ACS Publications. Most Trusted. Most Cited. Most Read
Machine Learning ADME Models in Practice: Four Guidelines from a Successful Lead Optimization Case Study
More

OR SEARCH CITATIONS

My Activity
Publications

Figure 1Loading Img
Download Hi-Res ImageDownload to MS-PowerPointCite This:ACS Med. Chem. Lett. 2024, 15, 8, 1169-1173
  • Open Access
Viewpoint

Machine Learning ADME Models in Practice: Four Guidelines from a Successful Lead Optimization Case Study
Click to copy article linkArticle link copied!

  • Alexander S. Rich*
    Alexander S. Rich
    Inductive Bio, Inc., 550 Vanderbilt Ave, #730, Brooklyn, New York 11238, United States
    *E-mail: [email protected]
    More by Alexander S. Rich
  • Yvonne H. Chan
    Yvonne H. Chan
    Nested Therapeutics, 1030 Mass Ave, Suite 410, Cambridge, Massachusetts 02138, United States
    More by Yvonne H. Chan
    Orcidhttps://orcid.org/0000-0002-3422-8164
  • Benjamin Birnbaum
    Benjamin Birnbaum
    Inductive Bio, Inc., 550 Vanderbilt Ave, #730, Brooklyn, New York 11238, United States
    More by Benjamin Birnbaum
  • Kamran Haider
    Kamran Haider
    Nested Therapeutics, 1030 Mass Ave, Suite 410, Cambridge, Massachusetts 02138, United States
    More by Kamran Haider
  • Joshua Haimson
    Joshua Haimson
    Inductive Bio, Inc., 550 Vanderbilt Ave, #730, Brooklyn, New York 11238, United States
    More by Joshua Haimson
  • Michael Hale
    Michael Hale
    Nested Therapeutics, 1030 Mass Ave, Suite 410, Cambridge, Massachusetts 02138, United States
    More by Michael Hale
  • Yongxin Han
    Yongxin Han
    Nested Therapeutics, 1030 Mass Ave, Suite 410, Cambridge, Massachusetts 02138, United States
    More by Yongxin Han
  • William Hickman
    William Hickman
    Inductive Bio, Inc., 550 Vanderbilt Ave, #730, Brooklyn, New York 11238, United States
    More by William Hickman
  • Klaus P. Hoeflich
    Klaus P. Hoeflich
    Nested Therapeutics, 1030 Mass Ave, Suite 410, Cambridge, Massachusetts 02138, United States
    More by Klaus P. Hoeflich
  • Daniel Ortwine
    Daniel Ortwine
    Nested Therapeutics, 1030 Mass Ave, Suite 410, Cambridge, Massachusetts 02138, United States
    More by Daniel Ortwine
  • Ayşegül Özen
    Ayşegül Özen
    Nested Therapeutics, 1030 Mass Ave, Suite 410, Cambridge, Massachusetts 02138, United States
    More by Ayşegül Özen
  • David B. Belanger
    David B. Belanger
    Nested Therapeutics, 1030 Mass Ave, Suite 410, Cambridge, Massachusetts 02138, United States
    More by David B. Belanger
Open PDF

ACS Medicinal Chemistry Letters

Cite this: ACS Med. Chem. Lett. 2024, 15, 8, 1169–1173
Click to copy citationCitation copied!
https://doi.org/10.1021/acsmedchemlett.4c00290
Published July 25, 2024

Copyright © 2024 The Authors. Published by American Chemical Society. This publication is licensed under

CC-BY 4.0 .

Abstract

Click to copy section linkSection link copied!

Optimization of the ADME properties and pharmacokinetic (PK) profile of compounds is one of the critical activities in any medicinal chemistry campaign to discover a future clinical candidate. Finding ways to expedite the process to address ADME/PK shortcomings and reduce the number of compounds to synthesize is highly valuable. This article provides practical guidelines and a case study on the use of ML ADME models to guide compound design in small molecule lead optimization. These guidelines highlight that ML models cannot have an impact in a vacuum: they help advance a program when they have the trust of users, are tuned to the needs of the program, and are integrated into decision-making processes in a way that complements and augments the expertise of chemists.

This publication is licensed under

CC-BY 4.0 .
  • cc licence
  • by licence
Copyright © 2024 The Authors. Published by American Chemical Society

Special Issue

Published as part of ACS Medicinal Chemistry Letters virtual special issue “Exploring the Use of AI/ML Technologies in Medicinal Chemistry and Drug Discovery”.

Optimization of ADME properties is a key challenge during the hit-to-lead and lead optimization phases of small molecule drug discovery. Machine learning (ML) models can be used to predict the outcomes of ADME-related assays such as permeability, solubility, or liver microsomal stability. They have been proposed as a tool to reduce the number of design-make-test cycles and accelerate programs. (1,2) However, building and using ML ADME models effectively can be challenging, particularly within the context of biotech companies. Without large in-house data sets, there can be a lack of sufficient data to build performant models, particularly in the critical early stages of a program. And without the right integration of ML models into design tools, there can be a disconnect between model builders and end users, leading to limited model use.
In this Viewpoint, we share four guidelines for using ML ADME models effectively to help drive forward a drug discovery program. In describing these guidelines, we draw upon a case study of a collaboration between Nested Therapeutics and Inductive Bio. In this collaboration, Nested Therapeutics used Inductive Bio’s ADME models for lead optimization in a best-in-class program. The models were integrated into interactive tools that were used by the medicinal chemistry team, enabling rapid iteration, enhanced ideation informed by predicted data, and elevated design quality. This allowed the team to efficiently resolve permeability and metabolic stability issues, resulting in the nomination of a development candidate with excellent cell potency and cross-species PK.

Guideline 1: Regular Time-Based and Series-Level Evaluation Gives a Realistic Picture of Model Performance and Builds Trust to Use ML Models As a Tool in the Design Process

Click to copy section linkSection link copied!
ML model evaluation is critical to earning user trust and ensures a model is fit for use. A key step in model evaluation is selecting the subset of compounds and measurements to evaluate, which must be withheld from model training. We follow two principles when choosing these evaluation sets: they should be separated temporally from training data, and they should be stratified by program and series.
Time-based splits simulate real usage, in which a model trained on all data up to a certain date is used prospectively. This is more rigorous than the commonly used techniques of random or scaffold splitting, which can overestimate how well a model will perform due to high similarity between training and evaluation sets. (3,4)
Stratifying evaluation metrics by program and series is important because ML models can vary in their performance across projects and chemotypes, in a way that can be hard to predict a priori. (5) Proactively measuring performance at project and series levels informs project teams on where and for what purpose models can be confidently used.
In our collaboration, we used these two principles to build initial trust in the ML models and to continually validate the models’ fitness for use. We performed a time-based split using data from an existing program that had completed lead optimization, evaluating performance across three distinct chemical series, to confirm the general suitability of the ML models to Nested’s data. We then performed an additional time-based evaluation on existing early data from the program of interest. From there, the ML models were used on an ongoing basis, and performance was re-evaluated weekly using a time-based split (see Guideline 3). We reported metrics for the project as a whole and by series where appropriate according to definitions crafted by the project team.

