A surrogate-assisted multi-objective optimization algorithm for antibody design with multi-source datasets

Abstract

Antibody design plays a crucial role in therapeutic development and can be formulated as a multi-objective optimization problem, in which multiple favorable biophysical properties with strong trade-offs must be optimized simultaneously. Since evaluating these properties is computationally expensive, newly generated antibodies during the optimization process cannot be evaluated and are instead estimated using surrogate models. However, the data available for surrogate training are limited and originate from multiple sources with varying experimental standards, making prediction accuracy difficult to guarantee. This issue may mislead the multi-objective optimization process. To address this issue, we propose a surrogate-assisted multi-objective optimization algorithm that fully leverages limited multi-source datasets. The proposed method integrates a global evolutionary search guided by a global surrogate model with a local refinement stage guided by a local surrogate model, thereby effectively enhancing optimization performance. Computational experiments demonstrate that the proposed method significantly outperforms state-of-the-art offline multi-objective baselines, achieving superior hypervolume values and consistently identifying a diverse set of high-quality antibody variants. Rosetta ddG simulations indicate that 93.8% of the optimized variants exhibit lower binding free energy values than the previously reported optimal antibody, while over 50% of the variants also maintain comparable specificity according to antibody–antigen binding site analysis. Structural modeling analysis further reveals the enhanced affinity and specificity of the optimized variants at the molecular level.