Steven V. Jerome
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Steven V. Jerome has a total of 82 co-authors. Here are the top co-authors by number of shared publications:
- Richard A. Friesner
Department of Chemistry, Columbia University, New York, USA
4 shared publications - Yuqi Zhang
Schrodinger (United States)
4 shared publications - Kun Yao
Schrödinger Inc, 1 Main St, 11th Floor, Cambridge, MA, 02142
3 shared publications - Matthew P. Repasky
Schrödinger, Inc., 101 SW Main Street, Suite 1300, Portland, Oregon 97204, United States
3 shared publications Da Shi
Schrödinger Inc, 9868 Scranton Rd Suite 3200, San Diego, CA 92121, United States
2 shared publications- Da Shi
Schrödinger Inc., 9868 Scranton Road, San Diego, California, 92121
2 shared publications
Latest Publications
- Open Access
- 4 June 2026
- Published by American Chemical Society (ACS) in ChemRxiv
Generative Glide: A De Novo Workflow for Ultra-large Library Screening
In the past few years, chemical vendors have embraced enormous virtual spacesoften exceeding 1010 molecules. The sheer size of these spaces presents a formidable computational challenge for traditional structure-based virtual screening. Exhaustive docking of such spaces is computationally prohibitive, and even modern active learning docking algorithms, which leverage fast ligand-based machine learning to identify the most promising candidates for docking, struggle to achieve an acceptable turnaround time. Methods based on de novo molecular generation using the reactions that define the virtual space can bypass the challenge of scoring every molecule in the space, but will often yield candidates where the synthesis as described will not work in the lab. To address these challenges, we present the Generative Glide workflow, an integrated framework that bridges the gap between in silico generation and physical acquisition. Generative Glide employs an ensemble of reinforcement learning agents optimized against a supervised GCN-based QSAR model to navigate toward chemical space regions with improved predicted binding likelihood. We implement a two-tiered strategy to ensure tractability: a high-throughput Bloom filter to rapidly identify molecules within ultra-large commercial catalogs and a retrosynthetic analysis engine to validate novel out-of-library candidates. Deployment across diverse protein targets and multiple ultra-large libraries demonstrates that the workflow consistently achieves a significant shift in docking score distributions toward better scoring compounds while maintaining a high level of scaffold diversity. Notably, the architecture is designed to be light-weight, enabling a relatively fast turnaround time compared to exhaustive enumeration. By incorporating active learning, Generative Glide further refines the scoring model’s predictive accuracy within high-potency domains, enabling the agents to effectively navigate and exploit the extreme negative tail of the docking score distribution that is poorly represented in the initial training data. This workflow provides a scalable, generative solution to augment and accelerate the virtual screening of modern chemical space.
- 4 June 2026
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessGenerative Glide: A De Novo Workflow for Ultra-large Library Screening
In the past few years, chemical vendors have embraced enormous virtual spacesoften exceeding 1010 molecules. The sheer size of these spaces presents a formidable computational challenge for traditional structure-based virtual screening. Exhaustive docking of such spaces is computationally prohibitive, and even modern active learning docking algorithms, which leverage fast ligand-based machine learning to identify the most promising candidates for docking, struggle to achieve an acceptable turnaround time. Methods based on de novo molecular generation using the reactions that define the virtual space can bypass the challenge of scoring every molecule in the space, but will often yield candidates where the synthesis as described will not work in the lab. To address these challenges, we present the Generative Glide workflow, an integrated framework that bridges the gap between in silico generation and physical acquisition. Generative Glide employs an ensemble of reinforcement learning agents optimized against a supervised GCN-based QSAR model to navigate toward chemical space regions with improved predicted binding likelihood. We implement a two-tiered strategy to ensure tractability: a high-throughput Bloom filter to rapidly identify molecules within ultra-large commercial catalogs and a retrosynthetic analysis engine to validate novel out-of-library candidates. Deployment across diverse protein targets and multiple ultra-large libraries demonstrates that the workflow consistently achieves a significant shift in docking score distributions toward better scoring compounds while maintaining a high level of scaffold diversity. Notably, the architecture is designed to be light-weight, enabling a relatively fast turnaround time compared to exhaustive enumeration. By incorporating active learning, Generative Glide further refines the scoring model’s predictive accuracy within high-potency domains, enabling the agents to effectively navigate and exploit the extreme negative tail of the docking score distribution that is poorly represented in the initial training data. This workflow provides a scalable, generative solution to augment and accelerate the virtual screening of modern chemical space.
