Repository logo
 

Towards generative models for de novo antibody design


Loading...
Thumbnail Image

Type

Change log

Abstract

Antibodies are among the most important tools in modern biology and medicine, with applications spanning therapeutics, diagnostics, and fundamental research. However, their discovery remains largely dependent on experimental screening of large candidate pools, limiting throughput and precise control over molecular properties. Computational approaches that can design antibodies de novo — specifying target epitopes and binding geometries directly — offer a promising route to overcoming these limitations. This thesis presents two complementary machine learning models for de novo antibody design: COSMO, a sequence-structure co-design model for epitope-specific CDR generation, and IgCraft, a Bayesian Flow Network for conditional antibody sequence generation. COSMO addresses the challenge of designing CDR loops to bind pre-specified epitopes, using a two-shot sampling approach in which separately trained designer and refiner networks decompose the all-atom design problem into tractable subtasks. To overcome the scarcity of antibody-antigen structural data, COSMO is pre-trained on a curated database of CDR- and epitope-like protein fragments identified through structural similarity searches of the Protein Data Bank. Evaluations demonstrate that this pre-training strategy transfers robustly to antibody CDR design, improving structural accuracy, sequence prediction, and epitope sensitivity. IgCraft generates full-length antibody and nanobody sequences using flexible conditioning sampling mechanisms. These can be applied independently or combined, enabling unconditional generation, sequence inpainting, inverse folding, and CDR scaffolding within a single framework without task-specific training. Evaluations across all four tasks demonstrate competitive or superior performance against state-of-the-art models purpose-built for each application. On its intended task of CDR scaffolding, IgCraft generates frameworks with strong nativeness and predicted binding retention comparable to wild-type sequences. Together, COSMO and IgCraft address complementary stages of the antibody design pipeline — CDR design and framework scaffolding — offering a path towards de novo antibody design with efficient models requiring only modest compute budgets.

Description

Date

2026-03-04

Advisors

Sormanni, Pietro

Qualification

Doctor of Philosophy (PhD)

Awarding Institution

University of Cambridge

Rights and licensing

Except where otherwised noted, this item's license is described as All rights reserved