Integrating artificial intelligence and molecular simulation for peptide discovery: limitations, validation and decision-centered workflows

Summary. Therapeutic peptides occupy a chemical and pharmacological space between smallmolecules and biologics, offering extended recognition surfaces and access to targets that resist conventionalmodalities. Their discovery, however, requires the simultaneous optimization of activity, selectivity, stability, permeability and synthetic feasibility across a conformationally complex and chemically heterogeneous search space. Artificial intelligence (AI) andmolecular simulation address complementary aspects of this problem: AI learns structure–property relationships and explores large candidate libraries, whereas simulation interrogates conformational ensembles, target recognition and membrane-associated behaviour within an explicit physical model. We organize the field around a single, more discriminating question: under which chemical, biological and decision-making conditions does adding molecular simulation to an AI workflow improve peptide discovery outcomes enough to justify its cost? To answer it, we introduce a decision framework linking peptide chemistry and discovery endpoint to the decision beingmade, the uncertainty limiting that decision, the evidence that computation or experiment can provide, the cost of obtaining that evidence and the criterion for proceeding or stopping. Recent advances in biomolecular structure prediction, protein language models, generative design andmacrocycle engineering supply powerful enabling capabilities, but their transfer to short, flexible, cyclic and chemically modified peptides remains unevenly validated. Critically, a structured survey of 15 integrated studies shows that none (0 of 15) included a matched-budget comparison against an AI-only or conventional baseline, a scarcity we treat as a central empirical finding rather than an inconvenience. We outline reporting, benchmarking and prospective-validation standards by which genuine progress should be judged, and define what ‘developable’ must mean at the hit, lead and candidate stages.