Functional Impact of Genomic Complexity on the Transcriptome of Multiple Myeloma

Abstract

Purpose:

Multiple myeloma is a biologically heterogenous plasma-cell disorder. In this study, we aimed at dissecting the functional impact on transcriptome of gene mutations, copy-number abnormalities (CNA), and chromosomal rearrangements (CR). Moreover, we applied a geno-transcriptomic approach to identify specific biomarkers for personalized treatments.

Experimental Design:

We analyzed 514 newly diagnosed patients from the IA12 release of the CoMMpass study, accounting for mutations in multiple myeloma driver genes, structural variants, copy-number segments, and raw-transcript counts. We performed an in silico drug sensitivity screen (DSS), interrogating the Cancer Dependency Map (DepMap) dataset after anchoring cell lines to primary tumor samples using the Celligner algorithm.

Results:

Immunoglobulin translocations, hyperdiploidy and chr(1q)gain/amps were associated with the highest number of deregulated genes. Other CNAs and specific gene mutations had a lower but very distinct impact affecting specific pathways. Many recurrent genes showed a hotspot (HS)-specific effect. The clinical relevance of double-hit multiple myeloma found strong biological bases in our analysis. Biallelic deletions of tumor suppressors and chr(1q)-amplifications showed the greatest impact on gene expression, deregulating pathways related to cell cycle, proliferation, and expression of immunotherapy targets. Moreover, our in silico DSS showed that not only t(11;14) but also chr(1q)gain/amps and CYLD inactivation predicted differential expression of transcripts of the BCL2 axis and response to venetoclax.

Conclusions:

The multiple myeloma genomic architecture and transcriptome have a strict connection, led by CNAs and CRs. Gene mutations impacted especially with HS-mutations of oncogenes and biallelic tumor suppressor gene inactivation. Finally, a comprehensive geno-transcriptomic analysis allows the identification of specific deregulated pathways and candidate biomarkers for personalized treatments in multiple myeloma.

Translational Relevance.

In this article, we studied the close interactions between the genomic architecture of multiple myeloma and its functional impact on the transcriptome. This comprehensive analysis highlighted how copy-number changes and chromosomal rearrangements impacted the most on gene expression. However, we showed the great importance that mutations play in the setting of hotspot variants and biallelic inactivation of tumor suppressor genes. Moreover, a widespread geno-transcriptomic analysis was able to infer the differential gene expression of specific druggable pathways and to identify biomarkers of sensitivity to novel treatments in multiple myeloma, suggesting this approach could drive personalized treatment decisions.

Introduction

Multiple myeloma is a plasma-cell neoplasm driven by recurrent trisomies- hyperdiploid multiple myeloma (HDMM)- or by immunoglobulin gene (IGH) locus rearrangements promoting overexpression of recurrent oncogenes. Classic PCR and FISH studies, and more recently next-generation sequencing (NGS) studies have highlighted a vast array of additional mutations, copy-number abnormalities (CNA) and structural variants (SV) which are thought to play a role in disease evolution and may be acquired in preclinical phases or at relapse after treatment (1–9).

Such abnormalities are predicted to influence the biological and clinical behavior of the tumor and yet, despite a clear driver role, only few carry prognostic value (2, 10–12). Additionally, as the genomic makeup of multiple myeloma is being comprehensively elucidated by NGS studies, classification efforts seem to have gained very little by the additional knowledge brought by NGS and initiating karyotypic events are still the mainstay of multiple myeloma classification. Indeed, most additional abnormalities discovered by NGS seem to be randomly distributed with few notable exceptions (13, 14) and their discovery has not led to the characterization of additional multiple myeloma categories so far. If not relevant from the point of view of a genomic classification then, a question is what their biological role really is, and whether they can serve as markers of specific transcriptional profiles.

Classically, gene-expression profiling arrays have shown how transcriptome profiling recapitulates the main karyotypic lesions in multiple myeloma (15–17). Few studies on RNA sequencing (RNA-seq) in multiple myeloma have been published (7, 14, 18–21). Most of them have reported that many DNA mutations are actually not expressed, and how the transcriptomic profile has very few genomic correlates aside from hyperdiploidy and IGH translocations. Laganà and colleagues have shown that RNA-seq from the public CoMMpass database can be mined through a network approach to identify multiple myeloma groups where annotation of genomic and clinical features highlights some commonalities (20). However, a systematic and comprehensive assessment of the functional impact of mutations, CNAs and SVs at the transcriptomic level is still lacking. This would have implications for disease pathogenesis and would allow prioritization of research on genomic abnormalities that have specific consequences on expression. Ideally, those would be the ones with the highest value as prognostic markers, biomarkers of drug response, and/or candidate drug targets. Indeed, while for example response to venetoclax is best predicted by assessing the expression levels of BCL2, BCL2L1, and MCL1 (22), this is unlikely to become routine clinical practice. DNA abnormalities that correlate with expression of target genes are much more easily amenable to clinical-grade diagnostics through NGS techniques (23–25).

