Score-Based Generative Models For Binding Peptide Backbones

ICML 2024 Workshop ML4LMS

Keywords: Diffusion, Machine Learning, Graph Neural Networks, Protein Design, Antibodies, Generative Models, Geometric Deep Learning

TL;DR: We develop a score-based generative model (SGM) and evaluation pipeline for binding peptide structures and use the framework to explore key design choices for SGMs and develop improved metrics for generated binder structures.

Abstract: Score-based generative models (SGMs) have emerged as powerful tools for protein design, capable of generating protein structures for a variety of biologically relevant design specifications. Among these, the ability to generate structures capable of binding a specified target holds particular relevance for a range of applications. Despite the success of SGMs in this domain, there has been little systematic exploration of the impact of model design choices for protein binder backbone generation, in part due to the lack of appropriate metrics for generated backbones and their complementarity to the target protein. Here we present LoopGen, a flexible SGM framework for the generation and evaluation of de novo binding protein backbones in the absence of inverse folding/folding models. This decoupling from existing inverse folding/folding models not only provides an orthogonal set of metrics but also enables the evaluation of protein structure SGMs in domains where such models are difficult to obtain (e.g. peptide design). We apply our framework to design antibody CDR loop structures, a class of peptides with notable structural diversity, and evaluate a variety of model design choices, showing that choices of structural representation and variance schedule have dramatic impacts on model performance. Furthermore, we propose three novel metrics for testing the dependency of a generated binder structure on its target protein, and demonstrate that LoopGen's generated backbones are indeed conditioned on the sequence, structure, and position of its input epitope. Our results identify promising avenues for further development of SGMs for protein design.

Submission Number: 31