Xu, J.; Li, Y.; Feng, J. D- and unnatural amino acid substituted antimicrobial peptides with improved proteolytic resistance and their proteolytic degradation characteristics. Front. Microbiol. 2020, 11, 563030  DOI: 10.3389/fmicb.2020.563030
      12
      D- and Unnatural Amino Acid Substituted Antimicrobial Peptides With Improved Proteolytic Resistance and Their Proteolytic Degradation Characteristics
      Lu Jianguang; Xu Hongjiang; Xia Jianghua; Li Yanan; Feng Jun; Lu Jianguang; Ma Jie; Xu Jun; Feng Jun; Xu Hongjiang; Li Yanan
      Frontiers in microbiology (2020), 11 (), 563030 ISSN:1664-302X.
      The transition of antimicrobial peptides (AMPs) from the laboratory to market has been severely hindered by their instability toward proteases in biological systems. In the present study, we synthesized derivatives of the cationic AMP Pep05 (KRLFKKLLKYLRKF) by substituting L-amino acid residues with D- and unnatural amino acids, such as D-lysine, D-arginine, L-2,4-diaminobutanoic acid (Dab), L-2,3-diaminopropionic acid (Dap), L-homoarginine, 4-aminobutanoic acid (Aib), and L-thienylalanine, and evaluated their antimicrobial activities, toxicities, and stabilities toward trypsin, plasma proteases, and secreted bacterial proteases. In addition to measuring changes in the concentration of the intact peptides, LC-MS was used to identify the degradation products of the modified AMPs in the presence of trypsin and plasma proteases to determine degradation pathways and examine whether the amino acid substitutions afforded improved proteolytic resistance. The results revealed that both D- and unnatural amino acids enhanced the stabilities of the peptides toward proteases. The derivative DP06, in which all of the L-lysine and L-arginine residues were replaced by D-amino acids, displayed remarkable stability and mild toxicity in vitro but only slight activity and severe toxicity in vivo, indicating a significant difference between the in vivo and in vitro results. Unexpectedly, we found that the incorporation of a single Aib residue at the N-terminus of compound UP09 afforded remarkably enhanced plasma stability and improved activity in vivo. Hence, this derivative may represent a candidate AMP for further optimization, providing a new strategy for the design of novel AMPs with improved bioavailability.
    13. 13
      Taechalertpaisarn, J.; Ono, S.; Okada, O.; Johnstone, T. C.; Lokey, R. S. A new amino acid for improving permeability and solubility in macrocyclic peptides through side chain-to-backbone hydrogen bonding. J. Med. Chem. 2022, 65, 5072– 5084,  DOI: 10.1021/acs.jmedchem.2c00010
      13
      A new amino acid for improving permeability and solubility in macrocyclic peptides through side chain-to-backbone hydrogen bonding
      Taechalertpaisarn, Jaru; Ono, Satoshi; Okada, Okimasa; Johnstone, Timothy C.; Lokey, R. Scott
      Journal of Medicinal Chemistry (2022), 65 (6), 5072-5084CODEN: JMCMAR; ISSN:0022-2623. (American Chemical Society)
      Despite the notoriously poor membrane permeability of peptides, many cyclic peptide natural products show high passive membrane permeability and potently inhibit a variety of "undruggable" intracellular targets. A major impediment to the design of cyclic peptides with good permeability is the high desolvation energy assocd. with the peptide backbone amide NH groups. While several strategies have been proposed to mitigate this deleterious effect, only few studies have used polar side chains to sequester backbone NH groups. We investigated the ability of N,N-pyrrolidinylglutamine (Pye), whose side chain contains a powerful hydrogen-bond-accepting C:O amide group but no hydrogen-bond donors, to sequester exposed backbone NH groups in a series of cyclic hexapeptide diastereomers. Analyses revealed that specific Leu-to-Pye substitutions conferred dramatic improvements in aq. soly. and permeability in a scaffold- and position-dependent manner. Therefore, this approach offers a complementary tool for improving membrane permeability and soly. in cyclic peptides.
    14. 14
      Geurink, P. P.; van der Linden, W. A.; Mirabella, A. C.; Gallastegui, N.; de Bruin, G.; Blom, A. E.; Voges, M. J.; Mock, E. D.; Florea, B. I.; van der Marel, G. A.; Driessen, C.; van der Stelt, M.; Groll, M.; Overkleeft, H. S.; Kisselev, A. F. Incorporation of non-natural amino acids improves cell permeability and potency of specific inhibitors of proteasome trypsin-like sites. J. Med. Chem. 2013, 56, 1262– 1275,  DOI: 10.1021/jm3016987
      14
      Incorporation of non-natural amino acids improves cell permeability and potency of specific inhibitors of proteasome trypsin-like sites
      Geurink, Paul P.; van der Linden, Wouter A.; Mirabella, Anne C.; Gallastegui, Nerea; de Bruin, Gerjan; Blom, Annet E. M.; Voges, Mathias J.; Mock, Elliot D.; Florea, Bogdan I.; van der Marel, Gijs A.; Driessen, Christoph; van der Stelt, Mario; Groll, Michael; Overkleeft, Herman S.; Kisselev, Alexei F.
      Journal of Medicinal Chemistry (2013), 56 (3), 1262-1275CODEN: JMCMAR; ISSN:0022-2623. (American Chemical Society)
      Proteasomes degrade the majority of proteins in mammalian cells by a concerted action of three distinct pairs of active sites. The chymotrypsin-like sites are targets of antimyeloma agents bortezomib and carfilzomib. Inhibitors of the trypsin-like site sensitize multiple myeloma cells to these agents. Here we describe systematic effort to develop inhibitors with improved potency and cell permeability, yielding azido-Phe-Leu-Leu-4-aminomethyl-Phe-Me vinyl sulfone (I), LU-102, and a fluorescent activity-based probe for this site. X-ray structures of I and related inhibitors complexed with yeast proteasomes revealed the structural basis for specificity. Nontoxic to myeloma cells when used as a single agent, I sensitized them to bortezomib and carfilzomib. This sensitizing effect was much stronger than the synergistic effects of histone acetylase inhibitors or additive effects of doxorubicin and dexamethasone, raising the possibility that combinations of inhibitors of the trypsin-like site with bortezomib or carfilzomib would have stronger antineoplastic activity than combinations currently used clin.
    15. 15
      Weininger, D. SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules. J. Chem. Inf. Comput. Sci. 1988, 28, 31– 36,  DOI: 10.1021/ci00057a005
      15
      SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules
      Weininger, David
      Journal of Chemical Information and Computer Sciences (1988), 28 (1), 31-6CODEN: JCISD8; ISSN:0095-2338.
      The SMILES (simplified mol. input line entry system) chem. notation system is described for information processing. The system is based on principles of mol. graph theory and it allows structure specification by use of a very small and natural grammar well suited for high-speed machine processing. The system is easy to use, has high machine compatibility, and allows many computer applications, including notation generation, const. speed database retrieval, substructure searching, and property prediction models.
    16. 16
      Siani, M. A.; Weininger, D.; Blaney, J. M. CHUCKLES: A method for representing and searching peptide and peptoid sequences on both monomer and atomic levels. J. Chem. Inf. Comput. Sci. 1994, 34, 588– 593,  DOI: 10.1021/ci00019a017
      16
      CHUCKLES: A method for representing and searching peptide and peptoid sequences on both monomer and atomic levels
      Siani, Michael A.; Weininger, David; Blaney, Jeffrey M.
      Journal of Chemical Information and Computer Sciences (1994), 34 (3), 588-93CODEN: JCISD8; ISSN:0095-2338.