In this article, we mined genomic and transcriptomic data from 514 cases from the CoMMpass dataset to dissect transcriptional effects of the most recurrent genetic abnormalities. In particular, starting from the comprehensive DNA exome-wide sequencing information, we performed a systematic and unbiased assessment of differentially expressed transcripts correlated to each recurrent gene mutation, CNA and SV. From here, we identified genetic predictors of specific differential gene expression levels to get insights into the functional relevance of recurrent genomic alterations and suggest therapeutic vulnerabilities.

Materials and Methods

Data

We interrogated 514 newly diagnosed patients enrolled in the CoMMpass study (IA12 release). The CoMMpass data were generated as part of the Multiple Myeloma Research Foundation (MMRF) Personalize Medicine Initiative (https://themmrf.org). We focused on mutations in multiple myeloma driver genes, SVs, copy number segments, and raw transcript counts. The CoMMpass trial is a prospective observational clinical trial (NCT01454297) that comprises whole genomic and transcriptomic data derived from newly diagnosed patients with multiple myeloma. The tumor samples were collected from several centers across the United States, Canada and Europe. Ethics committees or institutional review boards at the study sites approved the study. All the patients signed a written informed consent prior to enrollment. The study was conducted under the statement of the Helsinki declaration. Multiple myeloma samples were collected at diagnosis at enrollment sites and then shipped to the Translational Genomics Research Institute (TGen) to be processed and sequenced. Due to the lack of transcriptomic data from normal plasma cells in the CoMMpass dataset, all the statistical analyses were performed comparing genetic subsets of multiple myeloma samples.

Gene expression analysis

RNA-seq data were processed using the voom/limma pipeline (26, 27). Firstly, the dataset of raw-counts was filtered to remove genes with less than 10 reads in more than 95% of samples. Then, we performed the trimmed mean of M-values (TMM) normalization to estimate a scale factor used to decrease technical bias between samples, resulting from differences in library size (28). Finally, we applied the voom transformation to convert the raw counts in log2-counts per million (log₂-CPM) and calculated the respective observation-level weights to be used in differential expression analysis.

To assess the amount of differentially expressed transcripts a linear model was fitted to the expression data for each genomic feature, detecting the genes significantly associated with at least one genetic abnormality. This model evaluates the expression of each gene in each patient using a design matrix composed of information on genetic anomalies for each patient. From this, we obtained the coefficients that measure the gene expression change associated with the presence of each genetic anomaly. Then, to determine the association between gene expression and genomic alterations we used the F statistic obtained by the lmFit function of the R limma package. Finally, we applied the Benjamini–Hochberg correction for multiple testing. To evaluate the accuracy of the applied model, we calculated a random model in which all the values of each covariate (genetic anomaly) were randomly permuted, breaking all the correlations between the gene expression level and genotype (29). This model allowed to correct the subsequent differential expression analysis for the main karyotypic subtypes and segmental CNAs [hyperdiploidy, t(4;14), t(11;14), t(14;16), t(14;20) and chr(1q21)gains/amplifications (amp)]. Moreover, this approach allowed us to identify significant gene associations by calculating the coefficient of determination (R²), which tells how well the model fits the data, but also the portion of variance of the dependent variable (expression of gene in a patient) predicted by the independent variable (presence of genetic anomaly). This allowed us to measure the variance explained by the presence/absence of a given genetic alteration and to evaluate how much this alteration is associated with changes in expression of the candidate gene.

Hotspot and non-hotspot mutations

Mutational hotspots (HS) are amino acid positions in a protein-coding gene mutated more frequently than would be expected in the absence of selection (30), therefore conferring a fitness advantage. For this analysis, we selected 4 genes from a list of 53 mutations on known myeloma driver genes (13): KRAS, NRAS, IRF4, and BRAF. Then, we defined as HS the nonsynonymous mutations as follows: G12, G13, and Q61 were selected as HS for KRAS and NRAS; K123 for IRF4; and V600 and D594 codons for BRAF.

Identification of genes with a cumulative effect

Cumulative transcriptomic effect of genes was analyzed in regions of chr(1q21) amplifications (>3 copies) or biallelic inactivation of tumor suppressors (mutation of one copy and deletion of the other). To explore the functional significance of these events, we selected genes with a cumulative effect, i.e., genes significantly deregulated in biallelic versus wild-type (WT) [or amp vs. WT for chr(1q21)] and/or biallelic versus mono-allelic statuses [or amp vs. gain for chr(1q21)]. To perform this analysis, first we selected, for each analysis, the significant genes from the differential expression analysis contrasts, then for each selected gene we calculated with a linear regression their cumulative effect. Finally, we restricted the analysis to the significant genes (P value < 0.05) deregulated in biallelic (or amp) versus WT and/or biallelic (or amp) versus mono-allelic (or gain).

Gene ontology analysis

The Gene Ontology (GO) enrichment analysis was performed using ClusterProfiler R package to search for a biological interpretation of transcriptome deregulation. This algorithm implements methods to analyze and visualize functional profiles of genomic coordinates, genes, and gene clusters (31). The P values were corrected for multiple testing by using the Benjamini–Hochberg FDR method. In the HS analysis, we performed the pathway analysis with the genes from differential expression analysis contrasts (HS vs. WT, HS vs. non-HS, and non-HS vs. WT); whereas in the biallelic deletions and chr(1q21)gain/amp analysis, we executed the GO analysis with the cumulative effect genes. Gene set enrichment analysis (GSEA) was performed by the R package fgsea using the “t” statistics as ranking factor. The cancer hallmark list was inferred with a significant adjusted P value < 0.05 (32).