Guideline 2: Training on a Combination of “Global” Curated Data and “Local” Program Data Leads to the Best Model Performance

Click to copy section linkSection link copied!
When developing ML ADME models for a project team, one can create a local model trained solely on the program’s measured data, as in traditional QSAR approaches. (6) Alternatively, one might use a global model that has already been built using large external data sets to predict a given property. (7,8) An approach that balances these extremes is to train a model that combines nonproject global data with data from the project itself. This can be done by simply including all available data when training a model (2,9,10) or by using other more sophisticated fine-tuning approaches. (5) Studies have found that fine-tuned models trained with combined local and global data perform better than those trained with local or global data alone. (5,9)
The models used in our collaboration followed this best practice of combining global and local data for training. The models were initially developed using a curated global proprietary data set and were then fine-tuned by adding project data. Training was performed using a graph neural network model architecture. (11)
To explore how the fine-tuned global model compared to local-only or global-only alternatives, we analyzed performance of the Inductive models against versions of the models trained only on external curated data (global-only) and models trained locally using a QSAR software tool implementing AutoML (local-only). (12) For each of human liver microsomal stability (HLM), rat liver microsomal stability (RLM), Madin-Darby Canine Kidney (MDCK) permeability (MDCK AB), and MDCK efflux ratio (MDCK ER), we created a temporal evaluation split by choosing the first 100 compounds measured as local training data and the next 100 compounds as a test set.
As seen in Figure 1, the fine-tuned global modeling approach generally performed best across the assays. It achieved the lowest MAE (Mean Absolute Error) across all four properties, and the highest Spearman rank correlation across all assays except MDCK AB, where the global model correlation was slightly higher.

Figure 1

Figure 1. Performance of models trained on program data only (local only), nonprogram data only (global only), and on combined data (fine-tuned global), on temporally split test sets for HLM, RLM, MDCK AB, and MDCK ER. Error bars represent 68% bootstrapped confidence intervals. MAE (mean absolute error) units are in log 10 (mL/min/kg) for HLM and RLM, log 10 (μcm/s) for MDCK AB, and log 10 (ratio) for ER.

High Resolution Image
Download MS PowerPoint Slide
In this program, the global training seemed most helpful in predicting the MDCK measurements, and it was least helpful in RLM. This cannot be easily explained in terms of similarity of test compounds to those in the global training sets. For all four assays, no more than 2% of the Nested compounds had compounds in the global training set with a Tanimoto similarity of >0.3 (calculated with Morgan fingerprints, radius 2). However, we observed a surprising divergence in measured HLM and RLM intrinsic clearance in the Nested compounds, with RLM compounds exhibiting a median 8 times higher clearance. This difference between species was not anticipated by the global model, which predicted roughly equal clearance (leading to a high MAE for RLM), but was captured by the fine-tuned global model. This emphasizes the value of validating models on early program data, as well as that of training on local data.

Guideline 3: Frequent Model Retraining Enables ML Models to Learn Local SAR As a Program Shifts into New Chemical Space and Encounters Activity Cliffs

Click to copy section linkSection link copied!
Throughout lead optimization, new experimental data are collected that may improve model performance. Tested compounds may also move into new chemical space, potentially worsening performance. (13) These factors create an incentive for frequent model retraining so the models can learn the local SAR from the experimental data and maintain accuracy. (14,15)
Monthly retraining has been shown to provide a boost in performance compared to less frequent schedules across a variety of ADME end points. (16) Frequent retraining can be particularly useful for rapidly adjusting to activity cliffs. (14,17) Weekly retraining has been reported to be beneficial (2) and aligns well with the weekly cycle of design meetings common in drug programs. While weekly retraining may sometimes incorporate only a few new compounds, it also strengthens user trust by ensuring that model predictions are always informed by the most recent and relevant assay readouts.
Throughout the course of our collaboration, we applied weekly retraining to keep our microsomal stability and permeability models up to date. A retrospective analysis confirms that frequent training aided performance for HLM stability, the property for which data was collected most consistently throughout the time course of the program. When we split HLM measurements into periods of 1 month and evaluated predictions from the model deployed at the beginning of that month, we observed an average Spearman R of 0.65. If we instead used the model deployed 1 month previous, the Spearman R fell to 0.55. An additional month lag dropped it further to 0.49.
We also observed the helpfulness of model retraining for adjusting to activity cliffs. At one point, a change to a substitution position in a ring was discovered to cause a surprising several-fold jump in microsomal clearance. While the model did not predict this jump in advance, retraining weekly allowed it to rapidly adjust to the observed data and begin making appropriate predictions for additional compounds with the new motif.

Guideline 4: To Maximize Impact on the Design Process, ML Models Should Be Interactive, Interpretable, and Integrated with Other Tools

Click to copy section linkSection link copied!
The best ML ADME model will not have an impact unless it is actively used. (1) We have found that models have the best chance of being used effectively if they are integrated, interactive, and interpretable. Integrated models are available within software tools that computational and medicinal chemists are already using to guide decision-making. Interactive models provide real-time predictions as a chemist ideates new designs, rather than working only via bulk scoring or with a long computation time. And finally, interpretable models provide not only a predicted assay value but additional relevant information such as atom-level visualizations indicating important regions of the molecule for a given property. These three principles were brought together to make the Nested–Inductive collaboration successful.
At the start of the collaboration, the program was in early lead optimization with the goal of demonstrating in vivo target engagement at a projected low human dose. Compound 1 (Table 1) showed moderate cellular activity, good permeability, and HLM stability, yet in vitro/in vivo dog and rat clearance needed improvement.
Table 1. Key Compounds from the Case Study Campaign and Their Properties
compound #target engagement assay (nM)HLM T1/2 (min)RLM T1/2 (min)dog LM T1/2 (min)MDCK Papp (ER)projected human dose
17528337213.8 (0.8) 
21008244223.6 (2.6) 
32638232134.7 (2.2) 
41376565578.1 (0.9)4× higher than desired
51248372607.4 (0.8)desired
To address these goals, the computational team at Nested vetted and deployed docking, strain energy prediction, and atomistic simulations including free energy predictions. Frequently retrained ML ADME models from Inductive were integrated into the existing computational infrastructure via an application programming interface (API). This allowed ADME predictions to be presented in the context of complementary modeling output, reducing the number of compounds requiring evaluation through computationally intensive approaches before selection for synthesis.
As understanding of the SAR of the shallow, flexible target improved, chemists made increasing use of the interactive ML application (Figure 2). In contrast to a batch scoring interface, the interactive application enabled chemists to rapidly draw ideas, get instant feedback, and iterate to find the most promising candidates for synthesis. This interactive approach allowed scientists to explore and reason about the SAR of multiple design steps at once rather than primarily considering single point design changes. The interpretability tools, including reactivity-based prediction of likely sites of metabolism, (18) further helped guide design changes.

