Reading disease out of the genome — variants, networks, and the targets worth drugging.
I started in computer science at Islamic University, Bangladesh, and found my way into biology through code. An undergraduate project on COVID-19 and Mucormycosis co-infection turned into a genuine research direction: using computation to find where disease mechanisms are actually vulnerable.
That took me to Texas A&M–San Antonio for an M.S. in Biology, where I worked on antimicrobial resistance in Staphylococcus aureus and single-cell transcriptomics of HIV infection — and now to Purdue, where I study the shared genetic architecture of Cardiovascular-Kidney-Metabolic syndrome across ancestries.
The through-line is the same at every step: take a large, messy biological dataset and find the part that matters — the hub gene, the risk locus, the pathway, the druggable target.
In the era of big data, biology is not just an experimental science — it is a computational science. — Research statement
A transformer that predicts metabolomic profiles from microbial composition, then uses a signed Jacobian to call the direction of each interaction — production vs. consumption. Validated on synthetic ground truth, applied across six disease cohorts. See the interactive walkthrough ↓
Multivariate GWAS via Genomic SEM across 13 clinical traits, identifying 250+ shared risk loci. Post-GWAS pipelines — partitioned heritability, fine-mapping, and cross-ancestry polygenic risk scores — separated Stage 4 patients from Stage 0 controls in the All of Us cohort.
Analyzed 1,062 whole-genome sequences to catalogue AMR genes, then built interaction networks to isolate the genes actually driving resistance. Mapped resistance mechanisms — efflux pumps, target modification, target protection — onto drug classes including penams and tetracyclines.
Single-cell RNA-seq of immune cells from HIV patients, with differential expression called in Seurat and focused on CD4⁺ T cells. GO and KEGG enrichment classified the signal; molecular docking then nominated Hypericin as a candidate inhibitor of key viral-associated proteins.
At UT Health San Antonio, modelled miRNA correlates of immune resilience and vaccine response (Random Forest, SVM, gradient boosting) by integrating single-cell and bulk RNA-seq. Separately, built deep convolutional networks for colon cancer and osteosarcoma detection from medical images.
Microbiome sequencing tells you which organisms are present, not what they're doing. m2mBERT is a transformer that predicts a sample's full metabolomic profile from its microbial composition — then, crucially, is interrogated with a signed Jacobian to recover the direction of each microbe–metabolite interaction: production vs. consumption. The result isn't just a prediction, but a ranked set of testable biological hypotheses.
A positive Jacobian means: when microbe i's abundance rises slightly, the model predicts more of metabolite j — a production-like hypothesis. The magnitude ranks how strongly the two are coupled.
Validated on synthetic consumer–resource data where the true interaction signs are known, so the direction calls are defensible before being applied to real disease cohorts.