In silico analysis of drug sensitivity

To assess our findings, we performed an in silico analysis using the Cancer Dependency Map (DepMap) database and Celligner algorithm (detailed methods in the Supplementary Data S1; refs. 33–35).

Statistical analysis

The association tests have been performed with two-sided Wilcoxon rank–sum test. The P values were adjusted using the Benjamini–Hochberg FDR method (FDR<0.05). All analyses were performed in R, the language and environment for statistical computing (R Core Team, 2020).

Results

Genome-wide transcriptome patterns associated with multiple myeloma driver aberrations

To have an overview of the driver genomic lesions underlying global transcriptome deregulation in multiple myeloma, we first performed an unsupervised principal component analysis. Cases segregated mostly based on their HD versus IGH translocation status (Fig. 1A), in agreement with historical gene expression profiling data (15, 17, 36, 37). However, recent genomic efforts have greatly expanded the repertoire of driver genomic lesions in multiple myeloma, including CNAs and gene mutations. Therefore, we asked what the independent contribution of each to the derangement of the transcriptome could be. To this end, we implemented a linear regression model that assesses which transcripts are specifically associated to distinct genomic features (29). We plotted, for each driver lesion, the number of up- or downregulated transcripts and their fold change (Fig. 1B). IGH translocations and hyperdiploidy were associated overall with the most substantial differential gene expression patterns observed, along with amplifications and gains in chr(1q21). Other CNAs had fewer transcriptomic correlates, followed by specific gene mutations. Within the former group, CYLD deletions showed the highest number of differentially expressed transcripts compared with normal CYLD copy number (CN) status. Mutations of the 53 known driver genes in multiple myeloma (13) were characterized by less consistent transcriptomic patterns. However, interesting correlations emerged from this analysis. DIS3 and NRAS mutations were the two lesions whose occurrence was associated with the largest transcriptional modulation with respect to germline configuration. A high number of differentially expressed long noncoding RNA (lncRNA) was observed in DIS3-mutated versus WT patient samples in particular, confirming DIS3 may mostly regulate expression at the post-transcriptional level (Supplementary Data S2; ref. 38). As expected, GO analysis for NRAS and KRAS mutations showed differential expression of genes pertaining to the MAPK pathway (Supplementary Fig. S1A and S1B). Among less frequent mutations, TGDS and RB1 in 13q and the MYC transcriptional regulators IRF4 and MAX displayed correlation with the transcriptome. Conversely, most of the remaining gene mutations were not associated with any differentially expressed transcript.

Figure 1.

Transcriptomic profile. A, Principal component analysis based on the main karyotypic subtypes: samples are represented as dots in the space identified by the three principal components and are color-coded based on their karyotype. B, Stacked bar chart showing the differentially expressed genes per each genomic abnormality (FDR < 0.05, fold change cut-off 1.5). Bars indicate the contributions of upregulated genes (red) and downregulated genes (blue). FC, fold change. C, Scatterplot representing expression prediction for the CCND1 and CYLD genes versus observed expression values: samples are color-coded dots based on the significant genomic abnormality (FDR < 0.01), in blue t(11;14) samples, in green del(CYLD) samples and in black the other samples; R² represents the association between genomic alteration and expression changes.

IGH translocations exert their transforming activity by overexpressing target oncogenes. Indeed, modelling predicted versus observed expression of CCND1 showed that the t(11;14) accounted for most of the observed inter-patient heterogeneity in CCND1 transcript levels (Fig. 1C, left), as was the case for NSD2 and to a lesser extent FGFR3 expression in t(4;14), MAF in t(14;16), and MAFB in t(14;20) (Supplementary Fig. S2A–S2D). In the case of CNAs, the picture was not as clear cut: while low CYLD expression levels were mostly explained by locus deletions (Fig. 1C, right), not all cases of chr(17p13) or TRAF3 gene deletions resulted in downregulation of TP53 or TRAF3, respectively (Supplementary Fig. S2E and S2F). In the case of chr(1q21)gain/amp, heterogeneity of expression levels of the key target genes CKS1B and MCL1 could be attributed to the allelic status of the locus only to an extent (Supplementary Fig. S2G and S2H).

Therefore, structural, numeric variants and few mutations in driver genes were associated with most of the transcriptomic heterogeneity between multiple myeloma cases. For canonical IGH translocations this is likely mediated by overexpression of the target oncogene. The transcriptomic inter-patient variability observed in the case of CNAs is likely associated with the altered expression of more than one target gene and can be partially explained by the clonality of the event.

