Summary
Ensembling ESMFold2, ESMFold2-Fast, and Protenix-v2 confidence scores improves filtering of de novo protein-binder designs (1). On a public benchmark of 3,532 designs against 13 targets, including 391 experimentally confirmed binders, the per-target z-score ensemble reached a macro-average precision of 0.66, compared with 0.62 for ESMFold2-Fast, 0.61 for ESMFold2, and 0.55 for AlphaFold3. Adding self-consistency DockQ at one-quarter the weight of ipSAE did not improve discrimination.
See also
- Effectiveness of filtering metrics for de novo minibinder design vary by target
- Protein structure prediction and design confidence metrics do not correlate with binding affinity
1.
Claude Science, Shanehsazzadeh A. Autonomous de novo protein binder design with Claude. Anthropic; 2026. Available from: https://www-cdn.anthropic.com/30bf50e22a01388bb29bf077ee3f244531594b7a.pdf