ACS Publications. Most Trusted. Most Cited. Most Read
PlayMolecule Glimpse: Understanding Protein–Ligand Property Predictions with Interpretable Neural Networks
More

OR SEARCH CITATIONS

My Activity
Publications

Figure 1Loading Img
Download Hi-Res ImageDownload to MS-PowerPointCite This:J. Chem. Inf. Model. 2022, 62, 2, 225-231
  • Open Access
Application Note

PlayMolecule Glimpse: Understanding Protein–Ligand Property Predictions with Interpretable Neural Networks
Click to copy article linkArticle link copied!

  • Alejandro Varela-Rial
    Alejandro Varela-Rial
    Computational Science Laboratory, Universitat Pompeu Fabra, Barcelona Biomedical Research Park (PRBB), Carrer Dr. Aiguader 88, 08003 Barcelona, Spain
    Acellera Labs, Doctor Trueta 183, 08005 Barcelona, Spain
    More by Alejandro Varela-Rial
    Orcidhttps://orcid.org/0000-0002-6918-1765
  • Iain Maryanow
    Iain Maryanow
    Acellera Labs, Doctor Trueta 183, 08005 Barcelona, Spain
    More by Iain Maryanow
  • Maciej Majewski
    Maciej Majewski
    Computational Science Laboratory, Universitat Pompeu Fabra, Barcelona Biomedical Research Park (PRBB), Carrer Dr. Aiguader 88, 08003 Barcelona, Spain
    More by Maciej Majewski
    Orcidhttps://orcid.org/0000-0003-2605-8166
  • Stefan Doerr
    Stefan Doerr
    Acellera Labs, Doctor Trueta 183, 08005 Barcelona, Spain
    More by Stefan Doerr
  • Nikolai Schapin
    Nikolai Schapin
    Computational Science Laboratory, Universitat Pompeu Fabra, Barcelona Biomedical Research Park (PRBB), Carrer Dr. Aiguader 88, 08003 Barcelona, Spain
    Acellera Labs, Doctor Trueta 183, 08005 Barcelona, Spain
    More by Nikolai Schapin
  • José Jiménez-Luna
    José Jiménez-Luna
    Computational Science Laboratory, Universitat Pompeu Fabra, Barcelona Biomedical Research Park (PRBB), Carrer Dr. Aiguader 88, 08003 Barcelona, Spain
    More by José Jiménez-Luna
  • Gianni De Fabritiis*
    Gianni De Fabritiis
    Computational Science Laboratory, Universitat Pompeu Fabra, Barcelona Biomedical Research Park (PRBB), Carrer Dr. Aiguader 88, 08003 Barcelona, Spain
    Acellera Labs, Doctor Trueta 183, 08005 Barcelona, Spain
    Institució Catalana de Recerca i Estudis Avançats (ICREA), Passeig Lluis Companys 23, 08010 Barcelona, Spain
    *Email: [email protected]
    More by Gianni De Fabritiis
    Orcidhttps://orcid.org/0000-0003-3913-4877
Open PDFSupporting Information (1)

Journal of Chemical Information and Modeling

Cite this: J. Chem. Inf. Model. 2022, 62, 2, 225–231
Click to copy citationCitation copied!
https://doi.org/10.1021/acs.jcim.1c00691
Published January 3, 2022

Copyright © 2022 The Authors. Published by American Chemical Society. This publication is licensed under

CC-BY 4.0 .

Abstract

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

Deep learning has been successfully applied to structure-based protein–ligand affinity prediction, yet the black box nature of these models raises some questions. In a previous study, we presented KDEEP, a convolutional neural network that predicted the binding affinity of a given protein–ligand complex while reaching state-of-the-art performance. However, it was unclear what this model was learning. In this work, we present a new application to visualize the contribution of each input atom to the prediction made by the convolutional neural network, aiding in the interpretability of such predictions. The results suggest that KDEEP is able to learn meaningful chemistry signals from the data, but it has also exposed the inaccuracies of the current model, serving as a guideline for further optimization of our prediction tools.

This publication is licensed under

CC-BY 4.0 .
  • cc licence
  • by licence
Copyright © 2022 The Authors. Published by American Chemical Society

Subjects

what are subjects

Introduction

Click to copy section linkSection link copied!
Machine-learning methods have been widely applied in the field of chemoinformatics, ranging from simple, regressor-based QSAR models (1−4) to more complex neural networks. These latter methods have been reported to increase performance in some critical tasks for drug discovery, such as toxicity assessment, (5,6) pharmacokinetics, physicochemical property prediction, (7−10) and protein–ligand binding affinity prediction. (11−15)
In a previous work, we developed KDEEP–a 3D convolutional neural network (CNN) that accepts as input a voxelized representation of a protein–ligand complex and outputs a prediction of binding affinity with state-of-the-art accuracy. (11) However, it was unclear whether KDEEP was learning meaningful chemistry or just exploiting shortcuts such as the positive relationship between molecular weight and affinity. (16) Learning these shortcuts instead of the underlying nature of the problem is a topic of concern in the field. (17) It is then comprehensible for many machine learning methods to spark criticism regarding the difficulty to understand the rationale behind their predictions. It has been questioned whether a pharmaceutical company would promote a given molecule into a portfolio based only on an opaque prediction made by a neural network, without any clear explanation to support it. (18) Providing such an explanation would undoubtedly increase the value, trustworthiness, and usability of machine learning models in drug discovery.
Recently, advances in model interpretability, (19,20) as well as the availability of software libraries such as Captum (21) and Alibi, (22) have allowed researchers to get a first glimpse of what features of the input are more influential toward predictions made by neural networks (i.e., feature attribution assignment). One natural approach to measure this influence is to look at the gradients of the output neuron with respect to the input. In fact, in a CNN trained to discriminate accurate from inaccurate binding poses and to predict binding affinity, visually inspecting these gradients can reveal in which direction the atoms should move to improve the score that the network assigns it, (23) providing some degree of interpretability.
However, backpropagating the prediction relative to the input layer can produce very low gradients in the vicinity of the input vector, (19) a process which is known as gradient saturation. The Integrated Gradients (IG) feature attribution technique (19) helps to mitigate this problem, providing a better measure of how each input feature influences the prediction. Instead of evaluating the gradients at one particular input value (the image in a traditional 2D-CNN), gradients are computed for several variants of that image, ranging from a user-defined baseline (typically, an image with all its pixel-channel values set to zero) to the actual image. In each variant, the values of all its pixels are multiplied by a scalar α, ranging from the zeroed-out input to the original image. At low values of α, the resulting input vector is far from the usual input space the network has been exposed to during training, circumventing the gradient saturation issue.
In this article, we present an application to visualize the contribution of the input features for the prediction of KDEEP and similar CNNs. In addition to describing the methodology used herein, we also showcase several relevant examples of attributions which match with structural biology knowledge. We analyze the prediction of three distinct models: a clash detector, a docking pose classifier, and KDEEP. The clash detector provides a baseline to which we compare the other models and allowed us to validate the implementation of this application. The docking pose classifier and KDEEP models were evaluated to see if CNNs trained to perform chemically relevant tasks were learning meaningful chemistry. The application, called Glimpse, is available to use at https://www.playmolecule.org/Glimpse/.

Methods

Click to copy section linkSection link copied!

Model Training

KDEEP is a 3D-CNN which accepts as input a grid of size . This grid is generated by mapping the atom positions of the ligand and its surrounding protein residues to the corresponding voxel and channel in the grid. KDEEP uses 8 different channels: hydrophobic, aromatic, hydrogen bond donor, hydrogen bond acceptor, positive ionizable, negative ionizable, metals, and excluded volume (occupancy) for both protein and ligand. This gives a total of 16 channels and a grid of dimensions (see Jiménez et al. (11,24) for more details). The network was trained on the latest version of the refined set of PDBbind, (25) achieving a Pearson’s correlation coefficient of 0.79 in the test set. Details on the training and evaluation of the different models can be found in the SI.
We also trained a clash detector. The objective behind it is two-fold. First, there is a clear expectation in terms of what attributions should look like: clashing regions or close contacts would appear highlighted, while the remaining voxels would be of little importance. We were able to validate the implementation of Glimpse by checking if the computed attributions matched this expectation. Second, the computed attributions of this simple model served as a reference point to which we compare the other models, both visually and quantitatively. This model was trained to discriminate regular protein–ligand poses from clashed poses and achieved 0.97 classification accuracy and 0.98 precision in a held-out validation set.
We trained a third model–a docking pose classifier–for two reasons: (i) it is a challenging task, comparable to that of predicting binding affinity, and (ii) there is much more data available from which the model can learn. This model was trained on a large set of good (RMSD Å) and bad poses (RMSD Å), showing an accuracy and precision of 0.94 and 0.83, respectively, on the validation set.
It must be noted that the performance of these three models was evaluated on a random test and validation sets. In some cases, protein–ligand complexes in these sets might be similar to those in the training set, either in terms of protein structure or ligand composition. This yields overoptimistic results. In fact, when trained and evaluated in more strict splits which ensured sequence and ligand dissimilarity, the KDEEP performance ranged from r = 0.09 to r = 0.7 (see the SI for details).
The models analyzed in this work have been uploaded to Glimpse with the following names: “KDEEP”, “Pose classifier”, and “Clash detector”. Attributions for these models can be computed and visualized in the app for any valid protein–ligand complex.

Implementation

Click to copy section linkSection link copied!

Integrated Gradients

The IG method works by computing gradients of the output neuron with respect to the input layer along an interpolated path from a given input baseline () to the original input (xi) taking α infinitesimal steps as in (19)
(1)
where F denotes the forward pass of the neural network.
This effectively circumvents the issue of low gradients (gradient saturation) in the vicinity of the input by averaging the gradients along a range of different input values. Gradient saturation can occur if a given input value leads to a neuron being activated in a region of the activation function which is very flat, for instance, the extremes of a sigmoid. Hence, using the gradients of the prediction with respect to a given input could assign an importance of zero to it, regardless of its real importance.
The computed attributions provide a value for each voxel representing their importance toward the prediction. Glimpse uses the IG implementation from the Captum library. (21) Attributions are computed by approximating the integral as a series of discrete steps along the interpolated path from the selected baseline (an input vector in which all voxels are set to zero) to the evaluated, voxel map corresponding to the original protein–ligand complex. In this implementation, we used 100 steps, as it was shown to offer a good balance between computational expense and attribution quality.

Graphical User Interface

Glimpse provides a web-based graphical user interface (GUI) that helps to trace attributions to voxel maps. An overview of the GUI is provided in Figure 1. The computed attributions for the input channels are displayed as mesh isosurfaces whose isovalue can be tuned with a slider. A detailed description of the input channels is provided in the Model Training section of this manuscript. To offer a summary of the results, the interface displays by default only the most contributing regions. These regions are the result of identifying, for each of the 16 channels, the voxel with the highest absolute attribution value and the neighbors around it. In this summary, only the channels containing the best 5 voxels are displayed for simplicity. Additionally, the user can display the attributions for the different input channels individually, and the raw attribution maps can be downloaded as a Gaussian cube format file (.cube) and explored in VMD (26) or other molecular visualization software.

Figure 1

Figure 1. Main view of the graphical user interface. The protein–ligand complex is displayed, with the attributions of the most contributing voxels superimposed. The attributions for the different channels can be seen individually using the corresponding sliders in the menu on the right, which display isosurfaces at different isovalues. The full protein is shown in a cartoon representation, while residues in the binding site (defined by being within 5 Å to the ligand) are shown in a transparent ball–stick representation (only heavy atoms and polar hydrogens). The all-atom representation of the ligand is shown in a bold ball–stick. The region of space seen by the model (voxelization cube) is delimited by a transparent, gray box.

High Resolution Image
Download MS PowerPoint Slide

Usage

Glimpse requires a protein–ligand complex structure, either experimentally determined or predicted by a docking software. The protein must be correctly protonated and provided as a .PDB file. The PlayMolecule platform offers proteinPrepare (27) to protonate the protein. The ligands have to be provided a valid SDF file. If needed, Glimpse provides an option for protonation of these ligands. Only 100 ligands are allowed per job. Finally, users can select which model to use from a list, which by default is “KDEEP”. In terms of time, evaluating 10 protein–ligand complexes takes around 150 s. When inspecting the attributions, one would typically start looking at the visual summary, followed by an inspection of individual channels. It is worth paying particular attention to the voxels with the highest and lowest attribution values and checking if the nearby atoms are involved in an interaction. The occupancy channels offer a good overview of the whole picture and constitute a good place to start.

Analysis

For each model, we visually inspected the attributions computed for several protein–ligand complexes to evaluate how well they match with structural biology knowledge. We focused on interactions known to contribute toward binding free energy, e.g., hydrogen bonds or π-stacking. Another aspect we inspected was the reciprocity in the attributions, that is, whether the two parties involved in the protein–ligand interaction are reflected in their attribution values.
While visual inspection can provide valuable insights, it can also be misleading and prone to unintended biases. Therefore, we designed a quantitative analysis, in which we computed the IG for all the crystal structures (not clashed or docked) coming from the PDBbind database. For each of the 16 channels, the voxel with the highest, absolute value of the IG values was identified. Then, we measured the distance between the top contributing voxels in the complementary channels. We evaluated the following pairs of channels: protein hydrophobic and ligand hydrophobic, protein aromatic and ligand aromatic (π-stacking), protein acceptor and ligand donor, protein donor and ligand acceptor (hydrogen bonds), protein occupancy and ligand occupancy (steric component). As a baseline, we took the distance between two randomly selected voxels from the appropriate channels, whose occupancy value was over 0.75, ensuring that an atom was nearby the voxel.

Results

Click to copy section linkSection link copied!
As a leading example for the analysis, we selected a complex of a molecular chaperone, heat shock protein 90 kDa (Hsp90) sourced from the PDB (PDBid: 3D0B). (28) This well studied oncology target has been a subject of numerous structure based virtual screening campaigns. (29,30) In the analyzed example, HSP90 forms a potent complex with an analogue of benzamide tetrahydro-4H-carbazol-4-one (SNX), with an affinity of 290 nM. Additionally, the complex possesses few features that facilitate tight binding, mainly π-stacking and a hydrogen bond with a conserved aspartate (D93)–a very frequent interaction among HSP90 inhibitors. (31) This helped us to relate the predictions to structural features of the complex.

Clash Detector

As a sanity check, we started by evaluating the attributions of the simplest model, the clash detector. The visual inspection of multiple complexes revealed that, in all inspected cases, clashing regions or close contacts were highlighted, while residues far apart from the ligand remained ignored. The clashes were clearly indicated by occupancy channels of both protein and ligand, showing reciprocity (Figure 2.1A). An example of HSP90 with a clashed pose clearly highlights the clashed region between the ligand and a leucine (L92) in the pocket (Figure 2.1B). This confirmed that the protocol was working correctly and gave us a baseline for the analysis of the following models.

Figure 2

Figure 2. Comparison of computed attributions obtained for a complex of HSP90 with an analogue of benzamide tetrahydro-4H-carbazol-4-one (PDB code: 3D0B) by the three models: Clash detector (1A and 1B), Pose classifier (2A and 2B), and KDEEP (3A and 3B). Pictures on the top row show the attributions for the protein and ligand occupancy channels, in red and blue, respectively. The bottom row focuses on particular interactions. 1B shows a clash between the ligand and the leucine and the attributions for the occupancy channels of protein and ligand (red and blue). 2B and 3B show the hydrogen bond between the benzamide moiety in the ligand and the aspartate (D93) residue in the protein. Attributions for the ligand donor channel are shown in pink, while for the protein acceptor channel, the attributions are shown in blue.

