What question did this study set out to answer?

Develop a strategy to refine AI-based predictions of antibody-antigen interactions by integrating experimental data.

April 5, 2026

Abstract 6899: Mechanism-driven post-AI refinement enables accurate structure prediction of therapeutic antibody-antigen complexes for rational CAR-T and antibody design.

HYHsih‐Te YangLevine Cancer Institute JMJuan Jose Garcia MesaPhoenix (United States)JSJames T. SymanowskiAtrium Health Wake Forest Baptist

Key Points

Develop a strategy to refine AI-based predictions of antibody-antigen interactions by integrating experimental data.
Used AlphaFold3 and AlphaRED for initial structure predictions.
Implemented a post-AI refinement workflow with mechanism-driven constraints.
Conducted 42 independent trials using Rosetta-based replica exchange docking for optimization.
Performed in silico mutagenesis to validate critical residues for binding.
Corrected prediction failures in complex structures with experimental alignment.
Identified critical residues affecting binding affinity, with specific KD values reported.
Demonstrated algorithm generalizability through validation with a different antibody-EGFR system.

Abstract

Abstract Accurate prediction of antibody-antigen complexes remains a major challenge for AI-based structural modeling, with current methods failing to model over 50% of interactions. We developed a post-AI refinement strategy to correct AlphaFold predictions when they conflict with experimental observations, specifically for therapeutic antibodies exhibiting competitive binding mechanisms. We applied AlphaFold3 and AlphaRED to predict the structure of scFv47- IL13Rα2 complex, a therapeutic target for glioblastoma. When initial predictions failed to recapitulate experimentally validated competitive binding between scFv47 and IL13, we developed a computational post-AI refinement workflow incorporating mechanism-driven constraints. Using Rosetta-based replica exchange docking, we performed 42 independent trials (2,500 CPU-hours) with stepwise global-to-local docking refinement. Signal-to-noise filtering identified energetically favorable structures with minimal variance. The refined hypothetical 3D model successfully aligned with experimental observations. In silico mutagenesis recapitulated published data showing Y207, D271, Y315, and D318 residues are critical for IL13 binding but dispensable for scFv47 interaction. Computed interface scores correlated with experimentally-measured binding affinities (IL13-IL13Rα2±2: KD=500pM vs. scFv47-IL13Rα2±2: KD=1.39nM). Structural analysis revealed scFv47 heavy chain CDRs engage the D2 domain while light chain CDR1 primarily contacts D1, providing a rational template for affinity optimization. Validation using cetuximab-EGFR system (with available crystal structure PDB:1YY9) demonstrated 80% epitope overlap, confirming algorithm generalizability across competitive binding systems. This mechanism-driven post-AI refinement framework successfully corrects AlphaFold prediction failures for antibody-antigen complexes, enabling structure-guided therapeutic development when experimental structures are unavailable. The methodology is applicable to CAR-T targeting domains and FDA-approved antibodies, accelerating rational antibody engineering. Citation Format: Hsih-Te Yang, Juan Jose Garcia Mesa, James Symanowski, Richard Rovin. Mechanism-driven post-AI refinement enables accurate structure prediction of therapeutic antibody-antigen complexes for rational CAR-T and antibody design abstract. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 6899.

Ask AI

Helpful

Bookmark

View Full Paper