Extended Data Fig. 1. Theoretical comparison of minimum vs average aggregation.

(a) Schematic of the Gaussian model used to compare aggregation strategies: class-conditional LLR distributions with means (μ) and variances (σ²), and the resulting AUC landscape as a function of class separation |μ₁-μ₂| and variance ratio σ₁²/σ₂². Regions favoring minimum-LLR aggregation (max-confidence) are shaded accordingly. (b) Empirical ClinVar LLR distributions for individual ESM models (benign vs pathogenic), compared with the minimum (ESMIN) and average ensembles; the right panel shows per-class variances. The pathogenic class exhibits larger dispersion, aligning with the regime in which minimum aggregation is expected to outperform averaging.