Main

Chronic lymphocytic leukemia (CLL), the most common adult hematological malignancy in Western countries, is characterized by diverse treatment outcomes even in the era of targeted agents. The full complement of genomic events contributing to this clinical diversity have yet to be determined. Thus far, only mutations in TP53 influence clinical practice1,2,3,4,5,6,7. Other prognostic markers, including the immunoglobulin heavy chain variable (IGHV) region mutational status, and existing molecular classifications have limited predictive value in individual patients7,8,9,10.

Previous sequencing studies of CLL have focused largely on mutations affecting protein-coding genes7,8,9,10,11,12,13, and whole-genome sequencing (WGS) has been reported for only a small number of CLL patients, mostly with low-risk disease1,2,3,4,5,6. Hence, the association between clinical parameters and genomic alterations has largely been restricted to driver coding mutations and copy number changes.

Here, to provide the largest and most comprehensive analysis of the entire genomic landscape of CLL and its relationship to clinical outcome, we performed WGS of 485 clinical trial patients recruited to the United Kingdom’s 100,000 Genomes Project. The results of our study provide additional insights into coding and noncoding single nucleotide mutations. We then exploit WGS data to provide a detailed map of structural alterations and global features, including telomere length, mutational signatures and genomic complexity (GC). Finally, we integrate the different modes of genetic alterations to define five genomic subgroups (GSs) of CLL and relate these to clinical outcome. Our results provide a springboard to indepth functional validation of putative drivers and our integrated genome-wide approach could, after independent clinical validation, refine current clinical outcome prediction.

Results

We performed WGS of tumor and matched normal samples from 485 patients with treatment-naïve CLL enrolled in clinical trials to a median depth of 109× and 36×, respectively (Supplementary Tables 1–3). A second tumor sample was available for a subset of 25 patients at relapse. In addition, RNA sequencing (RNA-seq; n = 73) and assay for transposase-accessible chromatin with high-throughput sequencing (ATAC-seq; n = 24) data were generated for a subset of CLL samples with recurrent noncoding mutations (Supplementary Table 4).

Coding mutations and structural variants

We initially identified putative coding drivers by (1) screening for genes impacted by single nucleotide variants (SNVs) and small insertion/deletions (indels) and (2) integrating SNV/indels with copy number alterations (CNAs) (Fig. 1a; Methods). We identified 36 known and 22 putative driver genes (Fig. 1b and Supplementary Fig. 1), which were not found associated with CLL in the literature and also not prevalent above 1% in two landmark genomic studies in CLL3,7. These were classified as previously unknown putative drivers and included the immune checkpoint regulator IRF2BP2 (4.3%) (Supplementary Table 4).

Fig. 1: Identification of coding mutations and structural variants.
figure 1

a, Methodology used for the discovery of candidate coding drivers. Discovery method 1A selected genes with a FDR below significance threshold for two out of four algorithms. Discovery method 1B combined the P values of four algorithms using weighted Stouffer and weighted Harmonic mean. Genes with FDR below significance threshold for at least one result were selected. With discovery method 2, CNAs were used to define minimally affected regions (by copy number loss or gain). Then, genes included in these genomic regions were selected as candidate drivers if they presented at least five SNVs/indels impacting the coding sequence focality and recurrence scores greater than threshold and mechanism of action of gene in agreement with the type with CNAs (loss for TSG and gain for oncogenes). An additional list, not considered as candidate drivers, included genes fulfilling all requirements except the SNVs/indels count threshold. (Permissive list; see Methods for more details). b, Number of SNVs/indels (left axis) and proportion in the cohort (right axis) of the 58 candidate drivers. Other CLL cohorts used as comparators are described in (Supplementary Table 5; Methods). c, The 76 regions recurrently affected by CNAs. The y axis is shown in log10 scale. Known CLL drivers are indicated in blue and putative driver genes identified as hotspots are indicated in yellow. d, Candidate drivers found by integrating both CNAs and SNVs/indels. The score represents combined focality and recurrence scores derived from MutComFocal, integrating SNVs/indels data with CNA data (Discovery method 2; Methods); NS, not significant. Known CLL drivers are indicated in blue and putative driver genes identified as hotspots are indicated in yellow. e, All translocation breakpoint pairs found in more than three samples (out of 495), including those occurring in coding and noncoding regions.

Source data

Full size image

We identified 74 regions of the genome that were recurrently affected by CNAs in at least four samples (Fig. 1c, Extended Data Fig. 1a and Supplementary Table 6). Using DNA microarray data, 85% of CNAs could be validated (Supplementary Table 7). In addition to 14 well-known CNAs, including del13q14.2, del11q22.3 and del17p13.1, we identified a further 60 regions, of which 27 were previously not recognized. The breakpoints of the remaining 33 CNAs could be refined to a smaller minimally overlapping region14,15,16,17,18,19. By combining SNVs/indels with CNAs (discovery method 2; Methods), we predicted the most likely target gene for nine known regions, including TP53/del17p13.1, and seven additional regions including PCM1/del8p, IRF2BP2/del1q42.2q42.3 and SMCHD1/del18p11.32-p11.31 (Fig. 1d, Extended Data Fig. 1b,c and Supplementary Table 8). We also found 66 additional genes affected by recurrent CNAs using more permissive criteria (Methods). While these are potentially interesting, they were not considered to be putative CLL drivers and were not taken forward for downstream analyses (Supplementary Table 9).

A major advantage of WGS is the power to identify inversions and translocations. We identified 1,248 inversions (Extended Data Fig. 2a; Methods) with frequent breakpoints involving either the immunoglobulin light chain kappa (IGK) locus (n = 65, 13.4%), the immunoglobulin heavy chain (IGH) locus (n = 65, 13.4%) or chr13q14.2 (n = 40, 8.7%) (Extended Data Fig. 2b and Supplementary Tables 10 and 11). We detected 993 translocations, of which two occurred in more than ten samples and affected known genes with no previously documented role in CLL, including t(14;22) with breakpoints within WDHD1 (n = 12, 2.6%) and t(5;6) (CTNND2-ARHGAP18, n = 11, 2.4%) (Fig. 1e and Extended Data Fig. 2c).

The 22 potential coding driver genes were altered by truncating mutations or also affected by CNAs (Fig. 2a, Extended Data Fig. 3a–d, Supplementary Table 12 and Supplementary Figs. 2 and 3). Most mutations occurred in protein domains, and 62% of mutations were detectable in more than 50% of tumor cells (median cancer cell fraction (CCF) ≥0.5) and 89% in at least 20%. All previously unreported CNAs for which we could predict a target gene(s) were also clonal (median CCFs ≥0.8) (Fig. 2b and Extended Data Fig. 3e). Candidate driver mutations affected multiple biological pathways including the DNA damage/cell-cycle and RNA-ribosome processing (Fig. 2c).

