In this talk, we will discuss the development of predictive models to evaluate peptide target binding and passive diffusion across cell membranes. Application of AI-driven multi-objective optimization strategies to enhance both affinity and permeability simultaneously, and case examples demonstrating how these approaches accelerate peptide drug discovery, will also be highlighted.

Generative AI promises to revolutionize protein engineering, but are these tools genuinely useful for GPCR drug discovery? Here, we challenge generative models to design soluble analogs of GPCRs and evaluate their performance through experimental binding measurements and structural validation. Experimental validation of these soluble GPCR analogs translates to efficient antibody discovery against the native target, highlighting the rapid advancement and increasing practical utility of protein design technologies.

Accurately modeling biomolecular interactions is a central challenge in modern biology. Recent advances, such as AlphaFold3 and Boltz-1, have substantially improved our ability to predict biomolecular complex structures. With Boltz-2 we demonstrated the first AI model to approach the performance of free-energy perturbation (FEP) methods in estimating small molecule–protein binding affinity. On top of these advancements, I will present our most recent work in the space of structure-based small molecule and protein design.

Recent advances in protein structure prediction have transformed our ability to model individual proteins, yet predicting the structures and binding affinities of protein-ligand co-complexes remains limited due to a lack of experimental data and current modeling approaches. OpenBind seeks to enable a step-change in protein-ligand modeling by substantially expanding paired structure–affinity measurements. In this talk, I will discuss this effort as well as advances in computational methods for more accurate and reliable binding prediction.

We propose a new benchmarking framework for single-step retrosynthesis that evaluates both general-purpose and chemistry-specialized LLMs using ChemCensor, a novel metric for chemical plausibility. By emphasizing plausibility over exact match, this approach better aligns with human synthesis planning practices. We also introduce CREED, a novel dataset comprising millions of ChemCensor-validated reaction records for LLM training, and use it to train a model that improves over the LLM baselines under this benchmark.