Choose the analysis and check your panel
Match your question to the signature, HRD or MMRD/POLE analysis, and check that your files and panel can support it.
Before you begin, you need a description of your data: file format, panel, tumor type, genome build.
Match the question to the analysis
| Your question | Analysis | The call | Guide |
|---|---|---|---|
| Does this tumor carry Signature 3, APOBEC, SBS8, SBS18, UV or another signature of the catalog? | sg2.run(soi="SBS3"), or another signature | a score, a call at a 10% and at a 1% false-positive rate, a confidence word | 2. Detect a signature |
| Is this tumor homologous-recombination deficient? | sg2.run(soi="HRD") | HRD, HRD-low or HRP | 3. Detect HRD |
| Is mismatch repair deficient, or is POLE mutated? | sg2.predict_mmrd_ppd() | MMRD, POLE or MMRP | 4. Detect MMRD or POLE |
| Which signatures are in the spectrum at all, without a call? | sg2.run(mva=False) | the decomposition's exposures per signature | 2. Detect a signature |
| The same result as SigMA in R | sg2.run(model="legacy") | SigMA v1's Signature 3 score and call | Reproduce SigMA v1 |
A Signature 3 call and an HRD call answer different questions and rest on different readings of the tumor; neither replaces the other. What a call means states what each supports and what it does not establish.
Check what each analysis reads
| Analysis | Reads | Needs in the input | Optional |
|---|---|---|---|
| A signature | the SNV spectrum, 96 channels | SNVs: a MAF, VCFs or a 96-channel matrix | nothing |
| HRD | a Signature 3 score, deletions of five or more bases, copy-number scars | SNVs and indels: a MAF or VCFs, with the reference genome for the deletions' microhomology; a matrix only with mmej and nhej columns | copy-number segments (allele-specific, or a panel's log-ratio .seg), without which the variant trained without copy number runs; the repeat-locus files, for the MMRD/POLE step that runs first |
| MMRD or POLE | the spectrum's mismatch-repair and POLE content and the indels at microsatellites | SNVs and indels: a MAF or VCFs, and the repeat-locus files; a matrix only with indel counts in it | nothing |
The repeat-locus files are installed once, with sigma2 data fetch --all for a GitHub account the maintainers have given access while the SigMA2 repository is private, or from the public SigMA (R) repository; without them MMRD/POLE runs silently without its microsatellite counts and HRD without its MMRD/POLE step. Indels are read from a MAF's INS and DEL rows and from a VCF's PASS records; a matrix carries none, so HRD and MMRD/POLE on a matrix need the indel counts as columns (1. Prepare your input files).
Check that your panel has a classifier
A classifier exists per cell: tumor type, signature, platform. The platform is the capture design the tumors were sequenced with:
seq_type | Platform |
|---|---|
msk | MSK-IMPACT |
fo | FoundationOne |
op | OncoPanel (v3) |
seqcap | SeqCap EZ Exome v3 |
tcga_mc3 | TCGA MC3 exome (GENCODE v19 basic) |
wgs | whole genome |
A panel yields far fewer SNVs than an exome: on the example files, the fifty MSK-IMPACT panels have a median of 8 (30 of them fewer than ten), and the six TCGA exomes 33 to 169.
A panel that is not one of these has no classifier of its own: score it as the bundled platform closest in design and size, and name that choice in the report. seq_type="auto" picks, among the platforms with a classifier for the tumor type and signature, the one whose training cohort's median SNV count is closest to your cohort's, and reports the choice; it cannot tell panels of similar size (msk, fo, op) apart. Bundled classifiers lists every cell with its status; sg2.describe() prints the same tables.
Check that the tumors clear the floor
A tumor below the floor is not scored: five SNVs for most tumor types, four for prostate, three for osteosarcoma and pancreatic neuroendocrine tumor. A tumor that clears the floor can still sit below the limit of detection when few of its mutations come from the signature; when a call is possible explains the count that decides, and check what the panel delivered measures it on a cohort.
Check the genome build
SNV spectra are read on hg19 or hg38 (ref_genome=); a MAF's NCBI_Build is checked against it and a mismatch is refused. The repeat loci behind the microsatellite and deletion features, the msk HRD model's gene grid and the panel regions used in training are hg19 only; for hg38 indels pass repeat-locus files of that build with beds= (Prepare your input files).
Check the tumor type
tumor_type sets the signature set, the reference spectra and the classifier; the vocabulary lists the 39 names. A tumor type with no cell for your signature and platform is refused with the platforms that have one; HRD refuses a few tumor types it saw in training but has no Signature 3 signature set for (Detect HRD).