Fig. 2: Biological features of coding mutations and CNAs.
figure 2

a, Annotations of genes. CNAs, presence/absence of CNAs affecting the gene; COSMIC, proportion of variants reported in the COSMIC database; High impact, proportion of nonsense variants, Median CCF, median cancer cell fraction of variants; Prot domain: proportion of variants occurring in a protein domain from the Prot2HG database40. b, Distribution of cancer cell fractions in selected recurrent regions of CNAs (all regions in Extended Data Fig. 3e). The boxplot shows the minimum and maximum values and the interquartile range. c, Candidate drivers classified in ten main pathways described in CLL3,7. Genes in bold are present in more than 3% and genes in red font are candidate drivers. Other drivers are absent because not involved in these ten main pathways. d, Detection of variants of interest (n = 118) by RNA-seq (with minimum depth of five) in selected 73 samples. Difference of variant allele frequency (VAF) between RNA-seq methods and WGS methods shows allelic skew of variants. Ratio of expression in transcript per million (TPM) in sample with variant against all other samples reflects change in gene expression. Selected variants annotated with gene names, all data in Supplementary Table 13. DP, depth. e, Enrichment of genomic features in different CLL subgroups using two-sided Fisher’s exact test (plot showing the median, minimum and maximum values). The groups were (1) stage: relapsed/refractory (R/R), versus frontline (N = 443 frontline versus 30 R/R), (2) TP53: altered versus WT (N = 420 WT versus 65 disrupted), (3) IGHV mutational status: unmutated versus hypermutated (N = 197 hypermutated versus 288 unmutated), where an enrichment for the former is indicated by an odds ratio greater than one. Adjusted P values (FDR) are shown.

Source data

Full size image

Performing RNA-seq on representative CLL samples from 74 patients with known and potential coding mutations (for 40 of the 58 drivers, n variants = 118, Supplementary Table 4; Methods), we validated the expression of 73% of variants at the RNA level (Extended Data Fig. 4a and Supplementary Table 13). As expected, most (29/43) mutations that were either not detectable or were seen at low expression levels were truncating mutations consistent with nonsense-mediated decay (Supplementary Table 13). Additionally, allelic skewing and/or a reduction of mutant transcript expression compared with the mean expression of wild-type (WT) transcripts across the cohort was shown, notably for specific mutations in SPEN, SETD2, TP53 and IRF2BP2 (Fig. 2d). When considering all mutations, significantly reduced gene expression was demonstrated for TP53, ATM and SETD2 (refs. 20,21) (Extended Data Fig. 4b).

When we associated the 36 known and 22 putative drivers and regions of CNAs with other biological variables such as disease stage, TP53 alterations, IGHV mutation status (unmutated, u-IGHV; and hypermutated, m-IGHV) and stereotyped B cell receptor immunoglobulin subsets (BCR IG) including IGHV3-21 usage (Fig. 2e and Supplementary Table 14; Fisher’s exact test, false discovery rate (FDR) < 0.05), we found that SETD2/del3p21.31, del9p21.3 and gains of chr17q21.31 were associated with relapsed/refractory (R/R) disease and TP53 disruption, whereas MED12 and DDX3X mutations were associated with u-IGHV CLL. BCR IG subset 2, representing about 3% of all CLL, and known to be associated with poor prognosis22, was linked to the putative driver FAM50A. The IGHV3-21 rearrangement was also enriched for FAM50A and for ATM/del11q22, SF3B1 mutations and chr21q21.3-q22.3 gains.

Association of coding mutations with disease evolution

We examined the relationship between recurrent gene mutations and disease evolution in three different cohorts (Fig. 3a and Supplementary Table 4; Methods): (1) unpaired frontline-treated versus R/R (main cohort, unmatched, n = 443 versus 30—excluding the 12 early CLL); (2) paired samples from the CLL and Richter’s syndrome (RS) phases of the same patient (previously published cohort23, matched, n = 17) and (3) a second sample taken from a subset of the 485 patients at relapse who had already been profiled before frontline treatment: paired frontline-treated versus relapsed (main cohort, matched, n = 25/485).

Fig. 3: Associations of coding mutations and CNAs with disease progression.
figure 3

a, Three cohorts used for studying the presence of variants during disease evolution. Unpaired samples are taken from different patients; cohort (1) were samples from treatment-naïve patients and R/R patients; cohort (2) were paired samples of CLL and RS phase of the same patient; cohort (3) were paired samples taken at two different timepoints before treatment and at relapse. b, Distribution of cancer cell fractions in the three cohorts studied for selected genes. For cohort (1), figures are not shown if no R/R sample carried a mutation. Other genes are presented in Extended Data Fig. 4e,f. Boxplots showing results for unpaired samples and connected datapoints show results for paired samples (corresponding variants are connected by a dotted line). An asterisk indicates a candidate driver. c, Genomic features linked to patients’ PFS (left panel) and OS (right panel). Hazard ratio and FDR of each genomic feature tested against PFS using a Cox proportional-hazards model on the subset of patients for which clinical outcomes data were available (n = 243). Adjusted P values (FDR) are shown in different colors. (See Supplementary Table 14 for the full detailed list of genomic features tested and Supplementary Table 15 and 16 for full results of the statistical tests). d–f, candidate driver IRF2BP2 was recurrently affected by CN losses (d) and SNVs/indels, especially truncating ones (e), and was associated with increased CCF in variants for more advanced disease stage (f). Coloured rectangle in (e) represents protein domains. g–j, candidate driver SMCHD1 was recurrently affected by CN losses (g) SNVs/indels, especially truncating ones (h), presented evidence of increased CCFs in more advanced CLL in cohort (2) (no data available in R/R of cohort (1)) (1), and associated with more adverse overall survival as shown in the Kaplan–Meier plot where shaded areas show the 95% confidence intervals and P values were derived from a log-rank test (j). Coloured rectangle in e represent protein domains. Boxplots show the minimum and maximum values and interquartile range and each individual variant is represented with an individual datapoint.

Source data

Full size image

Recurrent coding gene mutations were linked to disease evolution in all three cohorts. They presented higher mutation counts and frequency in the RS compared with the CLL phase (P = 2.1 ×10−2; Extended Data Fig. 4c,d) and higher CCFs at the more advanced stages with a median CCF > 0.8 (Fig. 3b and Extended Data Fig. 4e–g).

Restricting analysis to patients with information on long-term survival outcome (n = 243 / 485), 13 known or putative drivers and recurrent CNAs were significantly associated with progression-free survival (PFS) and 11 with overall survival (OS) (FDR < 0.05) (Fig. 3c and Supplementary Tables 15 and 16).

Out of the 22 putative drivers, 21 were also related to disease progression (Extended Data Fig. 4c–f), including two of the most commonly mutated ones. IRF2BP2 (interferon regulatory factor 2 binding protein 2), located in the minimally deleted region of chr1q42.3 (Fig. 3d) was also affected by deleterious mutations and CNAs (Fig. 3e) (in total, n = 28/485, 5.8%) with high CCFs (Fig. 3f, left panel). Mutations showed evidence of clonal expansion in more advanced disease (Fig. 3f, right panel) and altered RNA expression (Fig. 2d). This gene contributes to the differentiation of immature B-cells and is associated with a familial form of common variable immunodeficiency disorder24.

Similarly, SMCHD1 (structural maintenance of chromosomes flexible hinge domain containing 1), previously reported as a candidate tumor suppressor in hematopoietic cancers25 was affected by copy number losses (del18p11.32-p11.31) (Fig. 3g) and truncating SNVs/indels with high CCFs (Fig. 3h) (n = 24/485, 5.0%). SMCHD1 mutations showed clonal expansion (Fig. 3i) and were associated with adverse OS (median = 48.2 months, P value < 1 × 10−4, log-rank test) (Fig. 3j).

