AbDD: Experimentally Validated Antibody Developability Optimization via Discrete Diffusion

Oliver M. Turnbull, Alexander Nicholas St John, Alec Nelson Thomas, Charlotte Deane

Keywords: antibody, diffusion, developability, language model, optmization

TL;DR: Antibody-specific discrete diffusion model (AbDD) combined with inference-time conditioning to optimize developability, experimentally validated 89% success rate in improved hydrophobicity.

Abstract: Therapeutic antibody development often fails due to poor developability issues identified late in discovery, such as aggregation, polyspecificity, poor expression, or low solubility. Elevated hydrophobicity is a common liability, contributing to aggregation and high viscosity. Here, we introduce AbDD, a 350M-parameter antibody-specific discrete diffusion model that jointly models amino acid sequences and Foldseek structural tokens. By combining AbDD with Reward-Guided Evolutionary Refinement in Diffusion (RERD), we provide a flexible, training-free framework for optimizing antibodies against arbitrary black-box property predictors while constraining structural deviation from the parent antibody. We experimentally validated this approach by optimising two clinical antibodies with known hydrophobicity liabilities, Galiximab and Rilotumumab. Across nine validated variants, we achieved an 89\% success rate in reducing hydrophobic interaction chromatography (HIC) retention times, with top variants reaching the therapeutically acceptable range with only two mutations. Orthogonal experimental assays (SEC, BVP, VIBE) confirmed no introduction of major new liabilities.

Submission Number: 47