- 23 April 2026
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 66 (10) , 5641-5649
DiffDock-Glide: A Hybrid Physics-Based and Data-Driven Approach to Molecular Docking
Recent years have seen a rise in applications of deep learning to problems in the molecular sciences. Among them, the diffusion model DiffDock stands out as a method for docking small molecules into protein binding sites. But DiffDock struggles to compete with conventional docking methods, especially for targets outside its training set. We develop a hybrid model called DiffDock-Glide which addresses some shortcomings of deep learning docking methods: it uses a modified generative process to generate samples within a binding pocket, and the confidence model is replaced with Glide's postdocking minimization pipeline. We evaluate DiffDock-Glide on the PoseBusters data set and show improved sampling of near-native poses, especially for sequences without homologues in the training set. We also evaluate DiffDock-Glide's performance in virtual screening of compounds from the DUD-E data set against receptor structures generated by AlphaFold2 and report enrichment values that broadly surpass those from traditional Glide.
- 23 April 2026
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 66 (10) , 5641-5649
DiffDock-Glide: A Hybrid Physics-Based and Data-Driven Approach to Molecular Docking
Recent years have seen a rise in applications of deep learning to problems in the molecular sciences. Among them, the diffusion model DiffDock stands out as a method for docking small molecules into protein binding sites. But DiffDock struggles to compete with conventional docking methods, especially for targets outside its training set. We develop a hybrid model called DiffDock-Glide which addresses some shortcomings of deep learning docking methods: it uses a modified generative process to generate samples within a binding pocket, and the confidence model is replaced with Glide's postdocking minimization pipeline. We evaluate DiffDock-Glide on the PoseBusters data set and show improved sampling of near-native poses, especially for sequences without homologues in the training set. We also evaluate DiffDock-Glide's performance in virtual screening of compounds from the DUD-E data set against receptor structures generated by AlphaFold2 and report enrichment values that broadly surpass those from traditional Glide.
- 3 December 2025
- Published by American Chemical Society (ACS) in Journal of Chemical Theory and Computation
- Vol. 21 (24) , 12696-12708
Glide WS: Methodology and Initial Assessment of Performance for Docking Accuracy and Virtual Screening
Powered by dramatic advances in computer hardware, the advent of ultralarge make-on-demand virtual libraries, and a shift in small-molecule discovery toward more challenging targets with limited known actives, there has been a growing interest in the development of performant virtual screening methods that can reliably deliver novel hits. We report on a new method called Glide WS, that builds on our earlier efforts (WScore) to introduce an explicit representation of water structure and dynamics to an empirical scoring function suitable for high-throughput docking. This scoring function has been carefully tuned using absolute binding free energy perturbation calculations (ABFEP). Compared with Glide SP, Glide WS offers significant gains in the two primary tasks for molecular docking in drug discovery, pose prediction and virtual screening enrichment. For docking accuracy, Glide WS achieves a self-docking accuracy of 92% on a diverse set of 1477 protein ligand complexes as compared to 85% for Glide SP, using a criterion of 2.5 Å. We also demonstrate significantly improved virtual screening enrichment using a diverse data set covering of 38 targets together with three different computationally generated libraries of decoys, combined with standard known ChEMBL actives. We focus on ligands ranked in the top few percent of the database (the subset that is relevant to practical virtual screening efforts) and demonstrate that, along with improved enrichment of ChEMBL actives, Glide WS achieves a remarkable reduction in the number of poorly scoring decoys (as calibrated by ABFEP calculations), across a high percentage of targets, as compared to Glide SP. These results suggest that considerably higher hit rates will be observed, as compared to conventional rigid receptor docking, in practical virtual screening applications.