      Dual representation of peptide and non-peptide structures in a chem. database as at.-level mol. graphs and sequence strings permits chem. substructure and similarity searches as well as sequence-based substring and regular expression searches. CHUCKLES interconverts monomer-based sequences with SMILES, which represent at.-level mol. graphs. Forward-translation maps peptide or other sequences into SMILES. Back-translation exts. monomer sequences from SMILES. This approach permits a generalized representation of monomers allowing user specification of any monomer. CHUCKLES allows mixing of atoms with user-defined monomer names; i.e., monomer representation is consistent with SMILES notation. In addn., oligomer branching and cyclization are handled.
    17. 17
      Siani, M. A.; Weininger, D.; James, C. A.; Blaney, J. M. CHORTLES: A method for representing oligomeric and template-based mixtures. J. Chem. Inf. Comput. Sci. 1995, 35, 1026– 1033,  DOI: 10.1021/ci00028a012
      17
      CHORTLES: A Method for Representing Oligomeric and Template-Based Mixtures
      Siani, Michael A.; Weininger, David; James, Craig A.; Blaney, Jeffrey M.
      Journal of Chemical Information and Computer Sciences (1995), 35 (6), 1026-33CODEN: JCISD8; ISSN:0095-2338. (American Chemical Society)
      Screening mixts. of synthetic oligomers or fixed templates (e.g., rings) with varying substituents is increasingly the focus of drug discovery programs. CHORTLES is designed and implemented to facilitate representation, storage, and searching of oligomeric and template-based mixts. of any size. Building upon the CHUCKLES method of representing oligomers as both monomer-based sequences and all-atom structures, CHORTLES compactly represents a mixt. without explicitly enumerating individual mols. This method lends itself to a hierarchy relating mixts. to submixts. and individual compds., as one finds when deconvoluting mixts. in drug lead discovery programs. In addn., we describe two methods of searching mixts. at the monomer level. We also present a simple pictorial representation for describing all components in a mixt., which becomes essential as the list of monomer names is expanded beyond common names (e.g., amino acids).
    18. 18
      Jensen, J. H.; Hoeg-Jensen, T.; Padkjær, S. B. Building a biochemformatics database. J. Chem. Inf. Model. 2008, 48, 2404– 2413,  DOI: 10.1021/ci800128b
      18
      Building a BioChemformatics Database
      Jensen, Jan H.; Hoeg-Jensen, Thomas; Padkjaer, Soeren B.
      Journal of Chemical Information and Modeling (2008), 48 (12), 2404-2413CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      The structural registration of chem. modified macromols. is vital for the development of biopharmaceuticals. However, registration and search of such complex mols. has so far posed formidable challenges performance-wise, since today's chem.-oriented databases do not scale well to macromols. As a practical consequence, macromols. tend to be stored in protein databases with a focus on protein sequence only, and salient chem. details are therefore lost. This article describes protein format extensions and the use of pseudoatoms for representing natural amino acids in chem. structures to allow high-performance registration and retrieval of large macromols. The representations include exact chem. modifications and enable lossless conversion between chem. and sequence formats. Registration is done in parallel in both sequence and chem. formats, and users can register and retrieve mols. in either format as they choose, resulting in what we call a BioChemformatics database. Having both sequence and chem. formats available on-demand allows for the construction of protein SAR tables with mixed sequence and chem. information. Likewise, searching may combine sequence and chem. terms and be performed in std. vendor applications like MDL's ISIS/Base or inhouse applications using std. SQL queries.
    19. 19
      Lin, T.-S.; Coley, C. W.; Mochigase, H.; Beech, H. K.; Wang, W.; Wang, Z.; Woods, E.; Craig, S. L.; Johnson, J. A.; Kalow, J. A.; Jensen, K. F.; Olsen, B. D. BigSMILES: A structurally-based line notation for describing macromolecules. ACS Cent. Sci. 2019, 5, 1523– 1531,  DOI: 10.1021/acscentsci.9b00476
      19
      BigSMILES: A Structurally-Based Line Notation for Describing Macromolecules
      Lin, Tzyy-Shyang; Coley, Connor W.; Mochigase, Hidenobu; Beech, Haley K.; Wang, Wencong; Wang, Zi; Woods, Eliot; Craig, Stephen L.; Johnson, Jeremiah A.; Kalow, Julia A.; Jensen, Klavs F.; Olsen, Bradley D.
      ACS Central Science (2019), 5 (9), 1523-1531CODEN: ACSCII; ISSN:2374-7951. (American Chemical Society)
      Having a compact yet robust structurally based identifier or representation system is a key enabling factor for efficient sharing and dissemination of research results within the chem. community, and such systems lay down the essential foundations for future informatics and data-driven research. While substantial advances have been made for small mols., the polymer community has struggled in coming up with an efficient representation system. This is because, unlike other disciplines in chem., the basic premise that each distinct chem. species corresponds to a well-defined chem. structure does not hold for polymers. Polymers are intrinsically stochastic mols. that are often ensembles with a distribution of chem. structures. This difficulty limits the applicability of all deterministic representations developed for small mols. In this work, a new representation system that is capable of handling the stochastic nature of polymers is proposed. The new system is based on the popular "simplified mol.-input line-entry system" (SMILES), and it aims to provide representations that can be used as indexing identifiers for entries in polymer databases. As a pilot test, the entries of the std. data set of the glass transition temp. of linear polymers (Bicerano, 2002) were converted into the new BigSMILES language. Furthermore, it is hoped that the proposed system will provide a more effective language for communication within the polymer community and increase cohesion between the researchers within the community.
    20. 20
      Zhang, T.; Li, H.; Xi, H.; Stanton, R. V.; Rotstein, S. H. HELM: A hierarchical notation language for complex biomolecule structure representation. J. Chem. Inf. Model. 2012, 52, 2796– 2806,  DOI: 10.1021/ci3001925
      20
      HELM: A Hierarchical Notation Language for Complex Biomolecule Structure Representation
      Zhang, Tianhong; Li, Hongli; Xi, Hualin; Stanton, Robert V.; Rotstein, Sergio H.
      Journal of Chemical Information and Modeling (2012), 52 (10), 2796-2806CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      When biol. macromols. are used as therapeutic agents, it is often necessary to introduce non-natural chem. modifications to improve their pharmaceutical properties. The final products are complex structures where entities such as proteins, peptides, oligonucleotides, and small mol. drugs may be covalently linked to each other, or may include chem. modified biol. moieties. An accurate in silico representation of these complex structures is essential, as it forms the basis for their electronic registration, storage, anal., and visualization. The size of these mols. (henceforth referred to as "biomols.") often makes them too unwieldy and impractical to represent at the at. level, while the presence of non-natural chem. modifications makes it impossible to represent them by sequence alone. Here we describe the Hierarchical Editing Language for Macromols. ("HELM") and demonstrate its utility in the representation of structures such as antisense oligonucleotides, short interference RNAs, peptides, proteins, and antibody drug conjugates.