High Resolution Image
Download MS PowerPoint Slide

Docking Pose Classifier

The next model was trained to discriminate good and bad docking poses. The true binding mode depends on an enthalpic factor which is determined by formation of strong and stable interactions between the ligand and the protein, like hydrogen bonds or π-stacking. (32) Hence, we expect models trained to perform such tasks to have learned these interactions.
We found several examples where the attributions correctly matched these expectations. Figure 2.2B shows a hydrogen bond being highlighted in the appropriate channels: hydrogen bond acceptor in the protein and donor in the ligand. An amide moiety in the ligand is establishing a hydrogen bond with the nearby aspartate (D93). It is indeed a key interaction frequently featured in HSP90 inhibitors. In addition to this hydrogen bond, a π-stacking interaction takes place between the aromatic ring of the ligand and a phenylalanine (F123) in the protein, as highlighted by the attributions for the occupancy (Figure 2.2A) and aromatic channels (not shown) of both the protein and the ligand.
After inspecting several examples, we saw that, as in the clash detector model, residues far from the ligand were ignored for the most part (Figure S7 shows one exception), and reciprocity between the ligand and the protein atoms was observed in the majority of inspected cases. However, the overall picture was less clear than in the clash detector model as can be seen by comparing the global view of the three models (sections 1A, 2A, and 3A of Figure 2). While in the clash detector model high attributions are well focused on the clashing regions and close contacts, the other two models exhibit a more disperse view.

KDEEP

Similarly to the previous model, KDEEP is expected to predict binding affinity by detecting and correctly weighting the molecular interactions between protein and ligand. In the majority of inspected cases, we saw reciprocity between the ligand and the protein atoms that formed interactions (Figure 2.3A), while protein residues distant from the ligand remained ignored (Figure S8 shows one exception). The overall attribution maps seem to be more disperse than for docking pose predictor.
For the example of HSP90, the predicted affinity value was 75 nM, reasonably close to the experimental value of 290 nM, making it a suitable example for attribution analysis. As in the case of the pose classifier, the network correctly identified the key hydrogen bond with D93 (Figure 2.3B), as well as π-stacking between the aromatic ring of the ligand and F123 (Figure 2.3A). In this case, however, only the phenylalanine ring is highlighted. Nonetheless, the attributions of the aromatic channels highlight the aromatic residues in close proximity to the ring system of the ligand, including Phe, ignoring all the other aromatic residues in the box.

Quantitative Analysis

The quantitative analysis confirmed, for the most part, the conclusions obtained by the visual inspection. Figure 3 shows the distribution of distances between the top voxels from protein occupancy and ligand occupancy channels. Figures S1–S6 show the distance distribution for the remaining relevant combinations of channels. We can see that models have learned that ligand and protein atoms close to each other are important, which is exemplified by the different distributions being shifted toward the contact range ( Å). This hints that the networks are learning relevant features of the complex: close contacts in the case of clash detector and interactions for the two remaining models. This is particularly clear for the clash detector, where the distance distribution is radically shifted toward the range under 3 Å. The pose classifier model follows and shows a similar, shifted distribution, although not as clear as in the clash detector. In these two models, the cloud of points describes a 3-line pattern at 0 and 1 and around 1.5 Å, showing that the most contributing voxels were in very close proximity. In fact, the high number of examples observed at 0 Å distance indicates that, in a large fraction of complexes, the same voxel in the two relevant channels was the most highlighted. The distance distribution for KDEEP is slightly shifted toward higher values but is still much better than the random baseline. This difference could be related to the fact that, during training, KDEEP is only exposed to crystal poses, in which the distances between ligand and protein atoms should be uniform across examples. This is obviously not the case in the clash detector nor in the pose classifier, as docking might generate poses which, in some areas, might be slightly too close or too far away from the surface, which might correlate with a bad pose. Hence, these two models would benefit more from paying attention to this low range of values, while KDEEP might not. A similar scenario occurs to the other combinations of channels (Figures S1–S6): The clash detector is usually the best, followed closely by the pose classifier and KDEEP, which show similar distributions. The distributions of all three models are significantly different from the random baseline in all channel combinations studied according to a two-sided Mann–Whitney U Test (all combinations had a p-value lower than the significance threshold 0.0001). These results provide evidence supporting the hypothesis that KDEEP and the pose classifier models have learned to focus on the interface between the ligand and the protein, as a trained chemist or biologist would do.

Figure 3

Figure 3. Distance distribution between the two voxels with the highest, absolute value in protein and ligand occupancy channels.

High Resolution Image
Download MS PowerPoint Slide
After this test, we checked if there was any correlation between the magnitude of the attribution of the two best voxels and the distance between them as a means to test if the quality of the explanation (reciprocity between the two parties) was correlated with the magnitude of the attribution. As in the prior experiment, we identified the voxels with the highest attribution values in the same pairs of channels and annotated the distance between them. We then evaluated the correlation between the sum of the attributions in those two voxels and their distance. For all models and combinations of channels studied, the Pearson’s correlation coefficient ranged between −0.48 and −0.15 (Figure S9). Hence, the higher the attribution value of the two voxels, the more likely it is that those two voxels are in proximity.
Furthermore, for the KDEEP model, we also evaluated if there was a correlation between the attribution values and the accuracy of the prediction (measured as an absolute difference between predicted and actual pKD values). Neither the maximum attribution value across all channels nor the sum of the absolute attribution values correlated with the accuracy of the prediction in the KDEEP model (Pearson’s r was −0.05 and −0.02, respectively). There was not any strong correlation either between far away residues being highlighted and prediction accuracy (r = 0.05, see the SI for details).
Finally, we evaluated how sensitive the attributions were to minor changes in the input, namely (1) rotations of the complex and (2) slight modifications of the protein–ligand pose. Ideally, the attributions should be consistent across different orientations and pose variants; hence, the same atoms should be highlighted in the different variations. All three models show a greater consistency than the expected by random, both for protein and ligand atoms. Although on average, the same atom was selected just around 2 times out of the 10 input variants (random baseline is close to 1.0); in all three models, in a great fraction of complexes, the same atom is picked more than 4 times, which is not the case in the random baseline (Figures S10–S13). These results show that the attributions tolerate some degree of input variability. Details on these experiments can be found in the SI.

Conclusion

Click to copy section linkSection link copied!
The results indicate that the trained networks are able to learn meaningful chemical interactions. However, for the pose classifier model and KDEEP, some cases were observed where the network had ignored strong interactions, highlighted residues far from the ligand (Figures S7 and S8) or highlighted ligand atoms whose interaction counterpart in the protein had low attributions. This can be the result of the difficulty of associating the occurrence of certain contacts or interactions with affinity or with the quality of the pose prediction, leading to shortcut learning. For instance, if all the complexes for kinases in the training set have a pKd of 5.0, the network might learn to identify this family by using a set of characteristic residues (which could be far from the ligand) and simply predict 5.0. In this sense, PDBbind is not an ideal training set, as the total number of examples contains few samples for deep-learning standards, and pKd values are distributed in a large range from 2 to 12, scarcely populated in both extremes. In the case of the clash detector model, we have more examples for each class, the two classes belong to very different distributions, and it is very easy to associate the occurrence of a given pattern in the input (a clash) to the correct class, discouraging shortcut learning, which manifests in the attributions of this model being much clearer.
Furthermore, voxelization is limited to only 8 properties and excludes crystallographic waters. Given that the latter are known to mediate certain protein–ligand interactions (e.g., water bridges), a fraction of the variability in the binding affinity can only be explained by the presence of these waters. Additionally, hydrogen-bond donors and acceptors have diverse strengths (thiol being a weak donor and hydroxyl being a strong one); however, in the featurization, they are grouped together in just two entities (donor and acceptor). The same reasoning applies for the positive and negative ionizable channels.
In this study, we have shown that Glimpse displays the capability to expose some of the flaws of the networks herein analyzed, suggesting that it can act as a useful diagnostic tool for structure-based 3D-CNN models. We were also able to identify atoms or regions of the protein–ligand complex that play a bigger role on the predictions made by the networks, which is key to improve the usability of CNNs in computational chemistry.
Data and Software Availability. Glimpse is available free of charge at https://www.playmolecule.org/Glimpse/. The three models studied in this article (“KDEEP”, “Pose classifier”, and “Clash detector”) can be found and used in the web interface. The protein–ligand complexes we used as input for generating the images are available to download in the “Examples” tab in the web interface. The databases used for training and validating the models (BindingMoad (33) and PDBbind (25)) are publicly available, as well as the docking software (rDock (34)).

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.1c00691.

  • “Model training” and “quantitative analysis”, additional information for these sections; Figures S1–S6, distance distribution between the two voxels with highest, absolute attribution value for different channel combinations studied; Figures S7 and S8, examples of protein residues far from ligand having high attribution values; Figure S9, correlation between magnitude of attributions of two best voxels and distance between them; and Figures S10–S13, attribution consistency distributions (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
    • Gianni De Fabritiis - Computational Science Laboratory, Universitat Pompeu Fabra, Barcelona Biomedical Research Park (PRBB), Carrer Dr. Aiguader 88, 08003 Barcelona, Spain;  Acellera Labs, Doctor Trueta 183, 08005 Barcelona, Spain;  Institució Catalana de Recerca i Estudis Avançats (ICREA), Passeig Lluis Companys 23, 08010 Barcelona, Spain;  Orcidhttps://orcid.org/0000-0003-3913-4877;  Email: [email protected]
  • Authors
    • Alejandro Varela-Rial - Computational Science Laboratory, Universitat Pompeu Fabra, Barcelona Biomedical Research Park (PRBB), Carrer Dr. Aiguader 88, 08003 Barcelona, Spain;  Acellera Labs, Doctor Trueta 183, 08005 Barcelona, Spain;  Orcidhttps://orcid.org/0000-0002-6918-1765
    • Iain Maryanow - Acellera Labs, Doctor Trueta 183, 08005 Barcelona, Spain
    • Maciej Majewski - Computational Science Laboratory, Universitat Pompeu Fabra, Barcelona Biomedical Research Park (PRBB), Carrer Dr. Aiguader 88, 08003 Barcelona, Spain;  Orcidhttps://orcid.org/0000-0003-2605-8166
    • Stefan Doerr - Acellera Labs, Doctor Trueta 183, 08005 Barcelona, Spain
    • Nikolai Schapin - Computational Science Laboratory, Universitat Pompeu Fabra, Barcelona Biomedical Research Park (PRBB), Carrer Dr. Aiguader 88, 08003 Barcelona, Spain;  Acellera Labs, Doctor Trueta 183, 08005 Barcelona, Spain
    • José Jiménez-Luna - Computational Science Laboratory, Universitat Pompeu Fabra, Barcelona Biomedical Research Park (PRBB), Carrer Dr. Aiguader 88, 08003 Barcelona, Spain
  • Notes
    The authors declare the following competing financial interest(s): Some of the authors are directly affiliated with Acellera.

Acknowledgments

Click to copy section linkSection link copied!

The authors thank Acellera Ltd. for funding. G.D.F. acknowledges support from PID2020-116564GB-I00/MICIN/AEI/10.13039/501100011033 Ministerio de Ciencia e Innovación. This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No. 823712 (CompBioMed2) and from the Industrial Doctorates Plan of the Secretariat of Universities and Research of the Department of Economy and Knowledge of the Generalitat of Catalonia.

References

Click to copy section linkSection link copied!

This article references 34 other publications.

