Protein phosphorylation encodes dynamic cellular signals, but tools for recognizing specific phosphorylation sites remain limited. I will present an all-atom flow-matching approach we developed to design phosphopeptide binders while simultaneously sampling flexible target conformations. Combining computational design with biochemical and structural characterization, I will show how phosphate recognition and sequence specificity can be programmed into new binding interfaces, enabling molecular tools for probing and ultimately controlling cellular signaling.

In this talk, I will present three deep learning methods that address successive stages of antibody engineering. TiDE-Ab is a conditional SE(3) flow matching framework for de novo epitope-specific antibody design, Jin-ddG predicts changes in binding affinity upon mutation, and our ProteinMPNN-based expression rescue approach recovers candidates with high affinity but low cellular productivity. The talk describes how these methods combine into a single design workflow and discusses their current limitations.

Automation lets a lab run experiments without people at the bench. A self-driving lab also decides which experimental conditions to run next in order to achieve a higher level scientific goal. That step is much harder. In this talk I share what we learned running an automated protein testing lab at Adaptyv: where automation works today, what still blocks the move to closed-loop, agent-driven design-build-test cycles, and how benchmarks can show when a self-driving lab actually performs better.

Designing highly active enzymes remains a frontier challenge in computational protein design. One obstacle is the lack of large-scale data on designed enzymes that can be used to investigate structural differences between active and inactive designs. We investigated the model single-turnover enzyme SpyCatcher by measuring reaction rates for over one million designed and natural SpyCatcher variants, revealing how sequence elements and predicted structural features influenced substrate binding (K_I) and intrinsic reactivity (k_inact). These mega-scale experiments over a uniquely large sequence space highlight how residues across the full SpyCatcher scaffold contribute to reactivity, how reactivity changes in altered scaffolds, and how folding stability influences reactivity.

AI agents are beginning to reshape how biomedical research is conducted—from answering scientific questions to autonomously executing increasingly complex research workflows. In this talk, we will first discuss how to build high-performing biomedical agents that can reason across scientific literature, data, computational tools, and specialized biological models to automate complex research tasks. We will look toward the frontier of agentic science: autonomous swarm of agents that curate and integrate scientific data, generate and test hypotheses, discover new insights, and train the next generation of AI-for-biology models.

We demonstrate that Chai's de novo platform designs epitope-specific, functional biologics with sub-nanomolar affinity from single 96-well plates. These designs exhibit outstanding agreement between their predicted and experimentally solved structures in complex with their targets, can be programmed with selectivity and cross-reactivity profiles across diverse modalities, and exhibit high hit rates and developability on par with clinical molecules. For challenging targets such as GPCRs and other multipass membrane proteins, Chai designs are able to drive cellular signaling, discriminate between conformational states, and activate their targets with potency and efficacy approaching native ligands. 

Recent advances in co-folding models demonstrated substantially improved performance in antibody-antigen complex structure prediction, making large-scale in silico epitope mapping a possibility within existing antibody screening platforms. Here, we benchmarked a representative set of open-source co-folding models for antibody-antigen complex structure prediction and calibrated their confidence metrics with empirical posterior success rate. We also demonstrate the capability and limitations of an integrated in silico epitope mapping workflow developed based on insights derived from our benchmarking.

Scaling antibody discovery to hundreds of targets per year requires every step to be efficient and focused on reducing failure. We will show how we built an end-to-end platform that generates thousands of research antibodies each year, and how the resulting data are helping us leverage AI and language models to predict developability properties, including polyreactivity and aggregation-related behavior, directly from antibody sequence.

Optimizing therapeutic antibodies requires improving binding affinity while maintaining expression and developability, often with limited experimental data and competing molecular properties. We introduce Property Enhancer (PropEn), a data-efficient generative framework that learns directly from local comparisons of experimentally characterized antibody sequences, enabling sequence optimization without a separate property predictor or reward model. We demonstrate PropEn in experimental affinity maturation across four therapeutic targets and eleven antibody leads. Trained on only a few hundred target-specific sequences, PropEn improved nine of eleven leads, achieving up to 38-fold affinity gains, with a 94.6% binding rate and 98% expression rate. We further extend the approach to multi-objective optimization (mPropEn), enabling simultaneous optimization of competing developability properties. In computational benchmarks, mPropEn increased the proportion of designs improving four properties simultaneously to 12.9%, compared with at most 4.5% for competing methods, while an experimental case study demonstrated favorable multi-property trade-offs. Together, these results highlight the potential of lightweight generative models to efficiently optimize therapeutic antibodies using sparse experimental data.

High-concentration monoclonal antibody (mAb) formulations are often required for subcutaneous delivery, but elevated viscosity limits manufacturability, syringeability, and injectability. Experimental viscosity measurements require substantial material, limiting early-stage assessment. We developed DeepViscosity, an ensemble of 102 artificial neural networks that classifies mAbs as low- or high-viscosity at 150 mg/mL using sequence-derived features alone, without requiring structural models or experimental measurements. Trained on sequence and high-concentration viscosity data from 229 mAbs, DeepViscosity achieves up to 89.5% accuracy on independent test sets. By assessing viscosity risk directly from sequence, DeepViscosity supports early-stage selection of developable antibody candidates for subcutaneous delivery.

FlashTAP is a sequence-structure-sequence model fine-tuned on TAP predictions, which enables ultra-fast developability predictions. We show how this can be used to screen antibody repertoires at scale for developable antibodies. Furthermore, we demonstrate how this can be integrated into design pipelines for developability optimisation in de novo antibody generation.