SigMA2 extends panel-based signature detection beyond SBS3. The signature of interest is a parameter of the model, so one framework returns a calibrated probability of presence for a broad set of COSMIC signatures; germline-aware training lets it run on tumour-only panels with no matched normal; and an indel module calls mismatch-repair and POLE status from the same data.
Targeted panels such as MSK-IMPACT, FoundationOne CDx and OncoPanel read one to three megabases of the genome, an order of magnitude fewer mutations per sample than whole-genome sequencing. At those counts, NMF recovers only one or two components, NNLS exposures are unstable, and cosine similarity ignores how many mutations actually support the call. SigMA showed a trained classifier can still detect SBS3 there. SigMA2 extends that in four directions.
Any signature, not just SBS3
The signature of interest is a model parameter. A classifier per signature, tumour type and assay returns a calibrated probability that the signature is present, including etiology groups scored on summed exposure: APOBEC (SBS2 and SBS13) and ultraviolet (SBS7a-d).
Tumour-only panels
Training injects Poisson-distributed germline variants into the panel-downsampled data, so the classifier tolerates residual germline contamination and can be applied to assays run without a matched normal.
Mismatch repair and POLE from indels
An indel module combines microsatellite indel counts with the ratio of deletions to insertions, separating mismatch-repair-proficient, mismatch-repair-deficient and POLE-mutant tumours from the same panel data.
Python, with v1 kept
Every model is trained natively in Python. The original SigMA v1 models are bundled and reimplemented, so earlier analyses reproduce without an R installation.
Trained classifiers ship with the package, so nothing has to be trained before a run. The models bundled in the current release cover breast (SBS3, SBS8, SBS18, APOBEC) and skin (SBS7) across five panel and exome platforms. For a combination that isn't covered, the Add workflow annotates, clusters and simulates your own cohort and tunes a classifier from it.
pip install --pre sigma2