Machine Learning in Drug Discovery: From QSAR Foundations to Deep Learning, Foundation Models, and Toxicity-Ready Prediction

Abstract

Drug discovery remains constrained by long development timelines, high capital requirements, and substantial attrition, particularly during clinical development and late-stage safety evaluation. Estimates of research and development (R&D) investment required per approved therapeutic agent vary widely across studies and assumptions, with industry-survey and public-data analyses reporting materially different central values. Beyond cost and time, the dominant methodological challenge for machine learning (ML) is not merely predictive accuracy on curated benchmarks, but decision-grade reliability under chemical space shift, endpoint heterogeneity, and label noise—conditions that are most acute in safety-sensitive absorption, distribution, metabolism, excretion, and toxicity (ADMET) prediction. This review critically examines how molecular representation choices, including descriptors, fingerprints, simplified molecular-input line-entry system (SMILES) strings, molecular graphs, and three-dimensional geometry, shape model inductive bias, data efficiency, and generalization; why classical quantitative structure–activity relationship (QSAR) modeling with strong baselines remains essential; and how deep learning—including graph neural networks, transformers, and self-supervised foundation models—has altered the landscape of molecular property prediction, drug–target interaction (DTI) modeling, and generative design. Using ADMET prediction, with particular emphasis on toxicity endpoints, as a stress test for the field, we analyze dataset and benchmark limitations, validation practices, including random, scaffold, and external splits, and the central role of uncertainty estimation and interpretability in safety-critical workflows. Unlike prior surveys that often treat model architectures, molecular representations, toxicity prediction, and validation practice as separate topics, this review integrates these elements under a unified toxicity-ready framework focused on decision-grade reliability. We conclude that the next phase of progress will be determined less by leaderboard improvements and more by robust evaluation, calibrated uncertainty, applicability-domain reasoning, and prospective validation aligned with real drug discovery decisions.