Learning Molecular Representations for Medicinal Chemistry

Abstract

The accurate modeling and prediction of small molecule properties and bioactivities depend on the critical choice of molecular representation. Decades of informatics-driven research have relied on expert-designed molecular descriptors to establish quantitative structure−activity and structure−property relationships for drug discovery. Now, advances in deep learning make it possible to efficiently and compactly learn molecular representations directly from data. In this review, we discuss how active research in molecular deep learning can address limitations of current descriptors and fingerprints while creating new opportunities in cheminformatics and virtual screening. We provide a concise overview of the role of representations in cheminformatics, key concepts in deep learning, and argue that learning representations provides a way forward to improve the predictive modeling of small molecule bioactivities and properties.

Could Graph Neural Networks Learn Better Molecular Representation for Drug Discovery? A Comparison Study of Descriptor-based and Graph-based Models

“…Here what we found is that the graph-based models can outperform the descriptor-based models on some lager or multi-task datasets such as the HIV, Tox21 and ToxCast datasets, which is well accord with the previous conclusions where DNN excel at larger amounts of data and multi-task learning [65,66]. However, to build such generalizable and robust deep models requires large-scale high-quality datasets and the datasets in the practical drug discovery campaigns routinely suffer from narrow chemical diversity and insignificant sample sizes [67]. On the ground, we believe that the descriptor-based models can be still widely used and give reliable predictions in the drug discovery campaigns.In conclusion, regardless of the statistical results on the same data folds used by Attentive FP or a more reliable 50 times independent runs, what we found is that the traditional descriptorbased models generally outperform the state-of-the-art graph-based models.…”

Artificial intelligence and machine learning‐aided drug discovery in central nervous system diseases: State‐of‐the‐arts and future directions

“…Here, we explain the molecular descriptors (i.e., target protein descriptors and compound descriptors) and compound fingerprints, and provide the highly used programs for generating them (i.e., sequence‐based tools and structure‐based tools) in the Supporting Information. Additionally, Chuang et al 24 comprehensively discussed how AI‐based methods (i.e., deep learning [DL]) could address limitations of molecular descriptors and fingerprints and thereby improve the predictive modeling of compound bioactivities.…”

Predicting Cellular Drug Sensitivity using Conditional Modulation of Gene Expression

“…We expect that combining structured representations of biological information with drug information will improve prediction performance of cellular viability. By designing models with stronger relational inductive biases, we can investigate methods of integrating learned representations [1, 7]. Traditional machine learning methods gain performance by integrating multiple independent datasets and annotated biological pathway information [8].…”

“…We hypothesize that models with stronger relational inductive biases, defined by a conditional formulation and expressed by architectural assumptions, will outperform a naive modeling approach. Neural networks are well suited for conditional model formulation due to their architectural flexibility, proven success integrating diverse data types, and most importantly, ability to learn hierarchical feature representations [4, 7].…”