PRIMO public inputs

Blind expression inputs for the PRIMO patient-representation benchmark. datasets.yaml lists opaque ids, modalities, paths, and dimensions; clinical targets, provenance, donor ids, and cohort metadata remain private.

Each <dataset_id>/expression.h5ad stores expression in X, NCBI Gene IDs in var_names, and symbols in var["gene_symbols"]. X is raw counts unless uns["input_normalization"] is set, which names the scale the source study published instead (TPM, for example); the same key appears in datasets.yaml. Check it before applying a count-based normalization. The preprocessing is yours to choose; the baseline applies log2(CPM + 1) to every dataset.

• Bulk RNA files have samples as rows and opaque sample ids in obs_names.
• Single-cell RNA files are sparse and have cells as rows. Cell ids in obs_names are opaque. Their only obs column is sample_id, an opaque collection-sample id shared by cells from the same prediction unit.

Each submission covers one selected modality. Submissions are sample-level: produce one embedding per unique obs["sample_id"], with repeated collection timepoints kept separate.

The standalone quickstart in the PRIMO evaluator Space shows the modality-specific download and submission flow. Its single-cell example applies per-cell log2(CP10K + 1), mean-pools by sample_id, then runs sample-level PCA.