SRA Expression Sources

Oncoref prefers an authoritative processed expression product when one is available for a public SRA study. A processed source avoids reproducing a large alignment or quantification workflow, while registry-pinned checksums preserve a reviewable connection to the exact upstream files.

The sra-ncbi-counts builder consumes NCBI's public Gene Feature RNA-seq count tables. It does not download raw reads and does not run Salmon or another aligner or quantifier. Raw-read sra-salmon support remains available for studies that have no suitable processed product, but it is a separate fallback contract.

MMNST Cohort

prjna1083972-mmnst provides the SARC_MMNST reference cohort. BioProject PRJNA1083972 / SRA study SRP493407 contains three primary malignant melanotic nerve sheath tumors and three normal-nerve controls. The controls are independent donors, not patient-matched specimens.

All six NCBI analysis accessions and count-file MD5 digests are pinned in expression_sources.yaml. All six runs appear in the run manifest and mapping coverage audit. Only records with both role: tumor and cancer_code: SARC_MMNST enter the released tumor matrix. A normal-control record cannot declare a cancer code, so an invalid route fails when the registry is parsed.

The cohort is described in PMID 41995777 and deposited under PRJNA1083972. The paper reports oligo(dT) mRNA enrichment, while the deposited INSDC records say RANDOM; the registry preserves this discrepancy instead of choosing one description silently.

VSCC Cohort

prjna994918-vscc provides the direct VSCC reference from BioProject PRJNA994918 / SRA study SRP449588. The paper reports 13 pathology-confirmed invasive vulvar squamous cell carcinomas; nine tumors had RIN above 6 and received poly(A), paired-end bulk RNA sequencing. The selected RNA libraries represent six tumors reported as primary and three as recurrent, and the Tumor 2 specimen was taken from an unspecified metastatic site.

Every NCBI analysis uses scf_rnaseq_gene_counts 0.5.6 against GCF_000001405.40, with paired RF libraries and reported alignment rates from 55.21% to 70.15%. The registry pins all nine run/analysis pairs and count-file MD5 digests. All nine profiles pass the expression QC policy and map 95.7% to 98.3% of source counts into the reference input. The resulting 44,446-gene matrix is a direct expression reference, but its small mixed-origin composition keeps VSCC non-classifying.

The source's public molecular-provenance rows retain all 13 study tumors. They record direct hybridization-capture evidence for three PCR-confirmed HPV16 integrations, two HPV coinfections without detected human-virus junctions, and eight HPV-DNA-negative tumors. HPV status is never inferred from diagnosis or expression. The cohort is described in PMID 38898085 and deposited under PRJNA994918.

Processed-Count Contract

NCBI publishes one unnormalized gene-level count table per SRA run in its SRA Gene Feature RNA-seq counts archive. For each run, the oncoref registry pins:

  • the run and NCBI analysis accessions;
  • the BioSample, sample title, and tumor/control role;
  • the immutable count-table path and MD5 digest.

The generic NCBI GEO pipeline page still names GRCh38.p13 and Human Annotation Release 109.20190905. These S3 analyses were generated in June 2026 and contain 6,574 GeneIDs absent from that release, including expressed rows. Their 59,068-row universe has length parity for all but 15 rows with GCF_000001405.40-RS_2025_08, so the registry pins that GFF and assembly report by URL and SHA-256 and records the discrepancy rather than silently dropping the newer genes.

The builder streams the compressed GFF and retains intervals only for GeneIDs present in the count tables. Gene length is the exon union across RefSeq accessions assigned to Primary Assembly, plus the mitochondrial non-nuclear accession. This selection reproduces NCBI's published raw-count-to-TPM output; alternate loci and patch scaffolds would incorrectly multiply some lengths. A gene or pseudogene without exon records uses its gene span. Retired NCBI GeneIDs use the current live GeneID's length when NCBI Gene history provides an exact replacement.

Only rows with both a canonical Ensembl mapping and a resolved gene length enter TPM normalization. The mapping audit retains every source row, and the per-run coverage sidecar records the source-count fraction that entered the reference. Reviewed translations reconcile NCBI mitochondrial symbols such as COX1 and RNR2 with canonical MT-CO1 and MT-RNR2; aggregate loci such as IGH, IGK, and IGL remain unresolved rather than being assigned to an arbitrary member gene.

Build

Run the processed-count builder:

python scripts/build_sra_ncbi_counts_source.py prjna1083972-mmnst
python scripts/build_sra_ncbi_counts_source.py prjna994918-vscc

The default cache is ~/.cache/oncoref/expression/<source-id>/. Each build downloads approximately 80 MB when the shared RefSeq annotation is not already cached; the count tables themselves are small.

Reviewed local inputs can be supplied without network access:

python scripts/build_sra_ncbi_counts_source.py prjna1083972-mmnst \
  --counts-dir /path/to/ncbi-counts \
  --annotation /path/to/GCF_000001405.40_GRCh38.p14_genomic.gff.gz \
  --assembly-report /path/to/GCF_000001405.40_GRCh38.p14_assembly_report.txt

--counts-dir must contain <run-accession>.ncbi-gene-counts.tsv for every declared run. Local inputs are still verified against the registry checksums.

Outputs

The derived directory contains:

  • SARC_MMNST_per_sample_tpm.parquet: the three tumor runs only.
  • SRP493407_MMNST_2024_run_manifest.csv: all six analyses, roles, routes, URLs, and checksums.
  • SRP493407_MMNST_2024_ncbi_count_mapping_coverage.csv: raw-count mapping and reference-input coverage for every run.
  • SRP493407_MMNST_2024_ncbi_gene_length_audit.csv: the length and fallback method for every NCBI GeneID.
  • The standard mapping, parse, sample-QC, and summary-row sidecars used by the expression artifact rebuild.

This is an ultra-rare cohort (n=3). Its dispersion and tail percentiles are necessarily thin even though all three tumor samples pass the expression QC policy.

For prjna994918-vscc, the analogous matrix and sidecars are named VSCC_per_sample_tpm.parquet and SRP449588_VSCC_2024_*. The matrix contains nine tumors, while the separate molecular-provenance API retains all 13 study tumors.

Raw-Read Fallback

SraSalmonSource and scripts/build_sra_salmon_source.py remain the generic fallback for an SRA study that has no reviewed processed matrix or NCBI count product. That contract pins paired-read checksums, a transcriptome checksum, Salmon arguments, and the Salmon version used by each index and quantification cache. Salmon is not imported by oncoref and is not a runtime dependency.

The MMNST source does not use this fallback.