We develop new generative models to design functional biologics and peptides from sequence. Our work has centered on language models that de novo design peptides to bind and modulate undruggable disease targets. Recently, we have pioneered new discrete diffusion models, such as PepTune and TR2-D2, to generate peptides that are "Pareto-optimal" across key ADMET properties. We have further developed discrete flow matching models, such as moPPIt and SOAPIA, that generate and refine highly specific binders under competing therapeutic objectives, and have extended these frameworks to peptide-drug conjugates, peptides that control target conformational dynamics and induce novel cell states.

Deep mutational scanning data used to train models conflate true binding with protein quality effects such as folding and expression. Using AlphaSeq data from over 7,000 mutations, we separate these signals via control VHHs binding non-overlapping epitopes. Most antigen mutations reduce apparent affinity through degraded protein quality, not altered interface energetics. We show that models including ESM-IF1 and ThermoMPNN primarily capture protein quality, with direct implications for model training.

Multispecific antibodies promise enhanced therapeutic precision by engaging multiple targets, yet complex architectures exacerbate chain mispairing and stability risks, complicating developability. Engineering robust formats demands early detection of problematic interfaces across diverse scaffolds and linkers. To de-risk development, we built a machine-learning model that predicts stability and flags binding modules prone to mispairing. By prioritizing compatible pairings, this approach increases engineering success rates while reducing cycles of experimental screening.

In Bits to Binders, a competition benchmarking de novo binder design in the context of chimeric antigen receptor (CAR) T cells, teams from 42 countries submitted 12,000 designs of 80-amino acid binders targeting human CD20 as CAR binding domains. Designs were screened by pooled CAR-T proliferation, with team hit rates ranging from 0.6% to 38.4%, and the top-performing candidates were validated across several other T cell functional assays. From this data we identified common design methodologies and factors correlated with DNA synthesis, expression, and target-specific T cell activation which nearly double the success rates when applied as a retrospective filter.

Antibody AI models often fail to generalize to novel sequences and underrepresented regions of sequence space. This reflects insufficient training data diversity and coverage, even as dataset sizes grow. This talk examines how federated data networks, including the AI Structural Biology Network with nine pharma partners and the Antibody Developability Network, improve model robustness and applicability across proprietary datasets, with examples in developability assessment, antibody–antigen co-folding, and binding affinity prediction.