Research

Long-form write-ups of my papers: the problem, the method, the evidence, and the caveats, with links to code and data.

Oct 23, 2025 · HONeYBEE

HONeYBEE: foundation model embeddings for multimodal oncology data

A publicly released framework that turns clinical text, slides, radiology and omics into patient embeddings, evaluated on 11,428 TCGA patients.

Dec 16, 2025 · CLEVER & pathology reports

Consensus reasoning with local LLMs for pathology report extraction

Eleven local LLMs extract, three reasoning models grade, one adjudicates: structured, auditable data from free-text surgical pathology reports.

Jun 12, 2025 · EAGLE

EAGLE: attention fusion and attribution for cancer survival

Fusing imaging, report and clinical embeddings with cross-modal attention, then asking which modality drove each patient's risk score.

Mar 2, 2024 · MINDS

MINDS: linking public cancer repositories into ML-ready cohorts

A metadata-first cloud framework that consolidates 41,499 open-access cancer cases into one queryable warehouse and fetches images and omics on demand.

Jul 30, 2025 · SeNMo

SeNMo: learning pan-cancer prognosis from sparse multi-omics data

A self-normalizing network trained on five omics layers plus clinical variables from TCGA, then adapted to predict survival, cancer type and TLS ratio.

Jul 10, 2026 · HetMoE

Gating, not averaging: HetMoE for transcription factor binding sites

A dense mixture of experts that weights DeepBIND, DeepSEA and DanQ embeddings per input, tested on held-out transcription factors with fair negatives.