Ovarian Cancers Harboring Inactivating Mutations in CDK12 Display a Distinct Genomic Instability Pattern Characterized by Large Tandem Duplications

Cyclin-dependent kinase 12 (CDK12) as a tumor suppressor gene was discovered in The Cancer Genome Atlas (TCGA) study of high-grade serous ovarian carcinoma (HGS-OvCa; ref. 1). It appeared among the 10 most recurrently mutated genes in HGS-OvCa, which, besides CDK12, included TP53, BRCA1, BRCA2, NF1, and RB1. CDK12 encodes a serine/threonine kinase involved in the regulation of RNA polymerase II (RNA pol II) and mRNA processing (2–4). CDK12 attracted particular attention as cells inactivated for the CDK12 display hypersensitivity to DNA-damaging agents and to PARP1/2 inhibitors (5, 6). This effect of CDK12 inactivation was associated with homologous recombination (HR) deficiency (HRD) due to decreased expression observed for some HR genes, such as BRCA1, FANCI, or FANCD2 (3, 5, 7, 8). HRD due to inactivation of BRCA1 or BRCA2, the two major genes implicated in ovarian cancer, leads to the large-scale chromosomal instability, readily observed in nearly half of HGS-OvCa (9–11). The essential question is what type of genomic instability accompanies CDK12 inactivation.

Here, we show that ovarian tumors inactivated for CDK12 have a particular genomic instability with genomic profiles characterized by frequent mega-sized gains scattered over the genome, and this remarkable genomic architecture is due to numerous tandem duplications (TD).

Controlled access data are from (i) serous ovarian carcinoma cohort (HGS-OvCa): 556 cases with SNP arrays including 527 cases with whole-exome sequencing (WES) or transcriptome (RNA-seq) sequencing data (1); (ii) breast carcinoma cohort: 760 cases with SNP arrays and WES/RNA-seq (12); (iii) prostate adenocarcinoma cohort: 407 cases with SNP arrays and WES/RNA-seq (12, 13). All human data were handled according to Data Access Request.

Cohort includes 95 primary ovarian tumors of mainly serous subtype assembled in Institut Curie (14). According to French regulation, patients were informed of the research performed on tissue specimens and did not express opposition.

Data from public repositories include (i) GEO (GSE18461 and GSE29398) and Sequence Read Archive (SRA ERP000703 and ERP000704; ref. 15); and (ii) European Genome-phenome Archive (EGA; controlled access EGAS00001000155; ref. 16).

SKOV3 cell line obtained from the ATCC was verified by reference genomic and mutational profiles at the time of experiment. SNP array was obtained from the Cancer Cell Line Encyclopedia (17).

SNP arrays were normalized using ChAS, Genotyping Console, or Genome Studio software (18, 19). Absolute copy number and allelic contents were obtained using the Genome Alteration Print (GAP; ref. 20). DNA index was calculated as the averaged copy number. Tumor ploidy was set to 2 (near-diploid tumors) if DNA index <1.3 or to 4 (near-tetraploid tumors) if DNA index ≥1.3. Relative copy-number profile was calculated as ratio of copy-number/DNA index. The interstitial gains were calculated after filtering microalterations (<50 SNPs in size) and following a step-wise procedure starting with the smallest segments between two consecutive breakpoints. The number of large-scale state transitions (LST), representing the number of breakpoints between large chromosome fragments, was calculated as previously described (9).

PCR and Sanger sequencing were performed using standard protocols available upon request.

MP2, MP3, and MP4 primary ovarian tumors and the cell line SKOV3 were sequenced using whole genome sequencing (WGS) mate-pair technique on the platform sequencer Illumina HiSeq 2500 with read length of 100 × 100 bp and approximately 40× physical whole-genome coverage. Mate-pair library was constructed according to the manufacturer's protocol: fragmentation by Covaris (80 seconds), 30 mn Tagmentation, no size selection, and approximately 3,000 bp distance between mates.

Mate-pair WGS and paired-end WGS were aligned using Bfast and/or BWA 0.7.0a (21, 22) and processed using Samtools (23), BedTools (24), Picard tools (25), and Integrative Genomic Viewer (26), including SVdetect (27) and DELLY (28) for structural and Control-FREEC (29) for copy-number alterations, visualized by Circos (30). Breakpoint junctions were extracted from split-reads. Mutations were annotated by Annovar (31).

