Assays and opportunity

COSMIC's signatures are defined on whole genomes. An exome or a panel sees each trinucleotide context at a different rate, so fitting its counts with genome signatures biases the exposures toward signatures rich in the contexts the territory over-represents. The studio reads which assay a sample came from and corrects the catalog to it, in every refit.

Which assay: read off the positions

When Tally reads a VCF or MAF, it checks where the SNVs fall. A genome puts about 0.1% of its SNVs in MSK-IMPACT's 2.3 Mb of targets and about 3% in the 95 Mb of exons, so:

• panel: at least 20% of SNVs inside MSK-IMPACT's targets;
• exome: otherwise, at least half inside exons (GENCODE v19, the TCGA MC3 territory);
• whole genome: otherwise.

Real panel calls do not all fall inside one panel's target file: 72% of MSK-IMPACT 2017's SNVs lie in the v2 targets (later genes and off-target calls account for the rest). Hence the rule reads enrichment, not containment. At least 3 SNVs are needed, and the territories are hg19 (lift an hg38 file over first).

| Real calls | Called right, by SNV count | |---|---| | MSK-IMPACT 2017, 10,464 samples (cBioPortal) | 3–5: 99% · 5–10: 100% · 10 and more: 100% (8,761 samples with 3+ SNVs) | | 300 PCAWG whole genomes | 100% at every count | | The same genomes restricted to exons | 100% (295 with 3+ SNVs) |

Not from a spectrum alone

A spectrum without positions does not say its assay. Fitting every PCAWG tumor's projection onto each territory with COSMIC corrected to each, and taking the territory with the highest likelihood (sigscape-core's assay_loglik), called whole genomes right above 300 mutations (95–100%) and exomes and panels near chance (exome or not: 44–69%; MSK-IMPACT: 15–50%). A fit with 79 signatures absorbs the small differences between territories. So the studio asks for positions, or for the assay, and never guesses it from a spectrum.

The correction

A catalog is corrected by multiplying each channel by the ratio of its context's count in the territory to its count in the genome, then renormalizing each signature (sigscape-core's correct_catalog). The counts are the 32 trinucleotides in each territory of hg19, built with the reference bundle:

| Territory | Trinucleotides | |---|---| | whole genome | 2,861,326,455 | | exome, GENCODE v19 (TCGA MC3) | 94,620,884 | | exome, Roche SeqCap EZ v3 | 63,959,813 | | panel, MSK-IMPACT v2 with introns | 2,288,851 |

The artifact screen and the de novo workbench's missing-signature refit apply it by default when the record's assay is known; choosing another territory is possible and says that the catalog no longer matches the record.