--- pretty_name: MetacellFormer evaluation inputs tags: - single-cell - benchmark --- # MetacellFormer evaluation inputs This repository provides available input matrices, fixed cell rosters, labels, data splits and evaluation settings **for the six main MetacellFormer benchmarks and the additional Gaussian-noise and cross-study PC9 analyses**. Some source datasets must be obtained from their original providers, as described below. Published scores and predictions are provided separately in the article’s Figure Source Data and Supplementary Tables. ## Layout |Path|Purpose| |-|-| |`inputs/cell_annotation/`|Seven archived cell-annotation matrices| |`inputs/perturbation/K562.h5ad`|Fixed K562 input for exact and functional perturbation retrieval| |`inputs/pc9/`|Three external PC9 study matrices| |`contracts/cell_type_annotation/`|Nine rosters, metric labels, Gaussian protocol| |`contracts/drug_response_classification/`|Cell metadata and fixed folds| |`contracts/perturbation_retrieval/`|K562 selected cells and functional annotations| |`contracts/pc9/`|Training/test cell IDs, comparison membership and metadata, gene mapping, protocol| |`contracts/tissue_specificity_prediction/`|GTEx labels, folds and shared gene support| |`contracts/protein_interaction_prediction/`|HuRI pairs, folds and gene-info snapshot| |`contracts/evaluation_roster.tsv`|Cohort and evaluation design for the retained analyses| |`contracts/model_representations.tsv`|Model dimensions, pooling and gene-mapping settings| |`input_manifest.tsv`|Included and externally obtained inputs, source terms, checksums and preparation notes| The six main benchmarks are cell-type annotation, drug-response classification, exact perturbation-state retrieval, functional perturbation retrieval, tissue-specificity prediction, and protein-interaction prediction. MetacellFormer-SC is the matched control. Gaussian and PC9 analyses are additional evaluations. ## Input preparation * Cell-annotation matrices retain their archived numerical values and feature order. `obs.split` uses the original `0` training and `2` test codes. `var.metacellformer_gene_id` and `var.geneformer_ensembl_id` retain model-specific feature mappings; do not collapse MetacellFormer aliases before Gaussian preparation. MCA and MHSP use the historical human one-to-one ortholog axes. * K562 preserves the fixed 262,998-cell selected population and 8,563-gene axis. Align embeddings with `contracts/perturbation_retrieval/selected_cells.tsv`; functional groups are in `functional_annotations.tsv`. * PC9 study inputs must be encoded separately and aligned to `contracts/pc9/test_cells.tsv.gz`. GSE282074 retains decimal RSEM expected counts; its controls are shared by two comparison memberships. The 2,393 training cells require authorized scDrugMap Data11 access. * HumanPBMC, Pancrm and the exact processed scDrugMap matrices are not redistributed. Obtain them from the sources in `input_manifest.tsv` and align to the released rosters, IDs and folds. GTEx and HuRI use released labels/pairs and fixed folds with gene embeddings The companion GitHub repository documents embedding extraction and scoring in `benchmarks/README.md`. Verify a download from this directory with `sha256sum -c checksums.sha256`. ## Terms Derived evaluation contracts are CC BY 4.0 where applicable. Input matrices and reference files retain their upstream terms; consult `input_manifest.tsv` and [LICENSES.md](LICENSES.md). This repository does not relicense upstream inputs.