Noncoding putative driver mutations

To gain insight into the significance of noncoding mutations, we first identified CLL-specific regulatory elements (REs) by integrating ATAC-seq and H3K27ac profiles26,27 as well as chromatin states28 from publicly available primary CLL (Fig. 4a; Methods). Out of the 29,224 promoters and 56,137 enhancers identified, 90% were present in CLL as a whole, whereas the remaining 10% were specific for IGHV subgroups and were used for the IGHV subtype-specific annotation (Methods). Mapping noncoding mutations to REs (Fig. 4b; Methods), we could identify 29 untranslated regions (UTRs), 25 enhancers (23 of them cataloged by the GeneHancer database29) and 72 promoters that had hotspot mutations or were recurrently mutated more frequently than expected (FDR < 0.1), defined as significantly mutated (Extended Data Fig. 5a and Supplementary Table 17).

Fig. 4: Significantly mutated noncoding REs.
figure 4

a, Methodology to localize noncoding REs in CLL primary cells. These REs were defined across the whole genome based on chromatin state data from CLL primary cells. We intersected H3K27ac peaks and open chromatin regions defined by ATAC-seq (derived from 104 and 106 primary CLL, respectively)27. Next, these regions were annotated using genome-wide segmentations of seven CLL samples (five mutated and two unmutated IGHV cases) with available chromatin immunoprecipitation followed by sequencing (ChIP–seq) data of six histone marks including H3K4me3, H3K4me1, H3K27ac, H3K36me3, H3K27me3 and H3K9me3. As our annotations of noncoding variants were based on CLL samples from different cohorts, chromatin states defined by ChIP–seq were considered only for regions that were seen in at least two samples. Common regions based on shared overlaps were used to define these REs. REs active exclusively in samples with m-IGHV and u-IGHV mutational status were also defined. REs were linked to target genes by correlating RNA expression (gene) and H3k27ac (REs) (Pearson correlation 0.3, FDR ≤ 0.05), within topologically associated domains of GM12878 defined by Hi-C30. For additional annotations and more details, see Methods. b, Candidate noncoding drivers including UTRs, promoters and enhancers affected by SNVs/indels, were revealed using several discovery algorithms and regions with FDR below the significance threshold were selected. The presence of single-site hotspots, and regions with high mutational density/kataegis were reported and regions with FDR below the significance threshold were selected. Annotations and postfiltering of somatic noncoding hits were including immunoglobulin loci and known false positive exclusion, AID and APOBEC signature annotations, and additional genomic and functional annotations from the literature. c,d, Significantly mutated REs for which target genes are CLL drivers or in the COSMIC database (c) or other genes (d). Upper panel, number of samples mutated; middle panel, proportion of variants with signature attributed to AID, APOBEC or other processes; lower panel, FDR of the likelihood these regions as mutated more frequently than expected. e, Gene set enrichment analysis based on the target genes of all noncoding candidate drivers for gene ontology terms biological process (GO:BP) and human phenotype ontology (HP). We applied a hypergeometric test and multiple testing correction of P value using the g:SCS algorithm41.

Source data

Full size image

Next, we defined the candidate target genes of these 126 mutated noncoding regulatory elements. Mutations within UTRs and promoters were annotated predominately according to proximity (Methods). For enhancers, we calculated the correlation between H3K27ac levels for each regulatory elements and the gene expression levels of surrounding genes located within the same topologically associated domain (TAD) of the B cell lymphoblastoid cell line GM1287830 (Methods). In total, 29 regulatory elements had target genes known to be CLL drivers or cancer drivers in the COSMIC database (Fig. 4c); 89 were linked to other genes (Fig. 4d) and 8 to none (Extended Data Fig. 5a and Supplementary Table 17). Four mutated regulatory elements were specific for u-IGHV (Extended Data Fig. 5b) and none for m-IGHV. Overall, genes targeted by mutated regulatory elements were enriched for gene ontology terms linked to the immune system, lymphocyte activation and cell death (Fig. 4e and Supplementary Table 18).

Of the 29 mutated UTRs, 58% (n = 17) had a median CCF ≥ 0.5, and 83% had a CCF > 0.2, thus indicating their selection during CLL pathogenesis (Extended Data Fig. 5c). These included the 3′ UTR mutations of NOTCH1 creating a splice site that leads to increased gene expression3,31 (n = 16; FDR = 4.57 × 10−2). The NF-κB signaling gene NFKBIZ (n = 8, FDR = 2.38 × 10−2) was also found significantly mutated, confirming previous findings6 and known to increase levels of mRNA and protein in lymphoma32,33. We observed clonal mutations in the 5′ UTR of IGLL5 (n = 28; FDR < 2.2 × 10−16), previously found to be associated with reduced expression4. Previously unreported significantly mutated UTRs included the 5′ UTR of BCL2 (n = 6; FDR = 1.01 ×10−6, Fig. 5a). We performed RNA-seq on samples carrying these mutations (Supplementary Table 4; Methods) demonstrating that 5′ UTR mutations were associated with BCL2 overexpression (P = 4.3 × 10−2; Fig. 5b), which is noteworthy given that BCL2 inhibitors are used therapeutically in CLL34.

Fig. 5: Noncoding mutations impacting BCL genes.
figure 5

a, Genome view of BCL2 5′ UTR. The significantly mutated region is indicated by a black rectangle. Individual somatic mutations are shown in blue. b, Gene expression of BCL2 in TPM determined by RNA-seq in samples with BCL2 5′ UTR mutations versus WT. Black dots are marks as outliers. P value was derived from a two-sided Welch’s t-test. c, Gene expression in TPM determined by RNA-seq of BCL6 in samples with BCL6 enhancer mutations versus WT. Expression levels were split in low (less than median expression; green), medium (between median expression and 100; orange) and high (≥100 TPM; purple). P value was derived from a two-sided Wilcoxon test. d, Genome view of the BCL6 gene and enhancers. Enhancers within these regions are annotated in blue font. eBCL6_2, which was the target of several variants is indicated in red. Annotation tracks of ATAC-seq and ChIP–seq are from publicly available RE annotation detailed above27 (references containing detailed of datasets and figure legends). The lower panel shows the individual mutations color coded as defined in c. All boxplots show the minimum and maximum values and interquartile range.

Source data

Full size image

A high clonality (>0.5) was also observed when considering the 97 significantly mutated promoters and enhancers; 72% had a median CCF >0.5 and 97% of a CCF >0.2 (Supplementary Fig. 4a). Six discrete regions spanning 117 kb contained 50 variants and were annotated in the previously reported PAX5 superenhancer3,6,35 (Extended Data Fig. 5a and Supplementary Fig. 4b). Another region spanning 325 kb on chr3q27.2 contained seven significantly mutated enhancers and linked to BCL6 (Extended Data Fig. 5a and Supplementary Table 17). RNA-seq of eight samples with mutations in this region showed overall increased expression of BCL6, although the effect was heterogenous (Fig. 5c), suggesting that some variants are more or less pathogenic than others and variants might exert a positional effect (Fig. 5d).

When considering the 72 significantly mutated promoters, we found mutations of known CLL drivers including BIRC3 (n = 31, 6.4%, FDR < 1.15 ×10−15), IKZF3 (n = 12, 2.5%, FDR = 8.16 × 10−13) and TP53 affecting splicing regions of noncoding exons/5′ UTR/promoter region (n = 2, 0.4%, FDR = 5.55 × 10−6). Next, we investigated mutations in these promoters further to identify those predicted to change chromatin state, using DeepHaem36, a deep neural network trained on chromatin feature data of 73 immune cell types. Seventy-four variants were predicted to lead to a loss of open chromatin (that is, loss-of-function variants), including those in the BACH2 promoter (Fig. 6a and Extended Data Fig. 6a). A recent study showed that decreased BACH2 expression in CLL is associated with adverse outcomes37. Notably, the mutations we detected in this promoter were mostly clonal (median CCF = 0.99). We therefore investigated this promoter further by performing ATAC-seq and RNA-seq (Fig. 6b) on mutated samples, when available (13 variants investigated, Supplementary Table 12; Methods) to understand the impact of these variants on chromatin accessibility and gene expression. Three variants within a 14-bp region were associated with allelic skew in the ATAC-seq compared with WGS data, demonstrating a preference for accessibility on the reference allele (Fig. 6c), which mirrored the decrease in chromatin accessibility in that region compared with WT samples (Fig. 6d). This allelic skew was also detected at the RNA level (Fig. 6e and Extended Data Fig. 6b). In addition, the same three samples also showed decreased BACH2 gene expression (Fig. 6f).

