projects

(Paraphrased from the abstract)

The prevailing computational approach for protein engineering uses a protein foundation model for two tasks: an oracle built on its representations scores mutation effects, and a search procedure proposes mutations through reinforcement learning.

We introduce an alignment-guided mutation proposer and oracle (AURORA), a protein engineering framework with two key components that optimize each of these tasks separately.

To search for optimal protein fitness, MSA Pairformer scores mutants while a separate lightweight policy trained with RL proposes mutations, enabling direct multi-site proposals rather than iterative single-site search. We then validate AURORA in vitro on green fluorescent protein, training on limited data to generate novel variants that demonstrate higher fluorescence than existing methods.

Accepted poster at ICML 2026 GenBio Workshop.

Multimodal reasoning based foundation models hold considerable promise for addressing key challenges in medical practice, yet their readiness for real-world deployment remains insufficiently explored. To bridge this gap, I contributed to developing a foundational model that excels in clinical generalizability and accuracy.

I've preprocessed EHR data from MIMIC-IV and used it in training for our foundational model, analyzed model performance before and after supervised finetuning, and evaluated model robustness to hallucinations and likelihood of perpetuating clinical inequities with LLM-as-a-judge.

Second author on a manuscript accepted by npj Digital Medicine.

Metabolic dysfunction-associated fatty liver disease (MAFLD) affects 25% of adults in the United States and affects those with Type 2 diabetes and class III obesity at disproportionately higher rates. We present a deep learning-based framework for early prognosis of MAFLD in adults using structured clinical data from Mass General Brigham. Our approach utilizes binary classification, neural network prediction, linear and logistic regression, and survival modeling, as well as experimentation with addressing class imbalance. The study supports early clinical risk stratification and reveals predictive biomarkers using SHAP interpretation.

MASH, the advanced stage of metabolic-associated steatosis liver disease, is characterized by severe accumulation of fat in the liver. The effort towards developing accurate in vitro liver models is crucial for better understanding disease progression and therapy development. A patient's disease severity is determined by a doctor's histological scoring of a liver biopsy, which produces two clinical scores: a fibrosis stage score and a NASH Activity Score (NAS). In this lab, I created various machine learning models (k-nearest neighbor, random forest, linear regression) to histologically score transcriptomic data with fibrosis stage score and NAS, allowing us to interpret in vitro liver-on-a-chip models using in vivo clinical language.