    21. 21
      David, L.; Thakkar, A.; Mercado, R.; Engkvist, O. Molecular representations in AI-driven drug discovery: a review and practical guide. J. Cheminf. 2020, 12, 56  DOI: 10.1186/s13321-020-00460-5
      21
      Molecular representations in AI-driven drug discovery: a review and practical guide
      David, Laurianne; Thakkar, Amol; Mercado, Rocio; Engkvist, Ola
      Journal of Cheminformatics (2020), 12 (1), 56CODEN: JCOHB3; ISSN:1758-2946. (SpringerOpen)
      A review. Abstr.: The technol. advances of the past century, marked by the computer revolution and the advent of high-throughput screening technologies in drug discovery, opened the path to the computational anal. and visualization of bioactive mols. For this purpose, it became necessary to represent mols. in a syntax that would be readable by computers and understandable by scientists of various fields. A large no. of chem. representations have been developed over the years, their numerosity being due to the fast development of computers and the complexity of producing a representation that encompasses all structural and chem. characteristics. We present here some of the most popular electronic mol. and macromol. representations used in drug discovery, many of which are based on graph representations. Furthermore, we describe applications of these representations in AI-driven drug discovery. Our aim is to provide a brief guide on structural representations that are essential to the practise of AI in drug discovery. This review serves as a guide for researchers who have little experience with the handling of chem. representations and plan to work on applications at the interface of these fields.
    22. 22
      Nguyen-Vo, T.-H.; Teesdale-Spittle, P.; Harvey, J. E.; Nguyen, B. P. Molecular representations in bio-cheminformatics. Memetic Comput. 2024, 16, 519– 536,  DOI: 10.1007/s12293-024-00414-6
      There is no corresponding record for this reference.
    23. 23
      Wigh, D. S.; Goodman, J. M.; Lapkin, A. A. A review of molecular representation in the age of machine learning. WIREs Comput. Mol. Sci. 2022, 12, e1603  DOI: 10.1002/wcms.1603
      There is no corresponding record for this reference.
    24. 24
      Sousa, T.; Correia, J.; Pereira, V.; Rocha, M. Generative deep learning for targeted compound design. J. Chem. Inf. Model. 2021, 61, 5343– 5361,  DOI: 10.1021/acs.jcim.0c01496
      24
      Generative Deep Learning for Targeted Compound Design
      Sousa, Tiago; Correia, Joao; Pereira, Vitor; Rocha, Miguel
      Journal of Chemical Information and Modeling (2021), 61 (11), 5343-5361CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      A review. In the past few years, de novo mol. design has increasingly been using generative models from the emergent field of Deep Learning, proposing novel compds. that are likely to possess desired properties or activities. De novo mol. design finds applications in different fields ranging from drug discovery and materials sciences to biotechnol. A panoply of deep generative models, including architectures as Recurrent Neural Networks, Autoencoders, and Generative Adversarial Networks, can be trained on existing data sets and provide for the generation of novel compds. Typically, the new compds. follow the same underlying statistical distributions of properties exhibited on the training data set. Addnl., different optimization strategies, including transfer learning, Bayesian optimization, reinforcement learning, and conditional generation, can direct the generation process toward desired aims, regarding their biol. activities, synthesis processes or chem. features. Given the recent emergence of these technologies and their relevance, this work presents a systematic and crit. review on deep generative models and related optimization methods for targeted compd. design, and their applications.
    25. 25
      Morgan, H. L. The generation of a unique machine description for chemical structures-a technique developed at chemical abstracts service. J. Chem. Doc. 1965, 5, 107– 113,  DOI: 10.1021/c160017a018
      25
      Generation of a unique machine description for chemical structures--a technique developed at Chemical Abstracts Service
      Morgan, H. L.
      Journal of Chemical Documentation (1965), 5 (2), 107-13CODEN: JCHDAN; ISSN:0021-9576.
      The description employed is a uniquely ordered list of the node symbols of the structure (or graph) in which the value (at. symbol) of each node and its attachment (bonding) to the other nodes of the total structure. When the entire structure has been numbered according to a given set of rules, the connection table is formed by recording the structural relation by a process of successive partial orderings.
    26. 26
      Durant, J. L.; Leland, B. A.; Henry, D. R.; Nourse, J. G. Reoptimization of MDL Keys for use in drug discovery. J. Chem. Inf. Comput. Sci. 2002, 42, 1273– 1280,  DOI: 10.1021/ci010132r
      26
      Reoptimization of MDL Keys for Use in Drug Discovery
      Durant, Joseph L.; Leland, Burton A.; Henry, Douglas R.; Nourse, James G.
      Journal of Chemical Information and Computer Sciences (2002), 42 (6), 1273-1280CODEN: JCISD8; ISSN:0095-2338. (American Chemical Society)
      For a no. of years MDL products have exposed both 166 bit and 960 bit keysets based on 2D descriptors. These keysets were originally constructed and optimized for substructure searching. We report on improvements in the performance of MDL keysets which are reoptimized for use in mol. similarity. Classification performance for a test data set of 957 compds. was increased from 0.65 for the 166 bit keyset and 0.67 for the 960 bit keyset to 0.71 for a surprisal S/N pruned keyset contg. 208 bits and 0.71 for a genetic algorithm optimized keyset contg. 548 bits. We present an overview of the underlying technol. supporting the definition of descriptors and the encoding of these descriptors into keysets. This technol. allows definition of descriptors as combinations of atom properties, bond properties, and at. neighborhoods at various topol. sepns. as well as supporting a no. of custom descriptors. These descriptors can then be used to set one or more bits in a keyset. We constructed various keysets and optimized their performance in clustering bioactive substances. Performance was measured using methodol. developed by Briem and Lessel. "Directed pruning" was carried out by eliminating bits from the keysets on the basis of random selection, values of the surprisal of the bit, or values of the surprisal S/N ratio of the bit. The random pruning expt. highlighted the insensitivity of keyset performance for keyset lengths of more than 1000 bits. Contrary to initial expectations, pruning on the basis of the surprisal values of the various bits resulted in keysets which underperformed those resulting from random pruning. In contrast, pruning on the basis of the surprisal S/N ratio was found to yield keysets which performed better than those resulting from random pruning. We also explored the use of genetic algorithms in the selection of optimal keysets. Once more the performance was only a weak function of keyset size, and the optimizations failed to identify a single globally optimal keyset. Instead multiple, equally optimal keysets could be produced which had relatively low overlap of the descriptors they encoded.
    27. 27
      Schissel, C. K.; Mohapatra, S.; Wolfe, J. M.; Fadzen, C. M.; Bellovoda, K.; Wu, C.-L.; Wood, J. A.; Malmberg, A. B.; Loas, A.; Gómez-Bombarelli, R.; Pentelute, B. L. Deep learning to design nuclear-targeting abiotic miniproteins. Nat. Chem. 2021, 13, 992– 1000,  DOI: 10.1038/s41557-021-00766-3
      27
      Deep learning to design nuclear-targeting abiotic miniproteins
      Schissel, Carly K.; Mohapatra, Somesh; Wolfe, Justin M.; Fadzen, Colin M.; Bellovoda, Kamela; Wu, Chia-Ling; Wood, Jenna A.; Malmberg, Annika B.; Loas, Andrei; Gomez-Bombarelli, Rafael; Pentelute, Bradley L.
      Nature Chemistry (2021), 13 (10), 992-1000CODEN: NCAHBB; ISSN:1755-4330. (Nature Portfolio)
      There are more amino acid permutations within a 40-residue sequence than atoms on Earth. This vast chem. search space hinders the use of human learning to design functional polymers. Here we show how machine learning enables the de novo design of abiotic nuclear-targeting miniproteins to traffic antisense oligomers to the nucleus of cells. We combined high-throughput experimentation with a directed evolution-inspired deep-learning approach in which the mol. structures of natural and unnatural residues are represented as topol. fingerprints. The model is able to predict activities beyond the training dataset, and simultaneously deciphers and visualizes sequence-activity predictions. The predicted miniproteins, termed Mach reach an av. mass of 10 kDa, are more effective than any previously known variant in cells and can also deliver proteins into the cytosol. The Mach miniproteins are non-toxic and efficiently deliver antisense cargo in mice. These results demonstrate that deep learning can decipher design principles to generate highly active biomols. that are unlikely to be discovered by empirical approaches.
