A practical guide for mutational signature analysis in hematological malignancies

Abstract

Analysis of mutational signatures is becoming routine in cancer genomics, with implications for pathogenesis, classification, prognosis, and even treatment decisions. However, the field lacks a consensus on analysis and result interpretation. Using whole-genome sequencing of multiple myeloma (MM), chronic lymphocytic leukemia (CLL) and acute myeloid leukemia, we compare the performance of public signature analysis tools. We describe caveats and pitfalls of de novo signature extraction and fitting approaches, reporting on common inaccuracies: erroneous signature assignment, identification of localized hyper-mutational processes, overcalling of signatures. We provide reproducible solutions to solve these issues and use orthogonal approaches to validate our results. We show how a comprehensive mutational signature analysis may provide relevant biological insights, reporting evidence of c-AID activity among unmutated CLL cases or the absence of BRCA1/BRCA2-mediated homologous recombination deficiency in a MM cohort. Finally, we propose a general analysis framework to ensure production of accurate and reproducible mutational signature data.

Fig. 1

Mutational signature de novo extraction vs. fitting. a, b The Alexandrov et al. NNMF framework^,. From the 96-mutational classes, NNMF extracted the signatures’ relative (a) and absolute (b) contribution among 30 MMs. c Running deconstructSig including all 30 COSMIC signatures several mutational processes were forced to be extracted (i.e., Signature 4). Furthermore, the new mutational process MM1 was not detected, being not included in the 30 COSMIC signatures. d Conversely, running the same fitting approach based on the mutational signature catalog extracted by NNMF, each signature contribution was better estimated for each patient. Sig. = signature

Fig. 2

HRDetect BRCA1/BRCA2 deficiency scores in MM. HRDetect was used to analyze the BRCA1/BRCA2 deficiency scores in MM samples a including only signature 8, b including both signatures 3 and 8, and c including only signature 3. In d, the same analysis was performed in 15 BRCA null and 15 BRCA wt breast cancers. Scores are ordered from highest to lowest and a classification threshold of 0.7 is used to classify samples as HRD-positive (see Davies et al.). Below each score, the contribution of the six features that are used by HRDetect is shown. Each contribution is given by the amount of a feature in a sample, log-transformed and standardized according to mean and standard deviation of the features in Davies et al. and finally multiplied by the corresponding HRDetect logistic regression coefficient. Thus, a positive contribution indicates a feature value higher than the average of the HRDetect original training set, and feature contributions are directly comparable. Sig. = signature

Fig. 3

Absence of BRCA-driven HRD in MM. a Pie charts showing the relative signature composition according to DeconstructSig in three MM cases, without a prior knowledge of which signatures are involved or detected by NNMF. Testing all 30 COSMIC mutational signatures, Signature 3 is extracted is all samples. b Circos plot of three MMs (ND = newly diagnosed; RR = relapsed/refractory; SMM = smoldering MM) where deconstructSig extracted a significant Signature 3 contribution. From the external ring to the internal: mutations, (vertically plotted according to their inter-mutational distance and where the color of each dot represents the mutation class), indels (dark green = insertion; and brown = deletion); copy number variants (red = deletions, green = gain), rearrangements (blue = inversion, red = deletions, green = ITD, black = translocations). PD26419a is the only patient with a slightly high HRDetect score when analyzed including Signature 3. c Circos plots of a breast cancer sample without BRCA deficiency (PD4069a), one with BRCA1 deficiency (PD6413a) and one with BRCA2 deficiency (PD4954a). The MM genomic landscape shows significant differences to the two BRCA-deficient breast cancers, in particular in terms of numbers of indels and SVs, suggesting BRCA-driven HRD is not present in the MM samples analyzed

Fig. 4

Mutational signature landscape of immunoglobulin loci. a The 96-mutational classes of all SNV within IGH/IGK/IGL loci. Canonical AID (c-AID) represented the main mutational process within these regions in all tested hematological malignancies, including U-CLLs as recently described^,,. b, c Mutational signature relative (b) and absolute (c) contribution within IGH/IGK/IGL loci for each sample tested by deconstructSig. d The Sanger-sequencing-based IGHV mutational status available for each CLL case. Sig. = signature

Fig. 5

Kataegis in hematological malignancies. a Example of a MM patient with a chromothripsis on chromosome 6 associated with APOBEC-mediated kataegis. The solid and dashed lines reflect the total ploidy and the copy number status of the minor allele, respectively. In these plots, the red arch represents a deletion, the green arch represents a tandem duplication and the blue arch represents an inversion. b Inter-mutational distance of all mutations in chromosome 6, color-coded by mutational class. c Ninety-six-mutational classes of all kataegis events on chromosome 6. d Chromothripsis event on chromosome 19 in a therapy-related AML. e Inter-mutational distance of all mutations across chromosome 19. f Ninety-six-mutational classes of all mutations involved in the chromosome 19 kataegis: APOBEC emerged as the dominant mutational process, despite its activity was not detectable across the genome (Supplementary Software File 3)

Fig. 6

Bleeding of signatures in CLLs. Summary of mutational signature analysis on 146 CLL cases. From the 96-mutational catalog (a) the Alexandrov et al.^, framework (NNMF) extracted different mutational processes. Signature 9 (nc-AID) was extracted also among U-CLL in contrast with their known pathogenesis (b). This is a typical example of inter-sample bleeding and it can be solved either running a fitting approach after the initial NNMF analysis using only the catalog of signatures extracted by NNMF (c), or analyzing M-CLL and U-CLLs in two different and independent runs (d, e). Using the 30 COSMIC signatures as reference, the first approach is usually the most appropriate in order to estimate the real contribution of each single mutational process. In fact, the NNMF extracted signatures may be over or under split, therefore preventing a precise estimation of their contribution. For example, in this analysis, Signature 1 and 5 were extracted as one single process and only by running a fitting approach we were able to differentiate these two processes (c). Sig. = signature. In b and c, red patient labels are used for U-CLL, green for M-CLL, and blue for unknown cases

Fig. 7

Bleeding of signatures in AMLs. Example of inter-sample bleeding among 52 AML WGSs. a, b Running NNMF on the entire cohort, we extracted two mutational signatures not currently included in COSMIC: one recently associated with platinum exposure and the second recently reported as a process specific to the hemopoietic stem cell (HPSC). c, d The inclusion of two t-AMLs (PD34280 and PD37515) affects the global signature extraction, with Platinum Signature extracted also in the primary AMLs. Removing the t-AMLs the inter-sample bleeding was corrected, and no Platinum Signature was extracted in primary AMLs. Sig. = signature

Fig. 8

Mutational Signature workflow. Our suggested workflow for mutational signature analysis for both genome-wide and clustered processes (a) and an example of its application on 30 MM WGSs (b)