Analysis was performed on Affymetrix geneChip HG-U133A(B) and RNA-seq gene-level RPKM (reads per kilobase per million mapped reads). Expression data were obtained from public repositories (TCGA: https://tcga-data.nci.nih.gov/tcga and in-house: http://microarrays.curie.fr/publications/biopathology/ovarian_carcinoma; ref. 14). After preprocessing (Supplementary Methods), the reference subset of tumors was compiled from five nearest neighbors to each CDK12-mutated case detected in the subspace of five principal components, where principal component analysis was applied to the subgroup of non-BRCA1/2 nonmesenchymal tumors.

Statistical analysis was performed using R (32).

SNP-array analysis of 556 HGS-OvCa cases from the TCGA (1) showed a subset of 17 tumors with extremely altered genomes characterized by a high number of gains over the chromosome baseline (Fig. 1A). GAP patterns (a two-dimensional plot of segmental copy number and allelic difference values; ref. 20) ensured the consistency and excluded noise artifacts (Fig. 1B). Besides the overall high number of breakpoints in copy-number profiles, this group of tumors was found to be best characterized by the high number of interstitial gains of 2 to 7 megabases (Mb) in size (Fig. 1C and D). Inspection of WES and/or RNA-seq data showed that 15 of 17 tumors with the intensive gain pattern were mutated for CDK12 (Table 1; Supplementary Fig. S1). CDK12 mutations were deleterious (nonsense mutations or frameshifts due to indels, 9 cases; 1 splice) or with unknown consequences (3 missense mutations and 2 in-frame deletions in the kinase domain; Fig. 1E). All mutations were associated with LOH in the gene locus and showed prevailing allele frequency. As for the two remaining cases, one tumor was not found mutated in CDK12 based on WES, and one tumor had no sequencing data. The mutational analysis of CDK12 in tumors not displaying the intensive interstitial gain pattern (511 HGS-OvCa cases with sequencing data available) showed five missense mutations and three rare SNPs. Thus, there is a significant association between CDK12 mutations and the highly altered genomes with intensive gain pattern (P value < 10⁻¹⁵, Fisher test).

Quantitatively, tumors with frameshift and nonsense mutations in CDK12 were characterized by more than 300 breakpoints (509 ± 72, median ± median absolute deviation) and more than 30 interstitial gains of 2 to 7 Mb (65 ± 12), which on average appeared to be 5 and 20 times more abundant compared with tumors not mutated for CDK12 (104 ± 44 breakpoints and 3 ± 2 interstitial gains of 2 to 7 Mb, respectively). Missense mutations in CDK12 that display the same phenotype as frameshift/nonsense mutations were further considered as deleterious (Table 1; Fig. 1E). Tumors with CDK12 missense mutations and low rate of interstitial gains were classified as not bearing a genomic phenotype associated with CDK12 inactivation, despite occasionally high level of genetic alterations (Supplementary Fig. S2). Importantly, the recently reported functional analysis of CDK12 mutants (8), which included four of missense variants mentioned above, was concordant with our conclusion: two CDK12 missense mutants not associated with an intensive regular gain pattern were shown to conserve their CCNK (cyclin K) interaction and/or kinase properties, whereas two CDK12 missense mutants associated with such a genomic pattern were demonstrated to be pathogenic (Table 1).

For validation, in-house cohort of 95 ovarian carcinomas was investigated, and 4 tumors displaying a genome profile characterized by the high number of interstitial gains were detected. All these 4 tumors were found mutated for CDK12 (three truncating mutations and one missense mutation in the protein kinase domain; all tumors displayed LOH in CDK12 locus; Fig. 1D; Supplementary Fig. S3).

It is worth noting that (i) one of the in-house CDK12-mutated cases was a low-grade ovarian tumor (supported by 10-year overall survival of the patient without tumor progression); (ii) two cases from the TCGA HGS-OvCa cohort were found with marginally low expression and increased level of CDK12 promoter methylation; these cases, however, did not display highly altered genomic pattern (Table 1); and (iii) inactivating mutations in CCNK (a key partner of CDK12) were not found in the TCGA cohorts (cBioPortal; ref. 13).