  1. 1
    Dudek, A. Z.; Arodz, T.; Gálvez, J. Computational methods in developing quantitative structure-activity relationships (QSAR): a review. Comb. Chem. High Throughput Screening 2006, 9, 213– 228,  DOI: 10.2174/138620706776055539
    Google Scholar
    1
    Computational methods in developing quantitative structure-activity relationships (QSAR): a review
    Dudek, Arkadiusz Z.; Arodz, Tomasz; Galvez, Jorge
    Combinatorial Chemistry & High Throughput Screening (2006), 9 (3), 213-228CODEN: CCHSFU; ISSN:1386-2073. (Bentham Science Publishers Ltd.)
    A review. Virtual filtering and screening of combinatorial libraries have recently gained attention as methods complementing the high-throughput screening and combinatorial chem. These chemoinformatic techniques rely heavily on quant. structure-activity relation (QSAR) anal., a field with established methodol. and successful history. In this review, we discuss the computational methods for building QSAR models. We start with outlining their usefulness in high-throughput screening and identifying the general scheme of a QSAR model. Following, we focus on the methodologies in constructing three main components of QSAR model, namely the methods for describing the mol. structure of compds., for selection of informative descriptors and for activity prediction. We present both the well-established methods as well as techniques recently introduced into the QSAR domain.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BD28XisVSqtbo%253D&md5=230cc86a6228a8f9a97c666e925a5efe
  2. 2
    Cherkasov, A.; Muratov, E. N.; Fourches, D.; Varnek, A.; Baskin, I. I.; Cronin, M.; Dearden, J.; Gramatica, P.; Martin, Y. C.; Todeschini, R. QSAR modeling: where have you been? Where are you going to?. J. Med. Chem. 2014, 57, 4977– 5010,  DOI: 10.1021/jm4004285
    Google Scholar
    2
    QSAR Modeling: Where Have You Been? Where Are You Going To?
    Cherkasov, Artem; Muratov, Eugene N.; Fourches, Denis; Varnek, Alexandre; Baskin, Igor I.; Cronin, Mark; Dearden, John; Gramatica, Paola; Martin, Yvonne C.; Todeschini, Roberto; Consonni, Viviana; Kuzmin, Victor E.; Cramer, Richard; Benigni, Romualdo; Yang, Chihae; Rathman, James; Terfloth, Lothar; Gasteiger, Johann; Richard, Ann; Tropsha, Alexander
    Journal of Medicinal Chemistry (2014), 57 (12), 4977-5010CODEN: JMCMAR; ISSN:0022-2623. (American Chemical Society)
    A review. Quant. structure-activity relationship modeling is one of the major computational tools employed in medicinal chem. However, throughout its entire history it has drawn both praise and criticism concerning its reliability, limitations, successes, and failures. In this paper, we discuss (i) the development and evolution of QSAR; (ii) the current trends, unsolved problems, and pressing challenges; and (iii) several novel and emerging applications of QSAR modeling. Throughout this discussion, we provide guidelines for QSAR development, validation, and application, which are summarized in best practices for building rigorously validated and externally predictive QSAR models. We hope that this Perspective will help communications between computational and exptl. chemists toward collaborative development and use of QSAR models. We also believe that the guidelines presented here will help journal editors and reviewers apply more stringent scientific stds. to manuscripts reporting new QSAR studies, as well as encourage the use of high quality, validated QSARs for regulatory decision making.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC3sXhvFCrsLnE&md5=a2d371ee33fbecf48ef59ecf914e4aa8
  3. 3
    Lo, Y.-C.; Rensi, S. E.; Torng, W.; Altman, R. B. Machine learning in chemoinformatics and drug discovery. Drug Discovery Today 2018, 23, 1538– 1546,  DOI: 10.1016/j.drudis.2018.05.010
    Google Scholar
    3
    Machine learning in chemoinformatics and drug discovery
    Lo, Yu-Chen; Rensi, Stefano E.; Torng, Wen; Altman, Russ B.
    Drug Discovery Today (2018), 23 (8), 1538-1546CODEN: DDTOFS; ISSN:1359-6446. (Elsevier Ltd.)
    Chemoinformatics is an established discipline focusing on extg., processing and extrapolating meaningful data from chem. structures. With the rapid explosion of chem. 'big' data from HTS and combinatorial synthesis, machine learning has become an indispensable tool for drug designers to mine chem. information from large compd. databases to design drugs with important biol. properties. To process the chem. data, we first reviewed multiple processing layers in the chemoinformatics pipeline followed by the introduction of commonly used machine learning models in drug discovery and QSAR anal. Here, we present basic principles and recent case studies to demonstrate the utility of machine learning techniques in chemoinformatics analyses; and we discuss limitations and future directions to guide further development in this evolving field.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXptFCnuro%253D&md5=d0bef55f60c8ee7bb2a31e42d4a85fcd
  4. 4
    Neves, B. J.; Braga, R. C.; Melo-Filho, C. C.; Moreira-Filho, J. T.; Muratov, E. N.; Andrade, C. H. QSAR-Based Virtual Screening: Advances and Applications in Drug Discovery. Front. Pharmacol. 2018, 9, 1275,  DOI: 10.3389/fphar.2018.01275
    Google Scholar
    4
    QSAR-based virtual screening: advances and applications in drug discovery
    Neves, Bruno J.; Braga, Rodolpho C.; Melo-Filho, Cleber C.; Moreira-Filho, Jose Teofilo; Muratov, Eugene N.; Andrade, Carolina Horta
    Frontiers in Pharmacology (2018), 9 (), 1275CODEN: FPRHAU; ISSN:1663-9812. (Frontiers Media S.A.)
    A review. Virtual screening (VS) has emerged in drug discovery as a powerful computational approach to screen large libraries of small mols. for new hits with desired properties that can then be tested exptl. Similar to other computational approaches, VS intention is not to replace in vitro or in vivo assays, but to speed up the discovery process, to reduce the no. of candidates to be tested exptl., and to rationalize their choice. Moreover, VS has become very popular in pharmaceutical companies and academic organizations due to its time-, cost-, resources-, and laborsaving. Among the VS approaches, quant. structure-activity relationship (QSAR) anal. is the most powerful method due to its high and fast throughput and good hit rate. As the first preliminary step of a QSAR model development, relevant chemogenomics data are collected from databases and the literature. Then, chem. descriptors are calcd. on different levels of representation of mol. structure, ranging from 1D to nD, and then correlated with the biol. property using machine learning techniques. Once developed and validated, QSAR models are applied to predict the biol. property of novel compds. Although the exptl. testing of computational hits is not an inherent part of QSAR methodol., it is highly desired and should be performed as an ultimate validation of developed models. In this minireview, we summarize and critically analyze the recent trends of QSAR-based VS in drug discovery and demonstrate successful applications in identifying perspective compds. with desired properties. Moreover, we provide some recommendations about the best practices for QSAR-based VS along with the future perspectives of this approach.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXhtVGmtbnP&md5=1aa305bf00e34d3ac592d19a7e107d7e
  5. 5
    Zhang, L.; Zhang, H.; Ai, H.; Hu, H.; Li, S.; Zhao, J.; Liu, H. Applications of Machine Learning Methods in Drug Toxicity Prediction. Curr. Top. Med. Chem. (Sharjah, United Arab Emirates) 2018, 18, 987– 997,  DOI: 10.2174/1568026618666180727152557
    Google Scholar
    5
    Applications of Machine Learning Methods in Drug Toxicity Prediction
    Zhang, Li; Zhang, Hui; Ai, Haixin; Hu, Huan; Li, Shimeng; Zhao, Jian; Liu, Hongsheng
    Current Topics in Medicinal Chemistry (Sharjah, United Arab Emirates) (2018), 18 (12), 987-997CODEN: CTMCCL; ISSN:1568-0266. (Bentham Science Publishers Ltd.)
    Toxicity evaluation is an important part of the preclin. safety assessment of new drugs, which is directly related to human health and the fate of drugs. It is of importance to study how to evaluate drug toxicity accurately and economically. The traditional in vitro and in vivo toxicity tests are laborious, time-consuming, highly expensive, and even involve animal welfare issues. Computational methods developed for drug toxicity prediction can compensate for the shortcomings of traditional methods and have been considered useful in the early stages of drug development. Numerous drug toxicity prediction models have been developed using a variety of computational methods. With the advance of the theory of machine learning and mol. representation, more and more drug toxicity prediction models are developed using a variety of machine learning methods, such as support vector machine, random forest, naive Bayesian, back propagation neural network. And significant advances have been made in many toxicity endpoints, such as carcinogenicity, mutagenicity, and hepatotoxicity. In this review, we aimed to provide a comprehensive overview of the machine learning based drug toxicity prediction studies conducted in recent years. In addn., we compared the performance of the models proposed in these studies in terms of accuracy, sensitivity, and specificity, providing a view of the current state-of-the-art in this field and highlighting the issues in the current studies.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXhvVejt7fO&md5=e98c7844a12ead052509399b9cf9c6b4
  6. 6
    Ma, H.; An, W.; Wang, Y.; Sun, H.; Huang, R.; Huang, J. Deep Graph Learning with Property Augmentation for Predicting Drug-Induced Liver Injury. Chem. Res. Toxicol. 2021, 34, 495,  DOI: 10.1021/acs.chemrestox.0c00322
    Google Scholar
    6
    Deep Graph Learning with Property Augmentation for Predicting Drug-Induced Liver Injury
    Ma, Hehuan; An, Weizhi; Wang, Yuhong; Sun, Hongmao; Huang, Ruili; Huang, Junzhou
    Chemical Research in Toxicology (2021), 34 (2), 495-506CODEN: CRTOEC; ISSN:0893-228X. (American Chemical Society)
    Drug-induced liver injury (DILI) is a crucial factor in detg. the qualification of potential drugs. However, the DILI property is excessively difficult to obtain due to the complex testing process. Consequently, an in silico screening in the early stage of drug discovery would help to reduce the total development cost by filtering those drug candidates with a high risk to cause DILI. To serve the screening goal, we apply several computational techniques to predict the DILI property, including traditional machine learning methods and graph-based deep learning techniques. While deep learning models require large training data to tune huge model parameters, the DILI data set only contains a few hundred annotated mols. To alleviate the data scarcity problem, we propose a property augmentation strategy to include massive training data with other property information. Extensive expts. demonstrate that our proposed method significantly outperforms all existing baselines on the DILI data set by obtaining a 81.4% accuracy using cross-validation with random splitting, 78.7% using leave-one-out cross-validation, and 76.5% using cross-validation with scaffold splitting.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXis1ehsrrI&md5=1eb843cf35938405c37027987f1daea4
  7. 7
    Montanari, F.; Kuhnke, L.; Ter Laak, A.; Clevert, D.-A. Modeling physico-chemical ADMET endpoints with multitask graph convolutional networks. Molecules 2020, 25, 44,  DOI: 10.3390/molecules25010044
    Google Scholar
    7
    Modeling physico-chemical ADMET endpoints with multitask graph convolutional networks
    Montanari, Floriane; Kuhnke, Lara; Ter Laak, Antonius; Clevert, Djork-Arne
    Molecules (2020), 25 (1), 44CODEN: MOLEFW; ISSN:1420-3049. (MDPI AG)
    Simple physico-chem. properties, like logD, soly., or m.p., can reveal a great deal about how a compd. under development might later behave. These data are typically measured for most compds. in drug discovery projects in a medium throughput fashion. Collecting and assembling all the Bayer inhouse data related to these properties allowed us to apply powerful machine learning techniques to predict the outcome of those assays for new compds. In this paper, we report our finding that, esp. for predicting physicochem. ADMET endpoints, a multitask graph convolutional approach appears a highly competitive choice. For seven endpoints of interest, we compared the performance of that approach to fully connected neural networks and different single task models. The new model shows increased predictive performance compared to previous modeling methods and will allow early prioritization of compds. even before they are synthesized. In addn., our model follows the generalized soly. equation without being explicitly trained under this constraint.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXjsFWqtLk%253D&md5=62240c8f1b151bab7a95d51e20b6c6a5
  8. 8
    Peng, Y.; Lin, Y.; Jing, X.-Y.; Zhang, H.; Huang, Y.; Luo, G. S. Enhanced Graph Isomorphism Network for Molecular ADMET Properties Prediction. IEEE Access 2020, 8, 168344– 168360,  DOI: 10.1109/ACCESS.2020.3022850
    Google Scholar
    There is no corresponding record for this reference.
  9. 9
    Feinberg, E. N.; Joshi, E.; Pande, V. S.; Cheng, A. C. Improvement in ADMET prediction with multitask deep featurization. J. Med. Chem. 2020, 63, 8835– 8848,  DOI: 10.1021/acs.jmedchem.9b02187
    Google Scholar
    9
    Improvement in ADMET Prediction with Multitask Deep Featurization
    Feinberg, Evan N.; Joshi, Elizabeth; Pande, Vijay S.; Cheng, Alan C.
    Journal of Medicinal Chemistry (2020), 63 (16), 8835-8848CODEN: JMCMAR; ISSN:0022-2623. (American Chemical Society)
    The absorption, distribution, metab., elimination, and toxicity (ADMET) properties of drug candidates are important for their efficacy and safety as therapeutics. Predicting ADMET properties has therefore been of great interest to the computational chem. and medicinal chem. communities in recent decades. Traditional cheminformatics approaches, using learners such as random forests and deep neural networks, leverage fingerprint feature representations of mols. Here, we learn the features most relevant to each chem. task at hand by representing each mol. explicitly as a graph. By applying graph convolutions to this explicit mol. representation, we achieve, to our knowledge, unprecedented accuracy in prediction of ADMET properties. By challenging our methodol. with rigorous cross-validation procedures and prognostic analyses, we show that deep featurization better enables mol. predictors to not only interpolate but also extrapolate to new regions of chem. space.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXntFCgsLs%253D&md5=845b33af28bfa7b817187c9e423a21eb
  10. 10
    Skalic, M.; Varela-Rial, A.; Jiménez, J.; Martínez-Rosell, G.; De Fabritiis, G. LigVoxel: inpainting binding pockets using 3D-convolutional neural networks. Bioinformatics 2019, 35, 243– 250,  DOI: 10.1093/bioinformatics/bty583
    Google Scholar
    10
    LigVoxel: inpainting binding pockets using 3D-convolutional neural networks
    Skalic, Miha; Varela-Rial, Alejandro; Jimenez, Jose; Martinez-Rosell, Gerard; Fabritiis, Gianni De
    Bioinformatics (2019), 35 (2), 243-250CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
    Motivation: Structure-based drug discovery methods exploit protein structural information to design small mols. binding to given protein pockets. This work proposes a purely data driven, structure-based approach for imaging ligands as spatial fields in target protein pockets. We use an end-to-end deep learning framework trained on exptl. protein-ligand complexes with the intention of mimicking a chemist's intuition at manually placing atoms when designing a new compd. We show that these models can generate spatial images of ligand chem. properties like occupancy, aromaticity and donor-acceptor matching the protein pocket. Results: The predicted fields considerably overlap with those of unseen ligands bound to the target pocket. Maximization of the overlap between the predicted fields and a given ligand on the Astex diverse set recovers the original ligand crystal poses in 70 out of 85 cases within a threshold of 2 Å RMSD. We expect that these models can be used for guiding structure-based drug discovery approaches.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXitVWgs73N&md5=e8bf4189689b73707a78dc94d58fa7af
  11. 11
    Jiménez, J.; Škalič, M.; Martínez-Rosell, G.; De Fabritiis, G. KDEEP: Protein-Ligand Absolute Binding Affinity Prediction via 3D-Convolutional Neural Networks. J. Chem. Inf. Model. 2018, 58, 287– 296,  DOI: 10.1021/acs.jcim.7b00650
    Google Scholar
    11
    KDEEP: Protein-Ligand Absolute Binding Affinity Prediction via 3D-Convolutional Neural Networks
    Jimenez, Jose; Skalic, Miha; Martinez-Rosell, Gerard; De Fabritiis, Gianni
    Journal of Chemical Information and Modeling (2018), 58 (2), 287-296CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    Accurately predicting protein-ligand binding affinities is an important problem in computational chem. since it can substantially accelerate drug discovery for virtual screening and lead optimization. We propose here a fast machine-learning approach for predicting binding affinities using state-of-the-art 3D-convolutional neural networks and compare this approach to other machine-learning and scoring methods using several diverse data sets. The results for the std. PDBbind (v.2016) core test-set are state-of-the-art with a Pearson's correlation coeff. of 0.82 and a RMSE of 1.27 in pK units between exptl. and predicted affinity, but accuracy is still very sensitive to the specific protein used. KDEEP is made available via PlayMol.org for users to test easily their own protein-ligand complexes, with each prediction taking a fraction of a second. We believe that the speed, performance, and ease of use of KDEEP makes it already an attractive scoring function for modern computational chem. pipelines.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXkslSitQ%253D%253D&md5=81943a6732be99e5439e1e3a25fa4414
  12. 12
    Ragoza, M.; Hochuli, J.; Idrobo, E.; Sunseri, J.; Koes, D. R. Protein-Ligand Scoring with Convolutional Neural Networks. J. Chem. Inf. Model. 2017, 57, 942– 957,  DOI: 10.1021/acs.jcim.6b00740
    Google Scholar
    12
    Protein-Ligand Scoring with Convolutional Neural Networks
    Ragoza, Matthew; Hochuli, Joshua; Idrobo, Elisa; Sunseri, Jocelyn; Koes, David Ryan
    Journal of Chemical Information and Modeling (2017), 57 (4), 942-957CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    Computational approaches to drug discovery can reduce the time and cost assocd. with exptl. assays and enable the screening of novel chemotypes. Structure-based drug design methods rely on scoring functions to rank and predict binding affinities and poses. The ever-expanding amt. of protein-ligand binding and structural data enables the use of deep machine learning techniques for protein-ligand scoring. We describe convolutional neural network (CNN) scoring functions that take as input a comprehensive three-dimensional (3D) representation of a protein-ligand interaction. A CNN scoring function automatically learns the key features of protein-ligand interactions that correlate with binding. We train and optimize our CNN scoring functions to discriminate between correct and incorrect binding poses and known binders and nonbinders. We find that our CNN scoring function outperforms the AutoDock Vina scoring function when ranking poses both for pose prediction and virtual screening.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC2sXlsVems7Y%253D&md5=9cae97167da0c93e1d896f85339cca7e
  13. 13
    Skalic, M.; Martínez-Rosell, G.; Jiménez, J.; De Fabritiis, G. PlayMolecule BindScope: large scale CNN-based virtual screening on the web. Bioinformatics 2019, 35, 1237– 1238,  DOI: 10.1093/bioinformatics/bty758
    Google Scholar
    13
    PlayMolecule BindScope: large scale CNN-based virtual screening on the web
    Skalic, Miha; Martinez-Rosell, Gerard; Jimenez, Jose; De Fabritiis, Gianni
    Bioinformatics (2019), 35 (7), 1237-1238CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
    Summary: Virtual screening pipelines are one of the most popular used tools in structure-based drug discovery, since they can can reduce both time and cost assocd. with exptl. assays. Recent advances in deep learning methodologies have shown that these outperform classical scoring functions at discriminating binder protein-ligand complexes. Here, we present BindScope, a web application for large-scale active-inactive classification of compds. based on deep convolutional neural networks. Performance is on a pair with current state-of-the-art pipelines. Users can screen on the order of hundreds of compds. at once and interactively visualize the results.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXitFWksb7K&md5=100cce3db2446853ed2f7c82b96d49df
  14. 14
    Yang, S.; Lee, K. H.; Ryu, S. A comprehensive study on the prediction reliability of graph neural networks for virtual screening. 2020, arXiv:2003.07611, arXiv preprint. https://arxiv.org/abs/2003.07611 (accessed 2021-10-29).
    Google Scholar
    There is no corresponding record for this reference.
  15. 15
    Sakai, M.; Nagayasu, K.; Shibui, N.; Andoh, C.; Takayama, K.; Shirakawa, H.; Kaneko, S. Prediction of pharmacological activities from chemical structures with graph convolutional neural networks. Sci. Rep. 2021, 11, 525,  DOI: 10.1038/s41598-020-80113-7
    Google Scholar
    15
    Prediction of pharmacological activities from chemical structures with graph convolutional neural networks
    Sakai, Miyuki; Nagayasu, Kazuki; Shibui, Norihiro; Andoh, Chihiro; Takayama, Kaito; Shirakawa, Hisashi; Kaneko, Shuji
    Scientific Reports (2021), 11 (1), 525CODEN: SRCEC3; ISSN:2045-2322. (Nature Research)
    Many therapeutic drugs are compds. that can be represented by simple chem. structures, which contain important determinants of affinity at the site of action. Recently, graph convolutional neural network (GCN) models have exhibited excellent results in classifying the activity of such compds. For models that make quant. predictions of activity, more complex information has been utilized, such as the three-dimensional structures of compds. and the amino acid sequences of their resp. target proteins. As another approach, we hypothesized that if sufficient exptl. data were available and there were enough nodes in hidden layers, a simple compd. representation would quant. predict activity with satisfactory accuracy. In this study, we report that GCN models constructed solely from the two-dimensional structural information of compds. demonstrated a high degree of activity predictability against 127 diverse targets from the ChEMBL database. Using the information entropy as a metric, we also show that the structural diversity had less effect on the prediction performance. Finally, we report that virtual screening using the constructed model identified a new serotonin transporter inhibitor with activity comparable to that of a marketed drug in vitro and exhibited antidepressant effects in behavioral studies.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXhsVemsLo%253D&md5=dfd62a89f9f705f4e740544f40ce705c
  16. 16
    Sieg, J.; Flachsenberg, F.; Rarey, M. Need of Bias Control: Evaluating Chemical Data for Machine Learning in Structure-Based Virtual Screening. J. Chem. Inf. Model. 2019, 59, 947– 961,  DOI: 10.1021/acs.jcim.8b00712
    Google Scholar
    16
    In Need of Bias Control: Evaluating Chemical Data for Machine Learning in Structure-Based Virtual Screening
    Sieg, Jochen; Flachsenberg, Florian; Rarey, Matthias
