Introduction

Large scale genomic analyses of localized and metastatic prostate cancers (PC) have identified a large spectrum of recurrent somatic alterations that involve the activation of oncogenic signaling pathways or the inactivation of tumor suppressor processes (1–5). For example, the androgen receptor (AR) serves as a key therapeutic target for most metastatic PCs, and recurrent somatic events that drive persistent AR activity promote treatment resistance and the emergence of metastatic castration resistant PC (mCRPC). A subset of other recurrent genomic alterations observed in mCRPC confer differential sensitivity to specific treatments: notably, mutations in genes involved in homology directed DNA repair (HR), such as BRCA2, confer responses to poly ADP ribose polymerase (PARP) inhibitors, and mutations in genes involved in DNA mismatch repair such as MSH2 and MSH6 associate with exceptional responses to immune checkpoint blockade (6–8). In addition to other frequent genomic aberrations that include gene fusions involving TMPRSS2 and ERG, mutations in TP53, and loss of PTEN, there is a ‘long tail’ of genes altered in 3-10% of mCRPCs (5). Though by definition, genes comprising the ‘long tail’ involve smaller subsets of patients, the high prevalence of PC in the population means that events occurring at low frequency will still affect thousands of individuals. The classification of tumors with both common and rarer genomic alterations may aid prognosis and prioritize the allocation of treatments.

The gene encoding cyclin dependent kinase 12 (CDK12) is functionally compromised by bi-allelic mutation or copy loss in about 5% of mCRPC cases (and in 1-2% of primary prostate cancers)(1, 9, 10). CDK12 is also lost in 3-4% of ovarian cancers (10, 11). CDK12 is a transcription-associated kinase that pairs with Cyclin K (CCNK) to form an active complex that phosphorylates the RNA Polymerase II (RNAP II) C-terminal tail (12–14). A germline Cdk12 knockout is embryonic lethal in mice (15). CDK12 has been reported to regulate mRNA splicing, suppress upstream intronic polyadenylation sites (IPAs), and maintain transcriptional elongation, especially for very large genes (16–20). Several genes involved in DNA repair, especially members of the HR pathway, are large (i.e. >50kb) and have been reported to be selectively downregulated in CDK12 loss models (15, 16, 18, 21). This has led to a proposed outcome whereby CDK12 loss in patient tumors may phenocopy HR deficiency (HRd). This is of notable clinical relevance as HRd tumors respond well to genotoxic platinum chemotherapy and PARP inhibitors (PARPi), which have proven to be effective across a range of cancers with underlying HR gene mutations, including mCRPC (6, 22–25).

Prior studies have evaluated the consequences of CDK12 deficiency in a variety of experimental models, though nearly always under acute loss conditions including: protein degraders (12, 26, 27), CDK12 knockdown or genetic knockout (15, 16), or treatment with small molecule CDK12/13 inhibitors (18, 21, 28). Acute CDK12 loss in cell models results in the down-regulation of HR associated genes, suppression of HR-mediated DNA repair, and synthetic lethality with PARP1/2 inhibitors (15, 21, 29). However, an HRd phenotype resulting from CDK12 loss has not been confirmed in assessments of patient tumors or clinically with therapeutics that exploit HRd, and there are conflicting reports on the presence of HRd-associated genomic ‘scars’ in CDK12 mutant cancers (9, 30–33). Crucially, mCRPC patients with CDK12-alterations have shown poor responses to PARPi, despite the fact that CDK12 mutation is a labeled indication for the PARPi olaparib (34–36). A notable deficiency in the field has been the lack of models with a stable CDK12^-/-genotype, as CDK12 loss is poorly tolerated and attempts to generate long-term stable cell lines models have failed with the exception of an engineered CDK12^-/- ovarian cancer line, which notably does not exhibit cisplatin or PARPi sensitivity (37–39). The objectives of this study were to address the molecular and phenotypic discordance between the preclinical studies associating CDK12 loss with HRd, and in vivo human pathobiology, and identify vulnerabilities in PCs with CDK12 loss that have potential applications for clinical management.

Results

Identification of genomic characteristics that associate with CDK12 loss in prostate cancer

To ascertain genomic and phenotypic alterations that associate with biallelic CDK12 loss (CDK12^BAL) in PC, we analyzed several large datasets where deep molecular assessments of tumors included analyses of genomic alterations by whole exome sequencing (WES) or whole genome sequencing (WGS), and metrics of gene expression by RNAseq. Four datasets with these criteria were evaluated: the SU2C/PCF International study of mCRPC comprising 442 tumors (SU2C-I), the SU2C/PCF US West Coast study of mCRPC comprising 101 tumors (SU2C-WC), the Hartwig Foundation study of metastatic PC comprising 168 tumors (HMF), and the University of Washington Autopsy study of mCRPC comprising 269 tumors from 127 patients (UW)(4, 40–42). Collectively, 39 tumors from 832 patients (4.7%) with at least 20% tumor cellularity were classified as CDK12^BAL. These grouped by: 1 with biallelic copy loss (2.5%), 15 with single copy loss with a pathogenic second allele mutation (38.5%), and 23 (31%) with biallelic pathogenic mutation (Fig 1a, Table S1). Of the pathogenic mutations, 31 (31%) were localized to the kinase domain (Fig 1b). In addition, monoallelic CDK12 pathogenic genomic alterations were identified in 92 tumors (11%) (Fig 1a).

Figure 1.

Having identified cohorts of PCs with and without CDK12^BAL, we next sought to determine if CDK12^BAL tumors exhibited evidence of HRd. Various mutational processes, including the loss of mechanisms that repair DNA damage through HR, produce characteristic signatures of residual structural alterations or mutations that can be classified according to the type of mutagen or compromised repair mechanism (33, 43, 44). Cancers with HRd are notable for genome instability resulting in large regions of copy loss and gain that can be scored based on metrics of loss of heterozygosity (LOH). LOH scores in PCs with BRCA2^BAL were significantly greater than PCs without biallelic loss of BRCA2, CDK12 or TP53 (‘All Intact’ group)(Mann-Whitney U p=2.8e^-3), whereas LOH scores in tumors with CDK12^BAL were not different than ‘All Intact’ tumors (p=1.0)(Fig 1c; Table S1, Methods).

COSMIC single base substitution (SBS) mutation signature 3 (CSig3) is associated with HRd and can be determined through WES or WGS (45). We determined that 10 of 39 tumors (26%) classified as CDK12^BAL had evidence CSig3 activity, a CSig3 proportion distribution which did not significantly differ compared to ‘All Intact’ tumors (Mann-Whitney U p=1.00)(Fig 1d, Table S1, Methods). In contrast, 46 of 69 (65%) PCs with BRCA2 biallelic loss exhibited CSig3 signatures, having a trend of higher CSig3 activity than in CDK12^BAL tumors (p<0.06)(Fig 1d).

CDK12 inactivation is documented to be associated with a tandem duplicator phenotype (TDP) classified by numerous copy gains of duplications across the genome (11, 46, 47). Of the 38 tumors with biallelic CDK12 loss that were evaluable for a TDP, 34 exhibited a TD genome (89%) (Fig 1a). Overall, we classified 46 tumors with a TDP across all cohorts, including nine with monoallelic CDK12 alteration and 3 with no alterations in CDK12 (Fig 1e, Methods). Notably, six different TDP groups have been described, based on the size of the duplicated segment and whether the size distributions are unimodal or bimodal (48). Prior reports determined that tumors with CDK12 loss generally categorize as Group 6 with TDs exhibiting a bimodal size of ∼230kb and ∼1.7Mb whereas tumors with BRCA1 loss classify as Group 1 TDP with a unimodal size of ∼10kb. Of the 46 mCRPCs with a TDP, 30 (65%) classified as Group 6, 11 (37%) in Group 3 (median 2.6 Mb), and 3 (7%) as Group 2 (median ∼380kb) (Fig S1a,b, Table S2). One of 30 tumors that exhibited a Class 6 TDP did not have CDK12 genomic alterations (Fig S1b). No tumors with biallelic BRCA2 loss exhibited a TDP though 8 of 31 tumors with monoallelic BRCA2 alterations classified as a Group 6 TDP. None classified as the Group 1 TDP associated with BRCA1 loss. Of tumors with TP53 alterations, 23 (3.7%) classified as TDP+, though of these, 17 also had a CDK12^BALevent (Fig S1b, Table S2). Collectively, these findings confirm prior reports detailing the unique TDP genomic structure associated with CDK12^BAL which is distinct from the types of genomic scars associated with HRd.

We next sought to determine the gene composition within tandem duplications (TDs) to assess whether consistent oncogenic drivers or tumor suppressor mechanisms accompanied CDK12 inactivation. The median number of TDs per tumor with a TDP was 103 (range 61-161) and 426 (range 159-781) for WES and WGS data, respectively. We compared the frequency of copy gain by TDs from TDP+ tumors (n=46) against the frequency in tumors without a TDP (TDP-; n=777) and identified 130Mb (2,601 50kb windows) of total regions enriched in TDP+ tumors (Fisher’s exact test; Bonferroni adjusted p-value < 0.001)(Fig 1f, Table S2, Methods). Of 433 genes annotated as Cancer Gene Census oncogenes, 29 were significantly enriched as altered in TDP+ tumors, and 5 of 394 Cancer Gene Census tumor suppressor genes were transected by a TD boundary (p < 0.01) (Fig 1f-h, Table S2).