- 3 December 2025
- Published by American Chemical Society (ACS) in Journal of Chemical Theory and Computation
- Vol. 21 (24) , 12696-12708
Glide WS: Methodology and Initial Assessment of Performance for Docking Accuracy and Virtual Screening
Powered by dramatic advances in computer hardware, the advent of ultralarge make-on-demand virtual libraries, and a shift in small-molecule discovery toward more challenging targets with limited known actives, there has been a growing interest in the development of performant virtual screening methods that can reliably deliver novel hits. We report on a new method called Glide WS, that builds on our earlier efforts (WScore) to introduce an explicit representation of water structure and dynamics to an empirical scoring function suitable for high-throughput docking. This scoring function has been carefully tuned using absolute binding free energy perturbation calculations (ABFEP). Compared with Glide SP, Glide WS offers significant gains in the two primary tasks for molecular docking in drug discovery, pose prediction and virtual screening enrichment. For docking accuracy, Glide WS achieves a self-docking accuracy of 92% on a diverse set of 1477 protein ligand complexes as compared to 85% for Glide SP, using a criterion of 2.5 Å. We also demonstrate significantly improved virtual screening enrichment using a diverse data set covering of 38 targets together with three different computationally generated libraries of decoys, combined with standard known ChEMBL actives. We focus on ligands ranked in the top few percent of the database (the subset that is relevant to practical virtual screening efforts) and demonstrate that, along with improved enrichment of ChEMBL actives, Glide WS achieves a remarkable reduction in the number of poorly scoring decoys (as calibrated by ABFEP calculations), across a high percentage of targets, as compared to Glide SP. These results suggest that considerably higher hit rates will be observed, as compared to conventional rigid receptor docking, in practical virtual screening applications.
- Open Access
- 11 August 2025
- Published by American Chemical Society (ACS) in ChemRxiv
Glide WS: Methodology and Initial Assessment of Performance for Docking Accuracy and Virtual Screening
Powered by dramatic advances in computer hardware, the advent of ultra-large make-on-demand virtual libraries, and a shift in small molecule discovery toward more challenging targets with limited known actives, there has been a growing interest in the development of performant virtual screening methods that can reliably deliver novel hits. We report on a new method called Glide WS, that builds on our earlier efforts (WScore) to introduce an explicit representation of water structure and dynamics to an empirical scoring function suitable for high-throughput docking. This scoring function has been carefully tuned using Absolute Binding Free Energy Perturbation calculations (ABFEP). Compared with Glide SP, Glide WS offers significant gains in the two primary tasks for molecular docking in drug discovery, pose prediction and virtual screening enrichment. For docking accuracy, Glide WS achieves a self-docking accuracy of 92% on a diverse set of 1477 protein ligand complexes as compared to only 85% for Glide SP, using a criterion of 2.5 Å. We also demonstrate significantly improved virtual screening enrichment using a diverse dataset covering of 38 targets together with three different computationally generated libraries of decoys, combined with standard known ChEMBL actives). We focus ligands ranked in the top few percent of the database (the subset that is relevant to practical virtual screening efforts) and demonstrate that, along with improved enrichment of ChEMBL actives, Glide WS achieves a remarkable reduction in the number of poorly scoring decoys (as calibrated by ABFEP calculations), across a high percentage of targets, as compared to Glide SP. These results suggest that considerably higher hit rates will be observed, as compared to conventional rigid receptor docking, in practical virtual screening applications.
- 11 August 2025
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessGlide WS: Methodology and Initial Assessment of Performance for Docking Accuracy and Virtual Screening
Powered by dramatic advances in computer hardware, the advent of ultra-large make-on-demand virtual libraries, and a shift in small molecule discovery toward more challenging targets with limited known actives, there has been a growing interest in the development of performant virtual screening methods that can reliably deliver novel hits. We report on a new method called Glide WS, that builds on our earlier efforts (WScore) to introduce an explicit representation of water structure and dynamics to an empirical scoring function suitable for high-throughput docking. This scoring function has been carefully tuned using Absolute Binding Free Energy Perturbation calculations (ABFEP). Compared with Glide SP, Glide WS offers significant gains in the two primary tasks for molecular docking in drug discovery, pose prediction and virtual screening enrichment. For docking accuracy, Glide WS achieves a self-docking accuracy of 92% on a diverse set of 1477 protein ligand complexes as compared to only 85% for Glide SP, using a criterion of 2.5 Å. We also demonstrate significantly improved virtual screening enrichment using a diverse dataset covering of 38 targets together with three different computationally generated libraries of decoys, combined with standard known ChEMBL actives). We focus ligands ranked in the top few percent of the database (the subset that is relevant to practical virtual screening efforts) and demonstrate that, along with improved enrichment of ChEMBL actives, Glide WS achieves a remarkable reduction in the number of poorly scoring decoys (as calibrated by ABFEP calculations), across a high percentage of targets, as compared to Glide SP. These results suggest that considerably higher hit rates will be observed, as compared to conventional rigid receptor docking, in practical virtual screening applications.
