Digital twin approaches for interpretable side effect prediction in drug discovery

Artificial intelligence increasingly supports preclinical drug development although accurate, early prediction of side effects remains limited. Leading approaches rely on expert-derived features, unavailable in early discovery, or on chemical similarity-based graph neural networks functioning as 'black boxes' and lacking actionable insight. We propose a paradigm shift: instead of directly mapping chemical features to side effects, derive biologically meaningful intermediate representations, such as predicted off-target proteins and their downstream effects simulated in a cellular digital twin. Simple models trained on these representations are interpretable and therefore actionable. Beyond safety assessment, this approach could inform rational polypharmacology strategies and guide the amendment of secondary pharmacology assays.