Digital twin approaches for interpretable side effect prediction in drug discovery
András Ecker, Gergely Szabó, János Szalma, Erzsébet Fichó, István Z. Reguly, Attila Csikász-Nagy · Drug Discovery Today · 2026
• No standard tool is available to predict side effects in early-phase drug discovery. • Because publicly available side effect data is scarce, only simple models should be trained. • Simple models trained on bio-realistic features, such as off-target proteins, are interpretable. • Interpreting predictions can highlight crucial off-targets and are therefore actionable. 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.