Recent breakthroughs in large language models have transformed artificial intelligence through scaling. Biology presents the next frontier. I will discuss how the principles behind foundation models can be extended to biomolecular systems and how scaling model capacity, biological data, and experimental feedback may fundamentally change drug discovery. I will outline our vision for AI-native drug discovery, where foundation models integrate sequence, structure, function, and experimental data into a unified learning system capable of accelerating therapeutic discovery.

De novo design promises intentional creation of functional therapeutics, yet translation into the clinic remains uncertain. Binding, though necessary, is an insufficient design objective. We discuss that function should instead guide the design process and be operationalized through an integrated design-build-test-learn paradigm. This way, experimental characterization actively informs and refines computational design—effectively transforming experimental readouts into active sources of learning instead of means of validation.

The specificity of alpha/beta T cell receptors (TCRs) for peptide epitopes underpins adaptive immunity. The ability to predict and design TCR-epitope recognition could open new therapeutic and diagnostic avenues with transformative potential. In this talk, I will describe work toward this goal, leveraging 3D structural modeling and deep neural networks to predict TCR-epitope pairing and to design TCRs and TCR-mimic antibodies.

Active learning offers an efficient strategy for improving predictive model performance in biologics design by prioritizing the most informative sequences for training. This talk will discuss how active learning can identify sequences expected to maximize model gain through uncertainty, diversity, and information-based criteria. Key practical considerations for implementing active learning in computational biologics applications will also be addressed, including how this approach can accelerate model refinement and reduce experimental burden.

The Chatterjee Lab at the University of Pennsylvania develops new generative models to design functional biologics. Our work has centered on language models that de novo design peptides to bind and modulate undruggable disease targets. To enable therapeutic translation, we have developed discrete diffusion models to generate peptides that are "Pareto-optimal" across key ADMET properties. Recently, we have extended these frameworks further to discrete flow matching models that generate and refine highly specific binders under competing therapeutic objective, and have extended these frameworks to peptide-drug conjugates, isoform, motif, and conformationally selective binders, and molecules capable of directing cellular state transitions,

Physics-based energetic scoring functions provide high-fidelity estimates of antibody-antigen binding but are too computationally expensive for high-throughput screening. We introduce a knowledge distillation framework that trains a lightweight neural model to inform binding affinity by learning from energetic scores generated by structural scoring tools. This approach preserves much of the predictive accuracy of the original physics-based scores while dramatically reducing inference cost. We demonstrate that the distilled model generalizes across diverse antibody-antigen pairs, offering a practical, scalable solution for integrating structure-informed binding prediction into early-stage antibody discovery pipelines.

Five major pharmaceutical companies are collaborating via federated computing to train models on antibody developability, a class of biologics properties hard to characterize, due to relatively low-throughput wet-lab assays. Federation provides access to collective data that no single company has, with better economics and speed than generation of such data. We will share early results and lessons from the consortium, highlighting the potential impact of advancing prediction in this space.

Predicting whether and where an antibody binds its antigen still depends on experimental measurement. We show how federated learning across pharma partners' proprietary antibody-antigen data trains co-folding models, a binder-confidence head, and an affinity-ranking capability tied to physically measured binding. We report how this was done operationally—from data harmonization to federated training across partner environments—and aim to share first results.