    Journal of Chemical Information and Modeling (2019), 59 (3), 947-961CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    A review. Reports of successful applications of machine learning (ML) methods in structure-based virtual screening (SBVS) are increasing. ML methods such as convolutional neural networks show promising results and often outperform traditional methods such as empirical scoring functions in retrospective validation. However, trained ML models are often treated as black boxes and are not straightforwardly interpretable. In most cases, it is unknown which features in the data are decisive and whether a model's predictions are right for the right reason. Hence, the authors reevaluated three widely used benchmark data sets in the context of ML methods and came to the conclusion that not every benchmark data set is suitable. Moreover, the authors demonstrate on two examples from current literature that bias is learned implicitly and unnoticed from std. benchmarks. On the basis of these results, the authors conclude that there is a need for eligible validation expts. and benchmark data sets suited to ML for more bias-controlled validation in ML-based SBVS. Therefore, the authors provide guidelines for setting up validation expts. and give a perspective on how new data sets could be generated.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXktVOgsrY%253D&md5=6cb48f2acd541c10c8565f202a6eb9e8
  17. 17
    DeGrave, A. J.; Janizek, J. D.; Lee, S.-I. AI for radiographic COVID-19 detection selects shortcuts over signal. Nat. Mach. Intell. 2021, 3, 610,  DOI: 10.1038/s42256-021-00338-7
    Google Scholar
    There is no corresponding record for this reference.
  18. 18
    Vamathevan, J.; Clark, D.; Czodrowski, P.; Dunham, I.; Ferran, E.; Lee, G.; Li, B.; Madabhushi, A.; Shah, P.; Spitzer, M.; Zhao, S. Applications of machine learning in drug discovery and development. Nat. Rev. Drug Discovery 2019, 18, 463– 477,  DOI: 10.1038/s41573-019-0024-5
    Google Scholar
    18
    Applications of machine learning in drug discovery and development
    Vamathevan, Jessica; Clark, Dominic; Czodrowski, Paul; Dunham, Ian; Ferran, Edgardo; Lee, George; Li, Bin; Madabhushi, Anant; Shah, Parantu; Spitzer, Michaela; Zhao, Shanrong
    Nature Reviews Drug Discovery (2019), 18 (6), 463-477CODEN: NRDDAG; ISSN:1474-1776. (Nature Research)
    A review. Drug discovery and development pipelines are long, complex and depend on numerous factors. Machine learning (ML) approaches provide a set of tools that can improve discovery and decision making for well-specified questions with abundant, high-quality data. Opportunities to apply ML occur in all stages of drug discovery. Examples include target validation, identification of prognostic biomarkers and anal. of digital pathol. data in clin. trials. Applications have ranged in context and methodol., with some approaches yielding accurate predictions and insights. The challenges of applying ML lie primarily with the lack of interpretability and repeatability of ML-generated results, which may limit their application. In all areas, systematic and comprehensive high-dimensional data still need to be generated. With ongoing efforts to tackle these issues, as well as increasing awareness of the factors needed to validate ML approaches, the application of ML can promote data-driven decision making and has the potential to speed up the process and reduce failure rates in drug discovery and development.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXosF2rtrY%253D&md5=211782aeea3d8b9f50368f89177a70d2
  19. 19
    Sundararajan, M.; Taly, A.; Yan, Q. Axiomatic Attribution for Deep Networks. 2017, arXiv preprint. https://arxiv.org/abs/1703.01365 (accessed 2021-10-29).
    Google Scholar
    There is no corresponding record for this reference.
  20. 20
    Henderson, R.; Clevert, D.-A.; Montanari, F. Improving Molecular Graph Neural Network Explainability with Orthonormalization and Induced Sparsity. Proceedings of the 38 th International Conference on Machine Learning ; 2021.
    Google Scholar
    There is no corresponding record for this reference.
  21. 21
    Kokhlikyan, N.; Miglani, V.; Martin, M.; Wang, E.; Alsallakh, B.; Reynolds, J.; Melnikov, A.; Kliushkina, N.; Araya, C.; Yan, S.; Reblitz-Richardson, O. Captum: A unified and generic model interpretability library for PyTorch. 2020, arXiv preprint. https://arxiv.org/abs/2009.07896 (accessed 2021-06-17).
    Google Scholar
    There is no corresponding record for this reference.
  22. 22
    Klaise, J.; Van Looveren, A.; Vacanti, G.; Coca, A. Alibi: Algorithms for monitoring and explaining machine learning models ; 2019. https://github.com/SeldonIO/alibi (accessed 2021-09-08).
    Google Scholar
    There is no corresponding record for this reference.
  23. 23
    Hochuli, J.; Helbling, A.; Skaist, T.; Ragoza, M.; Koes, D. R. Visualizing convolutional neural network protein-ligand scoring. J. Mol. Graphics Modell. 2018, 84, 96– 108,  DOI: 10.1016/j.jmgm.2018.06.005
    Google Scholar
    23
    Visualizing convolutional neural network protein-ligand scoring
    Hochuli, Joshua; Helbling, Alec; Skaist, Tamar; Ragoza, Matthew; Koes, David Ryan
    Journal of Molecular Graphics & Modelling (2018), 84 (), 96-108CODEN: JMGMFI; ISSN:1093-3263. (Elsevier Ltd.)
    Protein-ligand scoring is an important step in a structure-based drug design pipeline. Selecting a correct binding pose and predicting the binding affinity of a protein-ligand complex enables effective virtual screening. Machine learning techniques can make use of the increasing amts. of structural data that are becoming publicly available. Convolutional neural network (CNN) scoring functions in particular have shown promise in pose selection and affinity prediction for protein-ligand complexes. Neural networks are known for being difficult to interpret. Understanding the decisions of a particular network can help tune parameters and training data to maximize performance. Visualization of neural networks helps decomp. complex scoring functions into pictures that are more easily parsed by humans. Here we present three methods for visualizing how individual protein-ligand complexes are interpreted by 3D convolutional neural networks. We also present a visualization of the convolutional filters and their wts. We describe how the intuition provided by these visualizations aids in network design.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXhtF2ltb3P&md5=db697d859132f02c4a2cef4bfc00ea88
  24. 24
    Jiménez, J.; Doerr, S.; Martínez-Rosell, G.; Rose, A. S.; De Fabritiis, G. DeepSite: protein-binding site predictor using 3D-convolutional neural networks. Bioinformatics 2017, 33, 3036– 3042,  DOI: 10.1093/bioinformatics/btx350
    Google Scholar
    24
    DeepSite: protein-binding site predictor using 3D-convolutional neural networks
    Jimenez, J.; Doerr, S.; Martinez-Rosell, G.; Rose, A. S.; De Fabritiis, G.
    Bioinformatics (2017), 33 (19), 3036-3042CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
    Motivation: An important step in structure-based drug design consists in the prediction of druggable binding sites. Several algorithms for detecting binding cavities, those likely to bind to a small drug compd., have been developed over the years by clever exploitation of geometric, chem. and evolutionary features of the protein. Results: Here we present a novel knowledge-based approach that uses state-of-the-art convolutional neural networks, where the algorithm is learned by examples. In total, 7622 proteins from the scPDB database of binding sites have been evaluated using both a distance and a volumetric overlap approach. Our machine-learning based method demonstrates superior performance to two other competitive algorithmic strategies. Availability and implementation: DeepSite is freely available at www.playmol.org. Users can submit either a PDB ID or PDB file for pocket detection to our NVIDIA GPU-equipped servers through a WebGL graphical interface.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXhvFGju7nN&md5=fce2763a18192d7a7c4c98fcb2974a22
  25. 25
    Liu, Z.; Li, Y.; Han, L.; Li, J.; Liu, J.; Zhao, Z.; Nie, W.; Liu, Y.; Wang, R. PDB-wide collection of binding data: current status of the PDBbind database. Bioinformatics 2015, 31, 405– 412,  DOI: 10.1093/bioinformatics/btu626
    Google Scholar
    25
    PDB-wide collection of binding data: current status of the PDBbind database
    Liu, Zhihai; Li, Yan; Han, Li; Li, Jie; Liu, Jie; Zhao, Zhixiong; Nie, Wei; Liu, Yuchen; Wang, Renxiao
    Bioinformatics (2015), 31 (3), 405-412CODEN: BOINFP; ISSN:1367-4803. (Oxford University Press)
    Motivation:Mol. recognition between biol. macromols. and org. small mols. plays an important role in various life processes. Both structural information and binding data of biomol. complexes are indispensable for depicting the underlying mechanism in such an event. The PDBbind database was created to collect exptl. measured binding data for the biomol. complexes throughout the Protein Data Bank (PDB). It thus provides the linkage between structural information and energetic properties of biomol. complexes, which is esp. desirable for computational studies or statistical analyses. Results: Since its first public release in 2004, the PDBbind database has been updated on an annual basis. The latest release (version 2013) provides exptl. binding affinity data for 10 776 biomol. complexes in PDB, including 8302 protein-ligand complexes and 2474 other types of complexes. In this article, we will describe the current methods used for compiling PDBbind and the updated status of this database. We will also review some typical applications of PDBbind published in the scientific literature.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC28Xht1GjsrnI&md5=cb0debb7251bf8cd34385fb4b09e19bb
  26. 26
    Humphrey, W.; Dalke, A.; Schulten, K. VMD – Visual Molecular Dynamics. J. Mol. Graphics 1996, 14, 33– 38,  DOI: 10.1016/0263-7855(96)00018-5
    Google Scholar
    26
    VDM: visual molecular dynamics
    Humphrey, William; Dalke, Andrew; Schulten, Klaus
    Journal of Molecular Graphics (1996), 14 (1), 33-8, plates, 27-28CODEN: JMGRDV; ISSN:0263-7855. (Elsevier)
    VMD is a mol. graphics program designed for the display and anal. of mol. assemblies, in particular, biopolymers such as proteins and nucleic acids. VMD can simultaneously display any no. of structures using a wide variety of rendering styles and coloring methods. Mols. are displayed as one or more "representations," in which each representation embodies a particular rendering method and coloring scheme for a selected subset of atoms. The atoms displayed in each representation are chosen using an extensive atom selection syntax, which includes Boolean operators and regular expressions. VMD provides a complete graphical user interface for program control, as well as a text interface using the Tcl embeddable parser to allow for complex scripts with variable substitution, control loops, and function calls. Full session logging is supported, which produces a VMD command script for later playback. High-resoln. raster images of displayed mols. may be produced by generating input scripts for use by a no. of photorealistic image-rendering applications. VMD has also been expressly designed with the ability to animate mol. dynamics (MD) simulation trajectories, imported either from files or from a direct connection to a running MD simulation. VMD is the visualization component of MDScope, a set of tools for interactive problem solving in structural biol., which also includes the parallel MD program NAMD, and the MDCOMM software used to connect the visualization and simulation programs, VMD is written in C++, using an object-oriented design; the program, including source code and extensive documentation, is freely available via anonymous ftp and through the World Wide Web.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADyaK28Xis12nsrg%253D&md5=1e3094ec3151fb85c5ff05f8505c78d5
  27. 27
    Martínez-Rosell, G.; Giorgino, T.; De Fabritiis, G. PlayMolecule ProteinPrepare: a web application for protein preparation for molecular dynamics simulations. J. Chem. Inf. Model. 2017, 57, 1511– 1516,  DOI: 10.1021/acs.jcim.7b00190
    Google Scholar
    27
    PlayMolecule ProteinPrepare: A Web Application for Protein Preparation for Molecular Dynamics Simulations
    Martinez-Rosell, Gerard; Giorgino, Toni; De Fabritiis, Gianni
    Journal of Chemical Information and Modeling (2017), 57 (7), 1511-1516CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    Protein prepn. is a crit. step in mol. simulations that consists of refining a Protein Data Bank (PDB) structure by assigning titrn. states and optimizing the hydrogen-bonding network. In this application note, the authors describe ProteinPrepare, a web application designed to interactively support the prepn. of protein structures. Users can upload a PDB file, choose the solvent pH value, and inspect the resulting protonated residues and hydrogen-bonding network within a 3D web interface. Protonation states are suggested automatically but can be manually changed using the visual aid of the hydrogen-bonding network. Tables and diagrams provide estd. pKa values and charge states, with visual indication for cases where review is required. The authors expect the graphical interface to be a useful instrument to assess the validity of the prepn., but nevertheless, a script to execute the prepn. offline with the High-Throughput Mol. Dynamics (HTMD) environment is also provided for noninteractive operations.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC2sXps1Ghu7c%253D&md5=1a1adc5f0af564e9e9473cc5ebb16ab4
  28. 28
    Barta, T. E.; Veal, J. M.; Rice, J. W.; Partridge, J. M.; Fadden, R. P.; Ma, W.; Jenks, M.; Geng, L.; Hanson, G. J.; Huang, K. H. Discovery of benzamide tetrahydro-4H-carbazol-4-ones as novel small molecule inhibitors of Hsp90. Bioorg. Med. Chem. Lett. 2008, 18, 3517– 3521,  DOI: 10.1016/j.bmcl.2008.05.023
    Google Scholar
    28
    Discovery of benzamide tetrahydro-4H-carbazol-4-ones as novel small molecule inhibitors of Hsp90
    Barta, Thomas E.; Veal, James M.; Rice, John W.; Partridge, Jeffrey M.; Fadden, R. Patrick; Ma, Wei; Jenks, Matthew; Geng, Lifeng; Hanson, Gunnar J.; Huang, Kenneth H.; Barabasz, Amy F.; Foley, Briana E.; Otto, James; Hall, Steven E.
    Bioorganic & Medicinal Chemistry Letters (2008), 18 (12), 3517-3521CODEN: BMCLE8; ISSN:0960-894X. (Elsevier Ltd.)
    Hsp90 maintains the conformational stability of multiple proteins implicated in oncogenesis and has emerged as a target for chemotherapy. We report here the discovery of a novel small mol. scaffold that inhibits Hsp90. X-ray data show that the scaffold binds competitively at the ATP site on Hsp90. Cellular proliferation and client assays demonstrate that members of the series are able to inhibit Hsp90 at nanomolar concns.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BD1cXmvFSns74%253D&md5=f965e9a2cfd0d44a57bb1949cd87de63
  29. 29
    Erlanson, D. A. In Fragment-Based Drug Discovery and X-Ray Crystallography; Davies, T. G., Hyvönen, M., Eds.; Springer Berlin Heidelberg: Berlin, Heidelberg, 2012; pp 1– 32, DOI: 10.1007/128_2011_180 .
    Google Scholar
    There is no corresponding record for this reference.
  30. 30
    Ruiz-Carmona, S.; Schmidtke, P.; Luque, F. J.; Baker, L.; Matassova, N.; Davis, B.; Roughley, S.; Murray, J.; Hubbard, R.; Barril, X. Dynamic undocking and the quasi-bound state as tools for drug discovery. Nat. Chem. 2017, 9, 201,  DOI: 10.1038/nchem.2660
    Google Scholar
    30
    Dynamic undocking and the quasi-bound state as tools for drug discovery
    Ruiz-Carmona, Sergio; Schmidtke, Peter; Luque, F. Javier; Baker, Lisa; Matassova, Natalia; Davis, Ben; Roughley, Stephen; Murray, James; Hubbard, Rod; Barril, Xavier
    Nature Chemistry (2017), 9 (3), 201-206CODEN: NCAHBB; ISSN:1755-4330. (Nature Publishing Group)
    There is a pressing need for new technologies that improve the efficacy and efficiency of drug discovery. Structure-based methods have contributed towards this goal but they focus on predicting the binding affinity of protein-ligand complexes, which is notoriously difficult. The authors adopt an alternative approach that evaluates structural, rather than thermodn., stability. As bioactive mols. present a static binding mode, the authors devised dynamic undocking (DUck), a fast computational method to calc. the work necessary to reach a quasi-bound state at which the ligand has just broken the most important native contact with the receptor. This nonequil. property is surprisingly effective in virtual screening because true ligands form more-resilient interactions than decoys. Notably, DUck is orthogonal to docking and other 'thermodn.' methods. The authors demonstrate the potential of the docking-undocking combination in a fragment screening against the mol. chaperone and oncol. target Hsp90, for which the authors obtain novel chemotypes and a hit rate that approaches 40%.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC28XhvVGiur3I&md5=181235443427435bc9f3e8c17f8df9ee
  31. 31
    Hoxie, R. S.; Street, T. O. Hsp90 chaperones have an energetic hot-spot for binding inhibitors. Protein Sci. 2020, 29, 2101– 2111,  DOI: 10.1002/pro.3933
    Google Scholar
    31
    Hsp90 chaperones have an energetic hot-spot for binding inhibitors
    Hoxie, Reyal S.; Street, Timothy O.
    Protein Science (2020), 29 (10), 2101-2111CODEN: PRCIEI; ISSN:1469-896X. (Wiley-Blackwell)
    Although Hsp90-family chaperones have been extensively targeted with ATP-competitive inhibitors, it is unknown whether high affinity is achieved from a few highly stabilizing contacts or from many weaker contacts within the ATP-binding pocket. A large-scale anal. of Hsp90α:inhibitor structures shows that inhibitor hydrogen-bonding to a conserved aspartate (D93 in Hsp90α) stands out as most universal among Hsp90 inhibitors. Here we show that the D93 region makes a dominant energetic contribution to inhibitor binding for both cytosolic and organelle-specific Hsp90 paralogs. For inhibitors in the resorcinol family, the D93:inhibitor hydrogen-bond is pH-dependent because the assocd. inhibitor hydroxyl group is titratable, rationalizing a linked-protonation event previously obsd. by the Matulis group. The inhibitor hydroxyl group pKa assocd. with the D93 hydrogen-bond is therefore crit. for optimizing the affinity of resorcinol derivs., and we demonstrate that spectrophotometric measurements can det. this pKa value. Quantifying the energetic contribution of the D93 hotspot is best achieved with the mitochondrial Hsp90 paralog, yielding 3-6 kcal/mol of stabilization (35-60% of the total binding energy) for a diverse set of inhibitors. The Hsp90 Asp93→Asn substitution has long been known to abolish nucleotide binding, yet puzzlingly, native sequences of structurally similar ATPases, such as Topoisomerasese II, have an asparagine at this same crucial site. While aspartate and asparagine sidechains can both act as hydrogen bond acceptors, we show that a steric clash prevents the Hsp90 Asp93→Asn sidechain from adopting the necessary rotamer, whereas this steric restriction is absent in Topoisomerasese II.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXhslKhurfK&md5=1ab67c65a6e11c2cde6fd481ddef6f66
  32. 32
    Majewski, M.; Barril, X. Structural Stability Predicts the Binding Mode of Protein–Ligand Complexes. J. Chem. Inf. Model. 2020, 60, 1644– 1651,  DOI: 10.1021/acs.jcim.9b01062
    Google Scholar
    32
    Structural Stability Predicts the Binding Mode of Protein-Ligand Complexes
    Majewski, Maciej; Barril, Xavier
    Journal of Chemical Information and Modeling (2020), 60 (3), 1644-1651CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    The prediction of a ligand's binding mode into its macromol. target is essential in structure-based drug discovery. Even though tremendous effort has been made to address this problem, most of the developed tools work similarly, trying to predict the binding free energy assocd. with each particular binding mode. In this study, we decided to abandon this criterion, following structural stability instead. This view, implemented in a novel computational workflow, quantifies the steepness of the local energy min. assocd. with each potential binding mode. Surprisingly, the protocol outperforms docking scoring functions in case of fragments (ligands with MW < 300 Da) and is as good as docking for drug-like mols. It also identifies substructures that act as structural anchors, predicting their binding mode with particular accuracy. The results open a new phys. perspective for binding mode prediction, which can be combined with existing thermodn.-based approaches.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXivVaqsLk%253D&md5=e4f79145486f5552e96a9a7906160f83
  33. 33
    Hu, L.; Benson, M. L.; Smith, R. D.; Lerner, M. G.; Carlson, H. A. Binding MOAD (Mother of All Databases). Proteins: Struct., Funct., Genet. 2005, 60, 333– 340,  DOI: 10.1002/prot.20512
    Google Scholar
    33
    Binding MOAD (Mother of All Databases)
    Hu, Liegi; Benson, Mark L.; Smith, Richard D.; Lerner, Michael G.; Carlson, Heather A.
    Proteins: Structure, Function, and Bioinformatics (2005), 60 (3), 333-340CODEN: PSFBAF ISSN:. (Wiley-Liss, Inc.)
    Binding MOAD (Mother of All Databases) is the largest collection of high-quality, protein-ligand complexes available from the Protein Data Bank. At this time, Binding MOAD contains 5331 protein-ligand complexes comprised of 1780 unique protein families and 2630 unique ligands. We have searched the crystallog. papers for all 5000 + structures and compiled binding data for 1375 (26%) of the protein-ligand complexes. The binding-affinity data ranges 13 orders of magnitude. This is the largest collection of binding data reported to date in the literature. We have also addressed the issue of redundancy in the data. To create a nonredundant dataset, one protein from each of the 1780 protein families was chosen as a representative. Representatives were chosen by tightest binding, best resoln., etc. For the 1780 "best" complexes that comprise the nonredundant version of Binding MOAD, 475 (27%) have binding data. This significant collection of protein-ligand complexes will be very useful in elucidating the biophys. patterns of mol. recognition and enzymic regulation. The complexes with binding-affinity data will help in the development of improved scoring functions and structure-based drug discovery techniques. The dataset can be accessed at http://www.BindingMOAD.org.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BD2MXmvVyrtr4%253D&md5=146632a2f30fda98cb987e984beb7506
  34. 34
    Ruiz-Carmona, S.; Alvarez-Garcia, D.; Foloppe, N.; Garmendia-Doval, A. B.; Juhos, S.; Schmidtke, P.; Barril, X.; Hubbard, R. E.; Morley, S. D. rDock: A Fast, Versatile and Open Source Program for Docking Ligands to Proteins and Nucleic Acids. PLoS Comput. Biol. 2014, 10, e1003571  DOI: 10.1371/journal.pcbi.1003571
    Google Scholar
    34
    rDock: a fast, versatile and open source program for docking ligands to proteins and nucleic acids
    Ruiz-Carmona, Sergio; Alvarez-Garcia, Daniel; Foloppe, Nicolas; Garmendia-Doval, A. Beatriz; Juhos, Szilveszter; Schmidtke, Peter; Barril, Xavier; Hubbard, Roderick E.; Morley, S. David
    PLoS Computational Biology (2014), 10 (4), e1003571/1-e1003571/7, 7 pp.CODEN: PCBLBG; ISSN:1553-7358. (Public Library of Science)
    Identification of chem. compds. with specific biol. activities is an important step in both chem. biol. and drug discovery. When the structure of the intended target is available, one approach is to use mol. docking programs to assess the chem. complementarity of small mols. with the target; such calcns. provide a qual. measure of affinity that can be used in virtual screening (VS) to rank order a list of compds. according to their potential to be active. rDock is a mol. docking program developed at Vernalis for high-throughput VS (HTVS) applications. Evolved from RiboDock, the program can be used against proteins and nucleic acids, is designed to be computationally very efficient and allows the user to incorporate addnl. constraints and information as a bias to guide docking. This article provides an overview of the program structure and features and compares rDock to two ref. programs, AutoDock Vina (open source) and Schrodinger's Glide (com.). In terms of computational speed for VS, rDock is faster than Vina and comparable to Glide. For binding mode prediction, rDock and Vina are superior to Glide. The VS performance of rDock is significantly better than Vina, but inferior to Glide for most systems unless pharmacophore constraints are used; in that case rDock and Glide are of equal performance. The program is released under the Lesser General Public License and is freely available for download, together with the manuals, example files and the complete test sets, at online.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC2cXhsVGlsL%252FO&md5=e4cb786d6567fdc7f2a46f64955a9992

