Scaffold splits hide structural-frontier failures in ADMET models
Standard drug-property tests miss critical failure zones in AI models
Drug-discovery AI models perform much worse than expected when tested on chemically unusual molecules, even when using supposedly rigorous evaluation methods. Researchers found that standard testing approaches hide these "structural frontier" failures — inflating measured accuracy by 87% to 130% — and that adding penalty-based safeguards to training doesn't fix the underlying problem.
Drug companies rely on these AI models to screen millions of candidate molecules quickly. If models fail silently on unusual chemical structures, they could miss effective drugs or recommend dangerous compounds. This work reveals that current evaluation practices mask real weaknesses, meaning companies need better testing protocols before deploying these tools in actual drug discovery pipelines.