Impact of driver recurrent gene mutations on gene-expression regulation

Despite having a smaller correlation with the transcriptome, recurrent gene mutations have a clear driver role. Many genes are mutated in a HS pattern, and specific effects of HS versus non-HS variants as compared to their WT status are unknown. We therefore sought to analyze genotype–phenotype correlations for 4 of the main mutated genes with known HS: KRAS, NRAS, BRAF, and IRF4. The distribution of HS and non-HS mutations within the main cytogenetic categories was neutral except for kinase-dead BRAF mutations at codon D594, mostly occurring in t(14;16) cases (Supplementary Fig. S3). As a first step, we interrogated the Catalogue of Somatic Mutations in Cancer (COSMIC) to interrogate the codon usage of these gene mutations in CoMMpass as compared with different solid cancer types (colon, lung adenocarcinoma, and skin melanoma) and other hematologic malignancies. Interestingly, multiple myeloma showed a specific mutational spectrum for some of these genes. KRAS in particular was mostly mutated in the Q61 codon in multiple myeloma, while in solid cancers, diffuse large B-cell lymphomas, chronic lymphocytic leukemia (CLL), and acute myeloid leukemia (AML) showed mostly G12 and G13 mutations (Fig. 2A; Supplementary Fig. S4A). NRAS was mainly mutated at the Q61 codon. A similar picture was also observed for solid cancers. However, in this latter setting, G12 and G13 residues mutations were also enriched in comparison with multiple myeloma (Fig. 2B). Interestingly, other hematologic malignancies showed a predominance of NRAS G12 and G13 mutations, and only CLL showed a mutational pattern similar to multiple myeloma (Supplementary Fig. S4B). The IRF4 K123 HS was specific for multiple myeloma. Nevertheless, we identified other IRF4 mutated HS in B and in T-cell neoplasms, all within the DNA binding domain. In solid cancers, where IRF4 is not usually expressed, we did not identify any clustered mutational locus (Fig. 2C; Supplementary Fig. S4C). Finally, while BRAF mutations were mostly clustered on the V600, multiple myeloma and other hematologic malignancies showed a higher incidence of the “kinase-dead” D594 mutations as compared with solid tumors (Fig. 2D; Supplementary Fig. S4D). Notable exceptions were hairy cell leukemia and Langherans' histiocytosis, where only the V600 residue mutated (Supplementary Fig. S4D).

Figure 2.

Hotspot versus non-HS and WT analysis. A–D, Lollipop plots for the four main mutated genes; each plot shows in the top the hotspots in multiple myeloma, at the bottom the hotspots in the peculiar solid cancers. Lollipop plots have been generated applying the R package maftools. AI–DI, Venn plots for each differential expression analysis.

Next, we asked whether HS and non-HS mutations would carry different functional relevance as assessed by transcriptomic analysis: for KRAS and NRAS, we found transcriptomic correlates for HS gene mutations only. Interestingly, even if the two genes fuel the same proliferative pathway (Supplementary Fig. S1), NRAS HS mutations were associated with differential expression of more than 2,000 genes, as compared with 80 for KRAS HS mutations (Fig. 2AI, BI; Supplementary Fig. S5A and S5B). In the case of IRF4, again HS mutations correlated with the highest number of differentially expressed transcripts, while non-HS mutations were characterized by a less substantial and partly overlapping transcriptional profile (Fig. 2CI; Supplementary Fig. S5C). Interestingly, for BRAF mutations, the D594 codon was correlated with the highest level of differential gene expression, showing some overlap with mutations in the canonical codon V600. In particular, the kinase-dead mutations at D594 were associated with the differential expression of 183 genes, three times more than the V600 ones. This may be partly explained by the specific occurrence of D594 mutations in MAF-translocated multiple myeloma subgroup (4/7 cases) as previously described (14). Overall, non-HS mutations contributed very little in BRAF (Fig. 2DI; Supplementary Fig. S5D).

Therefore, transcriptomic correlates seem to correspond mostly to HS mutations, and the patterns of HS usage seem to be specific for multiple myeloma as compared with other solid and hematologic cancers.

Functional impact of biallelic events affecting tumor suppressor genes in multiple myeloma

Biallelic events carry an increasingly recognized prognostic effect in multiple myeloma. In particular, double-hit events occurring in tumor suppressor genes have been described to have a functional impact on cell fitness and survival and to represent the reservoir for disease relapse and drug resistance in more advanced stages (10–12, 39). Therefore, we sought to explore the transcriptomic consequences of biallelic events and how these are different from mono-allelic ones in key tumor suppressors: FAM46C, DIS3, RB1, TGDS, TRAF3, CYLD, and TP53.

To this end, we looked at genes with a cumulative effect, i.e., genes significantly deregulated in biallelic versus WT and/or biallelic versus mono-allelic statuses. Overall, among samples of patients with multiple myeloma the mono-allelic status was associated with little transcriptomic change as compared with WT, while biallelic events showed more profound changes (Supplementary Fig. S6; Supplementary Table S1 and Supplementary Data S3). Then, to better explain the putative functional impact of differential gene expression, we performed GO analysis on genes with a cumulative effect. As a common feature, we found a generalized upregulation of genes involved in cell-cycle regulation and cell proliferation. Interestingly, biallelic inactivations of TP53, TGDS, and RB1 upregulated similar cyclins (CCND2, CCNA2, CCNE2) and spindle checkpoints genes (E2F1 and MCMs). Furthermore, they shared also upregulation of aurora kinase-A (AURKA), cyclin-dependent kinases (CDK1 and CDK2) and Polo-kinase-1 (PLK1), genes for which targeted treatments are being trialed (Supplementary Fig. S7). These results were confirmed by a GSEA. Indeed, we observed that cases with inactivation of TP53, TGDS, RB1, or FAM46C shared many cancer hallmarks gene sets, all related to cell-cycle regulation and MYC (Supplementary Fig. S8).