Fig. 6: Mutations in the promoter of BACH2 associated with reduction of chromatin accessibility and RNA expression.
figure 6

a, Prediction of the impact of noncoding mutations in promoters on transcription factor binding from cell- and tissue-specific DNase footprints. Mutations in BACH2 promoters are annotated (blue), specific BACH2 promoters detailed later are in red. GoF/LoF, gain/loss of function; B cell type specific, prediction observed in dataset examined; multi-B cell type, prediction observed in several dataset examined (robust); open multi-B cell, open chromatin region predicted; none, no prediction. b, Methodology to explore the effect of BACH2 promoter mutations. (1) We compared VAF of WGS data and ATAC-seq data to find allelic skew, that is, a preference for accessibility on the reference or the mutant allele. TSS, transcription start site. (2) We examined the change in chromatin accessibility in regions of interest in mutated compared with WT samples. (3) We compared VAF of WGS data and RNA-seq data to find allelic skew, that is, a preference for RNA expression on the reference or the mutant allele. (4) We compared the gene expression in mutated versus WT samples by RNA-seq. c, Prioritization of noncoding variants based on sequencing depth at the loci in the ATAC-seq data and allelic skew between the ATAC-seq and WGS data. Datapoints with difference less then –0.1 or greater than 0.1 and sequencing depth of at least ten times are annotated in black font. The BACH2 promoter is indicated in red font. d, ATAC-seq signal at the promoter of BACH2. The blue track shows the combined signal from all 24 patient samples; overlaid is the signal from a sample (pink) with a variant in the center of the RE. The location of variants in the same RE from three other patients are highlighted. e, Fraction of mutant and WT read in three BACH2 promoter variants showing allelic skew in ATAC-seq and RNA-seq compared with WGS. Prediction and mean damage scores were calculated with DeepHaem. f, Gene expression distribution (the minimum and maximum values and interquartile range) of BACH2 in TPM determined by RNA-seq in samples with promoter mutations versus sample WT. The statistical test used was a two-sided Welch’s t-test.

Source data

Full size image

Finally, we analyzed 20 cases with paired WGS, ATAC-seq and RNA-seq data (Supplementary Table 4). We identified five recurrently mutated promoters with allelic skewing of chromatin accessibility and RNA expression. Three, BTG2, CCND1 and ST6GAL1, were associated with allelic skewing towards the mutant allele, whereas ATAD1 and BIRC3 showed the opposite effect (Extended Data Fig. 6c). In the case of ATAD1, which plays a role in mitochondria protein degradation, we additionally observed reduced expression in promoter-mutated samples (P = 7.0 × 10−4) (Extended Data Fig. 6d–f).

Collectively, these data suggest that a small subset of the noncoding mutations in CLL have characteristics indicative of a driver and target regulatory elements of genes that are critical for B cell development and function as well as cancer progression. However, the effects on chromatin accessibility and gene expression levels were subtle and require further indepth functional characterization.

Clinical impact of combined and global genome features

We recalculated the occurrence of mutations in each known or putative driver in CLL by combining coding mutations, noncoding mutations in regulatory elements and CNAs (Fig. 7a and Supplementary Table 19). In total, 33 of the 58 coding, known or putative driver genes were also affected by noncoding mutations in associated regulatory elements or by CNAs. Overall, 412 (29%) of all alterations in these genes were either CNAs or affected regulatory elements. ATM and BIRC3 were most frequently targeted by genetic lesions. The median number of mutated known or putative drivers in each tumor was 2 (0–7) or 5 (0–21) when excluding or including CNA/copy neutral loss of heterozygosity (cnLOHs) and noncoding variants, respectively (Fig. 7b). A higher number of mutated genes was associated with worse PFS, especially when noncoding variants were included (Extended Data Fig. 7a,b and Supplementary Tables 15 and 16). Furthermore, the number of samples containing mutations in particular pathways also increased (by 3.3%) (Fig. 7c and Supplementary Fig. 5), in particular for the NOTCH and the transcriptional regulations pathways.

Fig. 7: Data integration and genome-wide global lesions.
figure 7

a, Distribution of the type of alterations in CLL coding drivers affected only by coding mutations (top panel) and affected by coding, CNAs and/or mutations in their REs (bottom panels). b, Distribution of the number of mutations per sample when considering all functional mutations (blue shading), SNVs/indels in coding drivers (green shading) and coding and noncoding drivers and CNAs (purple shading). c, Proportion of samples with mutated pathways, when considering coding drivers only (green), coding drivers and other genes with high impact mutations (involving frameshift and stop coding mutations (yellow)) and all coding as well as noncoding drivers (red). d, Telomere lengths distribution (showing the minimum and maximum values and interquartile range) in normal samples and matched CLL samples. Lines link matched tumor-normal datapoints. Significance level shows two-sided paired Wilcoxon test of P value <0.001. e, Fraction of each mutational signature detected in different genomic scopes: the 58 coding drivers; exonic regions; promoters, enhancers and UTRs; and whole genome. f, Fraction of each mutational signature detected in each coding driver. DBS not shown as data were too scarce. g, Distribution of the eight GC groups, based on presence (dark gray) or absence (light gray) of the three variables selected as best predictor by MCA: CN losses (Loss), CN gains (Gain) and trisomy (Tri). TP53 alteration status and conventional GC status are indicated in the top panel.

Source data

Full size image

We explored whether global genomic features could also be associated with clinical outcome. Firstly, we evaluated telomere length and observed that it was reduced in CLL samples compared with paired germline (median length of 2.7 kb versus 3.8 kb, P < 2.2 × 10–16, median content of 405 versus 467, P = 3.9 × 10−6, paired Wilcoxon test) (Fig. 7d and Extended Data Fig. 8a,b). Shorter telomeres were significantly enriched in samples with p53 pathway alterations (P = 1.99 × 10−36; Fig. 7d), with R/R samples compared with frontline (FDR = 5.37 × 10−7; Supplementary Table 14) and were associated with poorer PFS (FDR = 4.39 × 10−4; Supplementary Table 15 and Extended Data Fig. 8c,d).

Secondly, we explored the clinical associations of mutation signatures including single base substitution (SBS), doublet base substitutions (DBS) and small insertions and deletions (ID)38 (Fig. 7e,f and Supplementary Table 20). Considering signatures with known or probable etiology, the most prevalent were SBS5 (clock-like), DBS11 (APOBEC activity) and ID2 followed by other clock-like signatures: SBS1 (deamination of 5-methylcytosines), SBS8, DBS2 and the AID signature SBS9. As previously documented, SBS9 was highly enriched in m-IGHV CLLs (FDR = 4.80 × 10−57, Fisher’s exact test; Supplementary Table 14), was mutually exclusive with TP53 alterations (2.29 × 10−3) and associated with good PFS (Supplementary Table 15 and Extended Data Fig. 8e). De novo signature ID83C was found associated with TP53 alterations (FDR = 2.53 × 10−2; Supplementary Table 14) and poorer PFS (1.57 × 10−2; Extended Data Fig. 8f and Supplementary Table 15). SBS1 was also associated with adverse outcome (3.70 × 10−2; Supplementary Table 15 and Extended Data Fig. 8g).