    28. 28
      Ji, X.; Nielsen, A. L.; Heinis, C. Cyclic peptides for drug development. Angew. Chem., Int. Ed. 2024, 63, e202308251  DOI: 10.1002/anie.202308251
      There is no corresponding record for this reference.
    29. 29
      Costa, L.; Sousa, E.; Fernandes, C. Cyclic peptides in pipeline: what future for these great molecules?. Pharmaceuticals 2023, 16, 996  DOI: 10.3390/ph16070996
      29
      Antibiotics, antifungals, anticancer, and immunosuppressants use of cyclic peptides in therapeutics for different diseases
      Costa, Lia; Sousa, Emilia; Fernandes, Carla
      Pharmaceuticals (2023), 16 (7), 996CODEN: PHARH2; ISSN:1424-8247. (MDPI AG)
      A review. Cyclic peptides are mols. that are already used as drugs in therapies approved for various pharmacol. activities, for example, as antibiotics, antifungals, anticancer, and immunosuppressants. Interest in these mols. has been growing due to the improved pharmacokinetic and pharmacodynamic properties of the cyclic structure over linear peptides and by the evolution of chem. synthesis, computational, and in vitro methods. To date, 53 cyclic peptides have been approved by different regulatory authorities, and many others are in clin. trials for a wide diversity of conditions. In this review, the potential of cyclic peptides is presented, and general aspects of their synthesis and development are discussed. Furthermore, an overview of already approved cyclic peptides is also given, and the cyclic peptides in clin. trials are summarized.
    30. 30
      Zhang, H.; Chen, S. Cyclic peptide drugs approved in the last two decades (2001–2021). RSC Chem. Biol. 2022, 3, 18– 31,  DOI: 10.1039/D1CB00154J
      30
      Cyclic peptide drugs approved in the last two decades (2001-2021)
      Zhang Huiya; Chen Shiyu
      RSC chemical biology (2022), 3 (1), 18-31 ISSN:.
      In contrast to the major families of small molecules and antibodies, cyclic peptides, as a family of synthesizable macromolecules, have distinct biochemical and therapeutic properties for pharmaceutical applications. Cyclic peptide-based drugs have increasingly been developed in the past two decades, confirming the common perception that cyclic peptides have high binding affinities and low metabolic toxicity as antibodies, good stability and ease of manufacture as small molecules. Natural peptides were the major source of cyclic peptide drugs in the last century, and cyclic peptides derived from novel screening and cyclization strategies are the new source. In this review, we will discuss and summarize 18 cyclic peptides approved for clinical use in the past two decades to provide a better understanding of cyclic peptide development and to inspire new perspectives. The purpose of the present review is to promote efforts to resolve the challenges in the development of cyclic peptide drugs that are more effective.
    31. 31
      Landrum, G. RDKit: Open-Source Cheminformatics Software. https://www.rdkit.org/.
      There is no corresponding record for this reference.
    32. 32
      The PyMOL Molecular Graphics System, version 3.0; Schrödinger, LLC.
      There is no corresponding record for this reference.
    33. 33
      Case, D. A.; Aktulga, H. M. A.; Belfon, K.; Ben-Shalom, I. Y.; Berryman, J. T.; Brozell, S. R.; Cerutti, D. S.; Cheatham, T. E., III; Cisneros, G. A.; Cruzeiro, V. W. D.; Darden, T. A.; Duke, R. E.; Giambasu, G.; Gilson, M. K.; Gohlke, H.; Goetz, A. W.; Harris, R.; Izadi, S.; Izmailov, S. A.; Kasavajhala, K.; Kaymak, M. C.; King, E.; Kovalenko, A.; Kurtzman, T.; Lee, T. S.; LeGrand, S.; Li, P.; Lin, C.; Liu, J.; Luchko, T.; Luo, R.; Machado, M.; Man, V.; Manathunga, M.; Merz, K. M.; Miao, Y.; Mikhailovskii, O.; Monard, G.; Nguyen, H.; O’Hearn, K. A.; Onufriev, A.; Pan, F.; Pantano, S.; Qi, R.; Rahnamoun, A.; Roe, D. R.; Roitberg, A.; Sagui, C.; Schott-Verdugo, S.; Shajan, A.; Shen, J.; Simmerling, C. L.; Skrynnikov, N. R.; Smith, J.; Swails, J.; Walker, R. C.; Wang, J.; Wang, J.; Wei, H.; Wolf, R. M.; Wu, X.; Xiong, Y.; Xue, Y.; York, D. M.; Zhao, S.; Kollman, P. A. Amber; University of California: San Francisco, 2022.
      There is no corresponding record for this reference.
    34. 34
      Jakalian, A.; Jack, D. B.; Bayly, C. I. Fast, efficient generation of high-quality atomic charges. AM1-BCC model: II. Parameterization and validation. J. Comput. Chem. 2002, 23, 1623– 4161,  DOI: 10.1002/jcc.10128
      34
      Fast, efficient generation of high-quality atomic charges. AM1-BCC model: II. parameterization and validation
      Jakalian, Araz; Jack, David B.; Bayly, Christopher I.
      Journal of Computational Chemistry (2002), 23 (16), 1623-1641CODEN: JCCHDD; ISSN:0192-8651. (John Wiley & Sons, Inc.)
      We present the first global parameterization and validation of a novel charge model, called AM1-BCC, which quickly and efficiently generates high-quality at. charges for computer simulations of org. mols. in polar media. The goal of the charge model is to produce at. charges that emulate the HF/6-31G* electrostatic potential (ESP) of a mol. Underlying electronic structure features, including formal charge and electron delocalization, are first captured by AM1 population charges; simple additive bond charge corrections (BCCs) are then applied to these AM1 at. charges to produce the AM1-BCC charges. The parameterization of BCCs was carried out by fitting to the HF/6-31G* ESP of a training set of >2700 mols. Most org. functional groups and their combinations were sampled, as well as an extensive variety of cyclic and fused bicyclic heteroaryl systems. The resulting BCC parameters allow the AM1-BCC charging scheme to handle virtually all types of org. compds. listed in The Merck Index and the NCI Database. Validation of the model was done through comparisons of hydrogen-bonded dimer energies and relative free energies of solvation using AM1-BCC charges in conjunction with the 1994 Cornell et al. forcefield for AMBER. Homo-dimer and hetero-dimer hydrogen-bond energies of a diverse set of org. mols. were reproduced to within 0.95 kcal/mol RMS deviation from the ab initio values, and for DNA dimers the energies were within 0.9 kcal/mol RMS deviation from ab initio values. The calcd. relative free energies of solvation for a diverse set of monofunctional isosteres were reproduced to within 0.69 kcal/mol of expt. In all these validation tests, AMBER with the AM1-BCC charge model maintained a correlation coeff. above 0.96. Thus, the parameters presented here for use with the AM1-BCC method present a fast, accurate, and robust alternative to HF/6-31G* ESP-fit charges for general use with the AMBER force field in computer simulations involving org. small mols.
    35. 35
      Bayly, C. I.; Cieplak, P.; Cornell, W.; Kollman, P. A. A well-behaved electrostatic potential based method using charge restraints for deriving atomic charges: the RESP model. J. Phys. Chem. A 1993, 97, 10269– 10280,  DOI: 10.1021/j100142a004
      35
      A well-behaved electrostatic potential based method using charge restraints for deriving atomic charges: the RESP model
      Bayly, Christopher I.; Cieplak, Piotr; Cornell, Wendy; Kollman, Peter A.
