PRIMO benchmark inputs

PRIMO evaluates whether embeddings from biological foundation models preserve clinically relevant information at the patient level.

This repository contains the data used to generate benchmark submissions. Dataset identities are hidden: files are grouped under IDs such as d001 and d002, without disclosing the disease, tissue, or prediction target. Models produce one embedding per sample, and the evaluation is run separately in the PRIMO Space.

Contents

datasets.yaml lists each dataset's ID, path, number of samples, number of genes, and gene identifier type.

Each dataset directory contains an expression.h5ad file with raw counts stored as AnnData. Rows are samples identified by sample_id. Columns use NCBI gene IDs, with gene symbols included in the metadata.

Submit embeddings

1. Download the data:

   hf download ScientaLab/primo --repo-type dataset --local-dir primo
   

2. Generate one embedding per sample for every dataset. Embedding dimensions may differ between datasets.

3. Combine the embeddings into one CSV, TSV, Parquet, or NPZ file. Tabular files must contain dataset_id, sample_id, and one column per embedding dimension (e0, e1, etc.). NPZ files must contain dataset_ids, sample_ids, and embeddings.

4. Upload the file to the PRIMO evaluation Space.

Scoring

For each task, PRIMO trains the same linear probe on fixed cross-validation folds. Classification tasks use AUROC and regression tasks use Pearson correlation. Dataset scores are converted to a 0 to 1 skill score and aggregated by medical specialty.

A submission is ranked only if it covers every dataset. Labels are kept private and evaluation runs server-side.

Each dataset is redistributed under its source's original open license.