Nonspecific binding remains a major developability liability for therapeutic antibodies, with potential consequences for pharmacokinetics and safety. Reliable sequence-based prediction could support earlier decision-making, but remains challenging. Here, we present machine learning models based on protein language model embeddings and biophysical descriptors to predict antibody non-specificity. The models identify the VH domain, heavy-chain CDRs, and isoelectric point as key determinants, achieving up to 71% accuracy.

Protein language models provide a powerful framework for learning sequence–property relationships in antibodies, but their industrial reliability is underexplored. We evaluate state-of-the-art PLMs on internal data from numerous therapeutic programs across three developability assays (PSR, HIC, AC-SINS). Domain-adaptive fine-tuning consistently outperforms pretrained representations, and pretrained sequence likelihoods offer useful unsupervised risk signals. Together, these results show PLMs provide robust, complementary signals for early-stage antibody developability assessment and candidate selection.

Developability plays a key role in modern antibody engineering but is costly and time-consuming to characterize in the lab. In this talk, we present a method for constructing datasets that are designed for training new machine learning models to predict features of antibody developability.

AI-driven antibody optimization is often developed in isolation, without experimental benchmarks, so apparent gains obscure real ones. With access to affinity maturation benchmarks for head-to-head evaluation, we investigate where AI can truly optimize existing experimental protocols, across novel target epitopes and distinct antibody backbones and paratopes. Automated curation builds training-ready datasets that continually fine-tune optimization models; variants are measured experimentally, so gains are confirmed and utilized for follow-up iterations.

Achieving developable multispecific antibody and NANOBODY therapeutics requires moving beyond retrospective assessment toward proactive, design-integrated prediction. We describe how machine learning models for key developability properties have been embedded into our multispecific design workflows, enabling earlier and more informed decision-making across therapeutic programs. We discuss large-scale experimental campaigns that underpin these models, sharing practical insights on data strategy, model building, and the unique challenges posed by complex multispecific formats. Project case studies illustrate how computational predictions translate into real design decisions, and we address the critical gaps that currently limit the field.

Foundation models trained across protein and molecular data are reshaping how we design therapeutics, not merely predicting properties but optimizing them. I'll show how protein language models power two AI/ML efforts: Raygun, which miniaturizes and redesigns natural proteins to improve developability, and PRECISE, a structure-based platform for drug-target screening and engagement. Together they show how learned representations enable more efficient and effective gene therapy and lead discovery.