Selected work

Projects

A few research directions I have worked on, spanning interpretable AI, RNA-targeted drug discovery, and AI-accelerated simulation.

TERP: model-agnostic post-hoc explanation of a black-box AI model Feature-importance comparison illustrating interpretation entropy

Thermodynamics-inspired explanations of AI

Interpretable AI TERP Nature Communications

Understanding why deep neural networks make specific predictions is essential for establishing trust. We propose TERP, which explains black-box behavior by fitting local linear models. Because some fitted explanations are far easier for humans to understand than others, we introduce interpretation entropy (S) as a principled measure of human interpretability — showing that model complexity alone is not a good descriptor of it.

Read the paper in Nature Communications ↗

Listen to an AI-generated podcast about this work ↗

Small-molecule ligand in an RNA binding pocket ZTP riboswitch RNA structure

AI-accelerated MD for RNA therapeutics

RNA therapeutics Drug dissociation kinetics Angewandte Chemie

Combining computational and experimental methods, we investigate how small-molecule ligands affect ZTP riboswitch ligand-dissociation kinetics. We identify distinct dissociation mechanisms for different ligand classes and confirm that the ligand on-rate determines efficacy — bridging the disconnect between binding affinity and biological activation.

Read the paper in Angewandte Chemie ↗

(Aib)9 polypeptide helical transition Benzoic acid permeating a DMPC phospholipid bilayer

Accelerating all-atom simulations with AI

Enhanced sampling SPIB JCTC

Using the AI-based State Predictive Information Bottleneck (SPIB) to learn reaction coordinates on the fly, we accelerate molecular dynamics for two systems: the helical transitions of the (Aib)9 polypeptide, and the permeation of benzoic acid through a DMPC phospholipid bilayer — gaining mechanistic understanding of both.

Read the paper in JCTC ↗