Data Science · Machine Learning · Drug Discovery
Data Scientist & Drug Discovery Scientist
I build statistical and machine-learning models on large, messy biological and clinical data, then turn what they find into decisions: which compound to advance, which patient is at risk, which signal is real. Ten-plus years across drug discovery, genomics, and applied ML, with a patent and peer-reviewed work behind it.
About
My PhD at Ohio State was in statistical genetics: small-sample and rare-event inference, work that ended up peer-reviewed in the American Journal of Epidemiology. That grounding shows up in everything since. I build models that are accurate, but also honest about their own uncertainty.
Since then I've worked across predictive modeling, biomarker analysis, and ML pipeline development in drug discovery: a Senior Scientist role at Amgen, data science and computational biology at Sensorium Therapeutics, and a research fellowship in HIV genomics.
I care about the unglamorous parts: reproducible pipelines, honest evaluation on imbalanced data, and code a colleague can read six months later. I work best sitting between the bench scientists and the cluster, translating what came out of an experiment into something a team can act on.
Experience
Apr 2025 to May 2026
Amgen
Sep 2024 to Apr 2025
Sensorium Therapeutics
Sep 2022 to Sep 2024
Sensorium Therapeutics
Mar 2022 to Aug 2022
Henry M. Jackson Foundation
Aug 2016 to Mar 2022
The Ohio State University
Selected projects
Each of these started as a modeling question and ended as a number I'm willing to defend, including the corrections that made the number smaller.
87%
AUC, corrected
Polygenic risk score from 182 sequenced exomes in preterm infants (26–31 weeks), combined with gestational age. Raw AUC of 96% dropped to 87% once I corrected for over-fitting. That is the honest number, and the one that matters. Published in Pediatric Research.
760
MoA classes
Random-forest model classifying compounds into 760 distinct mechanism-of-action classes from L1000 gene-expression signatures, at over 80% accuracy. Built to give chemists a first read on what a new compound is probably doing.
0.96
AUC
Random-forest classifier predicting BBB permeability from compound features, built to prioritize candidates earlier in the discovery pipeline. The features that carry the model (logP, molecular weight, TPSA) are exactly the ones medicinal chemists already reason about, which is what made it trustworthy to them.
~10%
true KO rate
CRISPR screen analysis with pertpy: Mixscape to separate true knockouts
from escapers, where only about one cell in ten was really perturbed, then E-distance to
rank effect size and an E-test for significance. Top hits recover the IFN-γ/JAK-STAT
axis with correct STAT1 specificity, plus a genetic-interaction table calling synergy
vs. suppression in double perturbations.
98%
cell-type accuracy
Benchmarked PCA vs. Harmony vs. scVI on the heart cell atlas using batch-mixing entropy and cell-type silhouette. Harmony wins on mixing, scVI preserves more biology, and both beat raw PCA. A genuine trade-off rather than a winner. Then scANVI for label transfer, reaching 98% on held-out cells, with errors concentrated in rare and similar types.
0.944
ROC-AUC
Scored AlphaMissense pathogenicity predictions against ClinVar labels across ten disease genes. The headline AUC is strong, but the interesting result is the failure mode: NPC1 I1061T, a common and firmly pathogenic Niemann-Pick C1 mutation, scores 0.303 and is called benign. Aggregate metrics hide the misses that matter clinically.
~90%
accuracy
ResNet transfer-learning model for binary gate classification, fine-tuned with early layers frozen to adapt a general-purpose image model to flow-cytometry data. A practical way to get a useful classifier out of a small labelled set.
2×
eval tiers
Retrieval-augmented QA over real PubMed abstracts, with cited answers and a hallucination guard that refuses when the retrieved context doesn't support a claim. Evaluated in two tiers. Recall@k and MRR for the retriever, then LLM-as-judge faithfulness and answer relevancy for the generator, because a good retriever and a good answer are different failures.
Skills