Guided Generation of B-cell Receptors with Conditional Walk-Jump Sampling

Taylor Joren, Sarah A Robinson, Homa MohammadiPeyhani, Sai Pooja Mahajan, Edith Lee, Stephen Lillington, Qixin Bei, Saeed Saremi, Jae Hyeon Lee, Richard Bonneau, Isidro Hotzel, Vladimir Gligorijevic, Simon Kelow, Nathan C. Frey

Track: Machine learning: computational method and/or computational results

Nature Biotechnology: Yes

Keywords: Conditional Generation, Antibodies, Lead discovery, B-cell receptors, Clonotypes

TL;DR: We present conditional walk-jump sampling (cWJS), a guided generative model that enables controlled antibody sequence generation by conditioning on clonotype variables, achieving high-quality and diverse sequences from thousands of distinct modes.

Abstract: Antibody drug discovery campaigns often leverage immune repertoires from antigen-exposed animals, which can be divided into clonotypes, subclasses of sequences derived from the same progenitor B-cell. In this work, we adapt discrete walk-jump sampling (dWJS) to condition generation on categorical variables like clonotype, extending both energy-based and score-based dWJS with predictor-free guidance during Langevin dynamics ("walking") and denoising ("jumping"). Categorical and numerical variables are learned during training and specified during sampling, producing diverse and novel sequences from target clonotype classes. We train conditional WJS models on datasets of over 1.5M sequences obtained from antigen-exposed rats and human patients post-vaccination. Surprisingly, increasing guidance improves both sample quality and sequence diversity, enabling controllable sampling from thousands of distinct modes.

Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.

Presenter: ~Taylor_Joren1

Format: Maybe: the presenting author will attend in person, contingent on other factors that still need to be determined (e.g., visa, funding).

Funding: No, the presenting author of this submission does *not* fall under ICLR’s funding aims, or has sufficient alternate funding.

Submission Number: 72