Though several notable genes with oncogenic functions were enriched in TDs, including CCND1, AKT1 and MDM4, there were no genomic regions comprising a TD that recurred with a frequency greater than 40% across the 46 TDP+ tumors. Genes involved in HRd were not significantly altered in TDP tumors. The androgen receptor (AR) and upstream AR enhancer locus were contained within a TD in 18 and 19 of the 46 TDP+ tumors, respectively, with a significantly higher median AR copy number of 3 compared to 2 in non-TDP tumors (p=0.004, two-tailed Mann-Whitney U test) (Fig 1i). CDK12^BAL was mutually exclusive with BRCA2^BAL (p=0.03, one-tailed Fisher’s exact test). Of CDK12^BAL tumors, 51% also harbored monoallelic (n=9) or biallelic (n=11) TP53 alterations (Fig 1j) while PTEN^BAL only occurred in 1 tumor with CDK12^BAL.

While the majority of the mCRPC cohorts were comprised of single tumors from an individual patient, the UW autopsy study included three patients with CDK12^BAL where multiple tumors were sampled, allowing for assessments of tumor heterogeneity with respect to CDK12 events. In each patient, all metastatic tumors shared the same CDK12 alteration, and all tumors exhibited a TDP with the majority of TDs shared across tumors within an individual (Fig 2a). These data suggest that CDK12 alterations are early events in tumorigenesis and support the monoclonal model of metastatic PC dissemination (42, 49). However, there was evidence for continued accumulation of TDs as individual metastasis also exhibited a number of unique TDs, a subset of which encompassed oncogenic drivers such as cMYC, ETS1, KRAS (Fig 2b). For two tumors, we were able to profile gene expression by whole transcriptome RNAseq. Both tumors expressed high levels of the AR and AR program activity with highly concordant proliferation rates as determined by cell cycle progression (CCP) scores (Fig 2c).

Figure 2.

Transcriptional alterations in mCRPCs with CDK12 Loss

While the genomic consequences of CDK12 loss in PC and other malignancies have been established with respect to features such as tandem duplications (11, 48, 50), the assessments of phenotypic alterations in mCRPCs that accompany CDK12 inactivation have not been evaluated extensively. As a first proxy for phenotype, we analyzed matched whole transcriptome RNAseq data from 332, 96, 135, and 269 tumors having at least 20% tumor cellularity in the SU2C-I, SU2C-WC, HMF and UW mCRPC datasets, respectively, and compared CDK12 intact tumors against those with CDK12^BAL for differential genes, pathways, and hallmarks that reflect relevant biological characteristics of mCPRC. We removed neuroendocrine (NEPC) and AR-negative/NE-negative (DNPC) samples due to their lack of representation in the CDK12^BALgroup and compared the remainder to a group CDK12 intact tumors lacking any CDK12 or BRCA2 alterations (n=296 tumors). Global comparisons of transcript abundance levels identified 344 genes differentially increased and 333 genes differentially decreased in CDK12 loss vs intact tumors (log2 FC and p<0.05 across 2 or more cohorts)(Fig 2d, Table S3). There was substantial overlap in these differentially expressed genes across the four mCRPC cohorts (Bonferroni adjusted p-values<0.0001, pairwise hypergeometric tests)(Fig S2b-c, Table S3). We observed concordance in these datasets with previously reported gene expression alterations resulting from CDK12 loss including the upregulation of ARID3C, TBX4, and downregulation of TSACC and CDNF in mCRPCs with CDK12^BAL, along with enrichment of a CDK12-loss transcriptional signature (9)(Fig S2a).

mCRPCs are now recognized to exhibit subtypes classified by differentiation states reflecting androgen receptor (AR), neuroendocrine (NE), and other lineage programs. Compared to CDK12 intact tumors, mCRPCs with CDK12^BAL exhibited significantly higher AR expression and AR activity (p=0.002, Wilcoxon rank-sum test) (Fig 2e; Fig S2d-g). Notably, of 29 tumors classified as NEPC in the cohorts, none harbored CDK12^BAL and overall NE activity scores were not different in CDK12^BAL tumors (Fig 2f). The cell cycle progression score (CCP), a metric of cell proliferation, did not distinguish CDK12 status (Fig 2g). The expression of the alternatively spliced ARv7 transcript was higher in tumors with CDK12^BAL (p=0.004, Wilcoxon rank-sum test), further supporting an AR-driven phenotype (Fig 2h). Several pathways were reproducibly altered in CDK12^BAL mCRPCs including alterations in cell cycle, androgen response, spliceosome and DNA replication (Fig 2l,m).

We considered several mechanisms that could explain the differential gene expression in CDK12^BAL mCRPCs. We confirmed a prior study reporting high rates of gene rearrangements and fusion transcripts that associate with CDK12^BAL (p=5e^-6, Wilcoxon rank-sum test)(Fig 2i). However, only 22 of 623 fusion transcripts were recurrent (Fig 2j), and overall did not explain the differential expression of specific genes recurrently altered across studies (Fig S2h). In contrast, several COSMIC-defined oncogenes and tumor suppressor genes located within regions of TDs were increased at the transcript level and were recurrent across mCRPCs with CDK12^BAL (Fig 2k). Genes involved in HR were not involved in gene fusion events and were not commonly transected by gene rearrangements or TDPs (Fig 1h).

A key function of CDK12 involves the regulation of gene transcription by complexing with Cyclin K to phosphorylate the C-terminal domain of RNA polymerase II which promotes transcriptional elongation and the synthesis of full-length mRNAs (14). Notably, acute loss of CDK12 in cell line models results in decreased expression of a small subset of genes that are long and comprise large numbers of exons (15). Genes with these characteristics shown to be influenced by acute CDK12 depletion include several involved in HR DNA repair such as ATM, BRCA1, ATR, FANCI, and FANCD2 (15). To determine if the expression of long genes or those with large numbers of exons were differentially altered in PCs with CDK12^BALwe analyzed the whole transcriptome RNAseq data from the 4 mCRPC datasets collectively and individually. We found no overall associations between differentially downregulated genes in CDK12^BAL vs CDK12-intact tumors based on gene length when integrating tumors from all cohorts (p=0.26) (Fig 3a). However, other gene parameters were associated with lower transcript levels in the context of CDK12^BALincluding longer coding sequence length, longer transcript length, and shorter 3’UTR length (Fig 3b, Fig S3a-c).

Figure 3.

In addition to gene length-dependent effects on transcriptional elongation, acute inhibition of CDK12 activity results in premature cleavage and polyadenylation (PCPA), with the use of alternative polyadenylation sites (APA), particularly in genes with features such as large gene length, greater exon numbers, large first intron, and more intronic poly(A) sites (IPAs) (18, 51). Several genes involved in DNA damage repair fit these criteria. CDK12 has been shown to globally repress the use of intronic polyadenylation sites and the consequent expression of full-length transcripts whereas cells with CDK12 depletion exhibit elevated numbers of truncated transcript isoforms resulting from IPA usage. Analysis of previously published RNA-seq data (16) using the APAlyzer (52) tool confirmed the reported selective increase in intronic APA usage in the TCGA-PRAD primary tumors (Fig S3d). In mCRPC, preferential upregulation of APA sites was observed tumors with CDK12^BAL in each of the mCRPC cohorts, though to variable degrees (Fig 3c,d and Fig S3e,f).

Several genes involved in sensing and repairing DNA damage via HR are large, and prior studies have reported downregulation of several, including ATM, BRCA1 and BRCA2 in the context of acute CDK12 loss which contributes to a ‘BRCAness’ phenotype with compromised HR. We confirmed prior studies demonstrating increased use of internal polyadenylation sites in ATM and a modest reduction in transcript reads derived from the distal 3’ exon (Fig 3e,f and Fig S2i-l). The expression of PALB2 and ATR was also modestly lower in tumors with CDK12^BAL (Fig 3j,l). However, other key HR genes were not affected as we observed no significant differential downregulation of BRCA1 or BRCA2 in mCRPCs with CDK12^BAL nor was there evidence of APA usage in these genes (Fig 3g-k) and data not shown).

