A Quiet Failure in Calibrated Virtual Screening: Marginal Conformal Prediction Under-Covers the Minority Class, and a Class-Conditional Fix Recovers It

Short summary

Standard marginal conformal prediction fails silently on imbalanced drug-discovery datasets, hitting global 90% coverage while minority-class coverage drops as low as 4.2% on clinical-trial toxicity data. The failure persists across random forests, graph networks, and chemical language models, with severity tracking baseline calibration on rare labels. Class-conditional (Mondrian) conformal prediction restores per-class coverage to target for a modest increase in prediction-set size, and a cost model shows abstaining on affected compounds flips screening campaigns from net-negative to net-positive utility.

• •Marginal conformal prediction under-covers minority classes on imbalanced drug-discovery datasets (as low as 4.2% vs 90% target)
• •Failure is architecture-independent and explained by a conservation identity linking minority shortfall to majority surplus
• •Mondrian class-conditional conformal prediction fixes the gap with modest prediction-set size increase

Generated with AI, which can make mistakes.