Raw SRA Expression Sources

Some public RNA-seq studies deposit raw reads without a processed expression matrix. Oncoref handles these as explicit build-time sources: the registry pins the complete run inventory, read checksums, sample roles, and transcriptome reference; scripts/build_sra_salmon_source.py turns those inputs into the same canonical source-matrix artifacts used by the other expression builders.

Salmon is required only to regenerate a raw-read source. It is not imported by oncoref and is not a runtime package dependency.

MMNST Source

prjna1083972-mmnst is the first source_type: sra-salmon entry. BioProject PRJNA1083972 / study SRP493407 contains three malignant melanotic nerve sheath tumors and three matched normal-nerve controls.

The build always quantifies and audits all six runs. Only the three runs whose registry records declare both role: tumor and cancer_code: SARC_MMNST enter the tumor matrix. Matched-normal records cannot declare a cancer code, so a configuration error fails during registry parsing instead of contaminating the reference cohort.

Build

Install Salmon, then run:

python scripts/build_sra_salmon_source.py prjna1083972-mmnst --threads 8

The default cache is ~/.cache/oncoref/expression/prjna1083972-mmnst/. Allow roughly 18 GB for the paired FASTQs, Ensembl transcriptome, Salmon index, quantifications, and derived artifacts.

Every FASTQ is downloaded from its pinned ENA INSDC mirror URL and verified against the deposited MD5. The Ensembl release 112 GRCh38 cDNA and ncRNA FASTAs are each checksum-pinned, then deterministically combined and verified against a third SHA-256. Salmon runs with automatic library detection, mapping validation, sequence bias correction, and GC bias correction. Transcript TPM is summed by the gene: identifier in that combined Ensembl FASTA before oncoref canonicalizes gene IDs and computes sample QC. Interrupted downloads retain their partial file and resume with an HTTP range request when the mirror supports it. The transcriptome checksum and selected Salmon version jointly key the index cache. A cached quant.sf is reused only when its transcriptome checksum, paired read checksums, automatic library type, and ordered Salmon arguments exactly match the current registry inputs and its recorded Salmon version matches the selected executable.

Reviewed Salmon output can be reused without rerunning quantification:

python scripts/build_sra_salmon_source.py prjna1083972-mmnst \
  --quant-dir /path/to/quant \
  --transcriptome /path/to/Homo_sapiens.GRCh38.cdna_ncrna.fa.gz

--quant-dir must contain <run-accession>/quant.sf for every declared tumor and control run. --transcriptome can also be used without --quant-dir; the supplied combined FASTA is checksum-verified and used directly to build the Salmon index, without downloading the registry FASTA components.

--force-download refreshes and checksum-verifies the FASTQs even when their existing quant.sf remains reusable. Use --force-quant separately when Salmon must run again. --quant-dir selects immutable external quantifications, so it cannot be combined with --force-quant or --force-index.

Outputs

The derived directory contains:

• SARC_MMNST_per_sample_tpm.parquet: the three tumor runs only.
• SRP493407_MMNST_2024_run_manifest.csv: all six roles, routes, URLs, and checksums.
• SRP493407_MMNST_2024_salmon_quantification_audit.csv: transcript mapping coverage, Salmon version, inferred library type, and TPM conservation for every run.
• The standard mapping, parse, sample-QC, and summary-row sidecars consumed by the expression artifact rebuild.

The source is an ultra-rare cohort (n=3), so its dispersion and tail percentiles are intrinsically thin. The registry preserves its salmon_gene_tpm source-scale class so consumers can distinguish it from alignment/count-based cohorts when absolute cross-source comparisons matter.