Differential effects of acute versus chronic CDK12 loss on gene expression and homology directed DNA repair

We next sought an explanation for the discrepancy between the reported mechanisms of CDK12- loss leading to HRd and clinical observations whereby patients with CDK12 loss exhibit poor responses to PARP inhibitors (PARPi). First, we chose to replicate acute loss conditions using approaches described in prior studies (16, 26, 53). We evaluated whether acute CDK12 inhibition caused long gene downregulation via transcript shortening, as would be expected from a role of CDK12 in suppressing APA usage. As there are no pharmacological agents that exclusively impede CDK12 activity, we used the dual CDK12/13 inhibitor SR4835 to acutely inhibit CDK12 function. Because some large DNA repair genes, including BRCA1 and BRCA2, show cell cycle linked expression (54), palbociclib (Palbo) was used as a control to demonstrate the extent of mRNA decrease solely due to G1 arrest. Actinomycin D (ActD) served as a control for non-specific RNA-Pol II inhibition. Genes down- regulated by SR4835 skewed longer (Students t-test p=3.2e^-4 for LNCaP; p=3.3e^-88 for LuCaP35_CL) than upregulated or unchanged genes, while such effects were non-significant or reversed (i.e. shorter genes downregulated) upon Palbo or ActD treatment (Fig 4a). This result is based on gene length (introns+exons) and not the spliced transcript length, in which case all treatments caused preferential downregulation of longer transcripts (Fig S4a). Though all three treatments led to some increases and decreases in APA site usage, SR4835 led to more selective enrichment for upregulated APA site usage (7,929 up, 11.0 up/down ratio) compared to ActD (3,257 up, 2.1 up/down ratio) and Palbo (1,700 up, 1.2 up/down ratio) in LNCaP with similar results in LuCaP35_CL (Fig 4b and S4b). After six hours of treatment, SR4835 resulted in substantial alterations in gene expression (1,128 down / 201 up in LNCaP; 2,933 down / 2,009 up in LuCaP35_CL) (Fig. 4c), including down regulation of multiple genes involved in the DNA repair pathways, in particular HR (13/31 in LNCaP and 22/31 in LuCaP_35CL) (Fig 4d). Pathway analysis showed that although SR4835 caused a significant decrease in the HR pathway activity by negative enrichment score (NES) (-1.1, FDR=0.9 for LNCaP, and -1.2, FDR=0.34 for LuCaP35_CL), palbociclib caused a much greater decrease (-1.9, FDR=0.001 in LNCaP, and -1.9, FDR=0.006 in LuCaP35_CL) (Fig S4c). In fact, approximately half (16/30 in LNCaP and 24/58 in LuCaP35_CL) of the downregulated KEGG pathways upon SR4835 treatment are also downregulated by palbociclib, leading to some difficulty in assigning which effects are due broadly to cell cycle arrest vs CDK12/13 specific effects (Fig. S4d). However, some genes, did show SR4835-selective downregulation (e.g. ATM; -0.88 log2(fold), p<0.0001) and others (e.g. RAD51D) were downregulated more dramatically with SR4835 (-0.77, p<0.0001) than Palbo (-0.33, p=0.002) (Fig 4e,f). Thus, while the transcriptional effects from CDK12/13i treatment may be partially confounded by cell arrest effects, several key DNA repair genes do show selective downregulation.

Figure 4.

To determine if levels of proteins involved in DNA repair pathways were concordantly diminished, LNCaP and 22Rv1 prostate cancer cells were treated for 6, 24 or 48h with SR4835 and proteins were analyzed by immunoblot (Fig 4f, Fig S4e,f). In agreement with previous studies, BRCA2, ATM, and ATR protein decreased at 24h and 48h post treatment with 200nM SR4835. Double strand DNA (dsDNA) breaks, as indicated by γH2A.X, increased by 48h (LNCaP) and 24h (22Rv1) but was largely ablated with the addition of Z-VAD, a pan-caspase inhibitor. Similar results were observed with an ovarian cancer line, Skov3 (Fig S4f). Thus, while SR4835 does cause moderate decreases in key DNA repair proteins, most of the corresponding γH2A.X appears due to caspase-dependent apoptosis and not impaired DNA repair directly.

To test if transcript shortening was responsible for the DNA repair gene downregulation with SR4835, qRT-PCR was performed with specific primers for 5’ and 3’ regions of each target. SR4835, but not ActD or Palbo, led to specific 3’ transcript loss in the long genes tested including BRCA1, BRCA2, ATM, ATR, HDAC1, and VCL (Fig 4g). Together, these results confirm mRNA shortening, the APA activation phenotype, and downregulation of transcripts and proteins involved in HR in prostate cancer models under acute CDK12 loss conditions.

CDK12 is classified as a ‘common essential’ gene (https://depmap.org/portal/) and CDK12/13 inhibitors cause apoptosis after 24-48h (Fig 4f and S4e,f). Despite this essentiality, a subset of human prostate cancers do tolerate the loss of CDK12 (Fig 1). We next investigated the possibility that cells adapted to CDK12 loss might show a different phenotype than cells undergoing acute CDK12 depletion. To carry out these studies, we developed a new in vitro model of de novo CDK12^BAL prostate cancer by generating a cell line (LuCaP189.4_CL) from the LuCaP189.4 patient derived xenograft (PDX) (55) that carries bi-allelic CDK12 frameshift mutations (p.S345Gfs*10, p.S521Qfs*89) (Fig S4g). LuCaP 189.4 does not express detectable CDK12 protein (Fig S4h) and exhibits a classic tandem duplicator phenotype (TDP) that is a hallmark of CDK12^BAL tumors (Fig S4i). We quantified the abundance of 5’ vs 3’ transcript levels by qRT-PCR in three CDK12 intact PC models and LuCaP189.4_CL and found that only the CDK12^BAL LuCaP189.4_CL displays a selective 3’ decrease (log2(fold) mean difference) in ATM (0.72, p<0.0001) and ATR (0.72, p=0.004), but non-significant or 3’ increases in BRCA1, BRCA2, HDAC4, and VCL (Fig 5a), indicating that these long genes are less affected by APA and splicing effects in cells naturally adapted to tolerate the absence of CDK12.

Figure 5.

To further study the transcriptional features of cells adapted to CDK12 loss, CRISPR-mediated knockout (KO) clones were generated in 22Rv1 (two clones: KO2 and KO5) and Skov3 cells (one clone: KO1), and LuCaP189.4_CL cells were engineered to re-express CDK12 (Fig 5b). Consistent with previous studies, very few clones tolerated CDK12 deletion, and those that did grow slower than the parental lines (Fig. S4j). At the protein level, cells with CDK12 deletion showed slight increases in CDK13 and decreases in CCNK/CyclinK but no obvious decrease in p-Ser2 RNA Polymerase II/RBP1/POLR2A levels (Fig 5b). The CDK12-KO clones did not show decreases in ATR, BRCA1, or BRCA2 protein but did show decreased ATM compared to the parental lines (Fig 5b). Interestingly, the CDK12^BAL LuCaP189.4_CL showed comparable levels of these DNA repair genes, and although the re-expression of CDK12 was not very high, ATM protein was slightly increased (Fig 5b). Consistent with the results in the de novo CDK12^BAL LuCaP189.4_CL, the 22Rv1 CDK12-KO clones also exhibited persistent 3’ vs 5’ transcript decreases (mean log2(fold)) in ATM (1.16, p <0.001 for KO2 and 1.71, p<0.0001 for KO5) and ATR (0.84, p<0.0001 for KO2 and 0.88 p<0.0001 for KO5), but minimal 5’/3’ difference in BRCA1, BRCA2, or VCL (Fig 5c). The 22Rv1 CDK12-KO clones did show lower overall BRCA1 (-1.46 KO2, -1.01 KO5) and BRCA2 (-1.79 KO2, -0.97 KO5) mRNA levels (log2(fold) vs DMSO, 3’ primer set) (Fig 5c). However, reduced transcripts of these genes may be due to cell-cycle linked expression and the slower proliferation of these clones (Fig S4i) and not a direct result of CDK12 loss.

RNAseq based assessments of the isogenic CDK12 intact versus knockout lines identified no enrichment for longer genes among those with lower expression in the stable 22Rv1-CDK12-KO clones (p=0.99 for KO2 and p=0.094 for KO5), though there was enrichment in the Skov3-CDK12-KO1 line (p=1.5e^-20) (Fig 5d). APA analysis found modest increases in IPA site usage (UP/DOWN ratio) in the CDK12-KO5 (1.2) and Skov3-CDK12-KO1 (2.7), but not 22Rv1-CDK12-KO2 (0.8) (Fig 5e and S5a), far less dramatic than under acute CDK12 inhibition which had UP/DOWN ratios of 11.0 in LNCaP and 6.1 in LuCaP35_CL (Fig 4b, S4b). These results show that, with the notable exception of ATM, most long genes (including BRCA1 and BRCA2) do not show downregulation in tumor cells that have adapted to CDK12 loss. Though some genes show IPA alterations, the phenotype is far less apparent than under acute loss conditions. Furthermore, though CDK12 inactivation in the Skov3 ovarian cancer cells showed some preferential downregulation of long genes overall (Fig 5d), BRCA1 and BRCA2 protein were not affected (Fig 5b).

Cells adapted to CDK12 loss are HR competent and lack exceptional sensitivity to PARPi or platinum chemotherapy

A key early step in HR involves BRCA2-mediated loading of RAD51 onto resected ssDNA. Loss of key HR pathway members, including BRCA1, BRCA2, or PALB2, all lead to loss of RAD51 loading and initiation of HR repair (56). Though CDK12-KO cells retain BRCA1 and BRCA2 protein expression (Fig 5b), it remains possible that HR function could still be altered by other means. To test this possibility, in addition to the CDK12-KO clones LNCaP, 22Rv1, and Skov3 cells were engineered with Tet-inducible shBRCA2 or shCDK12 (Fig S5b). Cells were exposed to ionizing radiation (IR) at 6Gy and immunostained for γH2A.X and RAD51 at 3h post irradiation (Fig 5f and S5c-e). 22Rv1-Tet-shBRCA2 cells functioned as an HRd control: shBRCA2 without dox went from 1.0 to 1.8 RAD51 foci per cell after IR, while dox treated cells went from 0.5 to 0.6 (Fig 5f). Following IR, the RAD51 foci per cell in 22Rv1-CDK12-KO lines increased from 1.2 to 2.9 in KO2 (p=0.04) and 0.9 to 2.4 in KO5 (p<0.0001) (Fig. 5f). Likewise, the CDK12^BAL LuCaP189.4_CL was competent at inducing RAD51 foci (1.4 to 5.4 foci per cell, p<0.0001) (Fig S5c). Since only G2/M cells can use HR repair, we also used an alternate quantification metric of % cells with 5 or more foci (i.e. % of population undergoing HR) which showed that BRCA2 knockdown greatly reduced the RAD51 population after IR (13.9% in -dox, 1.5% in +dox) while the 22Rv1-CDK12-KO clones and LuCaP189.4_CL showed a high proportion (25-28%) of HR-functioning cells (Fig 5g). Additional experiments with TetshBRCA2 and Tet-shCDK12 in LNCaP (Fig S5d) and Skov3 (Fig S5e) produced similar results, with CDK12 knockdown unable to prevent RAD51 foci induction (i.e. non-significant differences after IR between shCDK12 line -/+dox). These results with CDK12 knockdown and knockout, across multiple models, show that CDK12 deficient cells maintain the ability to load RAD51 and initiate HR repair.

