Research

I develop AI methods for understanding and controlling biomolecular systems and accelerating biochemical discovery.

Full publication list →

Molecular Dynamics

Learning, generating, and controlling biomolecular ensembles and transitions

How do biomolecules transition between functional states? I combine generative models, molecular simulations, and experimental measurements to study equilibrium ensembles and rare transitions.

BioEmu: generative protein ensembles
Generate equilibrium protein ensembles and finetune their ensemble statistics with property prediction fine-tuning (PPFT).
#Protein dynamics #Generative models

BioEmu-CV: learning collective variables
Learn slow collective variables through time-lagged generation, then use them for enhanced sampling and transition paths.
#Protein dynamics #Representations #Enhanced sampling

Guided diffusion for transition-state sampling
Guide diffusion models toward transition regions between molecular states.
#Generative models #Enhanced sampling #Transition states

Learning collective variables with geodesic interpolation
Use interpolated structures as training priors for collective variables that accelerate protein folding simulations.
#Protein dynamics #Representations #Enhanced sampling

Molecular Representations

Understanding what foundation models learn about structure, dynamics, and function

What do molecular models learn, and how can we make that knowledge useful? I study the structure of learned representations and use interpretable features to guide generation and simulation.

Converging representations of matter
Compare representations across scientific models (MLIPs) to understand shared structure and the limits of generalization.
#Representations #Foundation models #Materials Science

BioEmu-CV: representations of slow dynamics
Extract collective variables that capture slow protein motions from a generative foundation model.
#Protein dynamics #Representations #Enhanced sampling

IDiom and RL-SAE: interpretable protein design
Use sparse autoencoder features to control and combine function-associated patterns in disordered protein sequences.
#Representations  #Generative models #Molecular design

AI for Discovery

Reasoning over scientific evidence and deciding what experiments to perform next

How can AI help us ask better scientific questions and choose informative experiments? I study how models use prior knowledge and new observations, with the goal of connecting prediction, hypothesis generation, and experimental design.

LLM decision making under uncertainty
Measure how LLMs use chemical knowledge and observations, and why intended exploration can still lead to exploitative choices.
#Scientific reasoning #Experimental design #LLMs

Explorative fragment-based molecular design
Combine reinforcement learning and fragment assembly to find synthesizable molecules with favorable predicted docking scores.
#Molecular design #Reinforcement learning

Community Service

• Program chair and organizer of the integrating Generative and Experimental platforms for bioMolecular design (GEM) at ICLR 2024 in Vienna, ICLR 2025 in Singapore, ICLR 2026 in Rio, NeurIPS 2027 in Atlanta.  

• Program chair and organizer of the AI for Science workshop at ICML 2026 in Seoul.