Guidance of Diffusion-Based Conditional Generative Models for Antibody Design

Francesco Alesiani, Jonathan H Warrell, Henrik Christiansen, Matheus Ferraz

Keywords: Guidance, Diffusion model, Antibody design

TL;DR: We explore the guidance of continuous, discrete and SO(3) diffusion models with application to Antibody Design

Abstract:

Protein structure generation finds important applications in Drug Design and Antibody Design.

Diffusion models have quickly become one of the most prominent approaches for generative tasks. The diffusion process can be seen as applying the gradient of log-probability density functions to a time-varying sequence. With this interpretation, it becomes possible to control the diffusion process by manipulating the density function. This idea has motivated the introduction of both classifier-based and classifier-free guidance methods.

The score-based interpretation of diffusion models has been used to define alternative methods for conditioning, modifying, and reusing these models for tasks involving compositional generation and guidance. For Protein or drug structure prediction, SE(3)-equivariant message passing has been the predominant approach, while the atom types are typically modeled using discrete diffusion models.

We introduce a formal logical composition framework for conditional diffusion processes (AND and AND-NOT guidance), which respects Boolean De Morgan's laws, and demonstrate its application to antibody complementarity-determining region design.

Submission Number: 9