The paper examines the effect of sampling temperature on ESM3's zero-shot conformation generation for BPTI (n=100). At low temperatures (t=0.25), ESM3's validity drops substantially compared to t=1.0, indicating the pre-trained model produces more clash-free structures at higher temperatures. After fine-tuning (ESMdiff), this degradation is resolved. The paper notes that ESM3 and ESMdiff maintain high validity as temperature increases beyond 1.0, with minimal diversity improvement, suggesting limited sensitivity to high temperatures.
Evidence
correlational
Key metric
ESM3 validity drops at t=0.25 compared to t=1.0 on BPTI (n=100); after fine-tuning, validity improves significantly without compromising diversity
Caveat
Measured only on BPTI with n=100 samples; the paper does not report exact validity values at each temperature in the text (shown in Figure S1).