Figure 2

Figure 2. A screenshot of the interactive design environment with ML ADME predictions, comparison to a reference compound, and highlights of sites of likely metabolism.

High Resolution Image
Download MS PowerPoint Slide
Ultimately, the ML ADME models of permeability and metabolic stability were used in tandem with an existing structure-based design workflow to elevate the overall quality of the designs. Using this strategy, Nested co-optimized two regions of the compound to identify potent, permeable compounds 2 and 3. The team then used the ML application to fine-tune physiochemical properties and address metabolic soft-spots without sacrificing permeability or potency. Compound 4 advanced to cross-species pharmacokinetics (PK) and was soon followed, within a few weeks, by the identification of compound 5 with exquisite cell potency and cross-species PK.

Conclusion

Click to copy section linkSection link copied!
By following these guidelines, ML ADME models can be made an integral part of the lead optimization process within biotech. Using ML models does not negate the necessity of collecting experimental ADME data or eliminate the risk of synthesizing compounds with unfavorable properties. Nonetheless, it can meaningfully accelerate a program and improve a chemist’s ability to cooperatively optimize PK and potency to achieve a desired PD outcome. When an ADME challenge arises for a particular property, an effective ML model can be used to quickly generate and evaluate ideas that might fix the problem, which can then be experimentally validated. When ADME properties are not a current challenge, ML models can help maintain good properties while tuning other attributes and flag potential issues. Brought together, these uses help turn a lead into a development candidate faster and identify solutions to property tradeoffs that enable compounds with better PK–PD profiles.

Author Information

Click to copy section linkSection link copied!
  • Corresponding Author
    • Alexander S. Rich - Inductive Bio, Inc., 550 Vanderbilt Ave, #730, Brooklyn, New York 11238, United States;  Email: [email protected]
  • Authors
    • Yvonne H. Chan - Nested Therapeutics, 1030 Mass Ave, Suite 410, Cambridge, Massachusetts 02138, United States;  Orcidhttps://orcid.org/0000-0002-3422-8164
    • Benjamin Birnbaum - Inductive Bio, Inc., 550 Vanderbilt Ave, #730, Brooklyn, New York 11238, United States
    • Kamran Haider - Nested Therapeutics, 1030 Mass Ave, Suite 410, Cambridge, Massachusetts 02138, United States
    • Joshua Haimson - Inductive Bio, Inc., 550 Vanderbilt Ave, #730, Brooklyn, New York 11238, United States
    • Michael Hale - Nested Therapeutics, 1030 Mass Ave, Suite 410, Cambridge, Massachusetts 02138, United States
    • Yongxin Han - Nested Therapeutics, 1030 Mass Ave, Suite 410, Cambridge, Massachusetts 02138, United States
    • William Hickman - Inductive Bio, Inc., 550 Vanderbilt Ave, #730, Brooklyn, New York 11238, United States
    • Klaus P. Hoeflich - Nested Therapeutics, 1030 Mass Ave, Suite 410, Cambridge, Massachusetts 02138, United States
    • Daniel Ortwine - Nested Therapeutics, 1030 Mass Ave, Suite 410, Cambridge, Massachusetts 02138, United States
    • Ayşegül Özen - Nested Therapeutics, 1030 Mass Ave, Suite 410, Cambridge, Massachusetts 02138, United States
    • David B. Belanger - Nested Therapeutics, 1030 Mass Ave, Suite 410, Cambridge, Massachusetts 02138, United States
  • Author Contributions

    A.S.R. and Y.H.C. contributed equally.

  • Notes
    Views expressed in this Viewpoint are those of the author and not necessarily the views of the ACS.
    The authors declare the following competing financial interest(s): Alexander Rich, Benjamin Birnbaum, Joshua Haimson, and William Hickman report employment at Inductive Bio and equity ownership in Inductive Bio. Yvonne Chan, Kamran Haider, Michael Hale, Klaus Hoeflich, Aysegul Ozen, and David Belanger report past or current employment and/or equity ownership in Nested Therapeutics.

References

Click to copy section linkSection link copied!

This article references 18 other publications.

