ML Bioinformatics Scientist — Deep Learning for Genomics

Summary

5 years at the intersection of deep learning and genomics. Trained sequence-to-function transformer models predicting tissue-specific gene expression from DNA sequence (Enformer/Borzoi-class). Built production MLOps infrastructure at Insitro for 12 research scientists. Strong software engineering instincts — treats ML pipelines as software products with versioning, testing, and monitoring.

Experience

ML Scientist II — Insitro (2022–Present)

• Fine-tuned Borzoi-class sequence model on 400 GB of in-house functional genomics data (ATAC-seq, RNA-seq, Hi-C); model now standard for variant effect prediction across 3 disease programmes.
• Built end-to-end MLOps platform on AWS SageMaker + MLflow: automated training, evaluation, and model registration for 12 research scientists.
• Implemented JAX-based in-silico mutagenesis pipeline; identified 3 causal noncoding variants validated experimentally.

Graduate Researcher — Pfenning Lab, CMU (2015–2020)

• Developed HALGAN, a conditional GAN for generating cell-type-specific regulatory sequences; code open-sourced with 500+ GitHub stars.
• Benchmarked 8 sequence-to-function architectures across 218 ENCODE cell types; results published in Genome Research.

Selected Publications

• Park Y. et al. “Cell-type-aware regulatory sequence generation with conditional adversarial networks.” Genome Research, 2021.
• Park Y. & Pfenning AR. “Benchmarking deep learning models for regulatory genomics across 218 ENCODE cell types.” Bioinformatics, 2020.

Code Quality Notes

PyTorch and JAX modules ship with typed interfaces (mypy strict), property-based tests via hypothesis, and deterministic seed fixtures. Model training configs managed via Hydra; experiments fully reproducible from a single YAML. GPU-optimised dataloaders benchmarked with torch.profiler before merge.