Thirdly, we analyzed GC using unsupervised clustering (multiple correspondence analysis (MCA)) of 17 features related to CNAs (Extended Data Fig. 9a,b; Methods). These defined eight groups (GC1–GC8) (Extended Data Fig. 9c,d) with distinct genomic profiles (Fig. 7g and Extended Data Fig. 9e). GC4 (presenting CN losses only, n = 210) was enriched in del13q14.2 (FDR = 3.26 × 10−23). GC7 (presenting both CN gains and losses, n = 127) was associated with ten recurrent CNAs and seven known coding drivers including XPO1 (FDR = 3.98 × 10−11) and TP53 (FDR = 8.36 × 10−9). Together with GC8 (presenting trisomy, CN gains and losses, n = 15), GC7 comprised the most patients with conventional genomic complexity, defined by the presence of at least four CNAs (Extended Data Fig. 9f). None of the genomic complexity groups was significantly enriched in stereotyped subsets (Extended Data Fig. 9g). For the subset of samples with survival data (n = 243), we combined genomic complexity groups with copy number gains only, copy number losses only and both copy number gains and losses to increase statistical power. Interestingly, the eight groups were associated with different PFS and OS (Extended Data Fig. 10a,b), independent of TP53 status (Extended Data Fig. 10c,d). Furthermore, patients with both TP53 mutations and GC7/8 changes had ultrahigh-risk disease (median PFS = 8 months, median OS = 15 months) and fared worse compared with patients with TP53 mutations but no GC7/8 status (P = 0.03; Fig. 8a,b).

Fig. 8: Relationship between genomic features and patient outcome.
figure 8

a,b, Kaplan–Meier curve on PFS (a) and OS (b) of TP53 altered/WT in combination with GC7/8. The P value was derived from a log-rank test comparing the most two extreme curves (additional data in Extended Data Fig. 10). The dotted lines indicate the median survival for each subgroup. c,d, Genomic factors comprising the GS (cut-off 0.5) derived using non-negative matrix factorization, hypermutated subset (u-GS) (a), unmutated subset (m-GS) (b). The plot only shows features that split the data. e,f, Kaplan–Meier curves of PFS of samples divided by GS. Only samples with PFS data were included (n = 243). In e, the unmutated subset, del17p/TP53 mutated samples are plotted separately (black curve), all u-GS1 cluster 1 samples fell into this grouping; In f, the hypermutated samples, del17p/TP53 mutated samples are plotted separately (black curve). The P value was derived from a log-rank test comparing the most two extreme curves. The dotted lines indicate the median survival for each subgroup g, Confusion matrix showing agreement between true and predicted subgroup assignment. The true subgroup assignment was determined by applying the previously described NMF approach (Methods) to the whole set of genomic data. The predicted subgroup assignment was determined by first using 80% of the genomic data for subgroup assignment (training phase) followed by predicting the subgroup assignment in the remaining 20% of the data (testing). In all cases, sex and age were included to inform the model (Methods).

Full size image

Towards a patient classifier

To evaluate the potential clinical relevance of combining different genomic features, we first used penalized multivariate regression analysis for least absolute shrinkage and selection operator. This analysis led to the identification of 56 individual genomic features that predicted PFS and/or OS including SMCHD1/del18p11.32-p11.31, which retained significance as an independent predictor of OS (Extended Data Fig. 10e and Supplementary Fig. 6a).

Next, we applied non-negative matrix factorization (NMF) to identify robust subgroups of CLLs sharing subsets of the 186 different genetic alterations (Supplementary Table 21; Methods). Considering the profound clinical impact of the IGHV mutational status, we initially divided patients into m-IGHV and u-IGHV. Using this approach, we identified five distinct GS: three were u-IGHV (u-GS1, 2 and 3) and two m-IGHV (m-GS1 and 2) (Fig. 8c,d and Supplementary Table 22).

When considering u-IGHV CLL (Fig. 8c and Supplementary Table 23), u-GS1 was characterized by the presence of high-risk features including TP53 disruption, GC7, short telomeres and mutations in targetable pathways such as MAPK, PI3K and apoptosis, but there was no DNA damage response signature. By contrast, u-GS2 was defined by ATM/BIRC3/del11q22.2-22.3 alterations, as well as mutations in DNA damage response pathways, but without TP53 mutations or genomic complexity as defining features. Patients in u-GS2 were predominately male. u-GS3 had a high number of mutations in known and putative coding drivers, introns and UTRs, CN gains including trisomy 12, NOTCH1 mutations, and was enriched for older patients. All three subgroups included patients with BCR IG subsets 1 and 8, which are known to be associated with aggressive disease39 (Supplementary Fig. 6b). Although u-GS2 and u-GS3 were clearly distinct, they were associated with similar PFS after chemoimmunotherapy (Fig. 8e).

Regarding m-IGHV CLL (Fig. 8d), m-GS1 was similar to u-GS1 (cosine similarity of 0.81) and also to u-GS2 (cosine similarity of 0.7) (Supplementary Table 24). In contrast, m-GS1 was enriched for older men, BCR IG subset 2 (FDR = 2.96 × 10−6) and IGHV3-21 (FDR = 7.50 × 10−9) (Supplementary Fig. 6b), although most patients in m-GS1 did not have any defined CLL stereotype. m-GS2 had high mutation burden in enhancers, UTRs and promoters, was enriched for del13q4.2 but no other CNAs and had longer telomeres compared with the mean length in CLL. Additional clustering (Methods) further refined m-GS2 into distinct two clusters (Supplementary Fig. 6c). m-GS2 cluster 1 stood out by the high frequency of SBS9, the presence of GC4 and the absence of any other features. In comparison, m-GS2 cluster 2 had MYD88 mutations, trisomy 12 and other CN gains but no CN losses (Supplementary Table 25). Both clusters of m-GS2 had a very favorable PFS of 75% and showed a plateau of PFS, implying cure after chemoimmunotherapy (Fig. 8f). By contrast, patients belonging to m-GS1 had a shorter PFS than u-GS2/u-GS3 (median PFS = 38 versus 50 months; Fig. 8e) and there was no plateau.

In our analysis of patients treated with chemoimmunotherapy, NMF subgroups could not be defined without the different acquired local and global noncoding genomic changes, since combining all known coding drivers and the four common recurrent CNAs did not cluster patients into the GSs (Supplementary Figs. 6d and 7). Based on this observation, we examined whether the NMF method could be used to prospectively and precisely assign individual patients into their subgroup for individualized outcome prediction in the clinic. Our validation, performed by subsetting the dataset (Methods) showed that a total of 15/16 m-IGHV samples and 48/51 of u-IGHV samples were assigned correctly to their respective subgroup (Fig. 8g).

Discussion

Our study presents the first comprehensive WGS analysis of a large series of CLL patients requiring treatment. A main strength of our study is that it is based on patients enrolled into multicenter clinical trials, thereby reducing heterogeneity. This allowed us to not only define the genomic landscape of different stages of CLL3,4, but also to identify mutations associated with disease relapse and transformation.

