Decoding antibody-antigen recognition: the BL8 language and its application in calibrating deep learning predictions

BL8, an eight-view structural language, calibrates deep-learning antibody–antigen complex predictions by filtering 15.3%–20.1% of highly deceptive false positives in projected AF3/Protenix-v2/OpenDDE models.

• Why it matters: Antibody–antigen recognition is shaped by geometry, interactions, sequence context, and CDR conformation, but deep models can generate massive yet overconfident structural hypotheses that are hard to validate.
• What they did: Using 15,396 experimental interface records from SAbDab, the authors built a fixed atlas of 1,498 discrete BL8 signatures and projected 28,501 AF3/Protenix-v2/OpenDDE predicted complexes for PD-L1, SARS-CoV-2 RBD, and PD-1.
• The result: BL8 decoupled physical viability from native confidence illusions, exposed cross-target hallucinations, and enabled multi-model consensus binding paradigms that are reliable even when out of distribution.

AI summary of the abstract · may contain errors · not medical advice · read the paper.