Projects

Predicting 80+ molecular targets from routine H&E whole-slide images.

Problem

Molecular testing is expensive and slow. If routine H&E slides could predict molecular targets directly, that would give clinicians a fast, cheap signal for which tests to order.

Data

H&E whole-slide images with molecular labels, spanning three independent cohorts so the model could be tested on sites it never trained on.

Approach

Foundation-model embeddings for each tile feed an attention-based multiple-instance learning model that aggregates thousands of tiles into a slide-level prediction. It runs as containerized, reproducible environments on HPC.

Result

Predicts more than 80 molecular targets, reaching AUROC up to 0.90 on external validation across the three cohorts.

Next

Stronger calibration and per-target reliability reporting, and more testing of robustness across scanners and staining protocols.

Clinical project at LMU University Hospital. Data cannot be shared; happy to walk through the methods.

A Cox-regression tool that supports treatment planning for brain metastases.

Problem

Treatment planning for brain metastases weighs several risk factors at once, and clinicians need a signal they can read quickly and trust.

Approach

A Cox-regression model turned into a decision-support tool, built to stay interpretable at the point of care rather than act as a black box.

Result

A risk estimate that supports treatment-planning conversations in radiation oncology.

Clinical project at LMU University Hospital. Methods on request.

Morphometry for CT and MR segmentations, with a 3D viewer that fits in one HTML file.

Problem

Measuring shape and size from medical-image segmentations usually means stitching together several tools and a heavyweight viewer, which makes quick inspection painful.

Approach

A small Python library that computes morphometric features from CT and MR segmentations, plus a 3D viewer that runs from a single self-contained HTML file, so results open anywhere with no setup.

Result

A lightweight, reproducible way to compute and eyeball morphometry, useful in radiomics and imaging work.

Heuristic DMR caller for WGBS data using a grid approach.

Problem

Calling differentially methylated regions at genome scale across cytosine contexts, without a caller that overfits to any one context.

Approach

A grid-based method built from the algorithm up, handling the main contexts in one framework.

Result

Released as open source and used in peer-reviewed, co-first-authored work.

Benchmarking framework that scores methylation callers against known ground truth.

Problem

Differential-methylation comparisons usually report agreement between callers, which is not the same as being right.

Approach

Inject known methylation signals into real data to manufacture ground truth, then score callers against a known answer.

Result

An honest benchmark for methylation pipelines.

Not yet public. Happy to walk through it.

Multi-omics subtyping of the tumor microenvironment in head & neck cancer.

Problem

Patients with similar diagnoses respond differently to treatment, and some of that difference lives in the tumor microenvironment.

Approach

Unsupervised ML on multi-omics data to define microenvironment subtypes, then link them to outcomes across treatment cohorts.

Result

Stratification signals that could support treatment decisions.

Clinical project at LMU University Hospital. Methods on request.

Analyses of DNA methylation (5mC) and protein-DNA binding from CUT&RUN.

Problem

Connecting DNA methylation and protein-DNA binding to biological signal across epigenomics studies.

Approach

Processing and integrating 5mC methylation and CUT&RUN binding data, from raw sequencing through to interpretation.

Collaborative research. Details depend on the study.