Though cells adapted to survive CDK12 loss do not exhibit compromised HR, the possibility remains that CDK12 loss could sensitize to platinum chemotherapies or PARPi via other mechanisms (57–59). Dose response curves were performed with carboplatin (Fig. 5h) and the PARPi olaparib (Fig. 5i) using various prostate cancer lines, plus BRCA1-deficient UWB1.289 ovarian cancer cells as a bona fide HRd control (60). Though UWB1.289 showed the expected sensitivity to both carboplatin (EC50 0.36uM) and olaparib (0.71 uM) at 8 days, LuCaP189.4_CL was not sensitive (carboplatin EC50 = 19.15; olaparib EC50 >100uM). Inducible knockdown of BRCA2 in 22Rv1 cells altered the EC50 of carboplatin from 5.92 μM to 2.16 μM with dox, whereas the two 22Rv1 CDK12-KO clones showed either a greater or equivalent EC50 (Fig S6a). Skov3-CDK12-KO1 showed no difference in response to carboplatin (Fig. S6b) or olaparib (Fig. S6c) at 8 days treatment compared to control Skov3 cells. The 22Rv1-CDK12-KO clones and CDK12^BAL LuCaP189.4_CL showed mixed responses to PARPi: in a 12-day treatment, LuCaP189.4_CL (EC50 0.20μM) and the 22Rv1-CDK12-KO clones (0.88μM for KO2, 0.92μM for KO5) displayed some sensitivity to olaparib compared to 22Rv1 (17.37μM), but not to the same extent as UWB1.289 (0.08μM) (Fig S6d). LuCaP189.4_CL did not show enhanced sensitivity in an 8 day exposure to rucaparib (Fig S6e). In 14 day treatments with organoids harvested from PDX tumors, CDK12-intact LuCaP lines (23.1, 170.2, 86.2CR) had EC50 ranges of 2.23-6.33μM for olaparib while LuCaP189.4 was at 13.30μM, showing no apparent sensitization (Fig S6f) as was also the case with rucaparib (Fig S6g). Collectively, these results indicate that stable CDK12 loss does not sensitize to carboplatin, though the effect on PARPi sensitivity is less consistent with with clonal variation, but significantly less than the bona fide HRd line UWB1.289.

CDK13 is synthetic lethal in cells with biallelic CDK12 loss

CDK12 and CDK13 have overlapping and unique roles in regulating transcription and RNA metabolism, and both function in a heterodimer with Cyclin K (CCNK) (61). Analysis of CRISPR screen data from the Dependency Map project (https://depmap.org/portal/) shows that CDK13 depletion is generally tolerated in most lines, while CDK12 loss is detrimental in most cells (Fig 6a). Moreover, CCNK depletion has an even more negative fitness effect (<-1 CERES score). The LuCaP189.4_CL, which does not express CDK12, was the only natural line tested without negative growth effects upon sgCDK12 transduction (Fig. 6b), while LNCaP (Fig. 6c-d), C4-2B (Fig. 6e), Skov3 (Fig. 6f), and 22Rv1 (Fig. S6h,i) all showed significantly reduced growth (p<0.006) by confluence in sgCDK12 vs sgAAVS1, most dramatically in C4-2B (47.3% with sgAAVS1, 2.6% with sgCDK12). While LNCaP, C4-2B, and Skov3 tolerated CDK13 CRISPR lentivirus with no significant difference vs AAVS1 (Fig 6c-f), cells with stable CDK12^BAL loss: LuCaP189.4_CL and Skov3-CDK12-KO1, showed a marked growth inhibition with sgCDK13 vs control (LuCaP189.4_CL: 59.5% vs 88.3%, p<0.0001, Skov3KO1: 5.6% vs 12.9%, p=0.004) (Fig 6b,g). Of interest, the growth of 22Rv1 was repressed by either CDK12 or CDK13 sgRNAs (Fig S6i). However, 22Rv1-CDK12-KO5 appeared to have almost complete growth suppression with sgCDK13 from Day 5 to 14 (0.8% to 0.8%) compared to sgAAVS1 (0.6% to 5.5%) while parental 22Rv1 with sgCDK13 still had measurable growth (1.6% to 6.3%) (Fig S6i,j).

Figure 6.

Due to high protein conservation, all currently available pharmacological inhibitors of CDK12 also inhibit CDK13. We performed dose response curves with two different selective CDK12/13 inhibitors: SR4835, an ATP competitive inhibitor (21) (Fig 6h and S6k), and THZ531, a covalent binding inhibitor (28) (Fig 6i and S6l). LuCaP189.4_CL and the CDK12-KO 22Rv1 clones all showed higher sensitivity to THZ531 with EC50 of 88.6nM for 22Rv1 vs 17.3nM (KO2) and 33.3nM (KO5) and 37.8nM (LuCaP189.4-vec) vs 52.6nM for cells with re-expressed CDK12 (LuCaP189.4-CDK12) (Fig 6i). Surprisingly, the CDK12-KO clones did not show the same curve shift with SR4835 (Fig 6h), though LuCaP189.4_CL was quite sensitive with an EC50 of 38nM (LuCaP189.4-vec) (Fig 6h) and 29nM (LuCaP189.4-CDK12) (Fig S6k,m). This difference could be due to the fact that SR4835 is a non-covalent inhibitor, while THZ531 is a covalent modifier. Unfortunately, THZ531 has not been deemed usable for in vivo use while SR4835 has been used previously in mice (53).

To confirm whether the CDK13 vulnerability could exploited for in vivo treatment, we performed xenograft drug treatments using three LuCaP PDX lines (LuCaP35, LuCaP136, and LuCaP189.4) treated for 28 days with vehicle or SR4835. LuCaP 35 and 136 are both CDK12-intact. At the final 28 day timepoint, vehicle treated LuCaP 189.4 tumor volumes (mm³[95%CI]) measured 521[411-630] while SR4835 tumors were smaller at 310[196-425] with a significant (p=0.046) cumulative difference between growth curves (Kolmogorov-Smirnov, two-tailed). Tumors were harvested and weighed, with SR4835 treated tumor average weight being significantly smaller (0.59 vs 1.08 g, p=0.0015, Mann-Whitney test) (Fig 6j). There were no significant tumor volume or weight reductions in the CDK12- intact lines. Taken together, these results support the hypothesis that cells lacking CDK12 become dependent on CDK13 for their CCNK/CyclinK activity, thus presenting a potential targeted vulnerability with potential in vivo efficacy, even with dual CDK12/13 targeting compounds.

We performed RNA-seq on three of the harvested tumors per group after 28 days SR4835, plus a set of LuCaP189.4 PDX tumors treated acutely for 3 days with SR4835 or vehicle. Results were analyzed by APAlyzer and found large upregulation of APA sites in LuCaP189.4 after 3 day treatment (2,958 Up) but not LuCaP 35 after 28 days treatment (342 Up) (Fig. 6k). Furthermore, the UP/DOWN ratios of APA usage was only >1.0 for LuCaP189.4 at the 3 day treatment point (3.9 ratio) and at or below 1.0 for all 3 lines in the 28 day treated tumors (Fig. S6l). A selection of example long DNA repair gene transcripts were significantly decreased with the 3d treatment including: ATM (log2 FC -1.1), ATR (-1.2), BRCA1 (-2.6), BRCA2 (-2.0), and RAD51D (-0.8)(all p<0.01 in SR4835 vs vehicle) (Fig 6l). None of these genes were significantly decreased in the 28 day treated tumors, suggesting that cells able to survive and grow in the presence of the CDK12/13 inhibitor no longer show the long gene APA phenotype and downregulation.

CDK12 loss increases sensitivity to selected therapeutics targeting transcription

The most studied function of CDK12 is to maintain RNA polymerase II processivity and proper splicing and polyadenylation. Tumors adapted to CDK12 loss show modest if any alteration in the transcript lengths or APA usage of large genes involved in HR (Fig. 3g and 5b-d) and no substantial change in RNA-Pol II Ser2 phosphorylation (Fig. 5b), but these adapted cells may still have impaired transcriptional processes which could be exploited. We next tested if cells with CDK12 loss may exhibit enhanced sensitivity to therapeutics targeting transcription mechanisms. Dose response curves revealed a modest selective sensitivity in CDK12^BAL lines to α-amanitin, an RNA-Pol II poison with EC50 of 779nM for 22Rv1 compared to 506nM (KO2) and 339nM (KO5) (Fig 6m). Re-expression of CDK12 in the LuCaP189.4_CL promoted amanitin resistance to a small but not statistically significant degree (631nM to 735nM, p=0.4) (Fig 6m). These results suggest that cells adapted to CDK12 loss may continue to exhibit compromised transcriptional mechanisms, and that this may lead to a vulnerability to amanitin-class agents or other drugs that target mRNA synthesis or processing.