Biologic materials from 3 primary ovarian tumors with deleterious CDK12 mutations from the in-house cohort were investigated by mate-pair WGS (Supplementary Methods). The profiles of structural rearrangements in CDK12-inactivated tumors demonstrated overwhelming prevalence of TDs and consistent overlap between the regular gains in the SNP-array profile and the TDs detected by WGS (Fig. 2B; Supplementary Fig. S4; Supplementary Tables S1 and S2).

The lowest number of TDs, 186, was found in MP2 case with a near-diploid genomic content (DNA index = 0.98), where TDs covered 15% of the genome. The near-tetraploid cases MP3 and MP4 (DNA index = 1.78 and 1.92, respectively) had more abundant TDs detected: 797 and 384, with 35% and 30% of the reference genome affected by TDs, respectively. The size distribution of TDs was bimodal with the modes at approximately 0.3 and 3 Mb (Fig. 2C; Supplementary Fig. S5A). It is worth noting that numerous copy-number gains with complex shapes observed in SNP-array profiles mostly corresponded to several TDs affecting a common genomic region (Fig. 2B). When two intersecting TDs are located on the two different alleles, the sizes of TDs reported from mate-pair analysis correspond to the actual sizes of rearrangements. Intersecting TDs on the same allele could give rise to various patterns of copy-number alterations depending on their relative timing and positions. In this context, the size of reported TDs either fits to the actual size or underestimates it (Fig. 2D). Massive appearances of TDs on a near-tetraploid background (four alleles subjected to segmental TDs) form highly altered tumor genomic profiles characteristic of CDK12 inactivation.

The junction sequences covering the breakpoints (621 split-reads corresponding to ∼45% of all detected TDs) showed overlapping microhomologies with a mode equal to 2 nucleotides (P < 10⁻¹⁵, χ² test compared with random sampling; Fig. 2E; Supplementary Fig. S5B). Only a small fraction (<4%) of junctions showed the sequence homology in marginal DNA fragments (300 bp were tested using the Blat software and a minimal reported size of 30 bp).

To conclude, mate-pair sequencing showed that the regular gains observed in CDK12-inactivated ovarian tumors were almost exclusively due to TDs; TDs are characterized by a bimodal size distribution (with modes at ∼0.3 and ∼3 Mb), displaying overlapping microhomology at the breakpoints (with mode at 2 bp) and showing no significant homology between fragments in the vicinity of the breakpoint junctions.

After the description of TDs in one of the first WGS studies of breast cancers (33), two studies specifically addressed TDs in ovarian cancers (15, 16). We reanalyzed publicly accessible data from these studies focusing on CDK12 and on the size of TDs reported. Two cell lines (PEO14 and PEO23, derived from the same patient at diagnosis and relapse, respectively) described to have a tandem duplicator phenotype with a high frequency of TDs < 1 Mb (15) did not show any marked rate of interstitial gains of 2 to 7 Mb and were not found mutated in CDK12 (Fig. 1D; Supplementary Fig. S6). However, the primary tumor PD3722a (16) resembled the CDK12-inactivated cases (Supplementary Fig. S7). Amplification of the CDK12 locus allowed us to discover a nonsense CDK12 mutation (c.811A>T/p.R271X) in this tumor despite the overall low sequencing coverage. Thus, the frequent TDs described in PD3722a tumor were related to CDK12 inactivation (16).

In order to clarify the specificity of CDK12-inactivated tumors, we analyzed 25 TCGA HGS-OvCa cases with paired-end WGS available. Two tumors displayed the highest TDs rate: TCGA-24-1466 (CDK12 missense mutation; 346 TDs) and TCGA-13-1487 (CDK12-wildtype; 116 TDs; Supplementary Table S3; Supplementary Fig. S8A). The distribution of TD sizes in the CDK12-mutated tumor was bimodal, similarly to the in-house cases, whereas a unimodal distribution in the second case yielded >90% of TDs to be <1 Mb (Supplementary Fig. S8B and S8C). It is important to mention that the number of small-size TDs (1–30 kb) did not show any significant increase in the CDK12-inactivated case TCGA-24-1466 as compared with the 24 other cases (Supplementary Table S3).

