Mapping onto PCAWG

Map to PCAWG places a sample among the 2,703 PCAWG tumors of the reference bundle and says which tumor type its nearest neighbors carry. The type is a hint about tissue context read off the spectrum alone; it is not a diagnosis.

How

Every reference tumor's genome spectrum is projected onto the sample's assay (each channel scaled by how often its context occurs in the territory), and the sample's counts are scored under each projected profile by multinomial likelihood (sigscape-core's profile_loglik). Scoring by likelihood at the sample's own count puts the noise of a small count into the comparison instead of resampling the references. The 15 most likely tumors vote.

A PCAWG tumor mapped this way finds itself among its neighbors: the studio's references are all of PCAWG.

How well

Evaluated on held-out tumors: references from one half of PCAWG, queries from the other half as the bundle projects them onto each assay (their own mutations inside the territory), downsampled to each count. Top-1 tumor type, 27 types:

| Mutations | 10 | 20 | 50 | 200 | 1,000 | |---|---|---|---|---|---| | genome | 24% | 35% | 44% | 59% | 65% | | exome, matched / compared with genomes | 23% / 22% | 31% / 28% | 41% / 38% | 60% / 57% | 72% / 72% | | panel, matched / compared with genomes | 33% / 30% | 53% / 53% | 71% / 71% (34 queries) | | |

Matching the references to the assay helps a little at low counts and never hurts; few PCAWG tumors carry 50 or more mutations in a panel's targets, so the panel rows above 20 rest on few cases (hypermutated tumors).

The vote share is calibrated: on the same held-out tumors, over all counts, a vote of 80% or more named the right type 77% (exome), 86% (panel) and 87% (genome) of the time; 60–80%, 38–53%; 40–60%, 22–29%. The studio prints the figure for the vote it shows.

HR status, transferred

HR status transfers from neighbors too. Leaving one PCAWG breast tumor out at a time (208 with an HR call, 58 deficient), the share of HR-deficient tumors among its 15 nearest labeled neighbors separates deficient from proficient with an AUC of 0.85 at 20 mutations and 0.95 at 50 from the genome, 0.84 and 0.91 from the exome, and 0.73 at 5 mutations in a panel's targets (119 tumors with that many; a selection toward more mutated tumors). The HR calls come from the lab's breast list and are not in the site's reference files, so the studio does not show this transfer yet.

Not yet

Genome doubling, timing and topotype mixes transfer the same way once the bundle carries them. Map every sample maps a whole cohort and exports the calls as a table; the per-assay embedding is not drawn yet.