CDK12 loss does not consistently increase sensitivity to WEE1, ATR, or CHEK1 inhibitors

Among other potential targetable DNA repair pathway vulnerabilities considered, we tested two WEE1 inhibitors: (adavosertib/MK-775) (Fig S7a) and PD0166285 (Fig S7b). The EC50 values for LuCaP189.4_CL (with or without CDK12 re-expressed) and 22Rv1-CDK12-KO2 were all greater than parental 22Rv1 (Fig S7c). However, 22Rv1-CDK12-KO5 did show some increase in sensitivity (EC50) compared to the parental line to adavosertib (0.04μM vs 0.11μM) and PD0166285 (0.10μM vs 0.37μM) (Fig S7c). Considering that adapted CDK12^BAL cells do retain some ATM transcript shortening and protein decreases (Fig 5a,b), we evaluated ATR as a potential vulnerability. We tested two ATR inhibitors, berzosertib (Fig S7d) and elimusertib (Fig S7e), but found that LuCaP189.4_CL had the highest EC50 values to both inhibitors (Fig S7f). 22Rv1-CDK12-KO5 (but not KO2) showed a slight increase in sensitivity to berzosertib, and although both KO clones showed some increased sensitivity to elimusertib vs parental, the EC50 curves were similar to LNCaP and C4-2B reducing the like-lihood of a CDK12-specific effect (Fig S7f). Lastly, we tested two CHEK1 inhibitors, rabusertib (Fig S7g) and MK-8776 (Fig S7h), as potential ATR-related dependencies. However, the results show that LuCaP189.4_CL was the least sensitive line tested to both (Fig S7i), and only 22Rv1-CDK12-KO5 showed a clear increase in sensitivity vs parental to both rabusertib (0.53μM vs 1.81μM) and MK08776 (0.48μM vs 1.54μM). In summary, one of the tested CDK12 loss models (22Rv1-CDK12- KO5) showed modest CDK12-associated growth repression by WEE1, ATR, and CHEK1 inhibitors.

However, since differential treatment effects were not shared with 22Rv1-CDK12-KO2 or LuCaP189.4_CL it appears that these sensitivities are not consistent, but may relate to alternative mechanisms by which cells survive CDK12 loss.

Discussion

In the present work we sought to identify a molecular basis for the discrepancy between the poor clinical responses to PARPi in patients with CDK12 alterations and preclinical studies demonstrating that CDK12 loss or pharmacological inhibition compromises HR, phenocopies ‘BRCAness’ and results in synthetic lethality to drugs impeding PARP function. A central feature of most prior studies evaluating CDK12 and HR is the conduct of very short term in vitro experiments, with timepoints usually less than 72 hours after pharmacological inhibition of CDK12 activity or repressing CDK12 by genetic methods (12, 16, 19, 26, 28, 53, 62). A key reason for evaluating such acute timepoints is largely due to the fact that CDK12 inhibition leads to proliferation arrest and/or cell death, as CDK12 is a common essential gene in most cells (15, 16, 28, 63). Acute loss of CDK12 activity clearly results in increased APA site usage and the diminished expression of a cohort of large genes, including several involved in HR. This consequence can contribute to HRd in the immediate setting, and act in synergy with PARPi and DNA damaging agents. However, it should be noted that cell cycle arrest also results in the downregulation of many HR-associated genes, and it is not straightforward to attribute causality specifically to CDK12-mediated transcriptional compromise versus cell cycle-regulated expression (54). We note that the CDK4/6 inhibitor palbociclib caused more significant inhibition of DNA repair pathways than the CDK12/13 inhibitor SR4835.

In contrast to the acute effects of functional CDK12 loss, cells naturally adapted to survive and proliferate in the context of CDK12 absence, such as the LuCaP189.4 model and human tumors, show less differential down-regulation of long genes compared to acute CDK12 repression. In analyses of human tumors with CDK12^BAL, the expression of genes regulating HR are not reduced, and these tumors do not exhibit the genomic scars, such as high LOH, or mutation signatures, such as COSMIC mutation signature 3, that reflect compromised HR found in tumors with BRCA1 or BRCA2 loss. Further, using functional assays, such as RAD51 foci formation after radiation, both de novo CDK12^BAL cells, and those adapted to tolerate CDK12 deletion by CRISPR/Cas9, demonstrate HR competency.

Through literature searches and cell line database queries, we were not able to identify any cancer models with de novo CDK12^BAL, and only one prior study where a stable CDK12 deletion in a cancer cell line was generated (64). This supports the conclusion that CDK12 is a common essential gene in most cells, with a severe impact on viability that is challenging to overcome. In agreement with the results presented here, the above mentioned A2780 ovarian cancer cell line model with stable CDK12 deletion showed markedly slower proliferation, increased rates of apoptosis, and no significant reductions in HR-related proteins including ATM, BRCA1, ATR or FANCD2. The A2780 CDK12 knockout line was also not sensitive to platinum chemotherapy or olaparib, and was 5-fold more resistant to the ATR inhibitor VE822 and the WEE1 inhibitor MK1775 (64). We observed variable sensitivities to WEE1, ATR, and CHEK1 inhibitors in the context of CDK12 loss, suggesting actionable vulnerabilities in specific CDK12 loss-adapted contexts which have yet to be characterized.

Though we did see modest sensitivity to PARPi in CDK12^BAL cells in a few experiments, the lack of consistent HR gene downregulation, no effect on RAD51 foci formation, and low scoring of HRd genomic signatures, indicate that PARPi as a monotherapy is unlikely to benefit patients with CDK12^BAL tumors by exploiting HRd. However, other aspects of compromised CDK12 function could provide a therapeutic window for PARP inhibition as well as other targets. Recent studies revealed that PARPi can act beyond HR and affect other critical cell functions where CDK12 activity may be important including transcription, RNA splicing, R-loop repair, and replication (51, 57, 59, 65–68).

The identification of pharmacological vulnerabilities conferred by CDK12 loss is an important research priority. While we did not observe consistent effects using ATR or WEE1 antagonists, CDK13 appears to be an ideal synthetic lethal target for CDK12^BAL cells, as complete loss of cyclin K activity in tumors would be severe, and other tissues could likely tolerate CDK13 inhibition in the setting of functional CDK12 (12). However, available pharmacologic inhibitors have been unable to separate CDK12 and CDK13 antagonism due to high protein similarity. This can lead to some toxicity concerns, especially with covalent inhibitors (69). However, CDK12/13 dual inhibitors may still be useful due to the enhanced sensitivity of CDK12^BAL tumors to CDK13 inhibition. Furthermore, though less effective than THZ531 in culture assays, SR4835 was effective and generally tolerated well in mice and has been used in at least two published studies (53, 70). We observed modest differential sensitivity in several CDK12^BAL models to the transcriptional inhibitor α-amanitin. Amanitin-based compounds are extremely toxic, but there is interest in using these drugs in antibody drug conjugates (ADCs)(71), which would offer a way to deliver toxic payloads directly and selectively to the tumor.

In addition to transcription-related consequences of CDK12 loss, it is very possible that CDK12 could play additional roles beyond RNAP II that have yet to be explored. Known direct substrates of CDK12 are fairly limited, with RNAP II being the most well studied. However, there is a report that 4EBP1 is a direct CDK12 substrate and confers a role in translation (72), and LEO1, part of the transcription elongation complex, was also reported as a direct CDK12 substrate (20). CDK12 has also been pulled down in complex with various other factors, including splicing regulators, though it is unclear if they are direct substrates (13, 15).

Of importance, the direct oncogenic role(s) of CDK12 loss remains to be determined. Of interest is the limited organ site distribution of tumors with CDK12 alterations. While the genomic structures of TDs produce a substantial increase in gene rearrangements, we found that very few of these are recurrent, and thus unlikely to serve as drivers of neoplasia. In contrast, a large number of oncogenes are contained within regions of tandem duplication of which several were recurrently gained in copy number, and could individually or collectively represent pathways driving cancer development. Notable among these recurrent events is the AR, which is either gained or amplified within a TD in 46% percent of CDK12^BAL tumors. Of interest is the finding that no mCRPCs with a neuroendocrine phenotype had CDK12^BAL, suggesting that CDK12 loss is incompatible with neuroendocrine differentiation, or the AR locus copy gains serve to maintain AR signaling and prevent transdifferentiation to other lineages.

Methods

Somatic mutation analysis

For whole exome sequencing (WES) data, single nucleotide variant (SNV) calling was performed using MuTect2 (GATK version 4.1.8.1), Strelka2 version 2.9.2 and VarScan2 version 2.4.4. Insertion and deletion (Indel) calling was performed using SvABA (commit 9813e84, https://github.com/walaj/svaba). All SNV and Indel calls were annotated using ANNOVAR (release 20200607). SNVs were included if they were detected by two or more callers, were not labeled by ClinVar significance as ’benign’ or ’likely_benign’, and were annotated as either ’exonic’ or ’splicing.’ Indel calls were included if they were not labeled by CLNSIG as ’benign’ or ’likely_benign’, were annotated as either ’exonic’ or ’splicing,’ and passed the following filters: Alt_reads_in_Tumor >= 5, Ref_reads_in_Tumor >= 10, and a variant allele frequency (VAF) of > 0.1. For patients with discrepant SNV or Indel calls between samples, manual curation was performed using IGV (version 2.16.2). For SU2C-WC WGS data, we obtained mutation calls from the original published results by Quigley et al. 2018 (3). For the Hartwig Medical Foundation (HMF) WGS data, the mutation callset was obtained via controlled access as part of Data Access Request DR-250. The liftOver tool was used to convert mutations calls from GRCh37 to GRCh38 coordinates, and these results were annotated with ANNOVAR (release 20200607), and mutations were included if they were not labeled by ClinVar significance as ‘benign’ or ‘likely_benign’, and were annotated as either ‘exonic’ or ‘splicing.’

