Computational Biology · Scientific ML · Drug Discovery

From biological data to targets, mechanisms, and candidates.

I develop and evaluate quantitative and machine-learning methods, then build the analytical pipelines that connect perturbation screens, CRISPRi phenotypes, and molecular evidence to mechanism-of-action hypotheses and experimental decisions.

Computational Biology · Broad Institute

Functional genomics and perturbation screening at decision scale.

Current work is organized around the decisions it supports: which signals are reliable, what mechanism they suggest, and what should move forward.

PROSPECT · Functional genomics · M. tuberculosis

Owning the path from pooled compound screens to high-confidence candidates.

Large pooled screens require reliable separation of biological signal from barcode, contamination, replicate, and pool-structure artifacts. I lead computational development across quantification, QC, dose-response modeling, clustering, hit prioritization, and mechanism-of-action inference, translating the evidence with experimental scientists into go/no-go calls and follow-up experiments.

Scale: 100K+ small-molecule compounds
Decision: Top 1% prioritized
Delivery: Reproducible workflows · HPC · cloud

Public scientific context ↗

Virtual screening · MoA inference · Selection algorithms · Compound triage

Turning billion-scale predictions into compounds worth experimental review.

I built the analysis pipeline across 10B+ compounds, generating compound-level summary statistics and evidence for mechanism-of-action inference and candidate prioritization. I contributed to the design of the downstream compound-selection algorithm and evaluate alternatives for scaffold generalization and biological relevance. The workflow feeds a deterministic medicinal-chemistry triage system covering RDKit structure standardization and liabilities, drug-likeness and physicochemical properties, applicability-aware ADMET/toxicology, TB precedent and novelty, DuckDB-backed results, and Streamlit human review for final compound selection.

Scale: 10B+ virtual compounds
Evaluation: Scaffold-aware · profile-level
Downstream: MedChem/ADMET + scientist review

Target analysis · CRISPRi · Biological benchmarking

Quantitative methods for interpreting targets and mechanisms.

I develop gene-set curation and weighted, null-comparison, and threshold-free ranking approaches to benchmark biological signal against known antibiotic mechanisms. I also connect M. tuberculosis CRISPRi phenotypes with pathway and compound-response signatures to strengthen target interpretation and MoA hypotheses.

Broad Institute descriptions use public or aggregate information and exclude confidential research details.

Research & Publications · Multi-Omics

Exposure biology across transcriptomic and epigenetic evidence.

Integrated RNA-seq and DNA methylation across 250+ pediatric biospecimens, supported by a reproducible pipeline that increased processing speed by 50%.

Master's thesis · Publication 2024

Heavy-metal exposure leaves coordinated transcriptomic and epigenetic signals.

Integrated adolescent RNA-seq and cord-blood DNA methylation across 250+ pediatric biospecimens to identify exposure-associated signatures connected to immune and neurodevelopmental pathways, while examining epigenetic regulation and immune-cell composition.

Exposure biology
RNA-seq
DNA methylation
Immunotoxicology

DOI: 10.7302/27924 Deep Blue record

Scientific AI & Machine Learning

Representation learning and graph ML for structured biological data.

Selected academic projects show how I build and assess representation-learning and graph-based approaches while keeping class imbalance, validation, and biological interpretation visible.

Exploratory course project · Single-cell

Testing whether a latent model preserves biological time.

UMAP of single-cell expression states colored by measured capture time — True capture time ↔ posterior time estimate

UMAP of the same single-cell states colored by the model's inferred time — True capture time ↔ posterior time estimate

Built a VAE with a capture-time-centered KL objective across 14,717 cells and 3,815 expression features. The paired views test temporal preservation rather than reconstruction alone; ablation and held-out validation remain next steps.

Code and figures ↗

Collaborative course project · Spatial transcriptomics

Predicting metastatic potential while exposing class imbalance.

Spatial graph connecting neighboring breast-tissue observations by tissue position — 1,657-node spatial graph ↔ stratified test AUROC 0.91

ROC curve with class-level accuracy results for the spatial graph model — 1,657-node spatial graph ↔ stratified test AUROC 0.91

I co-implemented, optimized, and trained a three-layer graph convolutional network using spatial coordinates and a six-gene ECM signature. The report documents AUROC 0.91 alongside weaker minority-class accuracy and the need for external validation.

Full report ↗

Model Evaluation & Scientific Decision-Making

How I make quantitative methods decision-ready.

Useful methods respect the experiment, survive realistic evaluation, and leave an inspectable evidence trail.

Experiment

Match the model to the biological question.

Experimental design and assay behavior
Confounding and covariate structure
Effect sizes tied to mechanism

Validation

Evaluate under real data constraints.

Class imbalance and scaffold-aware splits
Calibration and profile-level benchmarks
Biological relevance beyond aggregate scores

Interpretation

Make uncertainty inspectable.

Decision-aware visualizations
Traceable analytical outputs
Explicit limitations and next experiments