Refit confidence
A refit gives every signature a number. Refit confidence says which of those numbers to believe, at this assay and this mutation count. Open it on any SBS-96 spectrum.
What each row says
The spectrum is fitted with COSMIC v3.2 (MuSiCal's catalog, the one the calibration below was built in), corrected to the sample's assay, by non-negative least squares with 200 bootstrap refits at the sample's own count. For each signature with a share:
-
Share and 90% interval, from the bootstrap.
-
Detected when the share is at least 5% and the 5th percentile of its bootstrap share is at least 1%.
-
Found when present and called when absent: how often that rule finds this signature in samples that carry it, and calls it in samples that do not, at this assay and about this count. These come from the calibration grid.
-
Nearest co-active signature: the most similar other signature with a share, their profiles' cosine, and how their bootstrap shares move together. Two similar signatures whose shares trade off (cosine ≥ 0.8, correlation ≤ −0.5) are a trade-off: the split between them is not determined at this count.
-
Right when called: the call's positive predictive value, from the detection rate, the false-positive rate and how often PCAWG tumors carry the signature at this assay and count.
The verdict is one of confident, trade-off, possible false positive (right when called less than 80% of the time, never carried by PCAWG at this assay and count, or called in more than 5% of samples without it), uncertain (5% or more, interval reaching zero) and minor. On the benchmark, unflagged calls were right 94–99% of the time at every assay and count (see The refit benchmark).
Fit all of COSMIC, or a tumor type's signatures
Choosing a PCAWG tumor type fits only the signatures that take at least 5% in at least 5% of that type's genomes. The summary line says which choice calls fewer absent signatures at this count. Restricting to the type's signatures roughly doubles how often a present signature is found at low counts:
| Mutations | 25 | 50 | 100 | 200 | 500 | 1,000 | 2,000 | 5,000 | |---|---|---|---|---|---|---|---|---| | found when present, all of COSMIC (genome) | 6% | 14% | 23% | 34% | 53% | 65% | 75% | 88% | | found when present, tumor type's signatures | 19% | 31% | 46% | 57% | 77% | 81% | 85% | 89% |
These are means over signatures; each signature's own rate is in its row.
The calibration grid
Synthetic samples are real PCAWG exposure profiles (MuSiCal's, 2,703 tumors), projected onto each assay's territory and drawn at 25 to 5,000 mutations, then refitted with the same core code the page runs. Truth in the territory: present when a signature's share is at least 5%, absent when MuSiCal gave it none. The grid is built on half of the tumors (split by a hash of the aliquot id, at most 40 per tumor type); the other half is held out for the benchmark. Rates informed by fewer than 10 cases show "–". Building it takes about a minute on 14 cores (sigscape-reference/evaluate/calibration.py).
Concordance on the runner
With the local runner connected, Refit with MuSiCal and SigProfilerAssignment runs both packages on the same spectrum with COSMIC v3.2 (exome signatures for an exome in SigProfilerAssignment). Their shares join the table, and Methods detecting counts how many of the three put the signature at 5% or more. Neither package corrects for a panel's territory. Both runs are kept and chosen like any other runner step.