However, each gene aberration also showed a specific transcriptomic profile associated with biallelic inactivation. CYLD double-hit events showed as expected upregulation of downstream effectors of the NF-kB pathway such as BIRC3, TRAF4, NFKB2, IKBKE, RELB, and TNFAIP3 (Fig. 3A, top; Supplementary Fig. S8). Moreover, we also detected an upregulation of genes involved in the interaction with microenvironment as ICAM1, EDN1, and FGF2 related to angiogenesis enhancement (40), or MIR146A implicated in the communication with mesenchymal stromal cells. Interestingly, CYLD inactivation correlated with a significant overexpression of BCL2, with a cumulative effect for biallelic events (Fig. 3A, bottom). In the case of TP53 double-hits, we noted an upregulation of cell-cycle genes as well as PHF19, a recently discovered poor prognostic marker in multiple myeloma (Fig. 3B; ref. 41). Therefore, we further explored this finding applying GSEA to dissect possible interactions between TP53 inactivation and PHF19 overexpression. Interestingly, the two abnormalities showed an overlap in upregulation of cell-cycle related and MYC controlled pathways (Supplementary Fig. S8), suggesting a convergence of mechanisms underlying poor prognosis in multiple myeloma. Furthermore, TP53 double-hits were associated with downregulation of SLAMF7, an immunotherapeutic target in multiple myeloma (Fig. 3B, bottom). Interestingly, this was only observed in double-hit events and not when TP53 was mutated or deleted only. FAM46C double-hit events were characterized by an upregulation of CD38 and a concomitant low expression of CD55 and CD59, recently associated with daratumumab resistance (Supplementary Fig. S7A) (42). Moreover, we found an upregulation of CD44, a gene involved in multiple myeloma cell homing and lenalidomide resistance (43). Then, we analyzed the impact of chr(13q) deletions and mutations of genes located within this region (Supplementary Fig. S7B–S7D). TGDS and RB1 biallelic events deregulated mostly genes involved in cell-cycle and cell proliferation. Double-hit events of TGDS were predominantly associated with gene upregulation, while patients with RB1 biallelic inactivations were characterized by differentially expressed genes in both directions compared with samples with normal/mono-allelic status of the gene. DIS3 alterations correlated with increased expression of lncRNA regulatory genes (Supplementary Fig. S7B–S7D) as expected given its function (38). TRAF3 inactivation was associated with NF-kB pathway activation and downregulation of signaling related to mRNA metabolic processes and protein catabolism (Supplementary Fig. S7E). Overall, these findings may help to explain the putative biological and prognostic impact of biallelic events in multiple myeloma.

Figure 3.

Biallelic inactivation analysis. A and B,CYLD and TP53 analysis. In the top left panel, the z-score value heatmaps of the cumulative effect genes, where columns represent the gene status and rows represent the genes. In the top right panel, the gene-network plot for cumulative effect genes, where big dots indicate the enriched pathways, small dots represent the deregulated genes, and the color-coding refers to the FC in biallelic versus WT contrast (log2CPM). In the lower panel, the box plots represent the cumulative gene expression trend in the three statuses, for each analysis: the dots specify the samples in each status, the red line symbolizes the trend of the gene expression, the red number denotes the incremental or decremental expression FC, the black number the P value.

