rpowalski/gsva_global

Global-reference GSVA scores

Precomputed GSVA scores for bulk RNA-seq and pseudobulk AnnData files. Each top-level directory is named after its original .h5ad file (without the extension).

All samples are scored against the same frozen empirical reference, built from 714,800 non-benchmark ARCHS4 and pseudobulk samples across 19,260 genes. Expression is transformed with CP10k + log1p. For every target sample, each gene is mapped to its percentile in the frozen per-gene reference distribution, then scored with the classical GSVA random walk against 6,245 GO Biological Process gene sets from MSigDB.

Benchmark samples were excluded from reference construction (11,829 samples).

Each dataset directory contains:

• —gsva_matrix.f16.npy: raw GSVA scores, shaped [samples, gene sets].
• —samples.parquet: matrix row index, sample ID, available annotations, and benchmark split flag.
• —sets.parquet: matrix column index, gene-set metadata and membership, plus global score mean and standard deviation.
• —reference/meta.json: reference identity and scoring parameters.

The global mean and standard deviation in sets.parquet were estimated by scoring a fixed random subset of 20,000 reference samples. They are provided for optional downstream calibration; gsva_matrix.f16.npy contains unscaled raw scores. The 20,000 samples are not an averaged expression vector and are not the empirical reference itself.

from huggingface_hub import hf_hub_download
import numpy as np
import pandas as pd

name = "archs4_tpm_counts_batch_cell_line_control"
repo = "rpowalski/gsva_global"
scores = np.load(hf_hub_download(repo, f"{name}/gsva_matrix.f16.npy", repo_type="dataset"))
samples = pd.read_parquet(hf_hub_download(repo, f"{name}/samples.parquet", repo_type="dataset"))
sets = pd.read_parquet(hf_hub_download(repo, f"{name}/sets.parquet", repo_type="dataset"))