From obstacle to design advantage: activity cliff aware modeling for small-molecule drug discovery

Activity cliffs (ACs), defined as near structural neighbors with large potency differences, provide high-value information on structure-activity relationship discontinuities that medicinal chemists have leveraged for years. Yet modern artificial intelligence-based discovery frameworks often smooth this signal, yielding confident but unrealistic proposals. We present a three-layer roadmap that treats ACs as a first-class design feature: representation is tuned to expose AC signal during pretraining and sampling; evaluation focuses on steep regions with early recognition, pair-level retrieval, and calibrated uncertainty under fair, series-aware splits; and de novo generation is steered toward the best candidates. Integrated as a closed, self-improving loop, this framework helps practitioners build AC-aware drug design tools.