For antibody sequence generative modeling, mixture models may be all you need

Abstract

Motivation Antibody therapeutic candidates must exhibit not only tight binding to their target but also good developability properties, especially low risk of immunogenicity. Results In this work, we fit a simple generative model, SAM, to sixty million human heavy and seventy million human light chains. We show that the probability of a sequence calculated by the model distinguishes human sequences from other species with the same or better accuracy on a variety of benchmark datasets containing >400 million sequences than any other model in the literature, outperforming large language models (LLMs) by large margins. SAM can humanize sequences, generate new sequences, and score sequences for humanness. It is both fast and fully interpretable. Our results highlight the importance of using simple models as baselines for protein engineering tasks. We additionally introduce a new tool for numbering antibody sequences which is orders of magnitude faster than existing tools in the literature Availability All tools developed in this study are available at https://github.com/Wang-lab-UCSD/AntPack. Supplementary information Supplementary data are available at Bioinformatics online.

RESP2: An uncertainty aware multi-target multi-property optimization AI pipeline for antibody discovery

“…Finally, we incorporate models for predicting developability. We make use of a fully human-interpretable generative model for human antibody sequences implemented in the AntPack package 21 , which can distinguish human from nonhuman sequences with better accuracy than any other comparator model currently available. As demonstrated in this paper, we can make use of humanness scores provided by AntPack at two stages in the pipeline.…”

“…If this filtering process yields few or no candidate sequences, the ISDE algorithm is re-run to generate more candidates. We use SAM 21 , a fully interpretable generative model with state of the art accuracy for distinguishing human from nonhuman antibodies, to score candidates to ensure low risk of immunogenicity; this model is available in the AntPack library. We also score selected sequences for solubility.…”

“…The model is fully interpretable and provides a simple yet powerful tool for dataset exploration. This model's simplicity makes it very easy to fit quickly; Parkinson et al for example fit it to 70 million light chains and 60 million heavy chains 21 . Despite its simplicity, this type of model has not previously been used for protein engineering to our knowledge.…”

“…To ensure good developability (low risk of immunogenicity, good solubility, stability etc. ), we filter the sequences generated by ISDE for solubility (using the CamSol model 41 ) and for humanness (using the SAM model of Parkinson et al 21 ).…”

“…The SAM model exhibited superior performance on a test set of 400 million sequences for distinguishing human heavy and light chains from those of other species 21 . Importantly, since it is fully interpretable, it can suggest modifications that would improve humanness for a tight binder with initially low scores.…”