Extended Data Fig. 4. Aggregation strategy ablation across all co-distillation rounds.

(a) Schematic of three co-distillation rounds and the aggregation schedules tested: avg–avg–avg, min–min–min, and min–avg–avg. Round 1 uses human proteins; round 2 repeats on the same set; round 3 uses non-human proteins (see Methods). (b) Performance on Balanced ClinVar (top) and ProteinGym DMS (bottom) for all four participating models (ESM1v-5, ESM2-650 M, ESM1b, ESM2-3B) across schedules. Using minimum aggregation in round 1 is required to achieve high performance; both aggregation strategies yield similar gains in subsequent rounds after minimum aggregation is employed in the first round.