  1. 1
    Ortwine, D. F.; Aliagas, I. Physicochemical and DMPK in silico models: facilitating their use by medicinal chemists. Mol. Pharmaceutics 2013, 10 (4), 1153– 1161,  DOI: 10.1021/mp3006193
    Google Scholar
    1
    Physicochemical and DMPK In Silico Models: Facilitating Their Use by Medicinal Chemists
    Ortwine, Daniel F.; Aliagas, Ignacio
    Molecular Pharmaceutics (2013), 10 (4), 1153-1161CODEN: MPOHBP; ISSN:1543-8384. (American Chemical Society)
    A review. It is known that the developability of drugs is related to their physicochem. and DMPK properties. Given the time and expense involved in discovering and developing new drugs, maximizing the chance of success by calcg. properties ahead of chem. synthesis and testing, and only acting on those candidates whose properties fall into a desired range, would seem to make sense. This paper provides an overview of calculable physicochem. and DMPK properties, an assessment of their relative difficulty of their calcn. and accuracy, and available software. Methods companies have employed to communicate results will be discussed, including the use of composite scoring functions and ranking schemes. Calcns. do no good if chemists will not use them to prioritize synthesis decisions. Strategies are presented for facilitating model usage. An approach adopted at Genentech for presenting results that involves the close coupling of property calcns. with 3D structure based drug design is described.
  2. 2
    Goller, A. H.; Kuhnke, L.; Montanari, F.; Bonin, A.; Schneckener, S.; ter Laak, A.; Wichard, J.; Lobell, M.; Hillisch, A. Bayer’s in silico ADMET platform: a journey of machine learning over the past two decades. Drug Discovery Today 2020, 25 (9), 1702– 1709,  DOI: 10.1016/j.drudis.2020.07.001
    Google Scholar
    2
    Bayer's in silico ADMET platform: a journey of machine learning over the past two decades
    Goller Andreas H; Bonin Anne; Lobell Mario; Kuhnke Lara; Ter Laak Antonius; Montanari Floriane; Schneckener Sebastian; Wichard Jorg; Hillisch Alexander
    Drug discovery today (2020), 25 (9), 1702-1709 ISSN:.
    Over the past two decades, an in silico absorption, distribution, metabolism, and excretion (ADMET) platform has been created at Bayer Pharma with the goal to generate models for a variety of pharmacokinetic and physicochemical endpoints in early drug discovery. These tools are accessible to all scientists within the company and can be a useful in assisting with the selection and design of novel leads, as well as the process of lead optimization. Here. we discuss the development of machine-learning (ML) approaches with special emphasis on data, descriptors, and algorithms. We show that high company internal data quality and tailored descriptors, as well as a thorough understanding of the experimental endpoints, are essential to the utility of our models. We discuss the recent impact of deep neural networks and show selected application examples.
  3. 3
    Wu, Z.; Ramsundar, B.; Feinberg, E. N.; Gomes, J.; Geniesse, C.; Pappu, A. S.; Leswing, K.; Pande, V. MoleculeNet: a benchmark for molecular machine learning. Chemical science 2018, 9 (2), 513– 530,  DOI: 10.1039/C7SC02664A
    Google Scholar
    3
    MoleculeNet: a benchmark for molecular machine learning
    Wu, Zhenqin; Ramsundar, Bharath; Feinberg, Evan N.; Gomes, Joseph; Geniesse, Caleb; Pappu, Aneesh S.; Leswing, Karl; Pande, Vijay
    Chemical Science (2018), 9 (2), 513-530CODEN: CSHCCN; ISSN:2041-6520. (Royal Society of Chemistry)
    Mol. machine learning has been maturing rapidly over the last few years. Improved methods and the presence of larger datasets have enabled machine learning algorithms to make increasingly accurate predictions about mol. properties. However, algorithmic progress has been limited due to the lack of a std. benchmark to compare the efficacy of proposed methods; most new algorithms are benchmarked on different datasets making it challenging to gauge the quality of proposed methods. This work introduces MoleculeNet, a large scale benchmark for mol. machine learning. MoleculeNet curates multiple public datasets, establishes metrics for evaluation, and offers high quality open-source implementations of multiple previously proposed mol. featurization and learning algorithms (released as part of the DeepChem open source library). MoleculeNet benchmarks demonstrate that learnable representations are powerful tools for mol. machine learning and broadly offer the best performance. However, this result comes with caveats. Learnable representations still struggle to deal with complex tasks under data scarcity and highly imbalanced classification. For quantum mech. and biophys. datasets, the use of physics-aware featurizations can be more important than choice of particular learning algorithm.
  4. 4
    Sheridan, R. P. Time-split cross-validation as a method for estimating the goodness of prospective prediction. J. Chem. Inf. Model. 2013, 53 (4), 783– 790,  DOI: 10.1021/ci400084k
    Google Scholar
    4
    Time-Split Cross-Validation as a Method for Estimating the Goodness of Prospective Prediction.
    Sheridan, Robert P.
    Journal of Chemical Information and Modeling (2013), 53 (4), 783-790CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    Cross-validation is a common method to validate a QSAR model. In cross-validation, some compds. are held out as a test set, while the remaining compds. form a training set. A model is built from the training set, and the test set compds. are predicted on that model. The agreement of the predicted and obsd. activity values of the test set (measured by, say, R2) is an est. of the self-consistency of the model and is sometimes taken as an indication of the predictivity of the model. This est. of predictivity can be optimistic or pessimistic compared to true prospective prediction, depending how compds. in the test set are selected. Here, we show that time-split selection gives an R2 that is more like that of true prospective prediction than the R2 from random selection (too optimistic) or from our analog of leave-class-out selection (too pessimistic). Time-split selection should be used in addn. to random selection as a std. for cross-validation in QSAR model building.
  5. 5
    Fluetsch, A.; Di Lascio, E.; Gerebtzoff, G.; Rodríguez-Pérez, R. Adapting Deep Learning QSPR Models to Specific Drug Discovery Projects. Molecular Pharmaceutics 2024, 21, 1817,  DOI: 10.1021/acs.molpharmaceut.3c01124
    Google Scholar
    There is no corresponding record for this reference.
  6. 6
    Muratov, E. N.; Bajorath, J.; Sheridan, R. P.; Tetko, I. V.; Filimonov, D.; Poroikov, V.; Oprea, T. I.; Baskin, I. I.; Varnek, A.; Roitberg, A.; Isayev, O.; Curtalolo, S.; Fourches, D.; Cohen, Y.; Aspuru-Guzik, A.; Winkler, D. A.; Agrafiotis, D.; Cherkasov, A.; Tropsha, A. QSAR without borders. Chem. Soc. Rev. 2020, 49 (11), 3525– 3564,  DOI: 10.1039/D0CS00098A
    Google Scholar
    6
    QSAR without borders
    Muratov, Eugene N.; Bajorath, Jurgen; Sheridan, Robert P.; Tetko, Igor V.; Filimonov, Dmitry; Poroikov, Vladimir; Oprea, Tudor I.; Baskin, Igor I.; Varnek, Alexandre; Roitberg, Adrian; Isayev, Olexandr; Curtalolo, Stefano; Fourches, Denis; Cohen, Yoram; Aspuru-Guzik, Alan; Winkler, David A.; Agrafiotis, Dimitris; Cherkasov, Artem; Tropsha, Alexander
    Chemical Society Reviews (2020), 49 (11), 3525-3564CODEN: CSRVBR; ISSN:0306-0012. (Royal Society of Chemistry)
    A review. Prediction of chem. bioactivity and phys. properties has been one of the most important applications of statistical and more recently, machine learning and artificial intelligence methods in chem. sciences. This field of research, broadly known as quant. structure-activity relationships (QSAR) modeling, has developed many important algorithms and has found a broad range of applications in phys. org. and medicinal chem. in the past 55+ years. This Perspective summarizes recent technol. advances in QSAR modeling but it also highlights the applicability of algorithms, modeling methods, and validation practices developed in QSAR to a wide range of research areas outside of traditional QSAR boundaries including synthesis planning, nanotechnol., materials science, biomaterials, and clin. informatics. As modern research methods generate rapidly increasing amts. of data, the knowledge of robust data-driven modeling methods professed within the QSAR field can become essential for scientists working both within and outside of chem. research. We hope that this contribution highlighting the generalizable components of QSAR modeling will serve to address this challenge.
  7. 7
    Xiong, G.; Wu, Z.; Yi, J.; Fu, L.; Yang, Z.; Hsieh, C.; Yin, M.; Zeng, X.; Wu, C.; Lu, A.; Chen, X.; Hou, T.; Cao, D. ADMETlab 2.0: an integrated online platform for accurate and comprehensive predictions of ADMET properties. Nucleic acids research 2021, 49 (W1), W5– W14,  DOI: 10.1093/nar/gkab255
    Google Scholar
    There is no corresponding record for this reference.
  8. 8
    Kawashima, H.; Watanabe, R.; Esaki, T.; Kuroda, M.; Nagao, C.; Natsume-Kitatani, Y.; Ohashi, R.; Komura, H.; Mizuguchi, K. DruMAP: A novel drug metabolism and pharmacokinetics analysis platform. J. Med. Chem. 2023, 66 (14), 9697– 9709,  DOI: 10.1021/acs.jmedchem.3c00481