Joint mutational analysis in PEO14/PEO23 cell lines with 10 ovarian and 3 breast cancer cases selected from the TCGA cohorts for accumulation of interstitial gains <1 Mb did not show any recurrent gene impairment explaining the TD phenotype (Supplementary Fig. S9). It is worth noting that among 760 breast tumors, only one case resembled CDK12-inactivated tumors but no CDK12 mutation was found (Supplementary Fig. S10A); in this case, one cannot exclude another inactivating mechanism, such as a translocation involving CDK12 (34).

We hypothesize that at least two different phenotypes are associated with frequent TDs in a tumor genome: the first one is the TD phenotype consisting of a high rate of TDs <1 Mb and the second one is associated with inactivation of CDK12 and is characterized by TDs of up to 10 Mb, which could be subsequently designated as TD-plus phenotype (or CDK12 TD-plus phenotype). Besides these phenotypes, TDs are frequent structural alterations observed in many tumor genomes at low/moderate frequencies. The CDK12 TD-plus phenotype occurs at a rate of 3% to 4% in serous ovarian tumors; it consists of hundreds of TDs of up to 10 Mb leading to the duplication of more than 10% of the genome, which is remarkably distinct from what is observed in other ovarian or breast cancer genomes.

Exploring the TCGA cohorts for CDK12 mutations, we found prostate adenocarcinoma cases with putative biallelic CDK12 inactivation (cBioPortal; ref. 13). Recurrent CDK12 mutations were indeed recently reported in advanced prostate cancer (35). Processing 407 SNP arrays from the TCGA prostate adenocarcinoma cohort revealed seven cases (1.7%) with CDK12 TD-plus phenotype, where CDK12 inactivation was confirmed in five cases (Supplementary Fig. S10B). Thus, CDK12 TD-plus phenotype is not restricted to the serous ovarian carcinoma, but it also occurs in 1% to 2% of prostate adenocarcinoma.

Genomic localization of TDs could indicate the origin of genomic instability in CDK12-mutated tumors, so numerous presumably relevant genomic features were tested for their association to TDs (Supplementary Methods). However, we found overall even distribution of TDs over the human genome, indicating a stochastic character of TD appearance. (i) The GC content at the TDs was similar to the neighboring genome segments as well as to the whole genome (Supplementary Fig. S11A). (ii) Duplicated segments were not found accumulating more somatic mutations compared with the not affected genomic segments (Supplementary Fig. S11B). (iii) TDs did not show any local concentrations or visible heterogeneity within chromosomes (Supplementary Fig. S12). The only TD's enrichment in 3q26 corresponds to the commonly gained region in ovarian carcinomas targeting MECOM (1) and is likely to reflect a selection process underlining a probable role of TDs in oncogenesis. (iv) TDs were not found associated to any of tested chromatin structure and organization marks available in the ENCODE project (Supplementary Fig. S13); however, it might be inconclusive in the absence of ovarian cancer cell line profile. (v) Breakpoints in CDK12-mutated tumors were not associated with CpG islands and not enriched for any repeated elements. (vi) The position of TD breakpoints was not directly associated with gene or expressed gene loci (∼44% breakpoint junctions found inside RefSeq genes did not differed significantly from that of randomly sampled TDs giving 41 ± 1.4%; Supplementary Fig. S14).

In conclusion, duplicated regions in CDK12-mutated tumors appear independent from the major genomic features tested, suggesting some general deregulation of the genome maintenance upon CDK12 inactivation during tumor development.

