Back to selected work

Team spatial-proteomics research

COHERE

Cross-cohort harmonization and representation transfer for cell-type annotation across heterogeneous cohorts.

PyTorchSet TransformersRepresentation Learning

Problem

Spatial-proteomics cohorts use different feature and annotation conventions, making transferable cell-type representation difficult.

Constraints

The project uses a held-out cohort and fold-local controls to limit leakage across heterogeneous datasets.

Approach

Standardize cohorts, engineer mask-aware cell representations, and evaluate an open-vocabulary Set Transformer with cosine-prototype transfer.

My contribution

My contribution covered cohort standardization pipelines, mask-aware representation engineering, fold-local leakage controls, and architecture/methodology diagrams.

Results

Team research remains subject to metric verification. No cross-cohort F1 or other quantitative claim is shown until the exact evaluation setup and result are confirmed.

What I would do next

Publish a verified result table, frozen-test protocol, and explicit limitations.

Artifacts & links