Research

A binding site need not be obvious in a protein's most commonly observed structure. I am interested in pockets that become accessible only in particular conformations or molecular environments. Chemoproteomic data—experimental evidence of small molecules engaging protein residues—provide a way to look for these opportunities across many proteins, including sites that structural datasets may miss.

I developed AI Protein Profiling (AiPP) to infer intrinsic covalent ligandability and LatentLift to turn heterogeneous experimental evidence into training signals. Building on this work, I aim to connect residue-level evidence with pocket hypotheses, then use molecular simulation to ask what makes those states accessible and how different ligands change their availability.

For an intrinsically disordered region, function depends on a range of conformations rather than one stable fold. My research asks how protonation, chemical modifications, and interaction partners alter that range—and expose or conceal features involved in molecular recognition. The goal is to explain how a local chemical change can influence behavior across a flexible protein.

My contributions include RIDAO for large-scale disorder analysis and molecular simulations of partner-dependent recognition. In collaborative work on p53, I interpreted simulations alongside thermodynamic and NMR measurements generated by experimental collaborators. This foundation motivates further study of how chemical context regulates access to functional states in disordered proteins, including those involved in biomolecular condensates.

A predicted binding site or conformational change raises a physical question: what makes that state accessible? I develop molecular-simulation methods to investigate how protonation, electrostatics, hydration, and molecular contacts reshape protein ensembles. Rare-state sampling and free-energy calculations connect structural possibilities with their thermodynamic costs.

I developed a second-generation explicit-solvent constant-pH implementation in Amber, with native asynchronous replica exchange and improved computational throughput. These advances make longer simulations of coupled protonation and conformational changes practical. I plan to use this foundation to examine both transient ligand-binding pockets and disordered ensembles, comparing predicted behavior with experimental evidence.

A sequence-based prediction cannot, on its own, distinguish every state of the same protein. I am interested in models that retain what is shared across contexts while adapting their predictions to a ligand, modification, interaction partner, or environment. A second challenge is observational: experimental datasets reflect not only molecular behavior, but also what each assay can detect.

This direction grows from KaML-ESM, which I co-developed to predict residue pKₐ values from protein language models, and my work on AiPP, LatentLift, and the PLMPG research framework. I aim to build models that distinguish molecular properties from assay effects, express uncertainty, and generalize to new proteins and conditions. Beyond individual applications, I am interested in what protein science can teach us about learning from context-dependent, imperfect observations.