Effect of CDK12 knockdown in cellular models has been consistently associated with HRD (3, 5, 7, 8). HRD in breast and ovarian cancers is mostly related to BRCA1/2 inactivation (9–11). The recently developed genomic signatures of HRD relying on the number of large-scale chromosomal breaks rather efficiently predict BRCA1/2 inactivation (36). With our discovery of CDK12 TD-plus phenotype, it became clear that genomic instability in CDK12-inactivated tumors differs from that of BRCA1/2-inactivated tumors. Moreover, the CDK12 TD-plus phenotype in the TCGA cohort is mutually exclusive with BRCA1/2 mutation (germline or somatic) and with BRCA1 promoter methylation (P < 0.001, Fisher test). To explore whether the genomic profiles of CDK12-inactivated tumors carry the footprint of HRD, we evaluated the genomic HRD signatures of these tumors (Supplementary Methods). To be noticed, HRD signatures appeared to be sensitive to TD-plus phenotype, and it was necessary to perform extensive filtering of interstitial gains for consistent predictions (Supplementary Fig. S15). According to the LST signature, TD-plus phenotype could display a genomic HRD (Fig. 3A; Table 1; Supplementary Fig. S16; ref. 9). However, the majority of cases with CDK12 TD-plus phenotype do not reach the threshold for genomic HRD (Fig. 3B). CDK12-inactivated tumors resemble the group of non-BRCA1/2 tumors and mostly (∼60%, 13/21) display the number of large-scale alterations (LST number) below the minimal level of BRCA1/2-inactivated tumors. Thus, genomic HRD associated with BRCA1/2 inactivation is not a general feature of the CDK12 TD-plus phenotype.

In addition, ovarian carcinomas with genomic HRD are known to be more sensitive to standard-of-care platinum therapy and to display better overall survival of the patients compared with non-HRD tumors (14, 37). Although in small number, CDK12-inactivated tumors did not show any tendency to better overall survival compared with ovarian cancers without genomic HRD (P > 0.3, log-rank test; Fig. 3C).

To conclude, genomic instability in CDK12-inactivated ovarian tumors is clearly different from the genomic HRD associated with BRCA1/2 inactivation. However, only direct evaluation of the HR pathway in cancer cell lines (or tumor samples) with CDK12 deleterious mutations will allow definitive conclusion about the status of the HR pathway.

Having defined tumors with the CDK12 TD-plus phenotype, we performed transcriptomic analysis in order to find genes and/or pathways affected by CDK12 inactivation. We first explored the molecular features of CDK12-inactivated tumors in order to minimize confounding factors in the context of highly heterogeneous HGS-OvCa cohort. Analysis of four molecular subtypes defined in HGS-OvCa (mesenchymal, proliferative, immunoreactive, differentiated; ref. 1) showed that tumors with CDK12 TD-plus phenotype do not cluster with the mesenchymal subtype (P < 0.02, Fisher test; Supplementary Fig. S17). Tumors with CDK12 TD-plus phenotype were not found inactivated for BRCA1/2 (as already mentioned above) or CDKN2A (found in 27% of nonmesenchymal HGS-OvCa, P < 0.01, Fisher test) and mostly expressed high level of CCNE1 (Supplementary Fig. S18A). Differential analysis of 17 CDK12 TD-plus tumors versus 57 matched cases from the reference group (see Material and Methods) showed 222 differentially expressed genes (P < 0.01, ∼25% false discovery rate; Supplementary Fig. S18B; see Supplementary Table S4 for the gene list). Sample clustering based on the top 34 deregulated genes (P < 0.001 and absolute log fold change > 1) supported the association of the gene list with CDK12 inactivation: 2 CDK12 methylated cases (without TD-plus phenotype and included in the reference group) clustered together with CDK12-mutated cases, whereas only one case with TD-plus phenotype and without CDK12 mutation did not (Fig. 3D). These results were largely consistent between platforms (RNA-seq and Affymetrix) and between cohorts (TCGA and in-house; Supplementary Fig. S18C). However, pathway analysis did not show any significant enrichment. In contrast with CCNK/CDK12 knockdown experiments, large genes were not markedly abundant among the downregulated ones; TERF2, FANCI, ORC3, and ERCC1 were the only intersection with DNA repair genes mentioned in refs. 3, 5. From the other hand, significant enrichment in large genes was observed for the set of genes positively correlated to CDK12 (P < 10⁻¹⁵, t test, TCGA RNA-seq; Supplementary Fig. S19). For example, ATM showed a moderate correlation to CDK12, and its marginal relative downregulation in CDK12-inactivated cases could be observed (P < 0.02; Supplementary Fig. S20). To be noticed, ATM also had a decreased expression in the three CDK12-inactivated cases from the in-house cohort (fold change > 2; P > 0.05).