Chromosome 1q gain/amplifications impacted transcriptome

Another recurrent CNA in multiple myeloma affects chr(1q21), where gains (3 copies) are thought to be early events and amplifications (4 or more copies) late events with a negative prognostic impact (11, 13). As for chromosomal deletions, also in this case we observed several genes characterized by a progressively increased change in expression in diploid versus gain versus amplifications of chr(1q21). While an intermediate differential expression was observed in the gain group compared with the disomic one, greater expression z-scores of these altered transcripts characterized the amplified cohort (Fig. 4A). As this was the single CNA with the largest number of transcriptomic correlates (Fig. 1B), we asked whether alteration of expression only impacted genes in the duplicated region or was genome-wide. Taking into account the number of differentially expressed genes in each chromosome corrected for the chromosomal length, we observed how small chromosomes (i.e., 21 or 22) contributed proportionally as much as chromosome 1 to transcriptomic changes between normal disomic configuration and 1q extra-copies. This suggests that chr(1q21)gain/amp may have profound implications on the transcriptome of the cell which are genome-wide (Fig. 4B). Subsequently, we applied a GO analysis on the genes with 1q CN-associated differential expression. Proliferation-related pathways, i.e., cell-cycle and spindle checkpoint genes were significantly upregulated (Fig. 4C). Chr(1q21)amp was associated with downregulation of CCND1 and upregulation of CCND2 along with other oncogenes (AURKA, AURKB, and PLK1) confirming the view that multiple myeloma is a disease of deregulated cyclins (Supplementary Fig. S9A and S9B; ref. 44). Given the known association between chr(1q21)gain/amp and t(4;14), we asked whether chr(1q21)amp effect on CCND1 and CCND2 could be mediated by t(4;14). Firstly, we noted that despite the presence of 1,102 shared differentially expressed genes between the two genetic abnormalities, chr(1q21)gain/amp was independently associated with 1,191 differentially expressed genes and t(4;14) with 3,841 (Supplementary Fig. S10A). Of note, while CCND1 and CCND2 were differentially expressed by both abnormalities, cases of chr(1q21)gain/amp without t(4;14) also showed CCND2 upregulation (Supplementary Fig. S10B), confirming this is a convergent independent downstream effect of the two abnormalities and not a confounding effect of their association. Other than this, a GO analysis of differentially expressed genes specific to each genetic abnormality did not show shared deregulated pathways (Supplementary Fig. S10C and S10D). Finally, several genes with a possible therapeutic impact were differentially expressed. In particular, PDL1 was downregulated, while the immunotherapy targets SLAMF7 and GPRC5D were significantly overexpressed (Fig. 4D; Supplementary Fig. S9C). Of note, the chromosome 1q amplification was also associated with a significant upregulation of MCL1 (Fig. 4E), whose overexpression is associated with a lower response to BCL2 inhibition. Altogether, these data confirmed the high impact that chromosome 1q alterations have on the transcriptome and of great importance are the highly significant expression changes on genes with possible therapeutical implication for new generation immunotherapies (SLAMF7 and GPRC5D) or targeted treatments (MCL1).

Figure 4.

Chromosome 1q gain/amplification analysis. A, The heatmap represents the z-score values for the cumulative effect genes. B, The pie chart represents the chromosomal distribution of the cumulative effect genes, corrected for chromosomal length. C, Gene-network plot for cumulative effect genes, related to the amplification versus WT contrast (log2CPM). D, Box plot for the SLAMF7 expression trend in the three statuses. E, Box plot for MCL1 expression trend.

Structural chromosomal alterations as possible predictors of response to BCL2 inhibitors

Venetoclax is becoming a commonly used therapeutic approach in hematological malignancies and shows great promise in the multiple myeloma field as well (22, 45, 46). Nevertheless, in multiple myeloma venetoclax use seems to only be beneficial for certain subgroups of patients and is detrimental for others, making it crucial to identify reliable biomarkers (47). Studies have highlighted the BCL2/MCL1 and BCL2/BCL2L1 expression ratios as predictors of response (48). However, in routine clinical practice these analyses are not performed and venetoclax use is confined to the t(11;14) setting, as these cases express high levels of BCL2 (22, 48). Since BCL2 overexpression seems not to be confined to t(11;14) cases, we mined our data to look for additional genomic predictors of response to venetoclax studying the expression levels of BCL2, MCL1, and BCL2L1. Interestingly, as described in mouse models (49), CYLD-deleted cases showed the most significant upregulation of BCL2, while t(11;14) cases did not cross our FDR threshold (Supplementary Table S2) for BCL2 expression unless associated with del(CYLD). In t(11;14), the most notable feature was a downregulation of BCL2L1. Conversely, chr(1q21)gain/amps alone or in association with chr(13) deletions were characterized by an overexpression of MCL1 (Supplementary Table S2 and Fig. 5), likely explaining resistance to venetoclax of chr(1q) cases, who, by contrast, are highly sensitive to MCL1 targeting (50). Indeed, among BCL2 overexpressing cases, there was an enrichment of t(11;14) and del(CYLD) (Supplementary Fig. S11A). MCL1 expression was characterized by an enrichment of chr(1q)gain/amps among high-expressors (Supplementary Fig. S11B). In the case of BCL2L1, it was clearly seen how most cases with low expression were indeed characterized by t(11;14) (Supplementary Fig. S11C). Concerning expression ratios, chr(1q)gain/amps were associated with a significant reduction of the BCL2/MCL1 ratio. On the other hand, the BCL2/BCL2L1 ratio was very high in the t(11;14) setting, explaining the positive effect of venetoclax in this subgroup (Supplementary Table S3).

Figure 5.

BCL2 family gene expression analysis. Box plots representing the specific BCL2 family gene expression in t(11;14), del(CYLD), and chr(1q21)gain/amps samples.

In conclusion, we observed genetic lesions that are associated with a peculiar modulation of expression of genes related to the BCL2 pathway which could translate into biomarkers for venetoclax treatment.