Cited By

Click to copy section linkSection link copied!

This article is cited by 11 publications.

  1. Yunjiang Zhang, Shuyuan Li, Kong Meng, Shaorui Sun. Machine Learning for Sequence and Structure-Based Protein–Ligand Interaction Prediction. Journal of Chemical Information and Modeling 2024, 64 (5) , 1456-1472. https://doi.org/10.1021/acs.jcim.3c01841
  2. Erin C. Day, Supraja S. Chittari, Matthew P. Bogen, Abigail S. Knight. Navigating the Expansive Landscapes of Soft Materials: A User Guide for High-Throughput Workflows. ACS Polymers Au 2023, 3 (6) , 406-427. https://doi.org/10.1021/acspolymersau.3c00025
  3. Vivek Kumar, Pawan Gupta, Vishnu Sharma, Anju Dhiman. Multi-target-based screening of phytochemicals found in aerial parts of Heliotropium indicum L. for identification of potential anti-urolithiatic agents using simulation methods. 2024https://doi.org/10.21203/rs.3.rs-5144782/v1
  4. Pawan Gupta, Umesh Panwar, Sanjeev Singh. Novel scaffolds identification against Mpro of SARS-CoV-2 using shape based screening and molecular simulation methods. Chemical Physics Impact 2024, 8 , 100496. https://doi.org/10.1016/j.chphi.2024.100496
  5. Chengcheng Zhang, Xiaoxue Zhao, Feng Li, Jingru Qin, Lu Yang, Qianqian Yin, Yiyi Liu, Zhiyao Zhu, Fei Zhang, Zhongqi Wang, Haibin Liang. Integrating single‐cell and multi‐omic approaches reveals Euphorbiae Humifusae Herba‐dependent mitochondrial dysfunction in non‐small‐cell lung cancer. Journal of Cellular and Molecular Medicine 2024, 28 (10) https://doi.org/10.1111/jcmm.18317
  6. Ignacio Ponzoni, Juan Antonio Páez Prosper, Nuria E. Campillo. Explainable artificial intelligence: A taxonomy and guidelines for its application to drug discovery. WIREs Computational Molecular Science 2023, 13 (6) https://doi.org/10.1002/wcms.1681
  7. Akey Krishna Swaroop, P.K. Krishnan Namboori, M. Esakkimuthukumar, T.K. Praveen, Palathoti Nagarjuna, Sunil Kumar Patnaik, Jubie Selvaraj. Leveraging decagonal in-silico strategies for uncovering IL-6 inhibitors with precision. Computers in Biology and Medicine 2023, 163 , 107231. https://doi.org/10.1016/j.compbiomed.2023.107231
  8. Peter G. Bolhuis, Z. Faidon Brotzakis, Bettina G. Keller. Optimizing molecular potential models by imposing kinetic constraints with path reweighting. The Journal of Chemical Physics 2023, 159 (7) https://doi.org/10.1063/5.0151166
  9. Vivek Kumar, Pawan Gupta, Vishnu Sharma, Anjana Munshi, Anju Dhiman. Multi-target based virtual screening of phytochemicals from Heliotropium indicum L. leaves for identification of potential anti-urolithiatic agent. 2022https://doi.org/10.21203/rs.3.rs-2387425/v1
  10. Sandhya Vivekanandan, Umashankar Vetrivel, Luke Elizabeth Hanna. Design of human immunodeficiency virus-1 neutralizing peptides targeting CD4-binding site: An integrative computational biologics approach. Frontiers in Medicine 2022, 9 https://doi.org/10.3389/fmed.2022.1036874
  11. Gergely Hajgató, Richárd Wéber, Botond Szilágyi, Balázs Tóthpál, Bálint Gyires-Tóth, Csaba Hős. PredMaX: Predictive maintenance with explainable deep convolutional autoencoders. Advanced Engineering Informatics 2022, 54 , 101778. https://doi.org/10.1016/j.aei.2022.101778
  12. Rocco Meli, Garrett M. Morris, Philip C. Biggin. Scoring Functions for Protein-Ligand Binding Affinity Prediction Using Structure-based Deep Learning: A Review. Frontiers in Bioinformatics 2022, 2 https://doi.org/10.3389/fbinf.2022.885983
  13. Jun Zhang. Atom typing using graph representation learning: How do models learn chemistry?. The Journal of Chemical Physics 2022, 156 (20) https://doi.org/10.1063/5.0095008
Download PDF
Go to Journal of Chemical Information and Modeling
Get e-Alerts
Get e-Alerts

Journal of Chemical Information and Modeling

Cite this: J. Chem. Inf. Model. 2022, 62, 2, 225–231
Click to copy citationCitation copied!
https://doi.org/10.1021/acs.jcim.1c00691
Published January 3, 2022

Copyright © 2022 The Authors. Published by American Chemical Society. This publication is licensed under

CC-BY 4.0 .