- Open Access
DiffDock-Glide: a hybrid physics-based and data-driven approach to molecular docking
Recent years have seen a rise in applications of deep learning to problems in the molecular sciences. Among them, the diffusion model DiffDock stands out as a method for docking small molecules into protein binding sites. But DiffDock struggles to compete with conventional docking methods, especially for targets outside its training set. We develop a hybrid model called DiffDock-Glide which addresses some shortcomings of deep learning docking methods: it uses a modified generative process to generate samples within a binding pocket and the confidence model is replaced with Glide’s post-docking minimization pipeline. We evaluate DiffDock-Glide on the Posebusters dataset and show improved sampling of near-native poses, especially for sequences without homologues in the training set. We also evaluate DiffDock-Glide’s performance in virtual screening compounds from the DUD-E dataset against receptor structures generated by AlphaFold2 and report enrichment values that broadly surpass those from traditional Glide.
Open AccessDiffDock-Glide: a hybrid physics-based and data-driven approach to molecular docking
Recent years have seen a rise in applications of deep learning to problems in the molecular sciences. Among them, the diffusion model DiffDock stands out as a method for docking small molecules into protein binding sites. But DiffDock struggles to compete with conventional docking methods, especially for targets outside its training set. We develop a hybrid model called DiffDock-Glide which addresses some shortcomings of deep learning docking methods: it uses a modified generative process to generate samples within a binding pocket and the confidence model is replaced with Glide’s post-docking minimization pipeline. We evaluate DiffDock-Glide on the Posebusters dataset and show improved sampling of near-native poses, especially for sequences without homologues in the training set. We also evaluate DiffDock-Glide’s performance in virtual screening compounds from the DUD-E dataset against receptor structures generated by AlphaFold2 and report enrichment values that broadly surpass those from traditional Glide.
- 20 May 2025
- Published by American Chemical Society (ACS) in Journal of Medicinal Chemistry
- Vol. 68 (11) , 10507-10519
Computational Hit Finding: An Industry Perspective
Computational hit finding, particularly virtual screening, has been a mainstay of drug discovery campaigns for decades, providing a cost-efficient complement to wet experiments. Innovation in this space slowed considerably as these approaches converged around mature software programs and stock chemical libraries up to ∼10 million in size. Recently, however, powered by massive increases in computational power, the emergence of ultralarge make-on-demand virtual libraries, the development of large capacity neural networks, the expansion of the domain of applicability of free energy calculations, and advances in protein structure prediction, the virtual screening field is currently seeing major change. We present a guide from industry practitioners summarizing key aspects on the changing computational hit finding landscape including practical recommendations for building a performant end-to-end screening workflow, strategies to mitigate risk by avoiding common pitfalls, determining success criteria, and a brief discussion of emerging technologies likely to impact drug discovery in the near future.
- 20 May 2025
- Published by American Chemical Society (ACS) in Journal of Medicinal Chemistry
- Vol. 68 (11) , 10507-10519
Computational Hit Finding: An Industry Perspective
Computational hit finding, particularly virtual screening, has been a mainstay of drug discovery campaigns for decades, providing a cost-efficient complement to wet experiments. Innovation in this space slowed considerably as these approaches converged around mature software programs and stock chemical libraries up to ∼10 million in size. Recently, however, powered by massive increases in computational power, the emergence of ultralarge make-on-demand virtual libraries, the development of large capacity neural networks, the expansion of the domain of applicability of free energy calculations, and advances in protein structure prediction, the virtual screening field is currently seeing major change. We present a guide from industry practitioners summarizing key aspects on the changing computational hit finding landscape including practical recommendations for building a performant end-to-end screening workflow, strategies to mitigate risk by avoiding common pitfalls, determining success criteria, and a brief discussion of emerging technologies likely to impact drug discovery in the near future.
- Open Access
- 23 April 2025
- Published by American Chemical Society (ACS) in ChemRxiv
Enabling in-silico Hit Discovery Workflows Targeting RNA with Small Molecules
Ribonucleic acids (RNAs) are emerging as important drug targets, yet in silico modeling of RNA-small molecule interactions presents unique challenges compared to protein targets. This study evaluates and improves computational tools within the Schrödinger Suite for hit discovery against RNA. We benchmarked and enhanced SiteMap for identifying RNA binding sites, introducing an RNA-specific scoring function that improved site prediction accuracy from 44.4% to 65.3% on a filtered HARIBOSS dataset. We also refined Glide, Schrödinger's molecular docking tool, by rebalancing scoring function weights for RNA systems, resulting in improved pose prediction accuracy to 73.7% on a recently published dataset. Furthermore, we expanded the applicability of absolute binding free energy perturbation (ABFEP) to nucleic acid receptors, demonstrating good correlation with experimental binding affinities (RMSE of 1.10 kcal/mol) on a diverse set of 5 RNA-ligand systems. These advancements provide a promising foundation for structure-based drug discovery targeting RNA, although further refinements are needed to achieve performance levels comparable to protein targets, particularly in addressing RNA's dynamic nature and docking into apo receptors.