Based on a strict pipeline for discovery of coding drivers, we selected the top ranked recurrently mutated genes, which comprised 36 known CLL drivers3,6,7 and 22 putative drivers. Only 32% of variants in those putative driver genes were missense variants, with most being truncating and stop-gain mutations. Although these putative drivers shared characteristics of known drivers (that is, damaging mutations in protein domains, impact on RNA expression, high CCF that further increased at disease progression, association with survival), we cannot exclude the possibility that some may simply represent passengers.

We defined recurrent translocations (with breakpoints in WDHD1; CTNND2-ARHGAP18) and 126 candidate noncoding drivers within REs pinpointing potentially druggable target genes (NOTCH1, DTX1, NFKBIZ, NTRK2 and BACH2). For a small subset of selected noncoding candidate mutations, we were able to demonstrate a modest impact on chromatin accessibility and/or target gene expression (5′ UTR of BCL2, enhancer of BCL6, promoter of BACH2 and promoter of ATAD1).

Exploring different layers of genomic data including coding, noncoding and genome-wide global changes allowed us to (1) derive a WGS-derived genomic complexity classification that further refines risk by identifying an independent ultrahigh-risk group associated with complex genomic alterations (GC7/8); (2) more precisely predict individual patients who achieve a plateau after chemoimmunotherapy (m-GS2) and are functionally cured, thereby clearly differentiating them from progressors in the m-GS1 subgroup.

Ideally, only genomic features experimentally validated as disease drivers should be included in any prognostic classification system, even if they were selected by very stringent criteria as those applied in this study (see above). However, it is well recognized that some genomic features are clearly not disease drivers, yet carry prognostic relevance. For example, in CLL, the IGHV mutation status representing the cell-of-origin or telomere length reflecting proliferative activity, are associated strongly with clinical outcome, but are not considered disease drivers.

In our NMF model using only the known coding drivers and recurrent CNAs did not allow us to recover the same level of discrimination as that afforded by inclusion of additional local and global noncoding information. This observation implies that the combination of coding and noncoding information in the classifier increases the precision of clinical risk prediction at least in our cohort of clinical trial patients.

Although treatment algorithms for CLL are shifting away from chemoimmunotherapy to targeted agents, the subgroups we define remain potentially clinically relevant as they reflect distinct biological entities. Collectively, our study provides a springboard for downstream functional analyses of putative coding and noncoding drivers. Robust testing on independent cohorts of patients undergoing targeted therapy will be required to further establish the clinical utility of this WGS-based classifier.

Methods

Patient cohorts, samples and ethics

All patients gave written informed consent and the study was approved under the 100,000 Genomes Project Ethics and the CLL Pilot ethics (MREC 09/H1306/54). A total of 485 patients with CLL were included in the study. A small subset was enrolled into CLEAR (CLL Empirical Antibiotic Regimen, early stage of the disease, n = 12, NCT01279252) and CLL210 (ref. 42) (relapsed/refractory patients, n = 30, EudraCT 2010-019575-29). All other patients were treatment-naïve and required treatment according to iwCLL criteria43. They were either fit patients receiving frontline treatment with fludarabine, cyclophosphamide, rituximab (FCR)-based treatment in ARCTIC44 (Attenuated dose Rituximab with ChemoTherapy In CLL, n = 61, EudraCT Number:2009-010998-20) or AdMIRe45 (Does the ADdition of Mitoxantrone Improve REsponse to FCR chemotherapy in patients with CLL, n = 65, EudraCT number: 2008-006342-25) or frail patients receiving ofatumumab with either bendamustine or chlorambucil chemoimmunotherapy in RIAltO (A Trial Looking at Ofatumumab for People With Chronic Lymphocytic Leukemia Who Cannot Have More Intensive Treatment, n = 92, NCT01678430). Patients recruited into FLAIR46 (Front-Line therapy in CLL: Assessment of Ibrutinib + Rituximab, n = 225, EudraCT 2013-001944-76) were randomized to ibrutinib alone or in combination with rituximab or venetoclax or standard first-line FCR treatment. In line with the studies’ data monitoring committees, baseline characteristics and clinical outcomes data were available only from studies once closed to recruitment (see Supplementary Table 1 for details of all patients recruited into the 100,000 Genomes Project). For patients recruited into the FLAIR study, these data are still awaited.

For a subset of 25 patients, we obtained a sample taken at relapse (Supplementary Table 4).

To investigate findings in more advanced disease, we reanalyzed WGS data coming from a cohort of 17 patients from whom two concurrent samples were collected: the CLL phase and the transformed phase (RS). This cohort includes samples and data generation as described in Klintman et al.23.

Only samples with a lymphocyte count of greater than 25 × 109 l–1 were included in the study ensuring a tumor purity greater than 80% and a median lymphocyte count of 80 × 109 l–1 (range, 33.9–166.5) (Supplementary Table 1).

Peripheral blood mononuclear cells (PBMCs) and a saliva sample were collected from each patient, which served as a source of tumor and germline DNA, respectively. DNA was extracted from PBMCs and saliva using QIAamp DNA mini kit (Qiagen) and the Oragene DNA saliva kit (DNA Genotek Inc) kits, respectively, according to the manufacturer’s instructions. DNA quality was assessed using Nanodrop (Thermo Fisher Scientific) and quantified using Qubit (Thermo Fisher Scientific) technology. RNA was extracted from PBMCs using the RNeasy Mini Kit (Qiagen) according to the manufacturer’s instructions. The quality of RNA was assessed using the Agilent 4200 Tapestation System, using High Sensitivity tapes. The concentration was assessed using the GeminiTM XPS Microplate Spectroflurometer from Molecular Devices and the Quant-iT HS RNA assay.

Whole-genome sequencing

Whole-genome 125 bp paired-end TruSeq PCR–free libraries were sequenced using Illumina HiSeq2500 technology. Raw sequencing data was aligned with using Isaac v.03.16.02.19 to GRCh38. Alignment and coverage metrics were calculated using Picard v.2.12.1 and Bwtool47 showing a mean read depth of 36× and 109× for normal and tumor samples, respectively. All downstream analysis of WGS data was performed on the whole dataset of 485 samples, unless otherwise stated.

RNA-seq

Libraries were prepared from samples of 74 patients using the Illumina Stranded Total RNA Prep, Ligation with Ribo-Zero Plus, with additional custom depletion probes, using 100 ng RNA. Libraries were sequenced on a NovaSeq 6000 system (Illumina) using 100 base paired-end chemistry (108–455 million read-pairs per sample). Sequencing reads were processed and aligned to Human Reference genome GRCh38 using the Illumina Dragen RNA pipeline v.3.8.4. Gentoyping was performed using bcftools mpileup48. Allele specific read counts were generated at sites of acquired SNVs determined by WGS.

ATAC-seq

ATAC-seq was performed as previously described49. Briefly 7.5 × 104 cells per technical replicate were resuspended in lysis buffer (10 mM Tris-HCl, pH 7.5, 10 mM NaCl, 3 mM MgCl2, 0.1% IGEPAL CA-630). Nuclei were pelleted (500g for 10 min), PBS was discarded and nuclei were resuspended in tagmentation buffer (25 µl 2× tagmentation DNA buffer, 2.5 µl Tn5 Transposase (Illumina) and 22.5 µl water) then incubated (37 °C for 30 min). DNA was extracted using the MinElute PCR Purification Kit (Qiagen), half the DNA was amplified (NEBNext High-Fidelity 2× PCR Master Mix (New England Biolabs)) and purified with the QIAquick PCR Purification Kit (Qiagen). Libraries were sequenced using 40-bp paired-end reads (Illumina NextSeq).