      Journal of Physical Chemistry (1993), 97 (40), 10269-80CODEN: JPCHAX; ISSN:0022-3654.
      The authors present a new approach to generating electrostatic potential (ESP) derived charges for mols. The major strength of electrostatic potential derived charges is that they optimally reproduce the intermol. interaction properties of mols. with a simple two-body additive potential, provided, of course, that a suitably accurate level of quantum mech. calcn. is used to derive the ESP around the mol. Previously, the major weaknesses of these charges have been that they were not easily transferably between common functional groups in related mols., they have often been conformationally dependent, and the large charges that frequently occur can be problematic for simulating intramol. interactions. Introducing restraints in the form of a penalty function into the fitting process considerably reduces the above problems, with only a minor decrease in the quality of the fit to the quantum mech. ESP. Several other refinements in addn. to the restrained electrostatic potential (RESP) fit yield a general and algorithmic charge fitting procedure for generating atom-centered point charges. This approach can thus be recommended for general use in mol. mechanics, mol. dynamics, and free energy calcns. for any org. or bioorg. system.
    36. 36
      Hornak, V.; Abel, R.; Okur, A.; Strockbine, B.; Roitberg, A.; Simmerling, C. Comparison of multiple amber force fields and development of improved protein backbone parameters. Proteins 2006, 65, 712– 725,  DOI: 10.1002/prot.21123
      36
      Comparison of multiple Amber force fields and development of improved protein backbone parameters
      Hornak, Viktor; Abel, Robert; Okur, Asim; Strockbine, Bentley; Roitberg, Adrian; Simmerling, Carlos
      Proteins: Structure, Function, and Bioinformatics (2006), 65 (3), 712-725CODEN: PSFBAF ISSN:. (Wiley-Liss, Inc.)
      The ff94 force field that is commonly assocd. with the Amber simulation package is one of the most widely used parameter sets for biomol. simulation. After a decade of extensive use and testing, limitations in this force field, such as over-stabilization of α-helixes, were reported by the authors and other researchers. This led to a no. of attempts to improve these parameters, resulting in a variety of "Amber" force fields and significant difficulty in detg. which should be used for a particular application. The authors show that several of these continue to suffer from inadequate balance between different secondary structure elements. In addn., the approach used in most of these studies neglected to account for the existence in Amber of two sets of backbone .vphi./ψ dihedral terms. This led to parameter sets that provide unreasonable conformational preferences for glycine. The authors report here an effort to improve the .vphi./ψ dihedral terms in the ff99 energy function. Dihedral term parameters are based on fitting the energies of multiple conformations of glycine and alanine tetrapeptides from high level ab initio quantum mech. calcns. The new parameters for backbone dihedrals replace those in the existing ff99 force field. This parameter set, which the authors denote ff99SB, achieves a better balance of secondary structure elements as judged by improved distribution of backbone dihedrals for glycine and alanine with respect to PDB survey data. It also accomplishes improved agreement with published exptl. data for conformational preferences of short alanine peptides and better accord with exptl. NMR relaxation data of test protein systems.
    37. 37
      Wang, J.; Wolf, R. M.; Caldwell, J. W.; Kollman, P. A.; Case, D. A. Development and testing of a general amber force field. J. Comput. Chem. 2004, 25, 1157– 1174,  DOI: 10.1002/jcc.20035
      37
      Development and testing of a general Amber force field
      Wang, Junmei; Wolf, Romain M.; Caldwell, James W.; Kollman, Peter A.; Case, David A.
      Journal of Computational Chemistry (2004), 25 (9), 1157-1174CODEN: JCCHDD; ISSN:0192-8651. (John Wiley & Sons, Inc.)
      We describe here a general Amber force field (GAFF) for org. mols. GAFF is designed to be compatible with existing Amber force fields for proteins and nucleic acids, and has parameters for most org. and pharmaceutical mols. that are composed of H, C, N, O, S, P, and halogens. It uses a simple functional form and a limited no. of atom types, but incorporates both empirical and heuristic models to est. force consts. and partial at. charges. The performance of GAFF in test cases is encouraging. In test I, 74 crystallog. structures were compared to GAFF minimized structures, with a root-mean-square displacement of 0.26 Å, which is comparable to that of the Tripos 5.2 force field (0.25 Å) and better than those of MMFF 94 and CHARMm (0.47 and 0.44 Å, resp.). In test II, gas phase minimizations were performed on 22 nucleic acid base pairs, and the minimized structures and intermol. energies were compared to MP2/6-31G* results. The RMS of displacements and relative energies were 0.25 Å and 1.2 kcal/mol, resp. These data are comparable to results from Parm99/RESP (0.16 Å and 1.18 kcal/mol, resp.), which were parameterized to these base pairs. Test III looked at the relative energies of 71 conformational pairs that were used in development of the Parm99 force field. The RMS error in relative energies (compared to expt.) is about 0.5 kcal/mol. GAFF can be applied to wide range of mols. in an automatic fashion, making it suitable for rational drug design and database searching.
    38. 38
      Zhou, C.-Y.; Jiang, F.; Wu, Y.-D. Residue-specific force field based on protein coil library. RSFF2: modification of AMBER ff99SB. J. Phys. Chem. B 2015, 119, 1035– 1047,  DOI: 10.1021/jp5064676
      38
      Residue-Specific Force Field Based on Protein Coil Library. RSFF2: Modification of AMBER ff99SB
      Zhou, Chen-Yang; Jiang, Fan; Wu, Yun-Dong
      Journal of Physical Chemistry B (2015), 119 (3), 1035-1047CODEN: JPCBFK; ISSN:1520-5207. (American Chemical Society)
      Recently, we developed a residue-specific force field (RSFF1) based on conformational free-energy distributions of the 20 amino acid residues from a protein coil library. Most parameters in RSFF1 were adopted from the OPLS-AA/L force field, but some van der Waals and torsional parameters that effectively affect local conformational preferences were introduced specifically for individual residues to fit the coil library distributions. Here a similar strategy has been applied to modify the Amber ff99SB force field, and a new force field named RSFF2 is developed. It can successfully fold α-helical structures such as polyalanine peptides, Trp-cage miniprotein, and villin headpiece subdomain and β-sheet structures such as Trpzip-2, GB1 β-hairpins, and the WW domain, simultaneously. The properties of various popular force fields in balancing between α-helix and β-sheet are analyzed based on their descriptions of local conformational features of various residues, and the anal. reveals the importance of accurate local free-energy distributions. Unlike the RSFF1, which overestimates the stability of both α-helix and β-sheet, RSFF2 gives melting curves of α-helical peptides and Trp-cage in good agreement with exptl. data. Fitting to the two-state model, RSFF2 gives folding enthalpies and entropies in reasonably good agreement with available exptl. results.
    39. 39
      Abraham, M. J.; Murtola, T.; Schulz, R.; Páll, S.; Smith, J. C.; Hess, B.; Lindahl, E. GROMACS: High performance molecular simulations through multi-level parallelism from laptops to supercomputers. SoftwareX 2015, 1–2, 19– 25,  DOI: 10.1016/j.softx.2015.06.001
      There is no corresponding record for this reference.
    40. 40
      Jorgensen, W. L.; Chandrasekhar, J.; Madura, J. D.; Impey, R. W.; Klein, M. L. Comparison of simple potential functions for simulating liquid water. J. Chem. Phys. 1983, 79, 926– 935,  DOI: 10.1063/1.445869
      40
      Comparison of simple potential functions for simulating liquid water
      Jorgensen, William L.; Chandrasekhar, Jayaraman; Madura, Jeffry D.; Impey, Roger W.; Klein, Michael L.