Germline mutation analysis

Germline mutation calling was performed for UW and SU2C-WC cohorts using GATK’s HaplotypeCaller using default settings. For WES data, the probe-set specific bait file was provided per sample as well. Output SNVs were annotated using ANNOVAR (release 20200607) and were included if they were annotated by CLNSIG as ‘pathogenic’ or ‘likely_pathogenic’. Manual curation was performed using IGV (version 2.16.2). Only somatic calls were provided in the provided data for HMF.

Somatic copy number alteration analysis

For WES samples (SU2C-I and UW cohorts), the standard tumor-normal paired workflow of TITAN (v1.15.0) for WES was used [https://github.com/gavinha/TitanCNA]. Briefly, read counts were computed at 50 kb bins overlapping the exome bait set intervals.

Centromeres were filtered based on chromosome gap coordinates obtained from UCSC for hg38. The read coverage in each bin across the genome was corrected for GC content and mappability biases independently for tumor and normal samples using ichorCNA v0.3.4. Heterozygous SNPs were identified from the matched germline normal sample using Samtools mpileup. Only SNPs overlapping HapMap3.3 (hg38) were retained. The reference and non-reference allele read counts at each heterozygous SNP were extracted from the tumor sample. SNPs were not analyzed in chromosome X. Copy number analysis was performed using TitanCNA R package v1.15.0. Automatically generated optimal solutions were used.

The following methods were used for analyses of whole genome sequencing data. For SU2C-WC WGS data, copy number alterations, tumor purity, and tumor ploidy was predicted using TITAN (v1.15.0) as originally reported in Zhou et al. 2022 (73) using the workflow provided in https://github.com/GavinHaLab/TitanCNA_SV_WGS. For HMF WGS data, the copy number results were obtained via controlled access as part of Data Access Request DR-250. The results were originally analyzed using GRCh37 (hg19) reference genome.

Tumor cellularity was estimated by TITAN for each sample, and only samples with ≥ 20% tumor cellularity were included in downstream analyses.

Mutational signature analysis

Mutational signature proportions were determined using the Analyze.cosmic_fit function from SigProfilerAssignment (v0.0.33) and Cosmic Version 3.4. Only SNV calls which passed our filtration were used for signature analysis. One input VCF was made per sample, with each input being reformatted to have five columns in the following order: chromosome, genomic coordinate, sample ID, reference base call, and alternative base call. When running Analyze.cosmic_file, the ‘build’ option was set to ‘GRCh38’, ‘input_type’ was set to ‘vcf’, and the ‘exome’ option was set to true for WES samples. All other options were left as default. To best capture HRD-related mutational signatures, we combined Cosmic Signatures SBS3 and SBS8 into an HRD-signature. Cosmic Signature 3 proportion was obtained by dividing the number of SBS3 SNVs by the total number of input SNVs per sample. The reported proportions were obtained by summing the counts of SBS3 and SBS8 per sample, then dividing by the total number of calls for each sample. We only considered SBS3 or SBS8 calls for samples with more than 50 passing SNVs and a combined SBS3 and SBS8 proportion of >0.05.

Structural variation analysis

For samples based on short-read WGS, SvABA was used in tumor-normal paired mode for SV detection with default parameters. Intra-chromosomal SV events with span > 1 kb were retained. The SvABA workflow can be accessed at https://github.com/GavinHaLab/TitanCNA_SV_WGS

Classification of structural variants in mCRPC

SV types were annotated based on orientations of breakpoints and bin-level copy-number around breakpoints, as described previously (73). The copy- number near each breakpoint was evaluated using 10 kb bins. For each SV event, copy-number values of the bins located to the upstream and downstream of breakpoint 1 were denoted as c1^up and c1^down, respectively; similarly, the copy-number values for breakpoint 2 were denoted as c2^up and c2^down. In addition, then mean copy-number c^mean of the 10 kb bins between the two breakpoints of the SV event were considered during SV classification. Intra-chromosomal SV events, i.e., both breakpoints were located on the same chromosome, were classified into the SV types: deletion, tandem duplication, inversion, balanced rearrangement (intra-chromosomal), unbalanced rearrangement (intra-chromosomal). Interchromosomal SV events are classified as translocations.

Tandem duplicator phenotype classification

Simple tandem duplications (TDs) were predicted from the copy number results genome-wide for each sample. For whole exome sequencing, duplication events must meet these criteria: (i) segment is shorter than 10 Mb; (ii) have flanking segments with lower copy number; (iii) the difference in copy number between the left and right flanking segment is ≤ 1. Out of the 632 WES samples, 581 were successfully analyzed by TITAN and having tumor purity >0.2 and MAD<0.25 were used to identify duplications. For WGS data, the tandem duplications were taken from the intersection of TITAN copy number segments and tandem duplication breakpoints from the final SV call set. As described previously (74), we used the Nearest-Neighbor Index (NNI) metric to distinguish the pattern of tandem duplications being dispersed (value near 1) as opposed to clustered. We applied a threshold for the NNI score as a guideline for manual curation of individual samples. For WES samples, TDP+ was defined as having NNI >1 and median segment length >100kb.

For WGS samples, TDP+ was defined as having NNI >1.25 and median segment length >100kb. Further manual inspection of the copy number profiles confirmed TDP+ and TDP-status.

Tandem duplication size group analysis

To categorize TDP+ cases into TD size (i.e. length) groups, we applied Gaussian mixture model fitting using the densityMclust (mclust v6.1 R package). Input TD segments consists of events meeting the criteria in the TDP classification analysis. For each sample, solutions of one to four possible Gaussian components were used to fit the log10-transformed TD lengths in kilobases. The optimal solution is selected via Bayesian information criteria (default in MClust), and components in the optimal solution with mixed weights ≤ 0.1 or variance. The estimated means for the remaining components determine the size groups: 1-100kb (Group 1), 100kb-1Mb (Group 2), 1-10Mb (Group 3). For solutions with > 1 group (i.e. > 1 component), additional groups were defined: Group 4 (Group 1 + Group 2), Group 5 (Group 1 + Group 3), and Group 6 (Group 2 + Group 3). TD Group size analysis is shown separately for WES and WGS due to differences in copy number segment size resolution – WGS is expected to have higher resolution and therefore smaller TD segments are better represented. Finally, the distribution of the estimated component means is shown for all WES (n=26) and WGS (n=20) samples.

Genomic regions and genes altered by TDs

Three different sets of genomic regions were analyzed to determine the frequency of alteration by TDs in TDP+ and TDP- samples. Simple TD segment events were defined based on criteria set in the TDP classification analysis: 1) Tiled 50kb windows across the genome. The number of TD events overlapping each 50kb window was computed across all TDP+ samples (n=46) and all TDP- samples (n=777) with >1 simple TD event meeting the TDP classification criteria. A two-sided c²-test of independence with Bonferroni multiple-test correction was used to determine enrichment of TD alteration between TDP+ and TDP- groups; 2) COSMIC Cancer Gene Census oncogenes. The Cancer Gene Census file was from a 1/27/2019 release; gene coordinates in this file were used. Genes that had a ‘Role in Cancer’ value containing the string “oncogene” or “fusion” were considered in this list of 433 oncogenes. A gene was considered to be altered by a TD if the gene coordinates were fully contained within a TD event (i.e. the TD event spans the entire gene). For each gene, a two-sided Fisher’s exact test was used to determine enrichment of TD alteration frequency between TDP+ and TDP-groups. Bonferroni correction was applied; 3) COSMIC Cancer Gene Census tumor suppressor genes. The same Cancer Gene Census file was used as above. Genes that had a ‘Role in Cancer’ value containing the string “TSG” or “fusion” were included in this list of 394 genes. A gene was considered to be altered by a TD if either “start” or “end” or both boundaries are located within the gene coordinates (i.e. either or both breakpoints transects the gene). A Fisher’s exact test, followed by Bonferroni correction was applied as above.

Analyses were performed using GRCh38, except for the HMF cohort which was GRCh37 and required liftOver or matching of gene symbols.

Multi-tumor tandem duplication and phylogenetic analysis of rapid autopsy samples

For UW patient 05-217, input TD segments consisted of events meeting the criteria in the TDP classification analysis. The complete set of TDs within the patient was defined as the union of TD events were taken across all samples within the patient. A matrix consisting of samples by TD event was constructed using 1 for presence of the TD in the sample or 0 if it was absent. A TD event was considered present if the proportion of the overlap by width was ≥ 0.9. This matrix was used to determine the number of overlapping events in pairwise, triplets, and all samples within the patient. This matrix was also used as input into neighbor-joining tree estimation (ape R package) using Manhattan distance to construct a phylogenetic tree with a rooted-tree configuration. The tree branching configuration and general branch length was used to produce a custom representation presented in Figure 2b.