Reads were mapped to GRCh38 using the PEPATAC pipeline with prealignment to the mitochondrial genome and default settings50. Gentoyping was performed using bcftools mpileup48. Allele specific ATAC-seq read counts were generated at sites of acquired SNVs determined by WGS.

Immunoglobulin gene characterization

To determine the IGHV status of our cohort, we prioritized data from Sanger sequencing, followed by WGS-derived data including IgCaller51 results and the presence of noncanonical AID mutational signature (SBS9). This prioritizing scheme resulted in 54% (264/485) cases classified by Sanger sequencing, 40% (194/485) by the IgCaller algorithm and 6% (27/485) by the mutational signature SBS9. The correlation between these three methodologies was high, as can be seen in Supplementary Table 26. In addition, the IgCaller algorithm was used to further characterize the IG genes, including to define the IGHV3-21 rearrangement in 10% (47/485) of cases and CLL stereotypy in 27% (132/485). To assign CLL stereotypes, the IgCaller output was used as input for AssignSubsets online tool52, which annotates the 19 main subsets, including subsets 1, 2, 4 and 8, as recommended by ERIC guidelines39. In cases more than one rearrangement were detected, we selected the rearrangement with the highest score to define the main CLL stereotype. In cases where a rearrangement was not assigned, but there was a proximal rearrangement reported, we included this rearrangement in our analysis.

Somatic variant calling and filtering

SNVs and indels were called using Strelka v.2.8.4 7 adopting default parameters. Filtering of SNVs/indels was performed as follow: depth required greater than ten and allele fraction (AF) greater than 0.05; the quality filter annotation should be ‘PASS’ and quality score greater than 30; variants with allele frequency less than 0.05 from 1KGP phase 3 1405.34_GRCh38.p8 and EXAC v.0.3 data (annotated from using Ensembl VEP GRCh38 release v.89.4 (ref. 53)). Additional filters according to the Illumina v.4 Genomics England annotation pipeline removed variants as follows: variants with a population germline frequency greater than 1% in either the Genomics England dataset or in the gnomAD v.3; recurrent somatic variants with a frequency greater than 5% in the Genomics England cohort; variants overlapping with LINE repeats or simple repeats found with Tandem Repeats Finder v.4.09 (ref. 54); calls within 50 bp either side of an indel where at least 10% of variants have been filtered due to quality; locus depth is greater than three times mean chromosomal depth in the germline sample; contains multiple alternate alleles; germline sample is not the homozygous reference or indel Q-score is less than 30; variant quality score recalibration (VQSR) score less than 2.75; most overlapping reads do not map uniquely to variant position; within ten bases of Genomics England inhouse database or Gnomad v.3 germline indel with frequency greater than 1%; SNVs resulting from systematic mapping and calling artefacts; fails somatic panel of normal Phred cut-off (< 80).

The Supplementary Notes include details on cancer cell fraction calculation as well as coding and noncoding variant annotations. In addition, it includes our approach for assigning target genes of regulatory elements, identifying of coding and noncoding candidate drivers.

Structural variant identification

The structural variant (SV) calling pipeline for detection of inversions and translocations was as follows. (1) Delly55 was used to call variants in each tumor–germline pair, with the following steps: complete somatic prefiltering, genotype all potentially somatic sites across all CLL germline samples, postfilter for somatic SVs using control samples. Variants with an alternative AF less than 0.05 were removed. (2) Lumpy v.0.2.13 (ref. 56) and (3) Manta 0.28.0 were also used to call SVs. Variants with an alternative AF < 0.05 or for which there was any evidence in the germline were removed for consistency. (4) The pcawg-merge-sv consensus calling pipeline57 was adapted for this analysis. SVs supported by two or more callers were reported.

Identification of CNAs

We used both DNA microarray (n = 109 samples) and WGS (n = 485 samples) to determine CNAs and observed high concordance between the two methods. Of 282 CNAs detected by WGS, 240 (85%) were also reported by DNA microarray with high confidence (Supplementary Table 6). In addition, we further reduced false positive signals using a combination of intersects between several variant callers and visual inspection as detailed below.

Samples from subset of 109 patients enrolled in ARCTIC and AdMIRe trials were genotyped using HumanOmni2.5-8 BeadChip arrays (Illumina Inc.). Genotypes were called using GenomeStudiov.2009.2 (Illumina Inc.). CN gains and losses greater than 50 kb and cnLOH less than 5 Mb were reported using Nexus Copy Number v.10 (BioDiscovery, Inc.), as previously described16,58, with the following settings (SNPRank Segmentation): significance threshold, 1 × 10–5; max contiguous probe spacing (kb), 1000.0; minimum number of probes per segment, 5; high gain, 0.6; gain, 0.2; loss, –0.2; big loss, –1.0; 3:1 sex chromosome gain, 1.2; homozygous frequency threshold, 0.95; homozygous value threshold, 0.8; heterozygous imbalance threshold, 0.4; minimum LOH length (kb), 20; percentage outliers to remove, 3%. We also inspected all genomes to scan visually for changes not identified using these analysis settings using Nexus visualization tool.

In the case of WGS, Canvas v.1.3.1 (ref. 59) and Manta v.0.28.0 were used to call CNAs, filtering out centromeric and telomeric regions as defined in the UCSC cytoband table. Variants reported by Canvas with a quality score less than ten were filtered out. Variants reported by Manta were filtered out as follows: (1) variants with a normal sample depth near one or both variant break-ends three times higher than the chromosomal mean, and (2) variants with somatic quality score of less than 30.

For each remaining CNA, its presence and type (gain or loss) were confirmed by visually inspecting the genome-wide mean coverage and B-allele frequency data, derived from the aligned reads in 100 kb windows. Calls with continuous copy number changes of length greater than 100 kb were kept. The Supplementary Notes include details on cancer cell fraction calculation.

Counts of number of drivers

We calculated the total number of drivers in each patient by the following methodologies: we established (1) the total mutational burden by counting the number of functional variants (that is, with the following exonic consequences splice acceptor variant, splice donor variant, stop gained, frameshift variant, stop lost, start lost, transcript amplification, in-frame insertion, in-frame deletion, missense variant, protein-altering variant or incomplete terminal codon variant), (2) the number of mutated coding drivers (out of 58) SNVs/indels and (3) the number of mutated coding (SNVs/indels and CNAs) and noncoding drivers.

Pathway analysis

Two pathway datasets were used: PANCANCER containing 14 pathways from The NanoString PanCancer Pathways Panel and KEGG containing 23 signaling pathways60. For the six pathways in common between the two lists, the PANCANCER pathway was selected, resulting in 31 unique pathways included in the analysis. We counted the number of patients with mutations per pathway considering (1) a gene panel of the coding drivers (n = 58); (2) the exome (coding drivers plus exonic mutation with high impact according to VEP annotations: splice_acceptor_variant, splice_donor_variant, stop_gained, frameshift_variant, stop_lost, start_lost); (3) a larger driver panel containing both coding drivers and regulatory candidate drivers (n = 58 + 126) and (4) all of the above combined (coding and noncoding drivers plus exonic mutation with high impact according to VEP annotations).

