Epik: pK_a and Protonation State Prediction through Machine Learning

Abstract

Epik version 7 is a software program that uses machine learning for predicting the pK a values and protonation state distribution of complex, druglike molecules. Using an ensemble of atomic graph convolutional neural networks (GCNNs) trained on over 42,000 pK a values across broad chemical space from both experimental and computed origins, the model predicts pK a values with 0.42 and 0.72 pK a 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.

Quantum Descriptors for Predicting and Understanding the Structure–Activity Relationships of Michael Acceptor Warheads

“…However, tertiary amines are expected to be protonated and charged at neutral or physiological pH, as their p K a ’s are higher than 8.5, e.g., the experimental p K a is 9.76 for trimethyl amine and 10.08 for N -methylpiperidine . Consistently, the DMAM group p K a ’s in 16 , 18 , and 27 were estimated as 8.64, 8.62, and 8.69 using Schrödinger’s Epikx program . Therefore, to emphasize the protonated form, we denote these compounds as compounds 16H , 18H , and 27H in Table .…”

Comment on the Definition and Labeling of pK50

“…For additional assurance in assigning the nature of the protonation/deprotonation process we also applied a separate empirical method Epik based on machine learning techniques, also developed by us . The accuracy of these methods’ predictions, as demonstrated in previous publications, − should be more than sufficient for distinguishing whether protonation or deprotonation was responsible for the p K a value measured in the experiment. The results presented in Table show not only good accuracy at reproducing the experimental values but also a strong agreement among three computational methods in interpreting the nature of the protonation/deprotonation processes.…”

“…For these calculations we used two related but independent computational workflows developed by us and known as Jaguar p K a , and Macro-p K a . For additional assurance in assigning the nature of the protonation/deprotonation process we also applied a separate empirical method Epik based on machine learning techniques, also developed by us . The accuracy of these methods’ predictions, as demonstrated in previous publications, − should be more than sufficient for distinguishing whether protonation or deprotonation was responsible for the p K a value measured in the experiment.…”

Accurate Computation of Aqueous pKas of Biologically Relevant Organic Acids: Overcoming the Challenges Posed by Multiple Conformers, Tautomeric Equilibria, and Disparate Functional Groups with the Fully Black-Box pK-Yay Method

“…In this scenario, we had recently developed the p K -Yay method to compute aqueous p K a values . This was done as part of our attempt to develop a black-box methodology based on error -cancellation to calculate aqueous p K a values which can be used easily even by a nonexpert (we note that commercially available software such as EpiK also aim to achieve the same goal–black-box determination of p K a values of drug like molecules). p K -Yay is aimed to be a general and widely applicable method useful to compute the p K a of any organic or mineral acid.…”