Widespread BRCA1/2-independent homologous recombination defects are caused by alterations in RNA-binding proteins

Summary

Defects in homologous recombination DNA repair (HRD) both predispose to cancer development and produce therapeutic vulnerabilities, making it critical to define the spectrum of genetic events that cause HRD. However, we found that mutations in BRCA1/2 and other canonical HR genes only identified 10%–20% of tumors that display genomic evidence of HRD. Using a networks-based approach, we discovered that over half of putative genes causing HRD originated outside of canonical DNA damage response genes, with a particular enrichment for RNA-binding protein (RBP)-encoding genes. These putative drivers of HRD were experimentally validated, cross-validated in an independent cohort, and enriched in cancer-associated genome-wide association study loci. Mechanistic studies indicate that some RBPs are recruited to sites of DNA damage to facilitate repair, whereas others control the expression of canonical HR genes. Overall, this study greatly expands the repertoire of known drivers of HRD, with implications for basic biology, genetic screening, and therapy stratification.

Graphical abstract

Highlights

• •Aberrations in known HR genes account for a small fraction of HR-deficient tumors
• •Integrated analysis identifies RNA-binding proteins as a driver of HR deficiency
• •RBP dysfunction can induce HR deficiency directly or via regulating HR repair genes

McGrail et al. discover that the majority of patient tumors with defects in DNA homologous recombination (HR) repair do not harbor aberrations in known drivers such as BRCA1/BRCA2. Leveraging computational and experimental approaches, they demonstrate that mutations in RNA-binding proteins are a large driver of HR defects.

Introduction

Genomic instability is a hallmark of cancer,¹ with implications both for treatment strategies as well as cancer screening and prevention. As normal, healthy cells have largely intact DNA damage response (DDR) pathways, therapeutic avenues that selectively target tumor cells with DNA repair defects have emerged as a promising treatment strategy.² For instance, poly (ADP-ribose) polymerase (PARP) inhibitors have emerged as a powerful approach to treat patients with defects in BRCA1 or BRCA2, genes involved in homologous recombination (HR) repair, a process that faithfully repairs DNA double-strand breaks (DSBs).^2,3,4 Similarly, microsatellite instability, caused by defects in DNA mismatch repair (MMR), was recently approved as a biomarker for response to immune checkpoint blockade, marking the first approval of tumor-type-agnostic biomarker by the US Food and Drug Administration (FDA).⁵

Although these DNA repair defects in either the HR or MMR pathways may provide treatment options, non-functional variants of BRCA1/BRCA2 or genes involved in the MMR pathway are also primary drivers in familial cancers.⁶ A large fraction (50%) of known drivers of hereditary cancers are genes involved in DNA repair and genome maintenance.⁶ In the case of breast cancer, women with a mother, sister, or daughter with breast cancer have a 2-fold higher risk of developing breast cancer⁷; however, only 15%–25% of patients with hereditary breast/ovarian cancers have BRCA1 or BRCA2 mutations, and the majority of hereditary drivers have not yet been identified.^8,9 Incomplete knowledge of genetic risk factors hinders approaches for effective cancer screening and prevention. A better understanding of these genetic risks could identify which patients require more aggressive risk reduction approaches and could minimize use of aggressive risk reduction interventions in patients who lack pre-disposition genes.¹⁰

Thus, there exists a critical need to understand molecular alterations associated with genomic instability and the specific genes that can drive DNA repair defects. To comprehensively identify tumors with HR defects, we utilized a genomic scar homologous recombination DNA repair (HRD) score. This score allows us to detect genomic lesions left by HRD and identify tumors with HRD regardless of the genetic event that caused the HRD.¹¹ HRD scores calculated across tumors from The Cancer Genome Atlas (TCGA) revealed numerous molecular alterations associated with HRD. Surprisingly, approximately 75% of tumors that scored positive for HRD exhibited no known molecular HRD driver. Using an integrated network-based approach, we identified nearly 100 undescribed candidate HRD drivers, with a particular enrichment for genes involved in RNA processing. Candidate HRD drivers had over a 90% experimental validation rate. Mechanistic studies indicated that that the RNA-binding proteins (RBPs) we identified may influence HR either by modulating expression of canonical DNA repair genes or by acting directly at sites of DNA damage.