      Journal of Chemical Physics (1983), 79 (2), 926-35CODEN: JCPSA6; ISSN:0021-9606.
      Classical Monte Carlo simulations were carried out for liq. H2O in the NPT ensemble at 25° and 1 atm using 6 of the simpler intermol. potential functions for the dimer. Comparisons were made with exptl. thermodn. and structural data including the neutron diffraction results of Thiessen and Narten (1982). The computed densities and potential energies agree with expt. except for the original Bernal-Fowler model, which yields an 18% overest. of the d. and poor structural results. The discrepancy may be due to the correction terms needed in processing the neutron data or to an effect uniformly neglected in the computations. Comparisons were made for the self-diffusion coeffs. obtained from mol. dynamics simulations.
    41. 41
      Essmann, U.; Perera, L.; Berkowitz, M. L.; Darden, T.; Lee, H.; Pedersen, L. G. A smooth particle mesh Ewald method. J. Chem. Phys. 1995, 103, 8577– 8593,  DOI: 10.1063/1.470117
      41
      A smooth particle mesh Ewald method
      Essmann, Ulrich; Perera, Lalith; Berkowitz, Max L.; Darden, Tom; Lee, Hsing; Pedersen, Lee G.
      Journal of Chemical Physics (1995), 103 (19), 8577-93CODEN: JCPSA6; ISSN:0021-9606. (American Institute of Physics)
      The previously developed particle mesh Ewald method is reformulated in terms of efficient B-spline interpolation of the structure factors. This reformulation allows a natural extension of the method to potentials of the form 1/rp with p ≥ 1. Furthermore, efficient calcn. of the virial tensor follows. Use of B-splines in the place of Lagrange interpolation leads to analytic gradients as well as a significant improvement in the accuracy. The authors demonstrate that arbitrary accuracy can be achieved, independent of system size N, at a cost that scales as N log(N). For biomol. systems with many thousands of atoms and this method permits the use of Ewald summation at a computational cost comparable to that of a simple truncation method of 10 Å or less.
    42. 42
      Frisch, M. J.; Trucks, G. W.; Schlegel, H. B.; Scuseria, G. E.; Robb, M. A.; Cheeseman, J. R.; Scalmani, G.; Barone, V.; Petersson, G. A.; Nakatsuji, H.; Li, X.; Caricato, M.; Marenich, A. V.; Bloino, J.; Janesko, B. G.; Gomperts, R.; Mennucci, B.; Hratchian, H. P.; Ortiz, J. V.; Izmaylov, A. F.; Sonnenberg, J. L.; Williams; ; Ding, F.; Lipparini, F.; Egidi, F.; Goings, J.; Peng, B.; Petrone, A.; Henderson, T.; Ranasinghe, D.; Zakrzewski, V. G.; Gao, J.; Rega, N.; Zheng, G.; Liang, W.; Hada, M.; Ehara, M.; Toyota, K.; Fukuda, R.; Hasegawa, J.; Ishida, M.; Nakajima, T.; Honda, Y.; Kitao, O.; Nakai, H.; Vreven, T.; Throssell, K.; Montgomery, J. A., Jr.; Peralta, J. E.; Ogliaro, F.; Bearpark, M. J.; Heyd, J. J.; Brothers, E. N.; Kudin, K. N.; Staroverov, V. N.; Keith, T. A.; Kobayashi, R.; Normand, J.; Raghavachari, K.; Rendell, A. P.; Burant, J. C.; Iyengar, S. S.; Tomasi, J.; Cossi, M.; Millam, J. M.; Klene, M.; Adamo, C.; Cammi, R.; Ochterski, J. W.; Martin, R. L.; Morokuma, K.; Farkas, O.; Foresman, J. B.; Fox, D. J. Gaussian 16, rev. C.01; Gaussian Inc.: Wallingford, CT, 2016.
      There is no corresponding record for this reference.
    43. 43
      Vanquelef, E.; Simon, S.; Marquant, G.; Garcia, E.; Klimerak, G.; Delepine, J. C.; Cieplak, P.; Dupradeau, F. Y. R.E.D. Server: a web service for deriving RESP and ESP charges and building force field libraries for new molecules and molecular fragments. Nucleic Acids Res. 2011, 39, W511– 517,  DOI: 10.1093/nar/gkr288
      43
      R.E.D. Server: a web service for deriving RESP and ESP charges and building force field libraries for new molecules and molecular fragments
      Vanquelef, Enguerran; Simon, Sabrina; Marquant, Gaelle; Garcia, Elodie; Klimerak, Geoffroy; Delepine, Jean Charles; Cieplak, Piotr; Dupradeau, Francois-Yves
      Nucleic Acids Research (2011), 39 (Web Server), W511-W517CODEN: NARHAD; ISSN:0305-1048. (Oxford University Press)
      R.E.D. Server is a unique, open web service, designed to derive non-polarizable RESP and ESP charges and to build force field libraries for new mols./mol. fragments. It provides to computational biologists the means to derive rigorously mol. electrostatic potential-based charges embedded in force field libraries that are ready to be used in force field development, charge validation and mol. dynamics simulations. R.E.D. Server interfaces quantum mechanics programs, the RESP program and the latest version of the R.E.D. tools. A two step approach has been developed. The first one consists of prepg. P2N file(s) to rigorously define key elements such as atom names, topol. and chem. equivalencing needed when building a force field library. Then, P2N files are used to derive RESP or ESP charges embedded in force field libraries in the Tripos mol2 format. In complex cases an entire set of force field libraries or force field topol. database is generated. Other features developed in R.E.D. Server include help services, a demonstration, tutorials, frequently asked questions, Jmol-based tools useful to construct PDB input files and parse R.E.D. Server outputs as well as a graphical queuing system allowing any user to check the status of R.E.D. Server jobs.
    44. 44
      Piana, S.; Laio, A. A bias-exchange approach to protein folding. J. Phys. Chem. B 2007, 111, 4553– 4559,  DOI: 10.1021/jp067873l
      44
      A Bias-Exchange Approach to Protein Folding
      Piana, Stefano; Laio, Alessandro
      Journal of Physical Chemistry B (2007), 111 (17), 4553-4559CODEN: JPCBFK; ISSN:1520-6106. (American Chemical Society)
      By suitably extending a recent approach [Bussi, G., et al., 2006] the authors introduce a powerful methodol. that allows the parallel reconstruction of the free energy of a system in a virtually unlimited no. of variables. Multiple metadynamics simulations of the same system at the same temp. are performed, biasing each replica with a time-dependent potential constructed in a different set of collective variables. Exchanges between the bias potentials in the different variables are periodically allowed according to a replica exchange scheme. Due to the efficaciously multidimensional nature of the bias the method allows exploring complex free energy landscapes with high efficiency. The usefulness of the method is demonstrated by performing an atomistic simulation in explicit solvent of the folding of a Triptophane cage miniprotein. It is shown that the folding free energy landscape can be fully characterized starting from an extended conformation with use of only 40 ns of simulation on 8 replicas.
    45. 45
      Tribello, G. A.; Bonomi, M.; Branduardi, D.; Camilloni, C.; Bussi, G. PLUMED 2: New feathers for an old bird. Comput. Phys. Commun. 2014, 185, 604– 613,  DOI: 10.1016/j.cpc.2013.09.018
      45
      PLUMED 2: New feathers for an old bird
      Tribello, Gareth A.; Bonomi, Massimiliano; Branduardi, Davide; Camilloni, Carlo; Bussi, Giovanni
      Computer Physics Communications (2014), 185 (2), 604-613CODEN: CPHCBZ; ISSN:0010-4655. (Elsevier B.V.)