Article Views

4234

Altmetric

-

Citations

11
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. Main view of the graphical user interface. The protein–ligand complex is displayed, with the attributions of the most contributing voxels superimposed. The attributions for the different channels can be seen individually using the corresponding sliders in the menu on the right, which display isosurfaces at different isovalues. The full protein is shown in a cartoon representation, while residues in the binding site (defined by being within 5 Å to the ligand) are shown in a transparent ball–stick representation (only heavy atoms and polar hydrogens). The all-atom representation of the ligand is shown in a bold ball–stick. The region of space seen by the model (voxelization cube) is delimited by a transparent, gray box.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 2

    Figure 2. Comparison of computed attributions obtained for a complex of HSP90 with an analogue of benzamide tetrahydro-4H-carbazol-4-one (PDB code: 3D0B) by the three models: Clash detector (1A and 1B), Pose classifier (2A and 2B), and KDEEP (3A and 3B). Pictures on the top row show the attributions for the protein and ligand occupancy channels, in red and blue, respectively. The bottom row focuses on particular interactions. 1B shows a clash between the ligand and the leucine and the attributions for the occupancy channels of protein and ligand (red and blue). 2B and 3B show the hydrogen bond between the benzamide moiety in the ligand and the aspartate (D93) residue in the protein. Attributions for the ligand donor channel are shown in pink, while for the protein acceptor channel, the attributions are shown in blue.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 3

    Figure 3. Distance distribution between the two voxels with the highest, absolute value in protein and ligand occupancy channels.

    High Resolution Image
    Download MS PowerPoint Slide

  • References

    This article references 34 other publications.