- 23 April 2025
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessEnabling in-silico Hit Discovery Workflows Targeting RNA with Small Molecules
Ribonucleic acids (RNAs) are emerging as important drug targets, yet in silico modeling of RNA-small molecule interactions presents unique challenges compared to protein targets. This study evaluates and improves computational tools within the Schrödinger Suite for hit discovery against RNA. We benchmarked and enhanced SiteMap for identifying RNA binding sites, introducing an RNA-specific scoring function that improved site prediction accuracy from 44.4% to 65.3% on a filtered HARIBOSS dataset. We also refined Glide, Schrödinger's molecular docking tool, by rebalancing scoring function weights for RNA systems, resulting in improved pose prediction accuracy to 73.7% on a recently published dataset. Furthermore, we expanded the applicability of absolute binding free energy perturbation (ABFEP) to nucleic acid receptors, demonstrating good correlation with experimental binding affinities (RMSE of 1.10 kcal/mol) on a diverse set of 5 RNA-ligand systems. These advancements provide a promising foundation for structure-based drug discovery targeting RNA, although further refinements are needed to achieve performance levels comparable to protein targets, particularly in addressing RNA's dynamic nature and docking into apo receptors.
- 11 May 2023
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 63 (10) , 3171-3185
Enhancing Hit Discovery in Virtual Screening through Absolute Protein–Ligand Binding Free-Energy Calculations
In the hit identification stage of drug discovery, a diverse chemical space needs to be explored to identify initial hits. Contrary to empirical scoring functions, absolute protein-ligand binding free-energy perturbation (ABFEP) provides a theoretically more rigorous and accurate description of protein-ligand binding thermodynamics and could, in principle, greatly improve the hit rates in virtual screening. In this work, we describe an implementation of an accurate and reliable ABFEP method in FEP+. We validated the ABFEP method on eight congeneric compound series binding to eight protein receptors including both neutral and charged ligands. For ligands with net charges, the alchemical ion approach is adopted to avoid artifacts in electrostatic potential energy calculations. The calculated binding free energies correlate with experimental results with a weighted average of R2 = 0.55 for the entire dataset. We also observe an overall root-mean-square error (RMSE) of 1.1 kcal/mol after shifting the zero-point of the simulation data to match the average experimental values. Through ABFEP calculations using apo versus holo protein structures, we demonstrated that the protein conformational and protonation state changes between the apo and holo proteins are the main physical factors contributing to the protein reorganization free energy manifested by the overestimation of raw ABFEP calculated binding free energies using the holo structures of the proteins. Furthermore, we performed ABFEP calculations in three virtual screening applications for hit enrichment. ABFEP greatly improves the hit rates as compared to docking scores or other methods like metadynamics. The good performance of ABFEP in rank ordering compounds demonstrated in this work confirms it as a useful tool to improve the hit rates in virtual screening, thus facilitating hit discovery.
- 11 May 2023
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 63 (10) , 3171-3185
Enhancing Hit Discovery in Virtual Screening through Absolute Protein–Ligand Binding Free-Energy Calculations
In the hit identification stage of drug discovery, a diverse chemical space needs to be explored to identify initial hits. Contrary to empirical scoring functions, absolute protein-ligand binding free-energy perturbation (ABFEP) provides a theoretically more rigorous and accurate description of protein-ligand binding thermodynamics and could, in principle, greatly improve the hit rates in virtual screening. In this work, we describe an implementation of an accurate and reliable ABFEP method in FEP+. We validated the ABFEP method on eight congeneric compound series binding to eight protein receptors including both neutral and charged ligands. For ligands with net charges, the alchemical ion approach is adopted to avoid artifacts in electrostatic potential energy calculations. The calculated binding free energies correlate with experimental results with a weighted average of R2 = 0.55 for the entire dataset. We also observe an overall root-mean-square error (RMSE) of 1.1 kcal/mol after shifting the zero-point of the simulation data to match the average experimental values. Through ABFEP calculations using apo versus holo protein structures, we demonstrated that the protein conformational and protonation state changes between the apo and holo proteins are the main physical factors contributing to the protein reorganization free energy manifested by the overestimation of raw ABFEP calculated binding free energies using the holo structures of the proteins. Furthermore, we performed ABFEP calculations in three virtual screening applications for hit enrichment. ABFEP greatly improves the hit rates as compared to docking scores or other methods like metadynamics. The good performance of ABFEP in rank ordering compounds demonstrated in this work confirms it as a useful tool to improve the hit rates in virtual screening, thus facilitating hit discovery.
