OngoingJun 2025 — Present
Liver Cancer Vascular & Immune Mapping
ML & Data Researcher · Stanford School of Medicine — Dhanasekaran Lab
Leading ML efforts to visualize circulatory and immune-cell patterns across 1.4 million cells in hepatocellular carcinoma tissue, uncovering vascular and patient-level structure invisible to manual review.
Methods & results
- Mapped microvessel density relative to immune-cell distribution across more than 100 tissue regions using spatial statistics (Ripley's L, Voronoi tessellation, Moran's I).
- Applied dimensionality reduction (t-SNE, UMAP) and clustering (DBSCAN, cellular neighborhood analysis) to identify recurring spatial neighborhoods linked to patient outcomes.
- Contributed to wet-lab procedures (IHC, cell culture) to ground computational findings in lab-verified biology.
- Work is in preparation for publication and lab co-authorship.
- Spatial Biology
- CODEX Imaging
- Python
- R
CompleteNov 2024 — Feb 2025
PD-L1 / TGFBR1 Dual-Inhibition Generative Model
Lead Researcher · Independent Research, consulted with Dhanasekaran Lab
Designed a generative drug-discovery model to find small molecules that block two pathways — PD-L1 and TGFBR1 — both implicated in HCC recurrence and minimal residual disease.
Methods & results
- Combined a graph-based variational autoencoder (GAT-VAE) with a fine-tuned molecular language model (ChemBERTa) to generate and score candidate inhibitors.
- Trained on a dataset of roughly 18,000 proteins and generated two validated, potential dual-pathway inhibitors.
- Placed 4th in Translational Medical Sciences at the Austin Science and Engineering Fair (GARSEF); manuscript in preparation for potential AASLD submission.
- GAT-VAE
- ChemBERTa
- Drug Discovery
- Generative AI
CompleteNov 2024 — Feb 2025
p53–CDKN1A Mutation Binding Model
Lead Researcher · Independent Research, mentored by Dhanasekaran Lab
Built a graph neural network to predict how p53 mutations affect DNA-binding to CDKN1A — a key tumor-suppressor interaction in cancer — replacing slow physical simulation with a fast, learned model.
Methods & results
- Achieved 91% accuracy modeling molecular binding behavior across mutation variants.
- Estimated to save roughly 2,500 hours of computation versus standard structural simulation methods.
- Surfaced novel structural insights, including interacting loop regions and especially destabilizing ('fatal') mutations.
- Placed 3rd at GARSEF, named a National STEM Festival finalist, invited speaker at AIMUG, and commended by MIT THINK (top 10% of 1,092 applicants).
- Graph Neural Networks
- Structural Biology
- PyTorch
CompleteNov 2025 — Jan 2026
PD-L1–TGFBR1 Pathway Extension (BioML Society)
Research Student · BioML Society, in consultation with a UT Austin researcher
Extended the dual-inhibition pathway work specifically toward cancer stem cells (CSCs) in HCC, under remote mentorship from a UT Austin researcher affiliated with the BioML Society.
Methods & results
- Focused on how dual-inhibition of PD-L1 and TGFBR1 affects cancer stem cell behavior specifically, rather than the bulk tumor population.
- Iterated on modeling approach based on structured external research feedback.
- Cancer Stem Cells
- HCC
- Mentored Research
Concept
AI Diagnostics
Screening tools that flag early signs of disease in routine imaging and lab data, built to explain their reasoning to clinicians.
- Medical imaging
- Explainable AI
Concept
Brain Health Analytics
Using longitudinal data to spot early changes in cognition and neurological health.
Concept
Healthcare Data Science
Finding where care is delayed or unequal by studying public health and hospital datasets.
Concept
Predictive Medicine
Models that estimate a patient's risk of recurrence or complications so follow-up can be personalized.
Concept
Biomedical Machine Learning
Reusable methods for protein, molecule and tissue data that other student researchers can build on.