Annotating gene alteration by copy-number

Copy-number segments were excluded if their cellular fraction was lower than 0.8, except for those which were determined as copy neutral or copy-number greater than 4. The gene annotation was based on known protein coding genes from GenCode release 30 (GRCh38.p12). For each gene, its copy-number was assigned to the copy-number value and LOH status of the segment that has the largest overlap with it. The gene-level copy-number was normalized based on ploidy of the corresponding sample, with autosomal genes normalized by the inferred ploidy rounded to nearest integer, and X-linked genes normalized by half such value. Then the copy-number status of each gene was categorized based on the following criteria: (i) Amplification. Normalized gene-level copy-number is greater than or equal to 2.5; (ii) Gain. Normalized gene-level copy-number is between 1.5 and 2.5; (iii) Homozygous deletion. Normalized gene-level copy-number is 0; (iv) Deletion with LOH. Normalized gene-level copy-number is between 0 and 1, and LOH status was found; (v) Copy neutral LOH. Normalized gene-level copy-number is 1 and LOH status was found.

Annotating gene alteration by structural variant

Gene coordinates were based on ENSEMBL v33 of hg38. Gene body region of one gene was defined as the widest region of all known isoforms collapsed. Gene flanking region was defined as the corresponding two 1 Mb regions next to the gene body region on 5’-end and 3’-end, respectively. Gene alteration status by genome rearrangements was defined based on the breakpoints and directions of involving structural variant events. A gene in one WGS sample (gene-sample pair) was considered having gene transecting events if any breakpoints of SV events were located within the gene body region. If the gene transecting status did not apply, then this gene-sample pair was examined for gene flanking status if the breakpoints of any intra-chromosomal SV events, including tandem duplications, deletions, and inversions, were located within the gene flanking regions. Additionally, translocation events including intra-chromosomal balanced and unbalanced events which spanned over 10 Mb, and inter-chromosomal translocation events were considered altering the gene flanking regions if any of their breakpoints was in the gene flanking region, and the direction of the SV was going towards the gene body region. The alteration status of rearrangements for each gene-sample pair was exclusive between gene transecting and gene flanking, with the former being prioritized in report.

Annotating gene allelic alteration status

For CDK12, BRCA1, BRCA2, TP53, CHD1, PTEN, and PABL2, allele status was defined using SNV, Indel, CNV, and SV results, and classified into three categories: Bi-allelic loss (BAL), Mono-allelic loss (MAL), and Intact. An event was only included in gene status calling if it passed all filters applied to the corresponding data type. For a given gene, BAL was defined as two or more events or a homozygous deletion within gene boundaries. MAL was defined as one detected loss-of-function event occurring within gene boundaries, while samples annotated as Intact had no recorded variants. Manual curation was performed to confirm gene statuses, with an emphasis being placed on confirming or amending discordant calls between samples from the same patient.

Loss of heterozygosity (LOH) analysis

The LOH score was defined as the proportion of the genome affected by LOH events (minor copy number = 0). To compute this value, we summed the genomic distance spanned by all segments reported by TitanCNA to have a minor allele copy number of 0, then divided this value by the genomic distance spanned by all segments reported by TitanCNA. Chromosome arms with LOH events spanning more than 75% of their length were excluded from our analysis, as these events are associated with non-HRD mechanisms.

Transcript Analyses

Sequencing reads were mapped to the hg38 human genome using STAR.v2.7.3a (75). Gene fusions were mapped and quantitated using STAR-Fusion. AR-V7 quantitation was performed as previously described (76). All subsequent analyses were performed in R. Gene level abundance was quantitated using GenomicAlignments (77). Differential expression between groups was assessed using limma (78) filtered for a minimum expression level using the filterByExpr function with default parameters prior to testing. Genome-wide gene expression results were ranked by their limma statistics and used to conduct Gene Set Enrichment Analysis (GSEA) to determine patterns of pathway activity utilizing the curated pathways from within the MSigDB (79). Single sample enrichment scores were calculated using GSVA (80) with default parameters using genome-wide log2 FPKM values as input. Intronic APA analyses were performed using APAlyzer (v1.2.0) using prebuilt hg38 intronic polyadenylation (IPA) and 3’-most exon regions (52). The read cutoff parameter “CUTreads” was set to 5 for analysis of IPA between groups.

Cell lines and culture

UWB1.289 were a gift from Dr. Elizabeth Swisher (60) and grown in 50:50 RPMI (Thermo 11875093) and MEGM (Lonza CC-3150) with 3% FBS (Thermo 16000044) and 0.5X Pen/Strep (Thermo 15140122). Skov3 (HTB-77), LNCaP (CRL-1740), C4-2B (CRL-3315), 22Rv1 (CRL-2505), and HEK293T (CRL-3216) were received from ATCC. LuCaP cell lines were generated by resecting the tumor implant, dissociating cells by enzymatic digestion and plating cells in DMEM medium with 10% FBS or with various additives used in organoid medium (81). Skov3, LNCaP, C42B, and 22Rv1 were grown in RPMI with 10% FBS and 0.5X Pen/Strep. LuCaP 35_CL, LuCaP 189.4_CL, and HEK293T were grown in DMEM (Thermo 11965092) with 10% FBS, 1X GlutaMAX (Thermo 35050061), 1mM additional sodium pyruvate (Thermo 11360070), and 0.5X Pen/Strep. Organoids were grown in customized media as previously described (82). Cells were grown at 37C with 5% CO2. Cell lines underwent DNA fingerprint (STR) confirmation and routine mycoplasma testing (R&D Systems CUL001B) via Fred Hutch Research Cell Bank Services.

Lentivirus production and transduction

HEK293T cells were seeded at 18 million per T75 flask and transfected the next day with 3μg pMISSION-VSVG, 6μg pMISSION-Gag-Pol, and 9μg transfer plasmid with 36μL of TransIT Leti reagent (Mirus MIR 6604). Cells were changed to fresh media the next day and viral media was collected on day 3 for concentration with Lenti-X concentrator (Takara 631232). Cells were transduced with 8μg/mL polybrene. Viral titer was determined by flow cytometry (for fluorescent vectors) or antibiotic selection titer with antibiotic started 48h after infection and viability assayed 6 days post transduction. For growth assay, cells were transduced at a target multiplicity of infection of 2.

Vectors and engineered lines

Tet-inducible shRNA vectors were cloned as previously described(83) into EZ-Tet-pLKO-Hygro (Addgene 85972). The shRNA sequences are in a supplementary table. We generated two customized sgRNA construct backbones: pLenti-sgStuffer-GFP-Puro (Addgene 208349) and pLenti-sgStuffer-mCherry-Puro (Addgene 208350). 22Rv1 CDK12-KO lines were generated by stable lentiviral transduction with two different sgRNA vectors: pLenti-CDK12(sg2)-GFP-Puro and pLenti-CDK12(sg3)-GFP-Puro. The stable CDK12 sgRNA lines were then transduced with pLentiCas9.mCherry (Addgene 208342) lentivirus and single cells were sorted by flow cytometer (Sony SH800S) into 96 well plates. Colonies were expanded and screened by western blot while mutations were confirmed by Sanger sequencing. Skov3 CDK12-KO1 was generated similarly, except that the sgRNA and Cas9 plasmids were transiently transfected before flow sorting for clonal expansion. The Skov3-CDK12-KO1 line has pre-existing puromycin resistance from previous transduction with a nonfunctional (frame-shifted) lentiviral vector. LuCaP189.4_CL were transduced with FUCGW(84) or FUCGW-CDK12 and flow enriched for GFP to create isogenic pools LuCaP189.4-vec and LuCaP189.4-CDK12. FUCGW was a gift from Dr. John K. Lee, which was modified into FUCGWDEST (Addgene 208408) and FUCRW-DEST (Addgene 208409) by inserting an attR1-CmR-ccdB-attR2 cassette. CDK12 CDS was amplified with Phusion HF (Thermo F530) and TOPO cloned into the pCR8 entry vector (Thermo K250020) to make pCR8-CDK12 (Addgene 208347). Silent sgRNA-resistant mutations were made via assembly cloning (NEB E2621) of fragments amplified with primers detailed in the supplemental table to generate pCR8-CDK12(sgR) (Addgene 208346), which was then recombined with LR clonase II (Thermo 11791020) into FUCGW-DEST to generate FUCGW- CDK12(sgR). All plasmid sequences were confirmed with full plasmid sequencing (Plasmidsaurus, Eugene, OR). For growth assays, a dual sgRNA vector system was made based on lentiCRISPRv2- Blast (Addgene 83480), generating pLCV2-AAVS1(hU6-sg1-mU6-sg2)-Blast (Addgene 208343), pLCV2-CDK12(hU6-sg2-mU6-sg3)-Blast (Addgene 208344), and pLCV2-CDK13(hU6-sg2-mU6- sg3)-Blast (Addgene 208348). Additional information on the cloning strategy, including the dual sgRNA cloning protocol, is available on the vector pages on addgene.org.

