Fig 1.
The SGZ pipeline is overviewed in panel A. Key components include fitting an optimal copy number model to the genome‐wide log‐ratio and minor allele frequency profiles (B), and modeling the expected allele frequencies of germline, somatic, and subclonal somatic mutations (C). In panel B, the dots in the top panel correspond to log ratios at each exon sequenced, segmented and fitted to discrete copy number levels, while the dots on the bottom panel are germline SNP minor allele frequencies. In panel C, examples of expected variant allele frequencies are shown for various scenarios of copy number and tumor purity. The expected allele frequencies are shown for germline (blue), somatic (red), and subclonal somatic (yellow).
Fig 2.
Copy number detection overview.
Aligned DNA sequences of the tumor specimen are normalized against a process‐matched normal, producing log‐ratio and minor allele frequency (MAF) data. Next, whole‐genome segmentation is performed using a circular binary segmentation (CBS) algorithm on the log‐ratio data. Then, a Gibbs sampler fitted copy number model and a grid‐based model are fit to the segmented log‐ratio and MAF data, producing genome‐wide copy number estimates. Finally, the degree of fit of candidate models returned by Gibbs sampling and grid sampling are compared and the optimal model is selected by an automated heuristic.
Table 1.
Validation of somatic and germline predictions.
Fig 3.
Breakdown of no-calls made by SGZ.
Reasons behind no-calls made by SGZ are shown for (left) all variants in 30 lung and colon samples and (right) 17 somatic hotspot mutations and 20 common germline variants within 20,182 clinical samples.
Table 2.
SGZ performance as a function of tumor purity in the cell line dataset.
Table 3.
Tumor zygosity predictions of somatic mutations in 20,182 clinical samples.
Table 4.
Likely somatic status mis-annotation in COSMIC, predicted by SGZ to be germline in multiple samples in Foundation Medicine sample set†.