Biomedical measurements often arrive in forms that hide their internal relationships: columns in a table, nodes in a graph, or genes in a sparse vector. I develop models that learn or preserve those relationships as spatial structure.
The result is a shared research direction across cancer, genomics, and single-cell biology: stronger prediction paired with a representation that scientists can inspect, perturb, and reason about.
Each project has an interactive site with the paper, central figures, demonstrations, and examples. These are research artifacts—not generic software demos.
A fully differentiable framework that jointly learns feature placement and prediction. The learned map exposes spatial neighborhoods of predictive variables while enabling vision models to operate on unordered biomedical tables.
A semantic-cartography framework that transforms networks into image sets. It gives image models multi-scale and global context without forcing billion-node graphs through conventional message-passing pipelines.
A reusable graph model that learns a common language for node structural roles from non-biological networks, then transfers that knowledge into unseen biological systems.
A knowledge-guided framework that builds a standalone molecular graph for each patient. Curated pathways define the edges while patient-specific expression defines node features, preserving pathway topology for prediction and interpretation.
A vision foundation model that uses optimal transport to give genes stable spatial homes. Co-expression becomes local texture, allowing one frozen encoder to support annotation, integration, gene-program discovery, perturbation, and disease-axis analysis.
A deterministic compiler that converts network topology into named structural roles, reports the measurements supporting each role, and describes structural counterfactuals for frozen language models.
The projects differ in data and scale, but they follow a common logic that keeps prediction and scientific interpretation connected.
Ask what the original representation removes: feature proximity, graph-wide context, continuous expression, or relationships across modalities.
Use learned cartography, semantic placement, or optimal transport so related variables form neighborhoods that vision models can exploit.
Project attribution, attention, masking, and perturbation back to locations with named genes, nodes, pathways, and measurable programs.