Machine Learning for Fast, Quantum Mechanics-Based Approximation of Drug Lipophilicity

Abstract

Lipophilicity, as measured by the partition coefficient between octanol and water (log P), is a key parameter in early drug discovery research. However, measuring log P experimentally is difficult for specific compounds and log P ranges. The resulting lack of reliable experimental data impedes development of accurate in silico models for such compounds. In certain discovery projects at Novartis focused on such compounds, a quantum mechanics (QM)-based tool for log P estimation has emerged as a valuable supplement to experimental measurements and as a preferred alternative to existing empirical models. However, this QM-based approach incurs a substantial computational cost, limiting its applicability to small series and prohibiting quick, interactive ideation. This work explores a set of machine learning models (Random Forest, Lasso, XGBoost, Chemprop, and Chemprop3D) to learn calculated log P values on both a public data set and an in-house data set to obtain a computationally affordable, QM-based estimation of drug lipophilicity. The message-passing neural network model Chemprop emerged as the best performing model with mean absolute errors of 0.44 and 0.34 log units for scaffold split test sets of the public and in-house data sets, respectively. Analysis of learning curves suggests that a further decrease in the test set error can be achieved by increasing the training set size. While models directly trained on experimental data perform better at approximating experimentally determined log P values than models trained on calculated values, we discuss the potential advantages of using calculated log P values going beyond the limits of experimental quantitation. We analyze the impact of the data set splitting strategy and gain insights into model failure modes. Potential use cases for the presented models include pre-screening of large compound collections and prioritization of compounds for full QM calculations.

Open-Source Machine Learning in Computational Chemistry

“…361 Focused on drug lipophilicity, rescoss_logp_ml was developed to predict log P for a given molecule. 362 Including liquid chromatography retention time as a molecular descriptor for training, the multilayer perceptron p_chem_prop_CEVR computes log P and the distribution coefficient log D. 363 Ionization Energies. The IonEner-Pred GitHub repository contains fourteen different conventional and GNN models that can compute ionization energies.…”

“…Using the DeepChem library, log_P_prediction uses convoluted graphs and an NN to predict the octanol–water partition coefficient (i.e., log P) . Focused on drug lipophilicity, rescoss_logp_ml was developed to predict log P for a given molecule . Including liquid chromatography retention time as a molecular descriptor for training, the multilayer perceptron p_chem_prop_CEVR computes log P and the distribution coefficient log D …”

Making the Case for Quantum Mechanics in Predictive Toxicology─Nearly 100 Years Too Late?

“…For example, Isert et al reported that ML can be learnt on calculated log K o/w values to obtain a computationally affordable, QM-based estimation of lipophilicity. 77 Thus, AI and QM have a shared future that can benefit in silico toxicology, so long as we keep in mind their fundamental differences and limitations. In this regard, while QM has evolved in the Darwinian sense from human knowledge and understanding of the natural world and has been extensively tested over the past nearly 100 years, AI is a difficult-to-verify construct of the machine world that requires large data to ensure robustness.…”

“…In the same vein as the two studies above, AI can be used to target physicochemical properties derived from molecular energy. For example, Isert et al reported that ML can be learnt on calculated log K o/w values to obtain a computationally affordable, QM-based estimation of lipophilicity . Thus, AI and QM have a shared future that can benefit in silico toxicology, so long as we keep in mind their fundamental differences and limitations.…”

The rise of automated curiosity-driven discoveries in chemistry

“…The resurgence of ML in chemistry has been fuelled by increased storage capacity, powerful hardware, and a wealth of data to delve on. Tasks that were once demanding to machines, e.g., retrosynthetic planning, 5-7 molecular design, 8-10 prediction of protein structure and function, [11][12][13] prediction of biological activity [14][15][16] and others, [17][18][19] have now become signicantly more attainable and generalizable, providing valuable assistance to bench chemists. 20 ML workows and pipelines are primarily implemented to accelerate chemical research, with the scope of prioritizing experiments with high condence and likelihood of success, while minimizing exploration towards preestablished objectives.…”