In silico drug sensitivity screening shows that t(11;14) is not the only predictor for venetoclax response

The above results highlighted some genotype–phenotype correlations that may predict response to BCL2 pathway inhibitors. To validate these findings in silico, we interrogated public datasets generated within the DepMap international collaboration, reporting on genetic dependencies and small molecule sensitivities of about 2,000 cancer cell lines (33, 34). To select the most appropriate cell lines for our analysis, we used the Celligner algorithm (35). This allowed the comparison of transcriptomic data from DepMap cell lines and multiple myeloma samples from CoMMpass, and therefore the selection of the cell lines most similar to our primary sample data. Multiple myeloma cell lines are characterized by specific IGH rearrangements, and Celligner correctly clustered them with multiple myeloma samples harboring the same translocation (Fig. 6A). Interestingly, even if HDMM cell lines are hardly reported, many cell lines were anchored with samples of patients with HDMM, which may have translational relevance. Analysis of genomic and expression data showed that this was not due to MYC translocation or overexpression. In the sample space defined by Celligner we selected cell lines harboring the three chromosomal abnormalities that we identified above as possible predictors of venetoclax response (features of cell lines are summarized in Supplementary Table S4). The cell line GRANTA519 and multiple myeloma cell lines SKMM2 and KMS12BM harbor a t(11;14) translocation, and a CYLD deletion in the latters. As expected, these three cell lines resulted highly sensitive to venetoclax treatment (Fig. 6B–D). Then, we interrogated the dataset for the U266 and MOLP8 multiple myeloma cell lines, both carrying a t(11;14). Notably, these two models showed refractoriness to BCL2 inhibition (Fig. 6E and F), most likely since they are both characterized by high expression of MCL1 due to chr(1q)gain/amp, and BCL2L1 gain in the case of U266. The only cell line in DepMap with an isolated del(CYLD) was the RI1, and it was sensitive to venetoclax (Fig. 6G). To gain further evidence that del(CYLD) could predict venetoclax response, we analyzed public data on venetoclax sensitivity reported in recent papers (50–52). By integrating these drug sensitivity data and genomic information retrieved from literature and publicly available repositories, we were able to report the characteristics of 53 additional multiple myeloma cell lines (Supplementary Table S5). Five of these harbored a del(CYLD), but this was isolated only in PCM6 and in the remaining 4 lines was associated with chr(1q21)gain/amps. Adding independent evidence to our hypothesis, the PCM6 cell line resulted highly sensitive to venetoclax (51), while the others were resistant likely due to MCL1 overexpression. In fact, the MCL1 inhibitor S63845 was tested in the AMO1 cell line harboring a chr(1q) amplification and resulted in decreased viability (50). However, the complex genotype of cell lines and the many genetic and epigenetic variables affecting expression represent major challenges to the discovery and wider applicability of candidate biomarkers of venetoclax sensitivity, and validation by functional studies will clearly be required. Extensive mining of geno-transcriptomic data and cancer dependencies may increase the spectrum of candidate biomarkers of venetoclax sensitivity in multiple myeloma.

Figure 6.

In silico drug sensitivity screen. A, UMAP 2D projection of Celligner-aligned sample and cell line expression data: dots represent the samples [HD in light blue, t(4;14) in yellow, t(11;14) in blue, t(14;16) in salmon, and t(14;20) in pink] and the lymphoma (dark blue) and myeloma (red) cell lines, samples missing karyotype in gray. B–G, Dose–response curves for cell lines of interest: orange dots represent tests and blue curves represent the fitted trends. drc, dr4pl R packages have been used to generate drug-sensitivity plots.

Discussion

In this study, we systematically dissected the impact of well known and more recently described genome-wide DNA abnormalities on the expression profile of 514 multiple myeloma cases to identify putative functional correlates and surrogates of gene expression that may have translational implications. As expected, and confirming the validity of our approach, we found that hyperdiploidy and recurrent IGH translocations were associated with the most substantial differential gene expression patterns between patients. The assessment of the transcriptomic profiles associated with any recurrent CNA and gene mutation, representing a hallmark of our analysis, highlighted novel correlates, by means of which we could also identify specific lesions worthy of investigation in clinical-grade diagnostics for their potential implications on patient management. Overall, gene mutations carried a lower transcriptomic impact. This may be explained by the fact that most show instances of convergent evolution (e.g., mutually exclusive mutations of KRAS and NRAS), do not segregate with specific genomic lesions, and are often subclonal. Therefore, their impact may be diffuse and redundant, and furthermore diluted when analyzing bulk sequencing data. Interestingly though, functional relevance seems to be only attributable to hotspot mutations for the two most common mutated genes KRAS and NRAS, and their hotspot usage is quite different between multiple myeloma and other solid cancers. This finding may have functional implications that need to be further studied. A questionable driver event is represented by IGLL5 mutations, as they are mostly seen as off-target AID-induced mutations (53). The fact that in our analysis IGLL5 mutations are not associated with the differential expression of any transcript strongly supports this view. However, they could be a marker of IGL translocations and poor prognosis (2).

The single CNA associated with the largest number of transcriptomic correlates was chr(1q21), where gains and, to a more extent, amplifications predicted downregulation of PDL1 and upregulation of SLAMF7, GPRC5D, and MCL1, with clear implications for treatment with immunotherapies and BCL2 inhibitors. Overall, chr(1q21)gain/amp corresponded to a transcriptional modulation that was genome-wide and not limited to that genomic region, and was associated with the upregulation of a proliferative program driven by CCND2 and other oncogenes. Indeed, a similar pattern has been reported by RNA-seq of proteasome inhibitors (PI) and immunomodulatory agents double-refractory cases (7), making chr(1q)gain/amp one of the worst prognostic markers even in the era of novel treatments (54). Of note, transcriptomic changes associated with chr(1q21) were much more profound in case of amplifications than gains. This was an overall very notable trend also for tumor suppressor lesions, where most of the transcriptomic changes were found associated with biallelic events (i.e., mutation of one allele and deletion of the other). Some of these had therapeutic correlates, such as NF-kB pathway upregulation in TRAF3 and CYLD inactivation, predicting good response to PI treatment (55, 56). Others had prognostic value, such as PHF19 upregulation that was seen upon TP53 biallelic loss (41).

