Fig 1.
A) Schematic of the SparseSignatures method. N represents the number of tumors in the dataset, K the number of signatures. B) Background signature derived from COSMIC SBS5. C) Background signature derived from the human germline mutation spectrum. Vertical bars represent the probability of mutation in each of 96 categories. These are based on 6 possible mutation types (upper gray labels) and 16 possible combinations of 5’ and 3’ flanking bases (x-axis labels). Source data are provided in S1 Table.
Fig 2.
Comparison between SparseSignatures and other methods on simulated data.
A) Bar and line plot showing, for each method, the number of simulations in which it selected each value of K (number of signatures). The x-axis shows values of K and the y-axis shows the number of times each value was selected. Each method was run on 50 simulated datasets. In all cases, the correct value of K was 4. B) Box plots showing the residual error for the solutions produced by each method, over 50 simulations. Residual error was measured as the mean squared error (MSE) in reconstructing the original count matrix. C) Box plots showing the fraction of variance in the count matrix explained by the solutions produced by each method, over 50 simulations. D) Box plots showing the cosine similarity in reconstructing the 3 non-background input signatures, over 50 simulations. E) Box plots showing the mean squared error in reconstructing the exposure values for the 3 non-background input signatures, over 50 simulations. F) Box plots showing the sparsity of the signatures produced by each method, over 50 simulations. Sparsity was measured as the fraction of cells in the signature matrix whose value is <10−3. SS: SparseSignatures. SP: SigProfiler. SA: SignatureAnalyzer. SR: signeR. SS-BG: SparseSignatures without fixed background. Source data are provided in S2 Table.
Fig 3.
A) The 9 mutational signatures obtained by applying SparseSignatures to a dataset of 147 pancreatic tumors. We report the number and correlation of the most similar (correlation higher than 0.70) corresponding signature from COSMIC (https://cancer.sanger.ac.uk/cosmic/signatures). Source data are provided in S13 Table. B) Boxplots showing fitted values for exposure to each of the 9 signatures obtained by SparseSignatures for the 147 pancreatic tumors. Boxplots represent the fraction of mutations per tumor (on the y-axis) contributed by the given signature (on the x-axis). Source data are provided in S14 Table. C) Relapse-free survival analysis of patients belonging to the 10 clusters. D) Clustering of patients based on their exposure values. Boxplots show the fraction of mutations per tumor contributed by each signature (x- axis) to each of 10 clusters. Source data are provided in S19 Table.
Table 1.
Signatures (including background) discovered by SparseSignatures in real cancer data, and their proposed etiology.
Table 2.
Comparison of signatures predicted by four signature discovery methods on real tumor sequencing data.
Sparsity is measured as the fraction of cells in the signature matrix with value < 10−4. Cross-signature similarity is measured as the mean cosine similarity between all pairs of predicted signatures. Background contamination is measured as the mean cosine similarity between the background signature and all non-background predicted signatures. Median per-patient correlation is measured as the median Pearson’s correlation coefficient between the observed mutation spectrum and the predicted mutation spectrum for each patient, indicating how well each method fits the observed mutations in individual patients.
Fig 4.
A) The 8 mutational signatures obtained by applying SparseSignatures to a dataset of 560 breast tumors. We report the number and correlation of the most similar (correlation higher than 0.70) corresponding signature from COSMIC (https://cancer.sanger.ac.uk/cosmic/signatures). Source data are provided in S23 Table. B) Boxplots showing fitted values for exposure to each of the 8 signatures obtained by SparseSignatures for BRCA1/2-mutated and non-mutated tumors. Boxplots represent the fraction of mutations per tumor (on the y-axis) contributed by the given signature (on the x-axis). Source data are provided in S24 Table.