Telomere analysis

Telomere analysis was carried out on all 485 CLL tumor-normal pairs. Telomere content was estimated using Telomere Hunter v.1.1.0 (ref. 61). Telomere content is normalized by the total number of overall reads that comprise a ‘telomere-like’ GC-content range (48–52%). Telomere length in basepairs was estimated using Telomerecat v.1.0 (ref. 62). We found that telomere content assessed using Telomere Hunter and telomere length assessed using Telomerecat were highly correlated (P = 0.84, P < 2.2 × 10–16, Extended Data Fig. 8a). We compared the telomere lengths and contents between CLL samples and matched saliva samples as germline63, considering that different cell types can naturally present different telomere lengths64.

Chromothripsis analysis

Chromothripsis was identified using Shatterseek65, which aims to detect candidate regions on the basis of oscillating copy number states (using CNAs as previously described), as well as intersection with clusters of interleaved structural variants (SVs; that is, deletions, duplications, inversions and translocations) identified from the SV consensus pipeline previously described. Potential regions of chromothripsis were classified as ‘high confidence’ or ‘low confidence’ using criteria as per Cortés-Ciriano et al.65.

Mutational signatures

Extraction of SBS, DBS and small ID signatures was performed using SigProfilerExtractor v.1.0.1810 (ref. 66). SigProfilerExtractor de novo signature extraction and decomposition were carried out according to default parameters, with potential de novo extracted signature solutions tested between 1 and 25 signatures. Signatures were referenced to the Catalogue of Somatic Mutations in Cancer (COSMIC) v.3; SigProfilerExtractor signatures were decomposed based on a cosine similarity greater than 0.9. Following decomposition to COSMIC signatures, SigProfilerExtractor estimated the overall signature contributions per tumor, as well as the per tumor signature estimates for each mutation context. Through associating these context estimates back to the original mutations, signature estimates were attributed to individual driver mutations, as well as genomic regions (exome, promoters, UTRs, and so on).

GC analysis

We investigated the presence of GC using an unsupervised multiple correspondence analysis with FactoMineR67. We included 17 genomic measures as binary data, including variables binned as less than median or greater than or equal to median: number of SNVs, number of indels, telomere lengths, telomere content and variables binned as presence/absence: SV breakpoint, CNA, CN gain, CN loss, cnLOH, trisomy, aneuploidy, CN gain excluding trisomy, CN loss excluding aneuploidy, cnLOH excluding whole chromosome cnLOH, inversion, translocation and chromothripsis.

Genomic alterations in known risk factors and disease states

All genomic alterations derived from WGS were combined and included as follows: noncoding candidate drivers mutated in more than 5% of samples; coding drivers were combined according to the presence of an SNVs/indels and CNAs (union); recurrent CNAs that significantly co-occurred (mean square contingency coefficient, mu > 0.3) and defined in the same chromosome were combined (union). In addition, only genomic alterations with at least five occurrences across all the samples were included in the analysis. In total, 186 genomic remained including 58 coding drivers, 36 recurrent CNAs, 44 noncoding drivers, 12 pathways affected by genetic alterations, 28 global genomic features and mutational signatures, and eight genomic complexity groups (Supplementary Table 21).

We tested for enrichment (two-sided Fisher’s exact test, FDR ≤ 0.05) of each genomic alteration in several known risk factor and disease state groups for samples with available data: age (195 samples < median age versus 216 samples ≥ median age); sex (338 male versus 136 female); disease stage (443 frontline versus 30 R/R); TP53 status (420 WT versus 65 disrupted); IGHV mutational status (197 hypermutated versus 288 unmutated); minimal residual disease (MRD; 59 negative versus 57 positive); BCR IG subset 2 (33 presenting 2 versus 450 others); IGHV3-21 rearrangement (47 with versus 436 without).

Relationship between genomic alterations and patient outcome

We examined the relationship between each of the 186 genomic features as detailed above (Supplementary Table 21) and patient outcomes using Cox proportional hazards models on 243 patients for PFS and 245 patients for OS. FDR-corrected P values were reported as significant if less than 0.05. In addition, several particular comparisons with more than two groups were performed using Kaplan–Meier curves and the log-rank test. These were: number of mutated drivers, the eight genomic complexity groups and the combination of different structural rearrangements. We also performed a multivariate analysis using penalized Cox regression, as implemented in the R package glmnet68, to find a minimal set of predictors with maximal predictive power. An optimal value of the penalization parameter λ was selected using leave-one-out cross-validation; specifically, the value of λ that minimizes the cross-validation error.

Patient stratification using non-negative matrix factorization

All 186 genomic features, as well as IGHV status including percent homology to germline (labeled MS), age and sex were selected for unsupervised clustering using non-negative matrix factorization69,70 using the NMF v.0.22.0 R package71 with the offset method72,73. Data were converted to a binary matrix using either presence or absence of a feature, or above or below the mean to avoid a mixture of binary and nonbinary data (Supplementary Note). After removal of samples without age information, samples were divided into m-IGHV (n = 168) and u-IGHV (n = 243) as defined above. The number of permitted NMF clusters in either the m- or u-IGHV subset was determined using a combination of rank estimation methods including the cophenetic correlation coefficient74,75,76. Data were randomized and the ranks estimated for comparison to avoid overfitting. NMF was carried out on each IG subset of samples separately to produce GSs.

DeconstructSigs v.1.9.0 (ref. 77) was designed to use the mutation catalog of a sample to define the linear combination of COSMIC signatures that best reconstruct that sample’s mutational profile. Here, we used this tool to define the linear combination of GSs calculated using the NMF method that best reconstruct the genomic features of a sample. The proportions of each GS within all patients were then clustered using mclust v.5.4.6 (ref. 78) and assigned a cluster that maximized parsimony whilst still producing an adequate prediction. The defined GSs were then compared with known subgroups such as BCR IG subsets and patients harboring an IGHV3-21 rearrangement.

Testing of the method was carried out as follows:

  1. (1)

    Data were randomly split into two trial groups each representing 50% of the dataset: and further divided into m-IGHV and u-IGHV CLL. The NMF was then performed on all genomic features on each group and evaluated using cosine similarity between group signature matrices (Supplementary Table 22);

  2. (2)

    all samples used for NMF were split into 80% (m-IGHV: n = 133, u-IGHV: n = 195) training and 20% (m-IGHV: n = 34, u-IGHV: n = 49) testing at random. The NMF was performed on the training data as described above to produce GS matrixes (m-GS, u-GS). The training data were then assigned to a GS using deconstructSigs to identify the combination of GSs that best reconstructed a sample’s genomic feature matrix and then assigning the signature that occurred at the highest percentage. The signature assigned to the test samples was then compared with the signatures assigned to those same individuals when 100% of data was used for both training and testing (Fig. 8g).

Data wrangling and plotting

Plotting of data was performed using tidyverse v.1.3.0 (refs. 79,80) in R v.3.6.2 (ref. 81). Mutation hotspot graphics were plotted using the package GenVisR v.1.18.1 (ref. 82). Lollipop plots were plotted with the MutationMapper from cbioportal accessible from https://www.cbioportal.org/mutation_mapper. Genomic views were prepared using the UCSC genome browser83.

Statistics and reproducibility

The sample size calculation was critical to the success of this program. Our power calculations considered the heterogeneity of CLL and a background somatic mutation frequency of 0.8 mutations per megabase. This means that, to reliably detect somatic mutations recurring in 2% of patients with CLL, we need to sequence approximately 500 CLL genomes (Supplementary Fig. 8). No data were excluded from the analyses. The experiments were not randomized. The investigators were not blinded to allocation during experiments and outcome assessment.

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.