    Google Scholar
    There is no corresponding record for this reference.
  9. 9
    Di Lascio, E.; Gerebtzoff, G.; Rodríguez-Pérez, R. Systematic evaluation of local and global machine learning models for the prediction of ADME properties. Mol. Pharmaceutics 2023, 20 (3), 1758– 1767,  DOI: 10.1021/acs.molpharmaceut.2c00962
    Google Scholar
    9
    Systematic evaluation of local and global machine learning models for the prediction of ADME properties
    Di Lascio, Elena; Gerebtzoff, Gregori; Rodriguez-Perez, Raquel
    Molecular Pharmaceutics (2023), 20 (3), 1758-1767CODEN: MPOHBP; ISSN:1543-8384. (American Chemical Society)
    Machine learning (ML) has become an indispensable tool to predict absorption, distribution, metab., and excretion (ADME) properties in pharmaceutical research. ML algorithms are trained on mol. structures and corresponding ADME assay data to develop quant. structure-property relationship (QSPR) models. Traditional QSPR models were trained on compd. sets of limited size. With the advent of more complex ML algorithms and data availability, training sets have become larger and more diverse. Most common training approaches consist in either training a model with a small set of similar compds., namely, compds. designed for the same drug discovery project or chem. series (local model approach) or with a larger set of diverse compds. (global model approach). Global models are built with all exptl. data available for an assay, combining compd. data from different projects and disease areas. Despite the ML progress made so far, the choice of the appropriate data compn. for building ML models is still unclear. Herein, a systematic evaluation of local and global ML models was performed for 10 different exptl. assays and 112 drug discovery projects. Results show a consistent superior performance of global models for ADME property predictions. Diagnostic analyses were also carried out to investigate the influence of training set size, structural diversity, and data shift in the relative performance of local and global ML models. Training set and structural diversity did not have an impact in the relative performance on the methods. Instead, data shift helped to identify the projects with larger performance differences between local and global models. Results presented in this work can be leveraged to improve ML-based ADME properties predictions and thus decision-making in drug discovery projects.
  10. 10
    Aliagas, I.; Gobbi, A.; Heffron, T.; Lee, M. L.; Ortwine, D. F.; Zak, M.; Khojasteh, S. C. A probabilistic method to report predictions from a human liver microsomes stability QSAR model: a practical tool for drug discovery. Journal of Computer-Aided Molecular Design 2015, 29, 327– 338,  DOI: 10.1007/s10822-015-9838-3
    Google Scholar
    There is no corresponding record for this reference.
  11. 11
    Wieder, O.; Kohlbacher, S.; Kuenemann, M.; Garon, A.; Ducrot, P.; Seidel, T.; Langer, T. A compact review of molecular property prediction with graph neural networks. Drug Discovery Today: Technologies 2020, 37, 1– 12,  DOI: 10.1016/j.ddtec.2020.11.009
    Google Scholar
    11
    A compact review of molecular property prediction with graph neural networks
    Wieder Oliver; Kohlbacher Stefan; Garon Arthur; Ducrot Pierre; Seidel Thomas; Kuenemann Melaine; Langer Thierry
    Drug discovery today. Technologies (2020), 37 (), 1-12 ISSN:.
    As graph neural networks are becoming more and more powerful and useful in the field of drug discovery, many pharmaceutical companies are getting interested in utilizing these methods for their own in-house frameworks. This is especially compelling for tasks such as the prediction of molecular properties which is often one of the most crucial tasks in computer-aided drug discovery workflows. The immense hype surrounding these kinds of algorithms has led to the development of many different types of promising architectures and in this review we try to structure this highly dynamic field of AI-research by collecting and classifying 80 GNNs that have been used to predict more than 20 molecular properties using 48 different datasets.
  12. 12
    Karmaker, S. K.; Hassan, M. M.; Smith, M. J.; Xu, L.; Zhai, C.; Veeramachaneni, K. AutoML to Date and Beyond: Challenges and Opportunities. ACM Computing Surveys (CSUR) 2022, 54 (8), 1– 36,  DOI: 10.1145/3470918
    Google Scholar
    There is no corresponding record for this reference.
  13. 13
    Yu, J.; Wang, D.; Zheng, M. Uncertainty quantification: Can we trust artificial intelligence in drug discovery?. iScience 2022, 25 (8), 104814,  DOI: 10.1016/j.isci.2022.104814
    Google Scholar
    There is no corresponding record for this reference.
  14. 14
    Sheridan, R. P.; Culberson, J. C.; Joshi, E.; Tudor, M.; Karnachi, P. Prediction accuracy of production ADMET models as a function of version: activity cliffs rule. J. Chem. Inf. Model. 2022, 62 (14), 3275– 3280,  DOI: 10.1021/acs.jcim.2c00699
    Google Scholar
    14
    Prediction Accuracy of Production ADMET Models as a Function of Version: Activity Cliffs Rule
    Sheridan, Robert P.; Culberson, J. Chris; Joshi, Elizabeth; Tudor, Matthew; Karnachi, Prabha
    Journal of Chemical Information and Modeling (2022), 62 (14), 3275-3280CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    As with many other institutions, our company maintains many quant. structure-activity relationship (QSAR) models of absorption, distribution, metab., excretion, and toxicity (ADMET) end points and updates the models regularly. We recently examd. version-to-version predictivity for these models over a period of 10 years. In this approach we monitor the goodness of prediction of new mols. relative to the training set of model version V before they are incorporated in the updated model V+1. Using a cell-based permeability assay (Papp) as an example, we illustrate how the QSAR models made from this data are generally predictive and can be utilized to enrich chem. designs and synthesis. Despite the obvious utility of these models, we turned up unexpected behavior in Papp and other ADMET activities for which the explanation is not obvious. One such behavior is that the apparent predictivity of the models as measured by root-mean-square-error can vary greatly from version to version and is sometimes very poor. One intuitively appealing explanation is that the obsd. activities of the new mols. fall outside the bulk of activities in the training set. Alternatively, one may think that the new mols. are exploring different regions of chem. space than the training set. However, the real explanation has to do with activity cliffs. If the obsd. activities of the new mols. are different than expected based on similar mols. in the training set, the predictions will be less accurate. This is true for all our ADMET end points.
  15. 15
    Aleksić, S.; Seeliger, D.; Brown, J. B. ADMET Predictability at Boehringer Ingelheim: State of the Art, and Do Bigger Datasets or Algorithms Make a Difference?. Molecular Informatics 2022, 41 (2), 2100113,  DOI: 10.1002/minf.202100113
    Google Scholar
    There is no corresponding record for this reference.
  16. 16
    Fang, C.; Wang, Y.; Grater, R.; Kapadnis, S.; Black, C.; Trapa, P.; Sciabola, S. Prospective validation of machine learning algorithms for absorption, distribution, metabolism, and excretion prediction: An industrial perspective. J. Chem. Inf. Model. 2023, 63 (11), 3263– 3274,  DOI: 10.1021/acs.jcim.3c00160
    Google Scholar
    There is no corresponding record for this reference.
  17. 17
    Sheridan, R. P. Stability of prediction in production ADMET models as a function of version: why and when predictions change. J. Chem. Inf. Model. 2022, 62 (15), 3477– 3485,  DOI: 10.1021/acs.jcim.2c00803
    Google Scholar
    17
    Stability of Prediction in Production ADMET Models as a Function of Version: Why and When Predictions Change
    Sheridan, Robert P.
    Journal of Chemical Information and Modeling (2022), 62 (15), 3477-3485CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    As with other pharma companies, we maintain prodn. QSAR models of ADMET end points and update them regularly. Here, for six ADMET end points, we examine the predictions of test set mols. on multiple versions of random forest models spanning a period of 10 years. For any given end point, the predictions for the majority of mols. are similar for all model versions. However, for a small minority of mols., the prediction shifts substantially over the span of a few versions. For most mols. that shift, the prediction becomes more accurate at later times. This Perspective investigates metrics that can help indicate which mols. will shift substantially in prediction and when the shift will occur.
  18. 18
    Rydberg, P.; Gloriam, D. E.; Zaretzki, J.; Breneman, C.; Olsen, L. SMARTCyp: a 2D method for prediction of cytochrome P450-mediated drug metabolism. ACS medicinal chemistry letters 2010, 1 (3), 96– 100,  DOI: 10.1021/ml100016x
    Google Scholar
    There is no corresponding record for this reference.