The absence of pathway enrichment in expression analysis could be partially explained by the significant input from the differential copy-number alterations: 25% to 30% of differentially expressed genes followed the direction of change in relative copy numbers (P < 0.05, t test; Supplementary Table S4). The only one exception displaying downregulation and significantly higher relative-copy number in CDK12-inactivated tumors is CSTF3 (3′ pre-RNA cleavage stimulation factor, chr11p13). CSTF3 has direct functional connection with CDK12 being involved in the 3′ end formation of pre-mRNA, a process coupled with CDK12-dependent Ser2 CTD phosphorylation (38). The top downregulated gene in CDK12-mutated tumors, validated in all cohorts, is RPRD1A (Regulation of Nuclear Pre-mRNA Domain Containing 1A), which interacts with phosphorylated CTD of the RNA pol II and participates in dephosphorylation of the CTD (39). Consistent downregulation of CSTF3 and RPRD1A may evidence a compensatory mechanism developed by tumors upon CDK12 inactivation, which would explain the different gene sets obtained in tumors and cellular models.

Microhomology at the breakpoint junctions of structural genomic rearrangements is a rather common phenomenon (40). Our analysis based on the 25 WGS cases from the TCGA HGS-OvCa cohort demonstrated that microhomology detected at the breakpoint junctions of TDs in the CDK12 TD-plus phenotype was in general similar to that found in TDs, translocations, and deletions in CDK12-wildtype ovarian tumors (Fig. 3E; Supplementary Table S3). However, microhomology size distribution in CDK12 TD-plus phenotype tumors showed significant difference from that in other structural alterations due to a more pronounced peak at 2 bp (P < 0.0014, χ² test).

Small overlapping microhomology at the TD breakpoint junctions in SKOV3 cell line (mate-pair WGS) displayed feeble bimodality with the most frequent microhomology being at 1 bp overlap, which also differed significantly from CDK12-mutated tumors (Supplementary Figs. S21 and S22).

Altogether, the abnormal junctions in ovarian tumors were found to have overall similar features, whether they arose from large or small TDs or from translocations and deletions; the results are concordant with those obtained previously (33).

It is not clear which pathway is employed for TD resolutions, as both nonhomologous end joining frequently operating with 1 to 4 bp microhomology (41) and microhomology-mediated end joining utilizing longer 5 to 25 bp stretches of homologous fragments are consistent with the observed microhomology size distribution (40).

Here, we reported a new genomic instability pattern consisting of a high number of TDs up to 10 Mb in size and associated it with inactivation of CDK12 gene. Three ovarian tumors mutated in CDK12 that were investigated by mate-pair WGS exhibited 200 to 800 TDs per tumor; these TDs were quasi-randomly distributed along the genomes and affected more than 10% of the genomic content. This genomic instability phenotype was denoted here as the CDK12 TD-plus phenotype. The CDK12 TD-plus phenotype was observed in around 4% of serous ovarian carcinomas and in 1% to 2% of prostate adenocarcinomas. Analysis of more than 1,000 breast cancers identified only one case resembling the TD-plus phenotype, which was not found mutated for CDK12.

This novel genomic instability observed in CDK12-inactivated ovarian and prostate cancers is the first described connection between massive genomic TDs and gene impairment in tumors. This finding gives a new perspective on the gene function and has direct connection to clinical application. The outstanding questions emerging from our study are which biologic mechanisms are responsible for the TDs formation and how impaired CDK12 leads to these genomic defects.

Analysis of a number of particular genomic features within and around TDs in the CDK12-mutated context gave no preference for TDs localization evidencing some general deregulation and stochastic process. The overwhelming prevalence of gains versus losses and absence of homologous fragments around the breakpoints excluded any “symmetric” mechanisms, such as misalignment of the sister chromatids or unbalanced translocations. Two processes leading to TDs have been described so far: DNA re-replication due to origin re-firing (42) or break-induced replication (BIR) associated to the strand invasion (43). The sizes of TDs resulted from BIR activated by cyclin E (CCNE1)–induced replication stress were measured to be hundreds of kb with the median around 200 kb and importantly, near one third of reported alterations were deletions (44). An increased rate of DNA double-strand breaks and release of DNA fragments consecutive to re-replicating fork collisions were shown in model systems of induced re-replication (45, 46). Resolution of these breaks could result in TDs (42). CDK12 TD-plus phenotype seems unlikely to be due to a BIR process, as the proportion of deletions in a CDK12-null context is much smaller than in the BIR-induced system. The re-replication hypothesis would seem more likely but does not account for the megabase size of CDK12-related TDs. The sizes of TDs are close to the size of the replication domains, which is too large to be replicated from the re-fired single origin and could suggest the re-initialization in the whole replication domains. This process has never been described in cellular models, whether yeast or mammalian.