Most importantly, we identified potential genomic correlates of response to BCL2 pathway inhibitors. While response to venetoclax is best associated with a high BCL2/MCL1 or BCL2/BCL2L1 ratio, this is rarely assessed in the clinic, and the easier to identify t(11;14) translocation is the typical biomarker used to select patients that may respond to venetoclax. Our data show how t(11;14) patients display high BCL2 levels and lack chr(1q21)gain/amp events thus showing low MCL1 levels. However, their sensitivity to venetoclax is best explained by their consistently low BCL2L1 levels. In our in silico analysis, t(11;14) was found associated with resistance to venetoclax therapy in 2 cell lines, where it cooccurred with amplification of 1q and high expression of MCL1. In this setting, Slomp and colleagues have shown how 1q amplifications could identify a high-risk patient subset suitable for therapy with specific MCL1 inhibitors. Interestingly, the authors have also shown how the combinatorial therapy with BCL2 and MCL1 antagonists could overcome venetoclax resistance in human myeloma cell lines as well as the newly diagnosed patients (50). Of note, the double inhibition leads to the formation of proapoptotic BAX-BAK hetero-complexes promoting apoptosis activation (52), and this could be the mechanism restoring apoptosis in these cells. Our systematic analysis also allowed to identify other DNA lesions that may predict venetoclax response, and among these are CYLD deletions. These are not significantly associated with t(11;14) (10, 14), and are associated with the highest BCL2 levels of the CoMMpass cohort. Consistently, our functional in silico validation confirmed that del(CYLD) cell lines respond to venetoclax and that this could represent a candidate biomarker to identify patients who could respond to BCL2 inhibition. Nevertheless, within multiple myeloma cell lines, CYLD deletions were more variably associated with venetoclax sensitivity, owing to complex karyotypes with additional chromosomal abnormalities involving BCL2 losses and chr(1q21)gain/amp associated with high MCL1 expression. Therefore, while our data from primary samples recapitulate in vivo mouse models where CYLD knockdown resulted in BCL2 upregulation (49), the translation of this evidence into a suitable biomarker of venetoclax sensitivity in the context of plasma cell neoplasms is hampered by the scanty cell line models with a clean genetic background, and likely by non-genetic mechanisms of resistance (57), and will require additional studies. In search for additional predictors of venetoclax sensitivity, Gupta and colleagues recently highlighted a new gene/antigen signature that is independent of t(11;14) (51), and whose clinical utility will need testing.

Altogether, our study highlights the impact of detailed molecular profiling partnered with clinical annotations and how, when combined with large datasets of patients, this can be leveraged to study outcomes of targeted drugs in well defined biological subgroups of patients with multiple myeloma.

Authors' Disclosures

S. Oliva reports personal fees from Celgene/BMS, Janssen, and Amgen outside the submitted work. M. D'Agostino reports personal fees from Sanofi and GSK outside the submitted work. O. Landgren reports grants and personal fees from Amgen, Janssen, Karyopharm, and Binding Site; grants from Celgene, Takeda, MMRF, Perelman Family Foundation, and NCI; personal fees from Adaptive Biotech, Bristol-Myers Squibb, Cellectis, Oncopeptides, and Pfizer; and other support from Janssen, Theadex, and Merck outside the submitted work. F. Iorio reports grants from Open Targets and other support from Joint CRUK-AstraZeneca Functional Genomics Center outside the submitted work. N. Bolli reports grants from European Research Council during the conduct of the study as well as personal fees from Celgene-BMS, Janssen, Amgen, and Takeda outside the submitted work. No disclosures were reported by the other authors.

Authors' Contributions

B. Ziccheddu: Data curation, software, formal analysis, validation, visualization, methodology, writing–original draft. M.C. Da Vià: Data curation, formal analysis, validation, visualization, methodology, writing–original draft, writing–review and editing. M. Lionetti: Data curation, formal analysis, validation, visualization, methodology, writing–review and editing. A. Maeda: Investigation, methodology. S. Morlupi: Investigation, methodology. M. Dugo: Software, formal analysis, methodology. K. Todoerti: Software, formal analysis, methodology. S. Oliva: Data curation, methodology. M. D'Agostino: Data curation, methodology. P. Corradini: Data curation. O. Landgren: Data curation. F. Iorio: Software, supervision, visualization. L. Pettine: Data curation. A. Pompa: Data curation. M. Manzoni: Software, formal analysis, investigation, methodology. L. Baldini: Supervision. A. Neri: Supervision, funding acquisition. F. Maura: Conceptualization, supervision, writing–review and editing. N. Bolli: Conceptualization, resources, supervision, funding acquisition, project administration, writing–review and editing.