    1. 1
      Dudek, A. Z.; Arodz, T.; Gálvez, J. Computational methods in developing quantitative structure-activity relationships (QSAR): a review. Comb. Chem. High Throughput Screening 2006, 9, 213– 228,  DOI: 10.2174/138620706776055539
      1
      Computational methods in developing quantitative structure-activity relationships (QSAR): a review
      Dudek, Arkadiusz Z.; Arodz, Tomasz; Galvez, Jorge
      Combinatorial Chemistry & High Throughput Screening (2006), 9 (3), 213-228CODEN: CCHSFU; ISSN:1386-2073. (Bentham Science Publishers Ltd.)
      A review. Virtual filtering and screening of combinatorial libraries have recently gained attention as methods complementing the high-throughput screening and combinatorial chem. These chemoinformatic techniques rely heavily on quant. structure-activity relation (QSAR) anal., a field with established methodol. and successful history. In this review, we discuss the computational methods for building QSAR models. We start with outlining their usefulness in high-throughput screening and identifying the general scheme of a QSAR model. Following, we focus on the methodologies in constructing three main components of QSAR model, namely the methods for describing the mol. structure of compds., for selection of informative descriptors and for activity prediction. We present both the well-established methods as well as techniques recently introduced into the QSAR domain.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BD28XisVSqtbo%253D&md5=230cc86a6228a8f9a97c666e925a5efe
    2. 2
      Cherkasov, A.; Muratov, E. N.; Fourches, D.; Varnek, A.; Baskin, I. I.; Cronin, M.; Dearden, J.; Gramatica, P.; Martin, Y. C.; Todeschini, R. QSAR modeling: where have you been? Where are you going to?. J. Med. Chem. 2014, 57, 4977– 5010,  DOI: 10.1021/jm4004285
      2
      QSAR Modeling: Where Have You Been? Where Are You Going To?
      Cherkasov, Artem; Muratov, Eugene N.; Fourches, Denis; Varnek, Alexandre; Baskin, Igor I.; Cronin, Mark; Dearden, John; Gramatica, Paola; Martin, Yvonne C.; Todeschini, Roberto; Consonni, Viviana; Kuzmin, Victor E.; Cramer, Richard; Benigni, Romualdo; Yang, Chihae; Rathman, James; Terfloth, Lothar; Gasteiger, Johann; Richard, Ann; Tropsha, Alexander
      Journal of Medicinal Chemistry (2014), 57 (12), 4977-5010CODEN: JMCMAR; ISSN:0022-2623. (American Chemical Society)
      A review. Quant. structure-activity relationship modeling is one of the major computational tools employed in medicinal chem. However, throughout its entire history it has drawn both praise and criticism concerning its reliability, limitations, successes, and failures. In this paper, we discuss (i) the development and evolution of QSAR; (ii) the current trends, unsolved problems, and pressing challenges; and (iii) several novel and emerging applications of QSAR modeling. Throughout this discussion, we provide guidelines for QSAR development, validation, and application, which are summarized in best practices for building rigorously validated and externally predictive QSAR models. We hope that this Perspective will help communications between computational and exptl. chemists toward collaborative development and use of QSAR models. We also believe that the guidelines presented here will help journal editors and reviewers apply more stringent scientific stds. to manuscripts reporting new QSAR studies, as well as encourage the use of high quality, validated QSARs for regulatory decision making.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC3sXhvFCrsLnE&md5=a2d371ee33fbecf48ef59ecf914e4aa8
    3. 3
      Lo, Y.-C.; Rensi, S. E.; Torng, W.; Altman, R. B. Machine learning in chemoinformatics and drug discovery. Drug Discovery Today 2018, 23, 1538– 1546,  DOI: 10.1016/j.drudis.2018.05.010
      3
      Machine learning in chemoinformatics and drug discovery
      Lo, Yu-Chen; Rensi, Stefano E.; Torng, Wen; Altman, Russ B.
      Drug Discovery Today (2018), 23 (8), 1538-1546CODEN: DDTOFS; ISSN:1359-6446. (Elsevier Ltd.)
      Chemoinformatics is an established discipline focusing on extg., processing and extrapolating meaningful data from chem. structures. With the rapid explosion of chem. 'big' data from HTS and combinatorial synthesis, machine learning has become an indispensable tool for drug designers to mine chem. information from large compd. databases to design drugs with important biol. properties. To process the chem. data, we first reviewed multiple processing layers in the chemoinformatics pipeline followed by the introduction of commonly used machine learning models in drug discovery and QSAR anal. Here, we present basic principles and recent case studies to demonstrate the utility of machine learning techniques in chemoinformatics analyses; and we discuss limitations and future directions to guide further development in this evolving field.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXptFCnuro%253D&md5=d0bef55f60c8ee7bb2a31e42d4a85fcd
    4. 4
      Neves, B. J.; Braga, R. C.; Melo-Filho, C. C.; Moreira-Filho, J. T.; Muratov, E. N.; Andrade, C. H. QSAR-Based Virtual Screening: Advances and Applications in Drug Discovery. Front. Pharmacol. 2018, 9, 1275,  DOI: 10.3389/fphar.2018.01275
      4
      QSAR-based virtual screening: advances and applications in drug discovery
      Neves, Bruno J.; Braga, Rodolpho C.; Melo-Filho, Cleber C.; Moreira-Filho, Jose Teofilo; Muratov, Eugene N.; Andrade, Carolina Horta
      Frontiers in Pharmacology (2018), 9 (), 1275CODEN: FPRHAU; ISSN:1663-9812. (Frontiers Media S.A.)
      A review. Virtual screening (VS) has emerged in drug discovery as a powerful computational approach to screen large libraries of small mols. for new hits with desired properties that can then be tested exptl. Similar to other computational approaches, VS intention is not to replace in vitro or in vivo assays, but to speed up the discovery process, to reduce the no. of candidates to be tested exptl., and to rationalize their choice. Moreover, VS has become very popular in pharmaceutical companies and academic organizations due to its time-, cost-, resources-, and laborsaving. Among the VS approaches, quant. structure-activity relationship (QSAR) anal. is the most powerful method due to its high and fast throughput and good hit rate. As the first preliminary step of a QSAR model development, relevant chemogenomics data are collected from databases and the literature. Then, chem. descriptors are calcd. on different levels of representation of mol. structure, ranging from 1D to nD, and then correlated with the biol. property using machine learning techniques. Once developed and validated, QSAR models are applied to predict the biol. property of novel compds. Although the exptl. testing of computational hits is not an inherent part of QSAR methodol., it is highly desired and should be performed as an ultimate validation of developed models. In this minireview, we summarize and critically analyze the recent trends of QSAR-based VS in drug discovery and demonstrate successful applications in identifying perspective compds. with desired properties. Moreover, we provide some recommendations about the best practices for QSAR-based VS along with the future perspectives of this approach.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXhtVGmtbnP&md5=1aa305bf00e34d3ac592d19a7e107d7e
    5. 5
      Zhang, L.; Zhang, H.; Ai, H.; Hu, H.; Li, S.; Zhao, J.; Liu, H. Applications of Machine Learning Methods in Drug Toxicity Prediction. Curr. Top. Med. Chem. (Sharjah, United Arab Emirates) 2018, 18, 987– 997,  DOI: 10.2174/1568026618666180727152557
      5
      Applications of Machine Learning Methods in Drug Toxicity Prediction
      Zhang, Li; Zhang, Hui; Ai, Haixin; Hu, Huan; Li, Shimeng; Zhao, Jian; Liu, Hongsheng
      Current Topics in Medicinal Chemistry (Sharjah, United Arab Emirates) (2018), 18 (12), 987-997CODEN: CTMCCL; ISSN:1568-0266. (Bentham Science Publishers Ltd.)
      Toxicity evaluation is an important part of the preclin. safety assessment of new drugs, which is directly related to human health and the fate of drugs. It is of importance to study how to evaluate drug toxicity accurately and economically. The traditional in vitro and in vivo toxicity tests are laborious, time-consuming, highly expensive, and even involve animal welfare issues. Computational methods developed for drug toxicity prediction can compensate for the shortcomings of traditional methods and have been considered useful in the early stages of drug development. Numerous drug toxicity prediction models have been developed using a variety of computational methods. With the advance of the theory of machine learning and mol. representation, more and more drug toxicity prediction models are developed using a variety of machine learning methods, such as support vector machine, random forest, naive Bayesian, back propagation neural network. And significant advances have been made in many toxicity endpoints, such as carcinogenicity, mutagenicity, and hepatotoxicity. In this review, we aimed to provide a comprehensive overview of the machine learning based drug toxicity prediction studies conducted in recent years. In addn., we compared the performance of the models proposed in these studies in terms of accuracy, sensitivity, and specificity, providing a view of the current state-of-the-art in this field and highlighting the issues in the current studies.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXhvVejt7fO&md5=e98c7844a12ead052509399b9cf9c6b4
    6. 6
      Ma, H.; An, W.; Wang, Y.; Sun, H.; Huang, R.; Huang, J. Deep Graph Learning with Property Augmentation for Predicting Drug-Induced Liver Injury. Chem. Res. Toxicol. 2021, 34, 495,  DOI: 10.1021/acs.chemrestox.0c00322
      6
      Deep Graph Learning with Property Augmentation for Predicting Drug-Induced Liver Injury
      Ma, Hehuan; An, Weizhi; Wang, Yuhong; Sun, Hongmao; Huang, Ruili; Huang, Junzhou
      Chemical Research in Toxicology (2021), 34 (2), 495-506CODEN: CRTOEC; ISSN:0893-228X. (American Chemical Society)
      Drug-induced liver injury (DILI) is a crucial factor in detg. the qualification of potential drugs. However, the DILI property is excessively difficult to obtain due to the complex testing process. Consequently, an in silico screening in the early stage of drug discovery would help to reduce the total development cost by filtering those drug candidates with a high risk to cause DILI. To serve the screening goal, we apply several computational techniques to predict the DILI property, including traditional machine learning methods and graph-based deep learning techniques. While deep learning models require large training data to tune huge model parameters, the DILI data set only contains a few hundred annotated mols. To alleviate the data scarcity problem, we propose a property augmentation strategy to include massive training data with other property information. Extensive expts. demonstrate that our proposed method significantly outperforms all existing baselines on the DILI data set by obtaining a 81.4% accuracy using cross-validation with random splitting, 78.7% using leave-one-out cross-validation, and 76.5% using cross-validation with scaffold splitting.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXis1ehsrrI&md5=1eb843cf35938405c37027987f1daea4
    7. 7
      Montanari, F.; Kuhnke, L.; Ter Laak, A.; Clevert, D.-A. Modeling physico-chemical ADMET endpoints with multitask graph convolutional networks. Molecules 2020, 25, 44,  DOI: 10.3390/molecules25010044
      7
      Modeling physico-chemical ADMET endpoints with multitask graph convolutional networks
      Montanari, Floriane; Kuhnke, Lara; Ter Laak, Antonius; Clevert, Djork-Arne
      Molecules (2020), 25 (1), 44CODEN: MOLEFW; ISSN:1420-3049. (MDPI AG)
      Simple physico-chem. properties, like logD, soly., or m.p., can reveal a great deal about how a compd. under development might later behave. These data are typically measured for most compds. in drug discovery projects in a medium throughput fashion. Collecting and assembling all the Bayer inhouse data related to these properties allowed us to apply powerful machine learning techniques to predict the outcome of those assays for new compds. In this paper, we report our finding that, esp. for predicting physicochem. ADMET endpoints, a multitask graph convolutional approach appears a highly competitive choice. For seven endpoints of interest, we compared the performance of that approach to fully connected neural networks and different single task models. The new model shows increased predictive performance compared to previous modeling methods and will allow early prioritization of compds. even before they are synthesized. In addn., our model follows the generalized soly. equation without being explicitly trained under this constraint.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXjsFWqtLk%253D&md5=62240c8f1b151bab7a95d51e20b6c6a5
    8. 8
      Peng, Y.; Lin, Y.; Jing, X.-Y.; Zhang, H.; Huang, Y.; Luo, G. S. Enhanced Graph Isomorphism Network for Molecular ADMET Properties Prediction. IEEE Access 2020, 8, 168344– 168360,  DOI: 10.1109/ACCESS.2020.3022850
      There is no corresponding record for this reference.
    9. 9
      Feinberg, E. N.; Joshi, E.; Pande, V. S.; Cheng, A. C. Improvement in ADMET prediction with multitask deep featurization. J. Med. Chem. 2020, 63, 8835– 8848,  DOI: 10.1021/acs.jmedchem.9b02187
      9
      Improvement in ADMET Prediction with Multitask Deep Featurization
      Feinberg, Evan N.; Joshi, Elizabeth; Pande, Vijay S.; Cheng, Alan C.
      Journal of Medicinal Chemistry (2020), 63 (16), 8835-8848CODEN: JMCMAR; ISSN:0022-2623. (American Chemical Society)
      The absorption, distribution, metab., elimination, and toxicity (ADMET) properties of drug candidates are important for their efficacy and safety as therapeutics. Predicting ADMET properties has therefore been of great interest to the computational chem. and medicinal chem. communities in recent decades. Traditional cheminformatics approaches, using learners such as random forests and deep neural networks, leverage fingerprint feature representations of mols. Here, we learn the features most relevant to each chem. task at hand by representing each mol. explicitly as a graph. By applying graph convolutions to this explicit mol. representation, we achieve, to our knowledge, unprecedented accuracy in prediction of ADMET properties. By challenging our methodol. with rigorous cross-validation procedures and prognostic analyses, we show that deep featurization better enables mol. predictors to not only interpolate but also extrapolate to new regions of chem. space.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXntFCgsLs%253D&md5=845b33af28bfa7b817187c9e423a21eb
    10. 10
      Skalic, M.; Varela-Rial, A.; Jiménez, J.; Martínez-Rosell, G.; De Fabritiis, G. LigVoxel: inpainting binding pockets using 3D-convolutional neural networks. Bioinformatics 2019, 35, 243– 250,  DOI: 10.1093/bioinformatics/bty583
      10
      LigVoxel: inpainting binding pockets using 3D-convolutional neural networks
      Skalic, Miha; Varela-Rial, Alejandro; Jimenez, Jose; Martinez-Rosell, Gerard; Fabritiis, Gianni De
      Bioinformatics (2019), 35 (2), 243-250CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
      Motivation: Structure-based drug discovery methods exploit protein structural information to design small mols. binding to given protein pockets. This work proposes a purely data driven, structure-based approach for imaging ligands as spatial fields in target protein pockets. We use an end-to-end deep learning framework trained on exptl. protein-ligand complexes with the intention of mimicking a chemist's intuition at manually placing atoms when designing a new compd. We show that these models can generate spatial images of ligand chem. properties like occupancy, aromaticity and donor-acceptor matching the protein pocket. Results: The predicted fields considerably overlap with those of unseen ligands bound to the target pocket. Maximization of the overlap between the predicted fields and a given ligand on the Astex diverse set recovers the original ligand crystal poses in 70 out of 85 cases within a threshold of 2 Å RMSD. We expect that these models can be used for guiding structure-based drug discovery approaches.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXitVWgs73N&md5=e8bf4189689b73707a78dc94d58fa7af
    11. 11
      Jiménez, J.; Škalič, M.; Martínez-Rosell, G.; De Fabritiis, G. KDEEP: Protein-Ligand Absolute Binding Affinity Prediction via 3D-Convolutional Neural Networks. J. Chem. Inf. Model. 2018, 58, 287– 296,  DOI: 10.1021/acs.jcim.7b00650
      11
      KDEEP: Protein-Ligand Absolute Binding Affinity Prediction via 3D-Convolutional Neural Networks
      Jimenez, Jose; Skalic, Miha; Martinez-Rosell, Gerard; De Fabritiis, Gianni
      Journal of Chemical Information and Modeling (2018), 58 (2), 287-296CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      Accurately predicting protein-ligand binding affinities is an important problem in computational chem. since it can substantially accelerate drug discovery for virtual screening and lead optimization. We propose here a fast machine-learning approach for predicting binding affinities using state-of-the-art 3D-convolutional neural networks and compare this approach to other machine-learning and scoring methods using several diverse data sets. The results for the std. PDBbind (v.2016) core test-set are state-of-the-art with a Pearson's correlation coeff. of 0.82 and a RMSE of 1.27 in pK units between exptl. and predicted affinity, but accuracy is still very sensitive to the specific protein used. KDEEP is made available via PlayMol.org for users to test easily their own protein-ligand complexes, with each prediction taking a fraction of a second. We believe that the speed, performance, and ease of use of KDEEP makes it already an attractive scoring function for modern computational chem. pipelines.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXkslSitQ%253D%253D&md5=81943a6732be99e5439e1e3a25fa4414
    12. 12
      Ragoza, M.; Hochuli, J.; Idrobo, E.; Sunseri, J.; Koes, D. R. Protein-Ligand Scoring with Convolutional Neural Networks. J. Chem. Inf. Model. 2017, 57, 942– 957,  DOI: 10.1021/acs.jcim.6b00740
      12
      Protein-Ligand Scoring with Convolutional Neural Networks
      Ragoza, Matthew; Hochuli, Joshua; Idrobo, Elisa; Sunseri, Jocelyn; Koes, David Ryan
      Journal of Chemical Information and Modeling (2017), 57 (4), 942-957CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      Computational approaches to drug discovery can reduce the time and cost assocd. with exptl. assays and enable the screening of novel chemotypes. Structure-based drug design methods rely on scoring functions to rank and predict binding affinities and poses. The ever-expanding amt. of protein-ligand binding and structural data enables the use of deep machine learning techniques for protein-ligand scoring. We describe convolutional neural network (CNN) scoring functions that take as input a comprehensive three-dimensional (3D) representation of a protein-ligand interaction. A CNN scoring function automatically learns the key features of protein-ligand interactions that correlate with binding. We train and optimize our CNN scoring functions to discriminate between correct and incorrect binding poses and known binders and nonbinders. We find that our CNN scoring function outperforms the AutoDock Vina scoring function when ranking poses both for pose prediction and virtual screening.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC2sXlsVems7Y%253D&md5=9cae97167da0c93e1d896f85339cca7e
    13. 13
      Skalic, M.; Martínez-Rosell, G.; Jiménez, J.; De Fabritiis, G. PlayMolecule BindScope: large scale CNN-based virtual screening on the web. Bioinformatics 2019, 35, 1237– 1238,  DOI: 10.1093/bioinformatics/bty758
      13
      PlayMolecule BindScope: large scale CNN-based virtual screening on the web
      Skalic, Miha; Martinez-Rosell, Gerard; Jimenez, Jose; De Fabritiis, Gianni
      Bioinformatics (2019), 35 (7), 1237-1238CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
      Summary: Virtual screening pipelines are one of the most popular used tools in structure-based drug discovery, since they can can reduce both time and cost assocd. with exptl. assays. Recent advances in deep learning methodologies have shown that these outperform classical scoring functions at discriminating binder protein-ligand complexes. Here, we present BindScope, a web application for large-scale active-inactive classification of compds. based on deep convolutional neural networks. Performance is on a pair with current state-of-the-art pipelines. Users can screen on the order of hundreds of compds. at once and interactively visualize the results.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXitFWksb7K&md5=100cce3db2446853ed2f7c82b96d49df
    14. 14
      Yang, S.; Lee, K. H.; Ryu, S. A comprehensive study on the prediction reliability of graph neural networks for virtual screening. 2020, arXiv:2003.07611, arXiv preprint. https://arxiv.org/abs/2003.07611 (accessed 2021-10-29).
      There is no corresponding record for this reference.
    15. 15
      Sakai, M.; Nagayasu, K.; Shibui, N.; Andoh, C.; Takayama, K.; Shirakawa, H.; Kaneko, S. Prediction of pharmacological activities from chemical structures with graph convolutional neural networks. Sci. Rep. 2021, 11, 525,  DOI: 10.1038/s41598-020-80113-7
      15
      Prediction of pharmacological activities from chemical structures with graph convolutional neural networks
      Sakai, Miyuki; Nagayasu, Kazuki; Shibui, Norihiro; Andoh, Chihiro; Takayama, Kaito; Shirakawa, Hisashi; Kaneko, Shuji
      Scientific Reports (2021), 11 (1), 525CODEN: SRCEC3; ISSN:2045-2322. (Nature Research)
      Many therapeutic drugs are compds. that can be represented by simple chem. structures, which contain important determinants of affinity at the site of action. Recently, graph convolutional neural network (GCN) models have exhibited excellent results in classifying the activity of such compds. For models that make quant. predictions of activity, more complex information has been utilized, such as the three-dimensional structures of compds. and the amino acid sequences of their resp. target proteins. As another approach, we hypothesized that if sufficient exptl. data were available and there were enough nodes in hidden layers, a simple compd. representation would quant. predict activity with satisfactory accuracy. In this study, we report that GCN models constructed solely from the two-dimensional structural information of compds. demonstrated a high degree of activity predictability against 127 diverse targets from the ChEMBL database. Using the information entropy as a metric, we also show that the structural diversity had less effect on the prediction performance. Finally, we report that virtual screening using the constructed model identified a new serotonin transporter inhibitor with activity comparable to that of a marketed drug in vitro and exhibited antidepressant effects in behavioral studies.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXhsVemsLo%253D&md5=dfd62a89f9f705f4e740544f40ce705c
    16. 16
      Sieg, J.; Flachsenberg, F.; Rarey, M. Need of Bias Control: Evaluating Chemical Data for Machine Learning in Structure-Based Virtual Screening. J. Chem. Inf. Model. 2019, 59, 947– 961,  DOI: 10.1021/acs.jcim.8b00712
      16
      In Need of Bias Control: Evaluating Chemical Data for Machine Learning in Structure-Based Virtual Screening
      Sieg, Jochen; Flachsenberg, Florian; Rarey, Matthias
      Journal of Chemical Information and Modeling (2019), 59 (3), 947-961CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      A review. Reports of successful applications of machine learning (ML) methods in structure-based virtual screening (SBVS) are increasing. ML methods such as convolutional neural networks show promising results and often outperform traditional methods such as empirical scoring functions in retrospective validation. However, trained ML models are often treated as black boxes and are not straightforwardly interpretable. In most cases, it is unknown which features in the data are decisive and whether a model's predictions are right for the right reason. Hence, the authors reevaluated three widely used benchmark data sets in the context of ML methods and came to the conclusion that not every benchmark data set is suitable. Moreover, the authors demonstrate on two examples from current literature that bias is learned implicitly and unnoticed from std. benchmarks. On the basis of these results, the authors conclude that there is a need for eligible validation expts. and benchmark data sets suited to ML for more bias-controlled validation in ML-based SBVS. Therefore, the authors provide guidelines for setting up validation expts. and give a perspective on how new data sets could be generated.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXktVOgsrY%253D&md5=6cb48f2acd541c10c8565f202a6eb9e8