- Top Cited Papers
- 6 April 2023
- Published by American Chemical Society (ACS) in Journal of Chemical Theory and Computation
- Vol. 19 (8) , 2380-2388
Epik: pKa and Protonation State Prediction through Machine Learning
Epik version 7 is a software program that uses machine learning for predicting the pKa values and protonation state distribution of complex, druglike molecules. Using an ensemble of atomic graph convolutional neural networks (GCNNs) trained on over 42,000 pKa values across broad chemical space from both experimental and computed origins, the model predicts pKa values with 0.42 and 0.72 pKa unit median absolute and root mean square errors, respectively, across seven test sets. Epik version 7 also generates protonation states and recovers 95% of the most populated protonation states compared to previous versions. Requiring on average only 47 ms per ligand, Epik version 7 is rapid and accurate enough to evaluate protonation states for crucial molecules and prepare ultra-large libraries of compounds to explore vast regions of chemical space. The simplicity and time required for the training allow for the generation of highly accurate models customized to a program's specific chemistry.
- 6 April 2023
- Published by American Chemical Society (ACS) in Journal of Chemical Theory and Computation
- Vol. 19 (8) , 2380-2388
Top Cited PapersEpik: pKa and Protonation State Prediction through Machine Learning
Epik version 7 is a software program that uses machine learning for predicting the pKa values and protonation state distribution of complex, druglike molecules. Using an ensemble of atomic graph convolutional neural networks (GCNNs) trained on over 42,000 pKa values across broad chemical space from both experimental and computed origins, the model predicts pKa values with 0.42 and 0.72 pKa unit median absolute and root mean square errors, respectively, across seven test sets. Epik version 7 also generates protonation states and recovers 95% of the most populated protonation states compared to previous versions. Requiring on average only 47 ms per ligand, Epik version 7 is rapid and accurate enough to evaluate protonation states for crucial molecules and prepare ultra-large libraries of compounds to explore vast regions of chemical space. The simplicity and time required for the training allow for the generation of highly accurate models customized to a program's specific chemistry.
- Open Access
- 27 March 2023
- Published by American Chemical Society (ACS) in ChemRxiv
Epik: pKa and Protonation State Prediction through Machine Learning
Epik version 7 is a software program that uses machine learning for predicting the pKa values and protonation state distribution of complex, drug-like molecules. Using an ensemble of atomic graph convolutional neural networks (GCNNs) trained on over 42,000 pKa values across broad chemical space from both experimental and computed origins, the model predicts pKa values with 0.42 and 0.72 log unit median absolute and RMS errors, respectively, across seven test sets. Epik version 7 also generates protonation states and recovers 95% of the most populated protonation states compared to previous versions. Requiring on average only 47 ms per ligand, Epik version 7 is rapid and accurate enough to evaluate protonation states for crucial molecules and prepare ultra-large libraries of compounds to explore vast regions of chemical space. The simplicity of and time required for the training allows for the generation of highly accurate models customized to a program’s specific chemistry.
- 27 March 2023
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessEpik: pKa and Protonation State Prediction through Machine Learning
Epik version 7 is a software program that uses machine learning for predicting the pKa values and protonation state distribution of complex, drug-like molecules. Using an ensemble of atomic graph convolutional neural networks (GCNNs) trained on over 42,000 pKa values across broad chemical space from both experimental and computed origins, the model predicts pKa values with 0.42 and 0.72 log unit median absolute and RMS errors, respectively, across seven test sets. Epik version 7 also generates protonation states and recovers 95% of the most populated protonation states compared to previous versions. Requiring on average only 47 ms per ligand, Epik version 7 is rapid and accurate enough to evaluate protonation states for crucial molecules and prepare ultra-large libraries of compounds to explore vast regions of chemical space. The simplicity of and time required for the training allows for the generation of highly accurate models customized to a program’s specific chemistry.