Can Folding Models Tell Binders from Bluffers? Evidence from POISK: The Patent-Derived Antibody Dataset

Protenix-v2 discriminates patent-validated antibody binders from sequence-similar decoys with 72% pairwise accuracy overall and 87% on the most confident half.

• Why it matters: Folding-confidence screening for antibody binding is widely used, but validation has been limited by small, narrow datasets and concerns about training-adjacent leakage.
• What they did: POISK was built from 10110 US patents into 89074 antibody–antigen records, then used to create a balanced benchmark of 22494 complexes with 11247 binders and putative non-binders. Six structure prediction models were benchmarked for binder–decoy discrimination, including consensus scoring.
• The result: Protenix-v2 identifies the true binder from 50 candidates in 29% of trials, and multi-model pose consensus reaches 73.1% pairwise accuracy (DockQ>0.23) over ~54.2% of pairs, while epitope localization is only 62.8% accurate.