    17. 17
      DeGrave, A. J.; Janizek, J. D.; Lee, S.-I. AI for radiographic COVID-19 detection selects shortcuts over signal. Nat. Mach. Intell. 2021, 3, 610,  DOI: 10.1038/s42256-021-00338-7
      There is no corresponding record for this reference.
    18. 18
      Vamathevan, J.; Clark, D.; Czodrowski, P.; Dunham, I.; Ferran, E.; Lee, G.; Li, B.; Madabhushi, A.; Shah, P.; Spitzer, M.; Zhao, S. Applications of machine learning in drug discovery and development. Nat. Rev. Drug Discovery 2019, 18, 463– 477,  DOI: 10.1038/s41573-019-0024-5
      18
      Applications of machine learning in drug discovery and development
      Vamathevan, Jessica; Clark, Dominic; Czodrowski, Paul; Dunham, Ian; Ferran, Edgardo; Lee, George; Li, Bin; Madabhushi, Anant; Shah, Parantu; Spitzer, Michaela; Zhao, Shanrong
      Nature Reviews Drug Discovery (2019), 18 (6), 463-477CODEN: NRDDAG; ISSN:1474-1776. (Nature Research)
      A review. Drug discovery and development pipelines are long, complex and depend on numerous factors. Machine learning (ML) approaches provide a set of tools that can improve discovery and decision making for well-specified questions with abundant, high-quality data. Opportunities to apply ML occur in all stages of drug discovery. Examples include target validation, identification of prognostic biomarkers and anal. of digital pathol. data in clin. trials. Applications have ranged in context and methodol., with some approaches yielding accurate predictions and insights. The challenges of applying ML lie primarily with the lack of interpretability and repeatability of ML-generated results, which may limit their application. In all areas, systematic and comprehensive high-dimensional data still need to be generated. With ongoing efforts to tackle these issues, as well as increasing awareness of the factors needed to validate ML approaches, the application of ML can promote data-driven decision making and has the potential to speed up the process and reduce failure rates in drug discovery and development.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXosF2rtrY%253D&md5=211782aeea3d8b9f50368f89177a70d2
    19. 19
      Sundararajan, M.; Taly, A.; Yan, Q. Axiomatic Attribution for Deep Networks. 2017, arXiv preprint. https://arxiv.org/abs/1703.01365 (accessed 2021-10-29).
      There is no corresponding record for this reference.
    20. 20
      Henderson, R.; Clevert, D.-A.; Montanari, F. Improving Molecular Graph Neural Network Explainability with Orthonormalization and Induced Sparsity. Proceedings of the 38 th International Conference on Machine Learning ; 2021.
      There is no corresponding record for this reference.
    21. 21
      Kokhlikyan, N.; Miglani, V.; Martin, M.; Wang, E.; Alsallakh, B.; Reynolds, J.; Melnikov, A.; Kliushkina, N.; Araya, C.; Yan, S.; Reblitz-Richardson, O. Captum: A unified and generic model interpretability library for PyTorch. 2020, arXiv preprint. https://arxiv.org/abs/2009.07896 (accessed 2021-06-17).
      There is no corresponding record for this reference.
    22. 22
      Klaise, J.; Van Looveren, A.; Vacanti, G.; Coca, A. Alibi: Algorithms for monitoring and explaining machine learning models ; 2019. https://github.com/SeldonIO/alibi (accessed 2021-09-08).
      There is no corresponding record for this reference.
    23. 23
      Hochuli, J.; Helbling, A.; Skaist, T.; Ragoza, M.; Koes, D. R. Visualizing convolutional neural network protein-ligand scoring. J. Mol. Graphics Modell. 2018, 84, 96– 108,  DOI: 10.1016/j.jmgm.2018.06.005
      23
      Visualizing convolutional neural network protein-ligand scoring
      Hochuli, Joshua; Helbling, Alec; Skaist, Tamar; Ragoza, Matthew; Koes, David Ryan
      Journal of Molecular Graphics & Modelling (2018), 84 (), 96-108CODEN: JMGMFI; ISSN:1093-3263. (Elsevier Ltd.)
      Protein-ligand scoring is an important step in a structure-based drug design pipeline. Selecting a correct binding pose and predicting the binding affinity of a protein-ligand complex enables effective virtual screening. Machine learning techniques can make use of the increasing amts. of structural data that are becoming publicly available. Convolutional neural network (CNN) scoring functions in particular have shown promise in pose selection and affinity prediction for protein-ligand complexes. Neural networks are known for being difficult to interpret. Understanding the decisions of a particular network can help tune parameters and training data to maximize performance. Visualization of neural networks helps decomp. complex scoring functions into pictures that are more easily parsed by humans. Here we present three methods for visualizing how individual protein-ligand complexes are interpreted by 3D convolutional neural networks. We also present a visualization of the convolutional filters and their wts. We describe how the intuition provided by these visualizations aids in network design.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXhtF2ltb3P&md5=db697d859132f02c4a2cef4bfc00ea88
    24. 24
      Jiménez, J.; Doerr, S.; Martínez-Rosell, G.; Rose, A. S.; De Fabritiis, G. DeepSite: protein-binding site predictor using 3D-convolutional neural networks. Bioinformatics 2017, 33, 3036– 3042,  DOI: 10.1093/bioinformatics/btx350
      24
      DeepSite: protein-binding site predictor using 3D-convolutional neural networks
      Jimenez, J.; Doerr, S.; Martinez-Rosell, G.; Rose, A. S.; De Fabritiis, G.
      Bioinformatics (2017), 33 (19), 3036-3042CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
      Motivation: An important step in structure-based drug design consists in the prediction of druggable binding sites. Several algorithms for detecting binding cavities, those likely to bind to a small drug compd., have been developed over the years by clever exploitation of geometric, chem. and evolutionary features of the protein. Results: Here we present a novel knowledge-based approach that uses state-of-the-art convolutional neural networks, where the algorithm is learned by examples. In total, 7622 proteins from the scPDB database of binding sites have been evaluated using both a distance and a volumetric overlap approach. Our machine-learning based method demonstrates superior performance to two other competitive algorithmic strategies. Availability and implementation: DeepSite is freely available at www.playmol.org. Users can submit either a PDB ID or PDB file for pocket detection to our NVIDIA GPU-equipped servers through a WebGL graphical interface.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXhvFGju7nN&md5=fce2763a18192d7a7c4c98fcb2974a22
    25. 25
      Liu, Z.; Li, Y.; Han, L.; Li, J.; Liu, J.; Zhao, Z.; Nie, W.; Liu, Y.; Wang, R. PDB-wide collection of binding data: current status of the PDBbind database. Bioinformatics 2015, 31, 405– 412,  DOI: 10.1093/bioinformatics/btu626
      25
      PDB-wide collection of binding data: current status of the PDBbind database
      Liu, Zhihai; Li, Yan; Han, Li; Li, Jie; Liu, Jie; Zhao, Zhixiong; Nie, Wei; Liu, Yuchen; Wang, Renxiao
      Bioinformatics (2015), 31 (3), 405-412CODEN: BOINFP; ISSN:1367-4803. (Oxford University Press)
      Motivation:Mol. recognition between biol. macromols. and org. small mols. plays an important role in various life processes. Both structural information and binding data of biomol. complexes are indispensable for depicting the underlying mechanism in such an event. The PDBbind database was created to collect exptl. measured binding data for the biomol. complexes throughout the Protein Data Bank (PDB). It thus provides the linkage between structural information and energetic properties of biomol. complexes, which is esp. desirable for computational studies or statistical analyses. Results: Since its first public release in 2004, the PDBbind database has been updated on an annual basis. The latest release (version 2013) provides exptl. binding affinity data for 10 776 biomol. complexes in PDB, including 8302 protein-ligand complexes and 2474 other types of complexes. In this article, we will describe the current methods used for compiling PDBbind and the updated status of this database. We will also review some typical applications of PDBbind published in the scientific literature.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC28Xht1GjsrnI&md5=cb0debb7251bf8cd34385fb4b09e19bb
    26. 26
      Humphrey, W.; Dalke, A.; Schulten, K. VMD – Visual Molecular Dynamics. J. Mol. Graphics 1996, 14, 33– 38,  DOI: 10.1016/0263-7855(96)00018-5
      26
      VDM: visual molecular dynamics
      Humphrey, William; Dalke, Andrew; Schulten, Klaus
      Journal of Molecular Graphics (1996), 14 (1), 33-8, plates, 27-28CODEN: JMGRDV; ISSN:0263-7855. (Elsevier)
      VMD is a mol. graphics program designed for the display and anal. of mol. assemblies, in particular, biopolymers such as proteins and nucleic acids. VMD can simultaneously display any no. of structures using a wide variety of rendering styles and coloring methods. Mols. are displayed as one or more "representations," in which each representation embodies a particular rendering method and coloring scheme for a selected subset of atoms. The atoms displayed in each representation are chosen using an extensive atom selection syntax, which includes Boolean operators and regular expressions. VMD provides a complete graphical user interface for program control, as well as a text interface using the Tcl embeddable parser to allow for complex scripts with variable substitution, control loops, and function calls. Full session logging is supported, which produces a VMD command script for later playback. High-resoln. raster images of displayed mols. may be produced by generating input scripts for use by a no. of photorealistic image-rendering applications. VMD has also been expressly designed with the ability to animate mol. dynamics (MD) simulation trajectories, imported either from files or from a direct connection to a running MD simulation. VMD is the visualization component of MDScope, a set of tools for interactive problem solving in structural biol., which also includes the parallel MD program NAMD, and the MDCOMM software used to connect the visualization and simulation programs, VMD is written in C++, using an object-oriented design; the program, including source code and extensive documentation, is freely available via anonymous ftp and through the World Wide Web.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADyaK28Xis12nsrg%253D&md5=1e3094ec3151fb85c5ff05f8505c78d5
    27. 27
      Martínez-Rosell, G.; Giorgino, T.; De Fabritiis, G. PlayMolecule ProteinPrepare: a web application for protein preparation for molecular dynamics simulations. J. Chem. Inf. Model. 2017, 57, 1511– 1516,  DOI: 10.1021/acs.jcim.7b00190
      27
      PlayMolecule ProteinPrepare: A Web Application for Protein Preparation for Molecular Dynamics Simulations
      Martinez-Rosell, Gerard; Giorgino, Toni; De Fabritiis, Gianni
      Journal of Chemical Information and Modeling (2017), 57 (7), 1511-1516CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      Protein prepn. is a crit. step in mol. simulations that consists of refining a Protein Data Bank (PDB) structure by assigning titrn. states and optimizing the hydrogen-bonding network. In this application note, the authors describe ProteinPrepare, a web application designed to interactively support the prepn. of protein structures. Users can upload a PDB file, choose the solvent pH value, and inspect the resulting protonated residues and hydrogen-bonding network within a 3D web interface. Protonation states are suggested automatically but can be manually changed using the visual aid of the hydrogen-bonding network. Tables and diagrams provide estd. pKa values and charge states, with visual indication for cases where review is required. The authors expect the graphical interface to be a useful instrument to assess the validity of the prepn., but nevertheless, a script to execute the prepn. offline with the High-Throughput Mol. Dynamics (HTMD) environment is also provided for noninteractive operations.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC2sXps1Ghu7c%253D&md5=1a1adc5f0af564e9e9473cc5ebb16ab4
    28. 28
      Barta, T. E.; Veal, J. M.; Rice, J. W.; Partridge, J. M.; Fadden, R. P.; Ma, W.; Jenks, M.; Geng, L.; Hanson, G. J.; Huang, K. H. Discovery of benzamide tetrahydro-4H-carbazol-4-ones as novel small molecule inhibitors of Hsp90. Bioorg. Med. Chem. Lett. 2008, 18, 3517– 3521,  DOI: 10.1016/j.bmcl.2008.05.023
      28
      Discovery of benzamide tetrahydro-4H-carbazol-4-ones as novel small molecule inhibitors of Hsp90
      Barta, Thomas E.; Veal, James M.; Rice, John W.; Partridge, Jeffrey M.; Fadden, R. Patrick; Ma, Wei; Jenks, Matthew; Geng, Lifeng; Hanson, Gunnar J.; Huang, Kenneth H.; Barabasz, Amy F.; Foley, Briana E.; Otto, James; Hall, Steven E.
      Bioorganic & Medicinal Chemistry Letters (2008), 18 (12), 3517-3521CODEN: BMCLE8; ISSN:0960-894X. (Elsevier Ltd.)
      Hsp90 maintains the conformational stability of multiple proteins implicated in oncogenesis and has emerged as a target for chemotherapy. We report here the discovery of a novel small mol. scaffold that inhibits Hsp90. X-ray data show that the scaffold binds competitively at the ATP site on Hsp90. Cellular proliferation and client assays demonstrate that members of the series are able to inhibit Hsp90 at nanomolar concns.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BD1cXmvFSns74%253D&md5=f965e9a2cfd0d44a57bb1949cd87de63
    29. 29
      Erlanson, D. A. In Fragment-Based Drug Discovery and X-Ray Crystallography; Davies, T. G., Hyvönen, M., Eds.; Springer Berlin Heidelberg: Berlin, Heidelberg, 2012; pp 1– 32, DOI: 10.1007/128_2011_180 .
      There is no corresponding record for this reference.
    30. 30
      Ruiz-Carmona, S.; Schmidtke, P.; Luque, F. J.; Baker, L.; Matassova, N.; Davis, B.; Roughley, S.; Murray, J.; Hubbard, R.; Barril, X. Dynamic undocking and the quasi-bound state as tools for drug discovery. Nat. Chem. 2017, 9, 201,  DOI: 10.1038/nchem.2660
      30
      Dynamic undocking and the quasi-bound state as tools for drug discovery
      Ruiz-Carmona, Sergio; Schmidtke, Peter; Luque, F. Javier; Baker, Lisa; Matassova, Natalia; Davis, Ben; Roughley, Stephen; Murray, James; Hubbard, Rod; Barril, Xavier
      Nature Chemistry (2017), 9 (3), 201-206CODEN: NCAHBB; ISSN:1755-4330. (Nature Publishing Group)
      There is a pressing need for new technologies that improve the efficacy and efficiency of drug discovery. Structure-based methods have contributed towards this goal but they focus on predicting the binding affinity of protein-ligand complexes, which is notoriously difficult. The authors adopt an alternative approach that evaluates structural, rather than thermodn., stability. As bioactive mols. present a static binding mode, the authors devised dynamic undocking (DUck), a fast computational method to calc. the work necessary to reach a quasi-bound state at which the ligand has just broken the most important native contact with the receptor. This nonequil. property is surprisingly effective in virtual screening because true ligands form more-resilient interactions than decoys. Notably, DUck is orthogonal to docking and other 'thermodn.' methods. The authors demonstrate the potential of the docking-undocking combination in a fragment screening against the mol. chaperone and oncol. target Hsp90, for which the authors obtain novel chemotypes and a hit rate that approaches 40%.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC28XhvVGiur3I&md5=181235443427435bc9f3e8c17f8df9ee
    31. 31
      Hoxie, R. S.; Street, T. O. Hsp90 chaperones have an energetic hot-spot for binding inhibitors. Protein Sci. 2020, 29, 2101– 2111,  DOI: 10.1002/pro.3933
      31
      Hsp90 chaperones have an energetic hot-spot for binding inhibitors
      Hoxie, Reyal S.; Street, Timothy O.
      Protein Science (2020), 29 (10), 2101-2111CODEN: PRCIEI; ISSN:1469-896X. (Wiley-Blackwell)
      Although Hsp90-family chaperones have been extensively targeted with ATP-competitive inhibitors, it is unknown whether high affinity is achieved from a few highly stabilizing contacts or from many weaker contacts within the ATP-binding pocket. A large-scale anal. of Hsp90α:inhibitor structures shows that inhibitor hydrogen-bonding to a conserved aspartate (D93 in Hsp90α) stands out as most universal among Hsp90 inhibitors. Here we show that the D93 region makes a dominant energetic contribution to inhibitor binding for both cytosolic and organelle-specific Hsp90 paralogs. For inhibitors in the resorcinol family, the D93:inhibitor hydrogen-bond is pH-dependent because the assocd. inhibitor hydroxyl group is titratable, rationalizing a linked-protonation event previously obsd. by the Matulis group. The inhibitor hydroxyl group pKa assocd. with the D93 hydrogen-bond is therefore crit. for optimizing the affinity of resorcinol derivs., and we demonstrate that spectrophotometric measurements can det. this pKa value. Quantifying the energetic contribution of the D93 hotspot is best achieved with the mitochondrial Hsp90 paralog, yielding 3-6 kcal/mol of stabilization (35-60% of the total binding energy) for a diverse set of inhibitors. The Hsp90 Asp93→Asn substitution has long been known to abolish nucleotide binding, yet puzzlingly, native sequences of structurally similar ATPases, such as Topoisomerasese II, have an asparagine at this same crucial site. While aspartate and asparagine sidechains can both act as hydrogen bond acceptors, we show that a steric clash prevents the Hsp90 Asp93→Asn sidechain from adopting the necessary rotamer, whereas this steric restriction is absent in Topoisomerasese II.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXhslKhurfK&md5=1ab67c65a6e11c2cde6fd481ddef6f66
    32. 32
      Majewski, M.; Barril, X. Structural Stability Predicts the Binding Mode of Protein–Ligand Complexes. J. Chem. Inf. Model. 2020, 60, 1644– 1651,  DOI: 10.1021/acs.jcim.9b01062
      32
      Structural Stability Predicts the Binding Mode of Protein-Ligand Complexes
      Majewski, Maciej; Barril, Xavier
      Journal of Chemical Information and Modeling (2020), 60 (3), 1644-1651CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      The prediction of a ligand's binding mode into its macromol. target is essential in structure-based drug discovery. Even though tremendous effort has been made to address this problem, most of the developed tools work similarly, trying to predict the binding free energy assocd. with each particular binding mode. In this study, we decided to abandon this criterion, following structural stability instead. This view, implemented in a novel computational workflow, quantifies the steepness of the local energy min. assocd. with each potential binding mode. Surprisingly, the protocol outperforms docking scoring functions in case of fragments (ligands with MW < 300 Da) and is as good as docking for drug-like mols. It also identifies substructures that act as structural anchors, predicting their binding mode with particular accuracy. The results open a new phys. perspective for binding mode prediction, which can be combined with existing thermodn.-based approaches.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXivVaqsLk%253D&md5=e4f79145486f5552e96a9a7906160f83
    33. 33
      Hu, L.; Benson, M. L.; Smith, R. D.; Lerner, M. G.; Carlson, H. A. Binding MOAD (Mother of All Databases). Proteins: Struct., Funct., Genet. 2005, 60, 333– 340,  DOI: 10.1002/prot.20512
      33
      Binding MOAD (Mother of All Databases)
      Hu, Liegi; Benson, Mark L.; Smith, Richard D.; Lerner, Michael G.; Carlson, Heather A.
      Proteins: Structure, Function, and Bioinformatics (2005), 60 (3), 333-340CODEN: PSFBAF ISSN:. (Wiley-Liss, Inc.)
      Binding MOAD (Mother of All Databases) is the largest collection of high-quality, protein-ligand complexes available from the Protein Data Bank. At this time, Binding MOAD contains 5331 protein-ligand complexes comprised of 1780 unique protein families and 2630 unique ligands. We have searched the crystallog. papers for all 5000 + structures and compiled binding data for 1375 (26%) of the protein-ligand complexes. The binding-affinity data ranges 13 orders of magnitude. This is the largest collection of binding data reported to date in the literature. We have also addressed the issue of redundancy in the data. To create a nonredundant dataset, one protein from each of the 1780 protein families was chosen as a representative. Representatives were chosen by tightest binding, best resoln., etc. For the 1780 "best" complexes that comprise the nonredundant version of Binding MOAD, 475 (27%) have binding data. This significant collection of protein-ligand complexes will be very useful in elucidating the biophys. patterns of mol. recognition and enzymic regulation. The complexes with binding-affinity data will help in the development of improved scoring functions and structure-based drug discovery techniques. The dataset can be accessed at http://www.BindingMOAD.org.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BD2MXmvVyrtr4%253D&md5=146632a2f30fda98cb987e984beb7506
    34. 34
      Ruiz-Carmona, S.; Alvarez-Garcia, D.; Foloppe, N.; Garmendia-Doval, A. B.; Juhos, S.; Schmidtke, P.; Barril, X.; Hubbard, R. E.; Morley, S. D. rDock: A Fast, Versatile and Open Source Program for Docking Ligands to Proteins and Nucleic Acids. PLoS Comput. Biol. 2014, 10, e1003571  DOI: 10.1371/journal.pcbi.1003571
      34
      rDock: a fast, versatile and open source program for docking ligands to proteins and nucleic acids
      Ruiz-Carmona, Sergio; Alvarez-Garcia, Daniel; Foloppe, Nicolas; Garmendia-Doval, A. Beatriz; Juhos, Szilveszter; Schmidtke, Peter; Barril, Xavier; Hubbard, Roderick E.; Morley, S. David
      PLoS Computational Biology (2014), 10 (4), e1003571/1-e1003571/7, 7 pp.CODEN: PCBLBG; ISSN:1553-7358. (Public Library of Science)
      Identification of chem. compds. with specific biol. activities is an important step in both chem. biol. and drug discovery. When the structure of the intended target is available, one approach is to use mol. docking programs to assess the chem. complementarity of small mols. with the target; such calcns. provide a qual. measure of affinity that can be used in virtual screening (VS) to rank order a list of compds. according to their potential to be active. rDock is a mol. docking program developed at Vernalis for high-throughput VS (HTVS) applications. Evolved from RiboDock, the program can be used against proteins and nucleic acids, is designed to be computationally very efficient and allows the user to incorporate addnl. constraints and information as a bias to guide docking. This article provides an overview of the program structure and features and compares rDock to two ref. programs, AutoDock Vina (open source) and Schrodinger's Glide (com.). In terms of computational speed for VS, rDock is faster than Vina and comparable to Glide. For binding mode prediction, rDock and Vina are superior to Glide. The VS performance of rDock is significantly better than Vina, but inferior to Glide for most systems unless pharmacophore constraints are used; in that case rDock and Glide are of equal performance. The program is released under the Lesser General Public License and is freely available for download, together with the manuals, example files and the complete test sets, at online.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC2cXhsVGlsL%252FO&md5=e4cb786d6567fdc7f2a46f64955a9992

  • Supporting Information

    Supporting Information

    The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.jcim.1c00691.

    • “Model training” and “quantitative analysis”, additional information for these sections; Figures S1–S6, distance distribution between the two voxels with highest, absolute attribution value for different channel combinations studied; Figures S7 and S8, examples of protein residues far from ligand having high attribution values; Figure S9, correlation between magnitude of attributions of two best voxels and distance between them; and Figures S10–S13, attribution consistency distributions (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.