Abstract
Foundation models pretrained on massive molecular and biological datasets are revolutionizing drug discovery by enabling transferable representations of small molecules, proteins, and cellular systems. However, translating predictive performance into real-world therapeutic impact demands mechanistic interpretability, that is, the capacity to link model outputs to chemical and biological causality rather than statistical correlations. This review synthesizes the current state of foundation models across modalities (Simplified Molecular Input Line Entry System strings, molecular graphs, 3D geometries, protein sequences, multi-omics profiles) and evaluates their mechanistic grounding through explainable AI probes, causal interventions, and concept-based interpretability. We critically assess architectures including transformers, equivariant geometric networks, diffusion models, and hybrid graph-transformer systems, highlighting recent advances in generative design with mechanistic constraints. Four major bottlenecks limiting biological translation are identified: data curation and provenance, causal validity of predictions, prospective experimental validation, and regulatory acceptability. To address these challenges, we propose a pragmatic roadmap combining mechanism-aware pretraining objectives, multi-scale causal representation learning, and closed-loop validation integrating synthesis and bioassay. Emphasis is placed on regulatory-grade documentation (datasheets, model cards, audit trails) and community benchmarks that quantify mechanistic plausibility, interventional predictivity, and transportability. This framework aims to transform foundation models from black-box predictors into interpretable, actionable tools for medicinal chemistry and translational drug development.