Cited By

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

This article is cited by 1 publications.

  1. Srijit Seal, Manas Mahale, Miguel García-Ortegón, Chaitanya K. Joshi, Layla Hosseini-Gerami, Alex Beatson, Matthew Greenig, Mrinal Shekhar, Arijit Patra, Caroline Weis, Arash Mehrjou, Adrien Badré, Brianna Paisley, Rhiannon Lowe, Shantanu Singh, Falgun Shah, Bjarki Johannesson, Dominic Williams, David Rouquie, Djork-Arné Clevert, Patrick Schwab, Nicola Richmond, Christos A. Nicolaou, Raymond J. Gonzalez, Russell Naven, Carolin Schramm, Lewis R Vidler, Kamel Mansouri, W. Patrick Walters, Deidre Dalmas Wilk, Ola Spjuth, Anne E. Carpenter, Andreas Bender. Machine Learning for Toxicity Prediction Using Chemical Structures: Pillars for Success in the Real World. Chemical Research in Toxicology 2025, 38 (5) , 759-807. https://doi.org/10.1021/acs.chemrestox.5c00033
Download PDF
Go to ACS Medicinal Chemistry Letters
Get e-Alerts
Get e-Alerts

ACS Medicinal Chemistry Letters

Cite this: ACS Med. Chem. Lett. 2024, 15, 8, 1169–1173
Click to copy citationCitation copied!
https://doi.org/10.1021/acsmedchemlett.4c00290
Published July 25, 2024

Copyright © 2024 The Authors. Published by American Chemical Society. This publication is licensed under

CC-BY 4.0 .

Article Views

9944

Altmetric

-

Citations

1
Learn about these metrics

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

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

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

  • Figure 1

    Figure 1. Performance of models trained on program data only (local only), nonprogram data only (global only), and on combined data (fine-tuned global), on temporally split test sets for HLM, RLM, MDCK AB, and MDCK ER. Error bars represent 68% bootstrapped confidence intervals. MAE (mean absolute error) units are in log 10 (mL/min/kg) for HLM and RLM, log 10 (μcm/s) for MDCK AB, and log 10 (ratio) for ER.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 2

    Figure 2. A screenshot of the interactive design environment with ML ADME predictions, comparison to a reference compound, and highlights of sites of likely metabolism.

    High Resolution Image
    Download MS PowerPoint Slide

  • References

    This article references 18 other publications.

