Interpretable Deep-Learning pKa Prediction for Small Molecule Drugs via Atomic Sensitivity Analysis DOI Creative Commons
Joseph A. DeCorte, Benjamin P. Brown, R. Brooke Jeffrey

et al.

Journal of Chemical Information and Modeling, Journal Year: 2024, Volume and Issue: 65(1), P. 101 - 113

Published: Dec. 30, 2024

Machine learning (ML) models now play a crucial role in predicting properties essential to drug development, such as drug's logscale acid-dissociation constant (pKa). Despite recent architectural advances, these often generalize poorly novel compounds due scarcity of ground-truth data. Further, lack interpretability. To this end, with deliberate molecular embeddings, atomic-resolution information is accessible chemical structures by observing the model response atomic perturbations an input molecule. Here, we present BCL-XpKa, deep neural network (DNN)-based multitask classifier for pKa prediction that encodes local environments through Mol2D descriptors. BCL-XpKa outputs discrete distribution each molecule, which stores and model's uncertainty generalizes well small molecules. performs competitively modern ML predictors, outperforms several generalization tasks, accurately effects common modifications on molecule's ionizability. We then leverage BCL-XpKa's granular descriptor set distribution-centered output sensitivity analysis (ASA), decomposes predicted value into its respective contributions without retraining. ASA reveals has implicitly learned high-resolution about substructures. further demonstrate ASA's utility structure preparation protein–ligand docking identifying ionization sites 93.2% 87.8% complex molecule acids bases. applied identify optimize physicochemical liabilities recently published KRAS-degrading PROTAC.

Language: Английский

When Do Quantum Mechanical Descriptors Help Graph Neural Networks to Predict Chemical Properties? DOI
Shih‐Cheng Li, Haoyang Wu, Angiras Menon

et al.

Journal of the American Chemical Society, Journal Year: 2024, Volume and Issue: 146(33), P. 23103 - 23120

Published: Aug. 6, 2024

Deep graph neural networks are extensively utilized to predict chemical reactivity and molecular properties. However, because of the complexity space, such models often have difficulty extrapolating beyond chemistry contained in training set. Augmenting model with quantum mechanical (QM) descriptors is anticipated improve its generalizability. obtaining QM requires CPU-intensive computational calculations. To identify when help properties, we conduct a systematic investigation impact atom, bond, on performance directed message passing (D-MPNNs) for predicting 16 The analysis surveys experimental targets, as well classification regression tasks, varied data set sizes from several hundred hundreds thousands points. Our results indicate that mostly beneficial D-MPNN small sets, provided correlate targets can be readily computed high accuracy. Otherwise, using add cost without benefit or even introduce unwanted noise degrade performance. Strategic integration unlocks potential physics-informed, data-efficient modeling some interpretability streamline

Language: Английский

Citations

10
Computational Tools for the Prediction of Site- and Regioselectivity of Organic Reactions DOI Creative Commons
Lukas M. Sigmund,

Michele Assante,

Magnus J. Johansson

et al.

Chemical Science, Journal Year: 2025, Volume and Issue: unknown

Published: Jan. 1, 2025

This article reviews computational tools for the prediction of regio- and site-selectivity organic reactions. It spans from quantum chemical procedures to deep learning models showcases application presented tools.

Language: Английский

Citations

0
Thermodynamics-informed neural networks and extensive data sets: key factors to accurate blind predictions of apparent pKa values in the euroSAMPL challenge DOI Creative Commons
Robert Fraczkiewicz

Physical Chemistry Chemical Physics, Journal Year: 2025, Volume and Issue: unknown

Published: Jan. 1, 2025

Microscopic and macroscopic p K a values for 35 compounds in the euroSAMPL 1 challenge were predicted with our thermodynamics-informed S + model which received first overall rank. We describe methodology discuss evaluation methods.

Language: Английский

Citations

0
pKa prediction in non‐aqueous solvents DOI Creative Commons
Jonathan W. Zheng, Emad Al Ibrahim, Ivari Kaljurand

et al.

Journal of Computational Chemistry, Journal Year: 2024, Volume and Issue: 46(1)

Published: Dec. 11, 2024

Acid dissociation constants (

Language: Английский

Citations

1
Widespread Misinterpretation of pKa Terminology for Zwitterionic Compounds and Its Consequences DOI
Jonathan W. Zheng, Ivo Leito, William H. Green

et al.

Journal of Chemical Information and Modeling, Journal Year: 2024, Volume and Issue: unknown

Published: Nov. 19, 2024

The acid dissociation constant (pKa), which quantifies the propensity for a solute to donate proton its solvent, is crucial drug design and synthesis, environmental fate studies, chemical manufacturing, many other fields. Unfortunately, terminology used describing acid–base phenomena sometimes inconsistent, causing large potential misinterpretation. In this work, we examine systematic confusion underlying definition of "acidic" "basic" pKa values zwitterionic compounds. Due confusion, some data are misrepresented in repositories, including widely highly trusted ChEMBL database. Such datasets frequently supply training prediction models, hence, errors make model performance worse. Herein, discuss intricacies issue. We suggestions phenomena, stewarding datasets, given high potentially impact downstream applications.

Language: Английский

Citations

1
Interpretable Deep-Learning pKa Prediction for Small Molecule Drugs via Atomic Sensitivity Analysis DOI Creative Commons
Joseph A. DeCorte, Benjamin P. Brown, R. Brooke Jeffrey

et al.

Journal of Chemical Information and Modeling, Journal Year: 2024, Volume and Issue: 65(1), P. 101 - 113

Published: Dec. 30, 2024

Machine learning (ML) models now play a crucial role in predicting properties essential to drug development, such as drug's logscale acid-dissociation constant (pKa). Despite recent architectural advances, these often generalize poorly novel compounds due scarcity of ground-truth data. Further, lack interpretability. To this end, with deliberate molecular embeddings, atomic-resolution information is accessible chemical structures by observing the model response atomic perturbations an input molecule. Here, we present BCL-XpKa, deep neural network (DNN)-based multitask classifier for pKa prediction that encodes local environments through Mol2D descriptors. BCL-XpKa outputs discrete distribution each molecule, which stores and model's uncertainty generalizes well small molecules. performs competitively modern ML predictors, outperforms several generalization tasks, accurately effects common modifications on molecule's ionizability. We then leverage BCL-XpKa's granular descriptor set distribution-centered output sensitivity analysis (ASA), decomposes predicted value into its respective contributions without retraining. ASA reveals has implicitly learned high-resolution about substructures. further demonstrate ASA's utility structure preparation protein–ligand docking identifying ionization sites 93.2% 87.8% complex molecule acids bases. applied identify optimize physicochemical liabilities recently published KRAS-degrading PROTAC.

Language: Английский

Citations

0