      Enhancing sampling and analyzing simulations are central issues in mol. simulation. Recently, we introduced PLUMED, an open-source plug-in that provides some of the most popular mol. dynamics (MD) codes with implementations of a variety of different enhanced sampling algorithms and collective variables (CVs). The rapid changes in this field, in particular new directions in enhanced sampling and dimensionality redn. together with new hardware, require a code that is more flexible and more efficient. We therefore present PLUMED 2 here-a complete rewrite of the code in an object-oriented programming language (C++). This new version introduces greater flexibility and greater modularity, which both extends its core capabilities and makes it far easier to add new methods and CVs. It also has a simpler interface with the MD engines and provides a single software library contg. both tools and core facilities. Ultimately, the new code better serves the ever-growing community of users and contributors in coping with the new challenges arising in the field.
    46. 46
      Damas, J. M.; Filipe, L. C.; Campos, S. R.; Lousa, D.; Victor, B. L.; Baptista, A. M.; Soares, C. M. Predicting the thermodynamics and kinetics of helix formation in a cyclic peptide model. J. Chem. Theory Comput. 2013, 9, 5148– 5157,  DOI: 10.1021/ct400529k
      46
      Predicting the thermodynamics and kinetics of helix formation in a cyclic peptide model
      Damas, Joao M.; Filipe, Luis C. S.; Campos, Sara R. R.; Lousa, Diana; Victor, Bruno L.; Baptista, Antonio M.; Soares, Claudio M.
      Journal of Chemical Theory and Computation (2013), 9 (11), 5148-5157CODEN: JCTCCE; ISSN:1549-9618. (American Chemical Society)
      The peptide, Ac-(cyclo-2,6)-R-[KAAAD]-NH2 (cyc-RKAAAD), is a short cyclic peptide known to adopt a remarkably stable single-turn α-helix in water. Due to its simplicity and the availability of thermodn. and kinetic exptl. data, cyc-RKAAAD poses as an ideal model for evaluating the aptness of current mol. dynamics (MD) simulation methodologies to accurately sample conformations that reproduce exptl. obsd. properties. Here, the authors extensively sampled the conformational space of cyc-RKAAAD using microsecond-timescale MD simulations. The authors characterized the peptide conformational preferences in terms of secondary structure propensities and, using Cartesian-coordinate principal component anal. (cPCA), constructed its free energy landscape, thus obtaining a detailed weighted discrimination between the helical and nonhelical subensembles. The cPCA state discrimination, together with a Markov model built from it, allowed the authors to est. the free energy of unfolding (-0.57 kJ/mol) and the relaxation time (∼0.435 μs) at 298.15 K, which were in excellent agreement with the exptl. reported values. Addnl., the authors presented simulations conducted using 2 enhanced sampling methods: replica-exchange mol. dynamics (REMD) and bias-exchange metadynamics (BE-MetaD). The authors compared the free energy landscape obtained by these 2 methods with the results from MD simulations and discussed the sampling and computational gains achieved. Overall, the results obtained attested to the suitability of modern simulation methods to explore the conformational behavior of peptide systems with a high level of realism.
    47. 47
      Nair, V.; Hinton, G. E. Rectified Linear Units Improve Restricted Boltzmann Machines, Proceedings of the 27th International Conference on Machine Learning (ICML’10), Haifa, Israel, June 21–24, Fürnkranz, J.; Joachims, T., Eds.; Omnipress: Madison, WI, 2010; pp 807– 814.
      There is no corresponding record for this reference.
    48. 48
      Prechelt, L. Automatic early stopping using cross validation: quantifying the criteria. Neural Networks 1998, 11, 761– 767,  DOI: 10.1016/S0893-6080(98)00010-0
      48
      Automatic early stopping using cross validation: quantifying the criteria
      Prechelt Lutz
      Neural networks : the official journal of the International Neural Network Society (1998), 11 (4), 761-767 ISSN:.
      Cross validation can be used to detect when overfitting starts during supervised training of a neural network; training is then stopped before convergence to avoid the overfitting ('early stopping'). The exact criterion used for cross validation based early stopping, however, is chosen in an ad-hoc fashion by most researchers or training is stopped interactively. To aid a more well-founded selection of the stopping criterion, 14 different automatic stopping criteria from three classes were evaluated empirically for their efficiency and effectiveness in 12 different classification and approximation tasks using multi-layer perceptrons with RPROP training. The experiments show that, on average, slower stopping criteria allow for small improvements in generalization (in the order of 4%), but cost about a factor of 4 longer in training time.
    49. 49
      Paszke, A. G. S.; Massa, F.; Lerer, A.; Bradbury, J.; Chanan, G.; Killeen, T. L.; Gimelshein, N.; Antiga, L.; Desmaison, A.; Kopf, A.; Yang, E. D. Z.; Raison, M.; Tejani, A.; Chilamkurthy, S.; Steiner, B. F.; Bai, J.; Chintala, S. An Imperative Style, High-Performance Deep Learning Library, Advances in Neural Information Processing Systems 32 (NeurIPS 2019); Wallach, H.; Larochelle, H.; Beygelzimer, A.; d’Alché-Buc, F.; Fox, E.; Garnett, R., Eds.; Curran Associates: Vancouver, Canada, 2019; pp 8024– 8035.
      There is no corresponding record for this reference.
    50. 50
      Hayashi, K.; Uehara, S.; Yamamoto, S.; Cary, D. R.; Nishikawa, J.; Ueda, T.; Ozasa, H.; Mihara, K.; Yoshimura, N.; Kawai, T.; Ono, T.; Yamamoto, S.; Fumoto, M.; Mikamiyama, H. Macrocyclic peptides as a novel class of NNMT inhibitors: A SAR study aimed at inhibitory activity in the cell. ACS Med. Chem. Lett. 2021, 12, 1093– 1101,  DOI: 10.1021/acsmedchemlett.1c00134
      50
      Macrocyclic Peptides as a Novel Class of NNMT Inhibitors: A SAR Study Aimed at Inhibitory Activity in the Cell
      Hayashi, Kyohei; Uehara, Shota; Yamamoto, Shiho; Cary, Douglas R.; Nishikawa, Junichi; Ueda, Taichi; Ozasa, Hiroki; Mihara, Kousuke; Yoshimura, Norito; Kawai, Taeko; Ono, Takashi; Yamamoto, Saki; Fumoto, Masataka; Mikamiyama, Hidenori
      ACS Medicinal Chemistry Letters (2021), 12 (7), 1093-1101CODEN: AMCLCT; ISSN:1948-5875. (American Chemical Society)
      Nicotinamide N-methyltransferase (NNMT), which catalyzes the methylation of nicotinamide, is a cytosolic enzyme that has attracted much attention as a therapeutic target for a variety of diseases. However, despite the considerable interest in this target, reports of NNMT inhibitors have still been limited to date. In this work, utilizing in vitro translated macrocyclic peptide libraries, we identified peptide 1 as a novel class of NNMT inhibitors. Further exploration based on the X-ray cocrystal structures of the peptides with NNMT provided a dramatic improvement in inhibitory activity (peptide 23: IC50 = 0.15 nM). Furthermore, by balance of the peptides' lipophilicity and biol. activity, inhibitory activity against NNMT in cell-based assay was successfully achieved (peptide 26: cell-based IC50 = 770 nM). These findings illuminate the potential of cyclic peptides as a relatively new drug discovery modality even for intracellular targets.
