Trustworthy AI for Computational Drug Discovery: A Realist Review of Explainability, Reproducibility, Validation, and Regulatory Confidence

Download PDF

Artificial intelligence is increasingly embedded in computational drug discovery, yet predictive capability alone does not establish whether an AI system is scientifically credible, reproducible, generalizable, or sufficiently governed to support consequential decisions. This realist review examines when, how, why, and under what conditions AI becomes trustworthy enough to contribute to computational drug-discovery workflows, with particular attention to explainability, reproducibility, validation, and regulatory confidence. Rather than treating these domains as independent technical attributes, the review develops context–mechanism–outcome configurations to explain how trust-building processes are activated, inhibited, or redirected by data provenance, chemical-space coverage, assay quality, workflow transparency, stakeholder expertise, validation design, decision consequence, and governance maturity. The synthesis shows that explainability may strengthen scrutiny, debugging, plausibility assessment, and human challenge when explanations are sufficiently faithful, stable, domain-relevant, and competently interpreted; under weaker conditions, the same explanatory layer may generate plausibility bias, automation bias, and false reassurance. Reproducibility can enable independent verification and error discovery when computational environments, preprocessing, versioning, randomness, and workflow provenance are reconstructable, but reproducible execution cannot transform invalid assumptions or leakage-contaminated evidence into valid science. Validation confidence depends on the distinctness and relevance of the challenge context, while uncertainty estimates require task-specific evaluation and decision-sensitive interpretation. Regulatory confidence is similarly conditional on traceability, intended-use clarity, evidence provenance, change control, accountable oversight, and risk-proportionate validation rather than transparency alone. The refined programme theory therefore conceptualizes trustworthy AI as a bounded and revisable outcome emerging from mutually checking mechanisms operating within supportive contexts. For computational drug discovery, implementation should prioritize independent challenge, transparent provenance, calibrated uncertainty awareness, and governance intensity proportionate to decision risk.