Extended Data Fig. 4. Extraction and assignment of mutational signatures.

a–d. Signatures extracted by the Hierarchical Dirichlet Process (HDP) (a), SignatureAnalyzer (b), SigProfiler (c), and sigfit (d). COSMIC signatures needed to reconstruct the extracted signatures are shown together with their contribution (in percentage). The cosine similarities between the extracted and reconstructed signatures are shown in brackets. e. Workflow of the mutational signature analysis. f. The 96-mutation profile of the RT sample of patient 839 (time point 2), which had marked evidence of APOBEC activity (SBS2 and SBS13). g. Comparison of the SBS-ganciclovir extracted by HDP and SignatureAnalyzer. Based on the high cosine similarity (0.996), we considered that both signatures represented the same mutational process and selected the one extracted by HDP for downstream analyses. h. Comparison of the SBS-ganciclovir extracted by HDP and the ganciclovir signature reported by de Kanter et al.³⁵. i. Comparison of the SBS-RT extracted by HDP and SignatureAnalyzer. Based on the high cosine similarity (0.941), we considered that both signatures represented the same mutational process and selected the one extracted by HDP for downstream analyses. j. Pairwise comparisons of the SBS-RT with known signatures from COSMIC and Kucab et al.³³. k. Decomposition of the SBS-RT in “n” known signatures using an expectation maximization approach. The low cosine similarity (<0.85) between SBS-RT and the best reconstituted signature obtained using any combination of known signatures suggests that SBS-RT represents a novel mutational signature.