The roles and consequences of CDK12 inactivation in ovarian cancers were previously addressed in several studies. DNA repair impairment and PARP inhibitor hypersensitivity observed upon CDK12 knockdown were explained by the indirect role of CDK12 affecting the transcription of key factors of the HR pathway (5, 6, 8). We addressed the CDK12 and HR connection by looking at the genomic footprint of HRD. Genomic HRD (as defined by the genomic signature) was found in approximately 50% of serous ovarian cancers and beside BRCA1/2 was linked only to RAD51C inactivation (11). We show here that after filtering interstitial gains, CDK12-inactivated tumors mostly display less large-scale chromosomal breaks than BRCA1/2-inactivated ovarian cancers. Thus, genomic instability associated with CDK12 inactivation does not display substantial genomic HRD in addition to massive TDs.

This raises further questions such as: does HR downregulation observed in cellular models upon CDK12 knockdown represent the major effect of CDK12 inactivation in tumors causing TD-plus phenotype (for example, TDs are shown to be abundant in HCC38 BRCA1 methylated cell line; ref. 33); or are there alternative/additional targets of CDK12 inactivation affecting genomic integrity? In depth, transcriptomic analyses did not show any candidate gene to explain the phenotype, but identified possible compensatory mechanisms (downregulation of CSTF3 and RPRD1A) that could have rescued gene expression changes since CDK12 inactivation.

Beside the role of CDK12 on gene expression, one could hypothesize that sudden CDK12 inactivation may cause transcription stalling, which may interfere with replication forks, generating the observed TDs. However, this hypothesis was not supported by our analyses. The majority of TDs exceeded in size the transcribed sequences, and no correlation between TD positions and gene loci was observed. Nevertheless, the general role of transcriptional stress on TD formation (due to BIR or re-replication) could not be excluded, and relationship between transcriptional stress and HR has recently been established (47).

Another intriguing observation is the two ovarian cancers found to have a total silencing of CDK12 by promoter methylation with concordant transcriptomic change and without TD-plus phenotype. The timing of CDK12 inactivation in the transformation process could be a critical parameter for the onset of genomic instability.

To conclude, our finding of a CDK12 TD-plus phenotype poses many direct questions and opens a new perspective in deciphering CDK12 function, its role in oncogenesis, and genomic instability, and the recently established analogue-sensitive CDK12 cell line will be instrumental in this regard (48).

T. Popova has expert testimony in patent licensed to Myriad Genetics. E. Manié reports receiving other commercial research support from Myriad Genetics. M.-H. Stern reports receiving other commercial research support from Myriad Genetics. No potential conflicts of interest were disclosed by the other authors

Conception and design: T. Popova, X. Sastre-Garau, M.-H. Stern

Development of methodology: T. Popova

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): E. Manié, N.K. Smith, O. Mariani, X. Sastre-Garau

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): T. Popova, E. Manié, V. Boeva, N.K. Smith, C.R. Mueller

Writing, review, and/or revision of the manuscript: T. Popova, E. Manié, V. Boeva, C.R. Mueller, M.-H. Stern

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): A. Battistella, O. Goundiam, V. Raynal, O. Mariani

Study supervision: T. Popova, M.-H. Stern

The authors thank Andrey Zinovyev, Benoit Le Tallec, Sarah Lambert, Arturo Londono, Alain Nicolas, Michèle Debatisse, Sophie Loeillet, Chunlong Chen, and Olivier Bensaude for helpful discussions and/or critical readings of the draft.

This study was supported by Institut National de la Santé et de la Recherche Médicale, Cancéropôle Ile-de-France, INCa, Institut Curie and its Translational Research Department. A. Battistella is supported by a grant from INCa.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.