Immunoblot

Lysates were prepared in RIPA buffer (150mM NaCl, 5mM EDTA, 50mM Tris, pH 8.0, 1% IGEPAL CA-630, 0.5% sodium deoxycholate, 0.1% SDS), passed through a 20G needle, pelleted to remove debris, then quantified by BCA assay (Thermo 23225). Loading samples were prepared in LDS sample buffer (Thermo NP0007) with 5% beta-mercaptoethanol and heated at 99C for 10m for denaturation. Samples were run on NuPAGE 4-12% bis-tris gels (Thermo NP0335BOX) with MOPS run buffer (Thermo NP000102) or 3-8% tris-acetate gels (Thermo EA0378BOX) with tris acetate SDS run buffer (Thermo LA0041). Gels were run and transferred in an XCell SureLock module (Thermo EI0002). Wet transfers were performed with NuPAGE transfer buffer (Thermo NP00061) containing 10% methanol for 3h at 30V fixed voltage in a cold room. Additional reagents include BLUEstain2 protein ladder (GoldBio P008-500) and low fluorescence PVDF (BioRad 1620260) and StartingBlock buffer (Thermo 37542). Blots were imaged on a ChemiDoc MP system (BioRad 12003154) with ImageLab software (v6.1).

RAD51+γH2A.X immunofluorescence and foci quantification

Cells were seeded on 12mm circle coverslips coated with poly-d-lysine (50ug/mL PBS, 400uL per 12-well, 1h at 37C) (Thermo A3890401). Cells were placed in a cesium-137 irradiator for a 6 Gy exposure or mock treated then returned to the tissue culture incubator and fixed 3h post irradiation with 4% paraformaldehyde/PBS (Thermo AAJ19943K2) for 10m. Fixed cells were then quenched with 125mM glycine/500mM tris pH 7.4 for 10m and permeabilized with 0.5% Triton X-100 in TBS (50mM tris, 150mM sodium chloride, pH 7.4) for 5m then washed once with TBS/T (TBS with 0.1% Tween 20). Next, cells were blocked with Dual Endogenous Enzyme Buffer (Agilent S200389-2) for 10m at room temp then washed twice with TBS/T. Cells were incubated overnight at 4C with RAD51 (Sigma ZRB1492) and γH2A.X (Sigma 05- 636) primary antibodies in 1% BSA(TBS/T) (see supplemental table for details). After primary incubation, cells underwent two washes in TBS/T and were incubated with HRP-anti-rabbit secondary (Leica PV6119) for 30m at room temp. Next, cells were washed three times with TBS/T and then incubated for tyramide amplification and Alexa 568 staining according to the vendor protocol (Thermo B40956). Cells were washed three times with TBS/T and then incubated 1h at room temp with Alexa Fluor 488 anti-mouse (Thermo A32723) or Alexa Fluor 633 anti-mouse (Thermo A-21052) for 22Rv1 lines. Following a TBS/T wash, cells were stained with DAPI (2ug/mL in TBS/T) for 10m at room temp, washed twice in TBS/T and once in water, then mounted onto glass slides with ProLong Gold (Thermo P36930). Slides were imaged on a Zeiss LSM800 confocal microscope running Zen software (v2.3) with a 20X/0.8NA objective, 40μm pinhole, and three 1.5μm z-steps. Image stacks were flattened with a maximum intensity projection and foci were quantified using imageJ (version 1.53q)(85) to outline nuclei and count foci using the FindMaxima tool. Data were plotted using GraphPad PRISM (v10.0.1) showing number of foci per nucleus with a line at the mean with at least 150 nuclei per group. Statistical significance was determined by ANOVA (Kruskal-Wallis).

Quantitative real time PCR (qPCR)

RNA was extracted with Qiagen RNeasy kit (Qiagen 74004) and quantified by Qubit (Thermo Q10211). Reverse transcription was performed with ProtoScript II (NEB M0368) using 4μM poly-dT plus 1μM random hexamer primers. qPCR was performed in 384-well plates (15μL reactions) with Power SYRB Green Mix (Thermo 4367660) and run on a QuantStudio 6 Flex (Thermo 4485691). Relative expression was standardized (ΔCT) to zero-centered average of four housekeeping primer sets (18S, RPL19, ACTB, GAPDH) and normalized (ΔΔCT) as described in each experiment, generating log2(fold change). Primer information can be found in the supplementary table.

Drug dose response curves

Cells were seeded at relatively low densities in 96-well plates (500-1,000 for Skov3 and UWB1.289; 1,000-4,000 for 22Rv1 and LNCaP; 10,000-20,000 for LuCaP189.4_CL).

Cells were seeded 24-48h before adding drug. There were no media changes until harvest unless otherwise noted. Cells were kept at room temp for 30m after seeding before transferring to incubator and were kept in a humidified chamber (e.g. plastic box with water reservoir) to promote even seeding and reduce edge effects.(86) Cells were harvested by adding 50uL of CellTiterGlo 2.0 reagent (Promega G9243) per well, incubating 20 minutes, and reading luminescence on a BioTek plate reader with Gen5 software. For organoid experiments, 5,000-10,000 cells were seeded per 15uL Matrigel droplet (Corning 354234) (2:1 gel:cell suspension ratio) in 96-well round bottom plates. At harvest, 50uL of CellTiterGlo was added per well, incubated 20m, then wells were mixed by pipette and transferred to solid black plates before reading luminescence. Data were normalized vs mean of media-only wells and plotted with PRISM showing mean-/+stdev and polynomial curve fitting (inhibitor vs response, variable slope, four parameter) for effective concentration 50 (EC50) calculation for concentration of drug at 50% relative viability.

Colony assays

Cells were seeded in 12-well plates at 100 (Skov3) or 200 (22Rv1, UWB1.289) cells per well. For shRNA lines, 200ng/mL doxycycline was included. One day after seeding drug was added to each well. Cells received full media/drug change every 6 days and were fixed at final timepoint with 4% paraformaldehyde/PBS (Thermo AAJ19943K2) for 10m. Cells were washed with PBS and stained with crystal violet (250μM in PBS) for 30m at room temp. Next, cells were washed with PBS then water and plates were air dried. Images were acquired on a Cytation5 (BioTek) imager using the TexasRed channel with a 1.25X objective and tile stitching to cover each well. Colonies were manually quantified with ImageJ.

sgRNA growth assays

GFP-tagged cells (transduced with FUCGW) were infected with sgRNA lentivirus in 6-well plates. 24-48h later 10ug/mL blasticidin was added for selection and on day 5 post-infection cells were seeded into 384-well plates (Corning 353962) which were read once per 24h on a Cytation5 (BioTek) with BioSpa using the GFP channel and 1.25X objective. GFP confluence was calculated in ImageJ setting a threshold and counting GFP positive pixels. LuCaP189.4_CL cells were seeded on low adherence plates (Perkin Elmer 6057800) and quantified with Gen5 software.

PDX tumor growth experiment

LuCaP PDX lines were implanted subcutaneously into male NSG mice (NOD scid gamma) (one tumor per mouse). Treatment began when tumors reached 150mm^3 volume as measured by calipers. SR4835 was purchased from SelleckChem (S8894). Fresh solutions were prepared by first solubilizing drug in DMSO (Sigma D2650) and then diluting the stock dropwise 1:10 into 30% cyclodextrin (w/v) in water (Sigma H107). Mice were given 20mg/kg SR4835 or vehicle by oral gavage 5 days on / 2 days off.(53) Tumor volume and mouse weights were measured 3 times/week using calipers, with volume calculated as 4/3*π*L*W*H/8. Animal use followed Fred Hutch IACUC approved protocol (#51077). Mice were sacrificed at 28 days treatment, maximum allowed tumor size (tumor avg. diameter >2cm), or if deemed necessary by comparative medicine staff. Tumors were excised and weighed upon sacrifice.

Code availability

Analysis code for custom analysis and pipeline configuration settings are accessible at https://github.com/GavinHaLab/CDK12-CRPC-paper.

Acknowledgements

We would like to thank the patients who generously donated tissue that made this research possible. We would also like to thank Lisa Ang and Talina Nunez for assistance with mouse studies and Pushpa Itagi for assistance with genomic analyses. We thank Evan Yu, Heather Cheng, Jessica Hawley, Michael Schweizer, Lawrence True, Meagan Chambers, Martine Roudier, and the rapid autopsy teams for their contributions to the University of Washington Medical Center Prostate Cancer Donor Rapid Autopsy Program and the development of the LuCaP PDX models. This research was supported by the Genomics and Bioinformatics core and Comparative Medicine Shared Resource, RRID:SCR_022610, of the Fred Hutch/University of Washington/Seattle Children’s Cancer Consortium, P30 CA015704.

This work was also supported by awards from the CDMRP PC200262, PC200262P1, PC180295, the Pacific Northwest Prostate Cancer SPORE P50CA97186, PO1CA163227, R21CA277368, F32CA243286, R50 CA274336-02 DP2CA280624, R01CA280056, the Institute for Prostate Cancer Research and the Prostate Cancer Foundation.

Conflicts of interest

PSN has received fees for advisory work from BMS, Pfizer, and Merck, and research support from Janssen for studies unrelated to the present work. MCH. served as a paid consultant/received honoraria from Pfizer and has received research funding from Merck, Novartis, Genentech, Promicell and Bristol Myers Squibb. EC served as a paid consultant to DotQuant, and received Institutional sponsored research funding unrelated to this work from AbbVie, Gilead, Sanofi, Zenith Epigenetics, Bayer Pharmaceuticals, Forma Therapeutics, Genentech, GSK, Janssen Research, Kronos Bio, Foghorn Therapeutics, K36, and MacroGenics. CM has received funds from Genentech and Novartis unrelated to the present work.