    51. 51
      Brousseau, M. E.; Clairmont, K. B.; Spraggon, G.; Flyer, A. N.; Golosov, A. A.; Grosche, P.; Amin, J.; Andre, J.; Burdick, D.; Caplan, S.; Chen, G.; Chopra, R.; Ames, L.; Dubiel, D.; Fan, L.; Gattlen, R.; Kelly-Sullivan, D.; Koch, A. W.; Lewis, I.; Li, J.; Liu, E.; Lubicka, D.; Marzinzik, A.; Nakajima, K.; Nettleton, D.; Ottl, J.; Pan, M.; Patel, T.; Perry, L.; Pickett, S.; Poirier, J.; Reid, P. C.; Pelle, X.; Seepersaud, M.; Subramanian, V.; Vera, V.; Xu, M.; Yang, L.; Yang, Q.; Yu, J.; Zhu, G.; Monovich, L. G. Identification of a PCSK9-LDLR disruptor peptide with in vivo function. Cell Chem. Biol. 2022, 29, 249– 258.e5,  DOI: 10.1016/j.chembiol.2021.08.012
      51
      Identification of a PCSK9-LDLR disruptor peptide with in vivo function
      Brousseau, Margaret E.; Clairmont, Kevin B.; Spraggon, Glen; Flyer, Alec N.; Golosov, Andrei A.; Grosche, Philipp; Amin, Jakal; Andre, Jerome; Burdick, Debra; Caplan, Shari; Chen, Guanjing; Chopra, Raj; Ames, Lisa; Dubiel, Diana; Fan, Li; Gattlen, Raphael; Kelly-Sullivan, Dawn; Koch, Alexander W.; Lewis, Ian; Li, Jingzhou; Liu, Eugene; Lubicka, Danuta; Marzinzik, Andreas; Nakajima, Katsumasa; Nettleton, David; Ottl, Johannes; Pan, Meihui; Patel, Tajesh; Perry, Lauren; Pickett, Stephanie; Poirier, Jennifer; Reid, Patrick C.; Pelle, Xavier; Seepersaud, Mohindra; Subramanian, Vanitha; Vera, Victoria; Xu, Mei; Yang, Lihua; Yang, Qing; Yu, Jinghua; Zhu, Guoming; Monovich, Lauren G.
      Cell Chemical Biology (2022), 29 (2), 249-258.e5CODEN: CCBEBM; ISSN:2451-9448. (Cell Press)
      Proprotein convertase subtilisin/kexin type 9 (PCSK9) regulates plasma low-d. lipoprotein cholesterol (LDL-C) levels by promoting hepatic LDL receptor (LDLR) degrdn. Therapeutic antibodies that disrupt PCSK9-LDLR binding reduce LDL-C concns. and cardiovascular disease risk. The epidermal growth factor precursor homol. domain A (EGF-A) of the LDLR serves as a primary contact with PCSK9 via a flat interface, presenting a challenge for identifying small mol. PCSK9-LDLR disruptors. We employ an affinity-based screen of 1013in vitro-translated macrocyclic peptides to identify high-affinity PCSK9 ligands that utilize a unique, induced-fit pocket and partially disrupt the PCSK9-LDLR interaction. Structure-based design led to mols. with enhanced function and pharmacokinetic properties (e.g., 13PCSK9i). In mice, 13PCSK9i reduces plasma cholesterol levels and increases hepatic LDLR d. in a dose-dependent manner. 13PCSK9i functions by a unique, allosteric mechanism and is the smallest mol. identified to date with in vivo PCSK9-LDLR disruptor function.
    52. 52
      Yoshida, S.; Uehara, S.; Kondo, N.; Takahashi, Y.; Yamamoto, S.; Kameda, A.; Kawagoe, S.; Inoue, N.; Yamada, M.; Yoshimura, N.; Tachibana, Y. Peptide-to-small molecule: a pharmacophore-guided small molecule lead generation strategy from high-affinity macrocyclic peptides. J. Med. Chem. 2022, 65, 10655– 10673,  DOI: 10.1021/acs.jmedchem.2c00919
      There is no corresponding record for this reference.
    53. 53
      Banerjee, R.; Basu, G.; Chène, P.; Roy, S. Aib-based peptide backbone as scaffolds for helical peptide mimics. J. Pept. Res. 2002, 60, 88– 94,  DOI: 10.1034/j.1399-3011.2002.201005.x
      53
      Aib-based peptide backbone as scaffolds for helical peptide mimics
      Banerjee, R.; Basu, G.; Chene, P.; Roy, S.
      Journal of Peptide Research (2002), 60 (2), 88-94CODEN: JPERFA; ISSN:1397-002X. (Blackwell Munksgaard)
      Helical peptides that can intervene and disrupt therapeutically important protein-protein interactions are attractive drug targets. In order to develop a general strategy for developing such helical peptide mimics, the authors have studied the effect of incorporating α-amino isobutyric acid (Aib), an amino acid with strong preference for helical backbone, as the sole helix promoter in designed peptides. Specifically, the focus is on the hdm2-p53 interaction, which is central to development of many types of cancer. The peptide corresponding to the hdm2 interacting part of p53, helical in bound state but devoid of structure in soln., served as the starting point for peptide design that involved replacement of noninteracting residues by Aib. Incorporation of Aib, while preserving the interacting residues, led to significant increase in helical structure, particularly at the C-terminal region as judged by NMR and CD. The interaction with hdm2 was also found to be enhanced. Most interestingly, trypsin cleavage was found to be retarded by several orders of magnitude. It is concluded that incorporation of Aib is a feasible strategy to create peptide helical mimics with enhanced receptor binding and lower protease cleavage rate.
    54. 54
      Karle, I. L. Controls exerted by the Aib residue: helix formation and helix reversal. Pept. Sci. 2001, 60, 351– 365,  DOI: 10.1002/1097-0282(2001)60:5<351::AID-BIP10174>3.0.CO;2-U
      There is no corresponding record for this reference.
    55. 55
      Lundberg, S. M.; Lee, S.-I. A Unified Approach to Interpreting Model Predictions; Proceedings of the 31st International Conference on Neural Information Processing Systems (NIPS’17), Long Beach, CA, USA, December 4–9, 2017; Von Luxburg, U.; Guyon, I.; Bengio, S.; Wallach, H.; Ferugs, R., Eds.; Curan Associates: Redhook, NY, 2017; pp 4768– 4777.
      There is no corresponding record for this reference.

  • Supporting Information

    Supporting Information

    The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.jcim.4c02102.

    • Example of a (ϕ, ψ) distribution binned using various grid sizes; comparison of backbone (BB) features of amino acids in the different amino acid libraries; comparison of the different normalization schemes applied to the BB and voxel (VOX) features; examples of learning curves from hyperparameter tuning; model performances (reporting R2) on the training and validation data sets for the cyclic pentapeptides; p-values from t-tests comparing different features; model performances (reporting R2) on the training and validation data sets for the cyclic hexapeptides; model performances (reporting weighted error, WE) on the various test data sets for the cyclic pentapeptides and cyclic hexapeptides; model performances (reporting R2) using combinations of features on the training and validation data sets for the cyclic pentapeptides and cyclic hexapeptides (PDF)


    Terms & Conditions

    Most electronic Supporting Information files are available without a subscription to ACS Web Editions. Such files may be downloaded by article for research use (if there is a public use license linked to the relevant article, that license may permit other uses). Permission may be obtained from ACS for other uses through requests via the RightsLink permission system: http://pubs.acs.org/page/copyright/permissions.html.