Research
I apply bioinformatics methods and machine learning to understand how
gene-regulatory programs drive cancer and other diseases. My research combines
transcriptomic analysis, network biology, and interpretable predictive modeling
to identify regulatory programs associated with disease phenotypes, with an
emphasis on population-aware analysis. I am now extending this work to
single-cell and spatial resolution. Throughout, my hands-on training in
molecular and cell biology keeps me close to how the data are actually generated.
Regulatory networks
Population-stratified regulatory networks in cancer
Tumor biology differs across ancestral populations, yet most regulatory
analyses pool patients together. I build population-specific ceRNA
(competing endogenous RNA) and gene-regulatory networks to expose those
differences. Across prostate (TCGA-PRAD) and testicular (TCGA-TGCT)
cohorts, I compared African American and European American patients —
integrating mRNA, lncRNA, and miRNA expression with somatic mutation
profiles — and built population- and tumor-type-specific networks that
reveal distinct hub genes and pathways. My current work prioritizes
regulatory axes with a multi-criterion composite score (statistical
robustness, experimental miRNA–target evidence, network topology, and
curated disease association) and links them to druggable targets through
a custom DGIdb GraphQL integration.
Methods — DESeq2 · starBase / HMDD · TCGAbiolinks · Cytoscape / MCODE · Fisher Z
Machine learning
Interpretable machine learning for biomarker discovery
I treat machine learning as a tool for discovery rather than a black box,
pairing predictive performance with interpretability. In early-onset
prostate cancer, I integrated WGCNA co-expression analysis with
differential expression to narrow candidate genes, then built a nested
leave-one-out pipeline with LASSO feature selection across SVM, Random
Forest, and ANN classifiers (AUC up to 0.91). This identified the lncRNA
PCSEAT as the single biomarker shared by all three models in a White
early-onset cohort, with robustness confirmed by permutation importance
and SHAP and validated in an independent dataset. Related work applies
the same philosophy to glioblastoma treatment response and tumor
classification.
Methods — WGCNA · LASSO · SVM / RF / ANN · SHAP · nested cross-validation
Emerging direction
Single-cell & spatial transcriptomics
I am extending network- and ML-based thinking to single-cell and spatial
data, where the goal is to connect regulatory programs to cell identity
and tissue organization. In an independent project on mouse-brain spatial
transcriptomics, I built a scanpy/squidpy pipeline for preprocessing,
clustering, and spatial-domain analysis, with a PyTorch cell-type
classifier benchmarked against a scikit-learn baseline and interpreted
through feature importance. Regulatory genomics at single-cell and spatial
resolution is the direction I most want to grow into.
Methods — scanpy · squidpy · PyTorch · spatial-domain analysis
Experimental grounding
A foundation at the bench
My computational work is anchored in hands-on molecular biology. Through
laboratory training I gained practical experience in nucleic-acid
extraction, conventional and real-time PCR, gel electrophoresis, and
mammalian cell culture with viability assays — experience that keeps my
analyses tied to how the data are actually generated.
Methods — DNA/RNA extraction · qRT-PCR · electrophoresis · cell culture · MTT / AO-PI assays