Decoding antibody-antigen recognition: the BL8 language and its application in calibrating deep learning predictions
- 1Genesius Lab, Shanghai Innovation Institute, Shanghai 200231, China; Key Laboratory of Medical Molecular Virology (MOE/NHC/CAMS), School of Basic Medical Scienc
Abstract
Antibodies are central to immunology and therapeutics, and understanding how they recognize antigens is essential to rational drug design. This recognition depends on the interplay of interface geometry, physicochemical interactions, sequence context, and complementarity-determining region (CDR) conformation. However, traditional analyses often oversimplify this complexity, which becomes particularly problematic when evaluating the massive and frequently overconfident structural hypotheses generated by modern deep learning models. Here, we introduce BL8, an interpretable, eight-view structural language that comprehensively parameterizes interfaces across topology, shape, interaction, CDR conformation, flexibility, interfacial area, amino-acid-pair composition, and sequence background. Using 15,396 experimental interface records from the Structural Antibody Database (SAbDab), we built a fixed reference atlas containing 1,498 discrete BL8 signatures. The atlas shows that natural interfaces occupy a diverse, long-tailed structural space in which similar geometric scaffolds can be associated with different chemical and sequence features. Furthermore, we projected 28,501 predicted complexes involving programmed death-ligand 1 (PD-L1), the SARS-CoV-2 receptor-binding domain (RBD), and programmed cell death protein 1 (PD-1), generated using AlphaFold 3 (AF3), Protenix-v2, and OpenDDE, into the reference space. Acting as a robust underlying calibration layer, BL8 successfully decoupled physical viability from native confidence illusions, revealing cross-target hallucinations and filtering out 15.3% to 20.1% of highly deceptive false-positive predictions. Crucially, by benchmarking distinct generative behaviors across models, it enabled the extraction of highly reliable, multi-model consensus out-of-distribution binding paradigms. Together, BL8 bridges foundational biophysics with large-scale computational predictions, providing an essential benchmarking framework for discovering novel mechanisms in the era of artificial intelligence.
Related
- Ourotide: decoding the hierarchical peptide recognition for generative design
- HLA-driven variability in human cytomegalovirus immune recognition: Epitope predictions from a kidney transplant Cohort
- Orientation detection for self-calibrating neuromorphic systems
- MAXWELL: Calibrating the probabilistic outputs of protein language models to the mutation-induced stability change landscape
- Sexing carnivores from faecal metagenomes: calibrating Rx for stone marten and red fox
BibTeX
@misc{miao2026decoding,
title = {{Decoding antibody-antigen recognition: the BL8 language and its application in calibrating deep learning predictions}},
author = {T. Miao and R. Zhu and R. Lin and J. Qian and Y. Feng and S. Chen and Y. Wang and D. Chen and Y. Wu and T. Ying},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.30.763264},
url = {https://www.biorxiv.org/content/10.64898/2026.10.30.763264v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Decoding antibody-antigen recognition: the BL8 language and its application in calibrating deep learning predictions
AU - T. Miao
AU - R. Zhu
AU - R. Lin
AU - J. Qian
AU - Y. Feng
AU - S. Chen
AU - Y. Wang
AU - D. Chen
AU - Y. Wu
AU - T. Ying
PY - 2026
DA - 2026/10/30
PB - bioRxiv
DO - 10.64898/2026.10.30.763264
UR - https://www.biorxiv.org/content/10.64898/2026.10.30.763264v1
AB - Antibodies are central to immunology and therapeutics, and understanding how they recognize antigens is essential to rational drug design. This recognition depends on the interplay of interface geometry, physicochemical interactions, sequence context, and complementarity-determining region (CDR) conformation. However, traditional analyses often oversimplify this complexity, which becomes particularly problematic when evaluating the massive and frequently overconfident structural hypotheses generated by modern deep learning models. Here, we introduce BL8, an interpretable, eight-view structural language that comprehensively parameterizes interfaces across topology, shape, interaction, CDR conformation, flexibility, interfacial area, amino-acid-pair composition, and sequence background. Using 15,396 experimental interface records from the Structural Antibody Database (SAbDab), we built a fixed reference atlas containing 1,498 discrete BL8 signatures. The atlas shows that natural interfaces occupy a diverse, long-tailed structural space in which similar geometric scaffolds can be associated with different chemical and sequence features. Furthermore, we projected 28,501 predicted complexes involving programmed death-ligand 1 (PD-L1), the SARS-CoV-2 receptor-binding domain (RBD), and programmed cell death protein 1 (PD-1), generated using AlphaFold 3 (AF3), Protenix-v2, and OpenDDE, into the reference space. Acting as a robust underlying calibration layer, BL8 successfully decoupled physical viability from native confidence illusions, revealing cross-target hallucinations and filtering out 15.3% to 20.1% of highly deceptive false-positive predictions. Crucially, by benchmarking distinct generative behaviors across models, it enabled the extraction of highly reliable, multi-model consensus out-of-distribution binding paradigms. Together, BL8 bridges foundational biophysics with large-scale computational predictions, providing an essential benchmarking framework for discovering novel mechanisms in the era of artificial intelligence.
ER -