Explainable AI methods for drug discovery: A survey of interpretability, metrics and mechanistic insight

Artificial intelligence (AI), including machine learning (ML) and deep learning (DL), is accelerating drug discovery by enhancing target identification, virtual screening, absorption, distribution, metabolism, excretion, and toxicity (ADMET), and lead optimization. The drug development lifecycle increasingly benefits from data-driven decision support enabled by these approaches. However, the opaque nature of deep models limits their adoption in pharmaceutical contexts, where interpretability is critical for compound prioritization, toxicity evaluation, and regulatory acceptance. Explainable AI (XAI) aims to bridge this gap by providing human-understandable explanations (interpretability) that support hypothesis generation and mechanistically plausible, testable rationales, while explicitly requiring subsequent experimental validation. This review introduces a novel multidimensional taxonomy of XAI approaches tailored to drug discovery, where methods are organized by input modality, degree of model transparency, and interpretability objectives, providing a task-centered framework for method selection across specific decision stages. We critically analyze core techniques including SHapley Additive exPlanations (SHAP), Local Interpretable Model-Agnostic Explanations (LIME), saliency maps, attention mechanisms, surrogate models, counterfactuals, and causal inference, together with evaluation metrics such as fidelity, stability, sparsity, and interpretability. In contrast to prior overviews largely centered on specific applications, our work emphasizes mechanistic plausibility, alignment with decision-making needs, and emerging hybrid frameworks that integrate symbolic reasoning with multimodal data to promote interpretability grounded in chemical and biological knowledge. By integrating method-agnostic tools with quantitative evaluation schemes and decision-focused case studies, this survey offers a structured roadmap for deploying XAI in drug discovery and advances the discussion on interpretable AI in high-stakes scientific domains. We further examine the limitations of current XAI approaches, including documented failure modes such as shortcut learning and Clever Hans–type effects, and analyze practical barriers that constrain the translation of XAI into real-world drug discovery workflows.