    1. 1
      Ortwine, D. F.; Aliagas, I. Physicochemical and DMPK in silico models: facilitating their use by medicinal chemists. Mol. Pharmaceutics 2013, 10 (4), 1153– 1161,  DOI: 10.1021/mp3006193
      1
      Physicochemical and DMPK In Silico Models: Facilitating Their Use by Medicinal Chemists
      Ortwine, Daniel F.; Aliagas, Ignacio
      Molecular Pharmaceutics (2013), 10 (4), 1153-1161CODEN: MPOHBP; ISSN:1543-8384. (American Chemical Society)
      A review. It is known that the developability of drugs is related to their physicochem. and DMPK properties. Given the time and expense involved in discovering and developing new drugs, maximizing the chance of success by calcg. properties ahead of chem. synthesis and testing, and only acting on those candidates whose properties fall into a desired range, would seem to make sense. This paper provides an overview of calculable physicochem. and DMPK properties, an assessment of their relative difficulty of their calcn. and accuracy, and available software. Methods companies have employed to communicate results will be discussed, including the use of composite scoring functions and ranking schemes. Calcns. do no good if chemists will not use them to prioritize synthesis decisions. Strategies are presented for facilitating model usage. An approach adopted at Genentech for presenting results that involves the close coupling of property calcns. with 3D structure based drug design is described.
    2. 2
      Goller, A. H.; Kuhnke, L.; Montanari, F.; Bonin, A.; Schneckener, S.; ter Laak, A.; Wichard, J.; Lobell, M.; Hillisch, A. Bayer’s in silico ADMET platform: a journey of machine learning over the past two decades. Drug Discovery Today 2020, 25 (9), 1702– 1709,  DOI: 10.1016/j.drudis.2020.07.001
      2
      Bayer's in silico ADMET platform: a journey of machine learning over the past two decades
      Goller Andreas H; Bonin Anne; Lobell Mario; Kuhnke Lara; Ter Laak Antonius; Montanari Floriane; Schneckener Sebastian; Wichard Jorg; Hillisch Alexander
      Drug discovery today (2020), 25 (9), 1702-1709 ISSN:.
      Over the past two decades, an in silico absorption, distribution, metabolism, and excretion (ADMET) platform has been created at Bayer Pharma with the goal to generate models for a variety of pharmacokinetic and physicochemical endpoints in early drug discovery. These tools are accessible to all scientists within the company and can be a useful in assisting with the selection and design of novel leads, as well as the process of lead optimization. Here. we discuss the development of machine-learning (ML) approaches with special emphasis on data, descriptors, and algorithms. We show that high company internal data quality and tailored descriptors, as well as a thorough understanding of the experimental endpoints, are essential to the utility of our models. We discuss the recent impact of deep neural networks and show selected application examples.
    3. 3
      Wu, Z.; Ramsundar, B.; Feinberg, E. N.; Gomes, J.; Geniesse, C.; Pappu, A. S.; Leswing, K.; Pande, V. MoleculeNet: a benchmark for molecular machine learning. Chemical science 2018, 9 (2), 513– 530,  DOI: 10.1039/C7SC02664A
      3
      MoleculeNet: a benchmark for molecular machine learning
      Wu, Zhenqin; Ramsundar, Bharath; Feinberg, Evan N.; Gomes, Joseph; Geniesse, Caleb; Pappu, Aneesh S.; Leswing, Karl; Pande, Vijay
      Chemical Science (2018), 9 (2), 513-530CODEN: CSHCCN; ISSN:2041-6520. (Royal Society of Chemistry)
      Mol. machine learning has been maturing rapidly over the last few years. Improved methods and the presence of larger datasets have enabled machine learning algorithms to make increasingly accurate predictions about mol. properties. However, algorithmic progress has been limited due to the lack of a std. benchmark to compare the efficacy of proposed methods; most new algorithms are benchmarked on different datasets making it challenging to gauge the quality of proposed methods. This work introduces MoleculeNet, a large scale benchmark for mol. machine learning. MoleculeNet curates multiple public datasets, establishes metrics for evaluation, and offers high quality open-source implementations of multiple previously proposed mol. featurization and learning algorithms (released as part of the DeepChem open source library). MoleculeNet benchmarks demonstrate that learnable representations are powerful tools for mol. machine learning and broadly offer the best performance. However, this result comes with caveats. Learnable representations still struggle to deal with complex tasks under data scarcity and highly imbalanced classification. For quantum mech. and biophys. datasets, the use of physics-aware featurizations can be more important than choice of particular learning algorithm.
    4. 4
      Sheridan, R. P. Time-split cross-validation as a method for estimating the goodness of prospective prediction. J. Chem. Inf. Model. 2013, 53 (4), 783– 790,  DOI: 10.1021/ci400084k
      4
      Time-Split Cross-Validation as a Method for Estimating the Goodness of Prospective Prediction.
      Sheridan, Robert P.
      Journal of Chemical Information and Modeling (2013), 53 (4), 783-790CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      Cross-validation is a common method to validate a QSAR model. In cross-validation, some compds. are held out as a test set, while the remaining compds. form a training set. A model is built from the training set, and the test set compds. are predicted on that model. The agreement of the predicted and obsd. activity values of the test set (measured by, say, R2) is an est. of the self-consistency of the model and is sometimes taken as an indication of the predictivity of the model. This est. of predictivity can be optimistic or pessimistic compared to true prospective prediction, depending how compds. in the test set are selected. Here, we show that time-split selection gives an R2 that is more like that of true prospective prediction than the R2 from random selection (too optimistic) or from our analog of leave-class-out selection (too pessimistic). Time-split selection should be used in addn. to random selection as a std. for cross-validation in QSAR model building.
    5. 5
      Fluetsch, A.; Di Lascio, E.; Gerebtzoff, G.; Rodríguez-Pérez, R. Adapting Deep Learning QSPR Models to Specific Drug Discovery Projects. Molecular Pharmaceutics 2024, 21, 1817,  DOI: 10.1021/acs.molpharmaceut.3c01124
      There is no corresponding record for this reference.
    6. 6
      Muratov, E. N.; Bajorath, J.; Sheridan, R. P.; Tetko, I. V.; Filimonov, D.; Poroikov, V.; Oprea, T. I.; Baskin, I. I.; Varnek, A.; Roitberg, A.; Isayev, O.; Curtalolo, S.; Fourches, D.; Cohen, Y.; Aspuru-Guzik, A.; Winkler, D. A.; Agrafiotis, D.; Cherkasov, A.; Tropsha, A. QSAR without borders. Chem. Soc. Rev. 2020, 49 (11), 3525– 3564,  DOI: 10.1039/D0CS00098A
      6
      QSAR without borders
      Muratov, Eugene N.; Bajorath, Jurgen; Sheridan, Robert P.; Tetko, Igor V.; Filimonov, Dmitry; Poroikov, Vladimir; Oprea, Tudor I.; Baskin, Igor I.; Varnek, Alexandre; Roitberg, Adrian; Isayev, Olexandr; Curtalolo, Stefano; Fourches, Denis; Cohen, Yoram; Aspuru-Guzik, Alan; Winkler, David A.; Agrafiotis, Dimitris; Cherkasov, Artem; Tropsha, Alexander
      Chemical Society Reviews (2020), 49 (11), 3525-3564CODEN: CSRVBR; ISSN:0306-0012. (Royal Society of Chemistry)
      A review. Prediction of chem. bioactivity and phys. properties has been one of the most important applications of statistical and more recently, machine learning and artificial intelligence methods in chem. sciences. This field of research, broadly known as quant. structure-activity relationships (QSAR) modeling, has developed many important algorithms and has found a broad range of applications in phys. org. and medicinal chem. in the past 55+ years. This Perspective summarizes recent technol. advances in QSAR modeling but it also highlights the applicability of algorithms, modeling methods, and validation practices developed in QSAR to a wide range of research areas outside of traditional QSAR boundaries including synthesis planning, nanotechnol., materials science, biomaterials, and clin. informatics. As modern research methods generate rapidly increasing amts. of data, the knowledge of robust data-driven modeling methods professed within the QSAR field can become essential for scientists working both within and outside of chem. research. We hope that this contribution highlighting the generalizable components of QSAR modeling will serve to address this challenge.
    7. 7
      Xiong, G.; Wu, Z.; Yi, J.; Fu, L.; Yang, Z.; Hsieh, C.; Yin, M.; Zeng, X.; Wu, C.; Lu, A.; Chen, X.; Hou, T.; Cao, D. ADMETlab 2.0: an integrated online platform for accurate and comprehensive predictions of ADMET properties. Nucleic acids research 2021, 49 (W1), W5– W14,  DOI: 10.1093/nar/gkab255
      There is no corresponding record for this reference.
    8. 8
