Research

Current Research (Postdoc @ Genentech)

Molecular glues represent a promising new drug modality. They are small molecules that induce productive protein–protein interactions (PPIs), expanding druggability to otherwise intractable targets. My postdoc research aims to make molecular glue discovery more rational and systematic rather than serendipitous. I develop computational frameworks that predict which protein pairs are amenable to glue-mediated interactions, design peptide-based glues using structure prediction models, and build AI-powered tools that accelerate early-stage target evaluation for induced proximity programs.

Graduate Research (PhD @ UMD)

My PhD work can be classified in two areas:

I. Physics-informed representation learning and modeling for dynamical systems.

To study the complex molecular systems, I capitalized on physics-inspired insights to learn useful but also meaningful low dimensional representations of the deluge of data generated from experiments and simulations.

1. Learn useful representations: Molecules such as proteins are dynamic entities capable of adopting a wide range of metastable states or conformations beyond their crystal structure. I have demonstrated that physics-informed representations effectively focus on these state-to-state transitions, facilitating automated partitioning of the configuration space and generating concise yet comprehensible descriptions of kinetic pathways using Markov models. These representations have proven to be valuable in providing mechanistic insights across various domains of chemical and biological physics.

2. Learn unique and meaningful representations: The lack of uniqueness in AI-learned representations poses a challenge for interpretation. I have addressed this issue by imposing constraints on latent representations, requiring them to adhere to a generic class of dynamics inspired by fundamental physics principles. These constraints are justified as different molecular systems follow similar fundamental dynamics, such as Newton’s or Schrödinger’s equations of motion. This physics-informed representation learning method consistently outperforms the competing method, yielding unique and highly interpretable representations across diverse datasets, including real-world fluorescent single-molecule movies.

II. AI augmented molecular dynamics (MD) that reach experimental timescales.

I have developed AI augmented molecular dynamics (MD) methods that allow for studying complex molecular systems across an extensive range of timescales, from femtoseconds to seconds, while preserving all-atom resolution. By leveraging the learned physics-informed representation as a blueprint, these methods significantly accelerate simulations of molecular systems which would otherwise be prohibitively slow. These techniques have been shown to speed up the MD simulations for a variety of systems, including permeation and dissociation of medically relevant ligands, conformational changes in proteins and nucleation of crystal polymorphs.