Fig. 1

Characterising local mutational processes with RM2. a Method overview. RM2 studies a set of genomic elements (i.e. sites) and somatic mutations in cancer genomes using a negative binomial regression model. Sites of constant genomic width (dark grey) and two control flanking sequences (light grey) are used (isSite). Sites and flanks are collapsed into unique nucleotides and grouped to ten bins using their megabase-scale mutation frequency (MbpRate). Mutations in sites and flanks (nMut) are grouped by trinucleotide type (triNucMutClass). Trinucleotide content corresponding to the potential genomic space for mutations is used as model offset (nPosits). Log-likehood tests are used to compare the mutation frequencies in sites and flanking regions by removing the model factor isSite. The optional factor coFac enables interaction analysis of genetic and clinical variables. b QQ-plot shows the observed and expected P values of true and simulated mutations from PCAWG. No significant signals were identified in simulated data (FDR < 0.05), indicating that our method is well-calibrated. c Comparison of model performance with and without MbpRate covariate. Analysis of true (left) and simulated mutations (right) shows the advantage of modelling megabase-scale mutation frequency. d Power analysis of RM2 using down-sampling of CTCF binding sites and liver cancer genomes. Fraction of significant results (left) and median P value (right) are shown. Panels b and c include total mutations, strands- and signature-specific mutations as in Fig. 3. Only total mutations were included for analyses in c, d