      Kawashima, H.; Watanabe, R.; Esaki, T.; Kuroda, M.; Nagao, C.; Natsume-Kitatani, Y.; Ohashi, R.; Komura, H.; Mizuguchi, K. DruMAP: A novel drug metabolism and pharmacokinetics analysis platform. J. Med. Chem. 2023, 66 (14), 9697– 9709,  DOI: 10.1021/acs.jmedchem.3c00481
      There is no corresponding record for this reference.
    9. 9
      Di Lascio, E.; Gerebtzoff, G.; Rodríguez-Pérez, R. Systematic evaluation of local and global machine learning models for the prediction of ADME properties. Mol. Pharmaceutics 2023, 20 (3), 1758– 1767,  DOI: 10.1021/acs.molpharmaceut.2c00962
      9
      Systematic evaluation of local and global machine learning models for the prediction of ADME properties
      Di Lascio, Elena; Gerebtzoff, Gregori; Rodriguez-Perez, Raquel
      Molecular Pharmaceutics (2023), 20 (3), 1758-1767CODEN: MPOHBP; ISSN:1543-8384. (American Chemical Society)
      Machine learning (ML) has become an indispensable tool to predict absorption, distribution, metab., and excretion (ADME) properties in pharmaceutical research. ML algorithms are trained on mol. structures and corresponding ADME assay data to develop quant. structure-property relationship (QSPR) models. Traditional QSPR models were trained on compd. sets of limited size. With the advent of more complex ML algorithms and data availability, training sets have become larger and more diverse. Most common training approaches consist in either training a model with a small set of similar compds., namely, compds. designed for the same drug discovery project or chem. series (local model approach) or with a larger set of diverse compds. (global model approach). Global models are built with all exptl. data available for an assay, combining compd. data from different projects and disease areas. Despite the ML progress made so far, the choice of the appropriate data compn. for building ML models is still unclear. Herein, a systematic evaluation of local and global ML models was performed for 10 different exptl. assays and 112 drug discovery projects. Results show a consistent superior performance of global models for ADME property predictions. Diagnostic analyses were also carried out to investigate the influence of training set size, structural diversity, and data shift in the relative performance of local and global ML models. Training set and structural diversity did not have an impact in the relative performance on the methods. Instead, data shift helped to identify the projects with larger performance differences between local and global models. Results presented in this work can be leveraged to improve ML-based ADME properties predictions and thus decision-making in drug discovery projects.
    10. 10
      Aliagas, I.; Gobbi, A.; Heffron, T.; Lee, M. L.; Ortwine, D. F.; Zak, M.; Khojasteh, S. C. A probabilistic method to report predictions from a human liver microsomes stability QSAR model: a practical tool for drug discovery. Journal of Computer-Aided Molecular Design 2015, 29, 327– 338,  DOI: 10.1007/s10822-015-9838-3
      There is no corresponding record for this reference.
    11. 11
      Wieder, O.; Kohlbacher, S.; Kuenemann, M.; Garon, A.; Ducrot, P.; Seidel, T.; Langer, T. A compact review of molecular property prediction with graph neural networks. Drug Discovery Today: Technologies 2020, 37, 1– 12,  DOI: 10.1016/j.ddtec.2020.11.009
      11
      A compact review of molecular property prediction with graph neural networks
      Wieder Oliver; Kohlbacher Stefan; Garon Arthur; Ducrot Pierre; Seidel Thomas; Kuenemann Melaine; Langer Thierry
      Drug discovery today. Technologies (2020), 37 (), 1-12 ISSN:.
      As graph neural networks are becoming more and more powerful and useful in the field of drug discovery, many pharmaceutical companies are getting interested in utilizing these methods for their own in-house frameworks. This is especially compelling for tasks such as the prediction of molecular properties which is often one of the most crucial tasks in computer-aided drug discovery workflows. The immense hype surrounding these kinds of algorithms has led to the development of many different types of promising architectures and in this review we try to structure this highly dynamic field of AI-research by collecting and classifying 80 GNNs that have been used to predict more than 20 molecular properties using 48 different datasets.
    12. 12
      Karmaker, S. K.; Hassan, M. M.; Smith, M. J.; Xu, L.; Zhai, C.; Veeramachaneni, K. AutoML to Date and Beyond: Challenges and Opportunities. ACM Computing Surveys (CSUR) 2022, 54 (8), 1– 36,  DOI: 10.1145/3470918
      There is no corresponding record for this reference.
    13. 13
      Yu, J.; Wang, D.; Zheng, M. Uncertainty quantification: Can we trust artificial intelligence in drug discovery?. iScience 2022, 25 (8), 104814,  DOI: 10.1016/j.isci.2022.104814
      There is no corresponding record for this reference.
    14. 14
      Sheridan, R. P.; Culberson, J. C.; Joshi, E.; Tudor, M.; Karnachi, P. Prediction accuracy of production ADMET models as a function of version: activity cliffs rule. J. Chem. Inf. Model. 2022, 62 (14), 3275– 3280,  DOI: 10.1021/acs.jcim.2c00699
      14
      Prediction Accuracy of Production ADMET Models as a Function of Version: Activity Cliffs Rule
      Sheridan, Robert P.; Culberson, J. Chris; Joshi, Elizabeth; Tudor, Matthew; Karnachi, Prabha
      Journal of Chemical Information and Modeling (2022), 62 (14), 3275-3280CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      As with many other institutions, our company maintains many quant. structure-activity relationship (QSAR) models of absorption, distribution, metab., excretion, and toxicity (ADMET) end points and updates the models regularly. We recently examd. version-to-version predictivity for these models over a period of 10 years. In this approach we monitor the goodness of prediction of new mols. relative to the training set of model version V before they are incorporated in the updated model V+1. Using a cell-based permeability assay (Papp) as an example, we illustrate how the QSAR models made from this data are generally predictive and can be utilized to enrich chem. designs and synthesis. Despite the obvious utility of these models, we turned up unexpected behavior in Papp and other ADMET activities for which the explanation is not obvious. One such behavior is that the apparent predictivity of the models as measured by root-mean-square-error can vary greatly from version to version and is sometimes very poor. One intuitively appealing explanation is that the obsd. activities of the new mols. fall outside the bulk of activities in the training set. Alternatively, one may think that the new mols. are exploring different regions of chem. space than the training set. However, the real explanation has to do with activity cliffs. If the obsd. activities of the new mols. are different than expected based on similar mols. in the training set, the predictions will be less accurate. This is true for all our ADMET end points.
    15. 15
      Aleksić, S.; Seeliger, D.; Brown, J. B. ADMET Predictability at Boehringer Ingelheim: State of the Art, and Do Bigger Datasets or Algorithms Make a Difference?. Molecular Informatics 2022, 41 (2), 2100113,  DOI: 10.1002/minf.202100113
      There is no corresponding record for this reference.
    16. 16
      Fang, C.; Wang, Y.; Grater, R.; Kapadnis, S.; Black, C.; Trapa, P.; Sciabola, S. Prospective validation of machine learning algorithms for absorption, distribution, metabolism, and excretion prediction: An industrial perspective. J. Chem. Inf. Model. 2023, 63 (11), 3263– 3274,  DOI: 10.1021/acs.jcim.3c00160
      There is no corresponding record for this reference.
    17. 17
      Sheridan, R. P. Stability of prediction in production ADMET models as a function of version: why and when predictions change. J. Chem. Inf. Model. 2022, 62 (15), 3477– 3485,  DOI: 10.1021/acs.jcim.2c00803
      17
      Stability of Prediction in Production ADMET Models as a Function of Version: Why and When Predictions Change
      Sheridan, Robert P.
      Journal of Chemical Information and Modeling (2022), 62 (15), 3477-3485CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      As with other pharma companies, we maintain prodn. QSAR models of ADMET end points and update them regularly. Here, for six ADMET end points, we examine the predictions of test set mols. on multiple versions of random forest models spanning a period of 10 years. For any given end point, the predictions for the majority of mols. are similar for all model versions. However, for a small minority of mols., the prediction shifts substantially over the span of a few versions. For most mols. that shift, the prediction becomes more accurate at later times. This Perspective investigates metrics that can help indicate which mols. will shift substantially in prediction and when the shift will occur.
    18. 18
      Rydberg, P.; Gloriam, D. E.; Zaretzki, J.; Breneman, C.; Olsen, L. SMARTCyp: a 2D method for prediction of cytochrome P450-mediated drug metabolism. ACS medicinal chemistry letters 2010, 1 (3), 96– 100,  DOI: 10.1021/ml100016x
      There is no corresponding record for this reference.