Antibody Library Design by Seeding Linear Programming with Inverse Folding and Protein Language Models

Abstract

We propose a novel approach for antibody library design that combines deep learning and multi-objective linear programming with diversity constraints. Our method leverages recent advances in sequence and structure-based deep learning for protein engineering to predict the effects of mutations on antibody properties. These predictions are then used to seed a cascade of constrained integer linear programming problems, the solutions of which yield a diverse and high-performing antibody library. Operating in a cold-start setting, our approach creates designs without iterative feedback from wet laboratory experiments or computational simulations. We demonstrate the effectiveness of our method by designing antibody libraries for Trastuzumab in complex with the HER2 receptor, showing that it outperforms existing techniques in overall quality and diversity of the generated libraries.