Results

HRD scores vary across tumor types and patient demographics

HRD leaves a quantifiable genomic scar allowing the calculation of an HRD score, defined as the combination of three measures of genomic instability: loss of heterozygosity (LOH), telomeric allelic imbalance (TAI), and large-scale transitions (LST) (Figure 1A).¹¹ We determined the HRD score across tumors from all patients within TCGA, and we found high HRD scores for basal-like breast cancer and ovarian cancer, both of which are known to have high levels of HRD.^9,12 Luminal androgen receptor and luminal A and luminal B breast cancers had low HRD scores (Figures S1A and S1B). In addition to basal-like/ovarian cancer, numerous other cancer types not typically associated with HRD exhibited high HRD scores, including lung squamous carcinoma, bladder cancer (BLCA), and gastric cancer (Figure 1B). Consistent with early onset of basal/triple-negative breast cancers (TNBCs), we found that in tumors from patients with basal breast carcinoma, HRD score was negatively associated with patient age. There were similar trends observed in lung adenocarcinoma (LUAD), head and neck squamous carcinoma, and mesothelioma. Nonetheless, when quantified across all patients, HRD score generally showed a positive relationship with age (p = 5.3 × 10⁻⁷, Figures 1B and S1C). When compared across all cancer types, tumors from male patients tended to have higher HRD scores than female patients (p = 0.04, Figures 1B and S1D), and tumors from Asian patients had statistically higher HRD scores than other patients (p = 4.3 × 10⁻³, Figures 1B, S1E, and S1F).

Figure 1.

Pan-cancer analysis of HR defects

(A) Schematic describing the HR defect score, defined as the sum of scores for loss of heterozygosity (LOH), telomeric allelic imbalance (TAI), and large-scale state transitions (LSTs).

(B) HRD scores across tumor types (top) and associations with various demographic features (bottom). HRD score is plotted as median value with error bars representing interquartile range. Demographic features scale from red (positive association with HRD score) to blue (negative association with HRD score). Black dots represent significant relationships within a given cancer type. Red dots next to demographic categories represent a significant positive association across all cancer types.

(C) Gene set enrichment based on relationship with HRD score. A generalized linear mixed model was used to determine the association between RNA-seq-derived gene expression levels and HRD scores, taking tumor type as a random effect. The resulting coefficients were used for gene set enrichment analysis.

(D) Association between HRD score and protein levels. A generalized linear mixed model was used to determine the association between reverse phase protein array-derived protein levels and HRD scores, taking tumor type as a random effect. False discovery rate (FDR) was determined using the Benjamini-Hochberg procedure.

(E) Pathway enrichment determined from whole-proteome mass spectrometry in breast cancer. Spearman correlation coefficients were determined between proteins and HRD scores. The resulting correlation coefficients were used for gene set enrichment analysis. The five most significantly enriched pathways are shown. NES, normalized enrichment score.

(F) Association between HRD score and microRNA levels. A generalized linear mixed model was used to determine the association between microRNA levels and HRD scores, taking tumor type as a random effect. FDR was determined by the method of Storey.

(G) Gene set enrichment of microRNA target genes. For each gene, a score was quantified as the sum of all coefficients for significant (FDR < 1%) microRNAs that could target that gene. The resulting list of scores was used for GSEA. The five most significantly enriched pathways are shown, demonstrating pathways predicted to be suppressed by miRNAs associated with HRD score. NES, normalized enrichment score. See also Figure S1.

HRD scores are associated with suppression of cell death pathways and activation of DNA damage response checkpoints

To begin to define the molecular changes associated with HRD across cancer types, we performed gene set enrichment analysis (GSEA)¹³ and found that HRD scores were associated with increased expression of cell-cycle checkpoint genes and decreased expression of genes associated with caspase activation (Figure 1C). As shown in Figure 1D, protein-level analysis using reverse-phase protein array (RPPA) data indicated HRD score was positively associated with numerous cell-cycle regulators, including the DNA damage checkpoint marker phospho-CHK2, as well as MSH6 previously implicated in non-homologous end-joining outside of its canonical role in DNA mismatch repair.¹⁴ We confirmed the association of HRD score with an increased expression of cell-cycle checkpoint proteins in breast tumors using an orthogonal whole-proteome mass spectrometry-based dataset (Figure 1E).²⁹ Analysis of microRNAs (miRNAs) again identified numerous miRNAs associated with HRD score (Figure 1F). We desired to assess which pathways these miRNAs might modulate, but the analysis was complicated: each miRNA can target multiple genes, and each gene can be targeted by multiple miRNAs. To address this issue, we calculated a gene-wise miRNA suppression score for each gene, defined as the sum of the miRNA coefficients predicted to target each specific gene. Final miRNA scores for all genes were used for GSEA. We found that miRNAs associated with high HRD scores preferentially suppressed oncogene-induced senescence and apoptotic pathways (Figure 1G). Taken together, these results indicate that HRD-positive tumors tend to have suppressed tumor suppressor pathways and activated DNA damage checkpoint pathways.

The majority of HRDs are of unknown etiology

To bifurcate tumors into HRD-positive and HRD-negative groups, we took a multi-step approach. We began by analyzing tumors of patients with known deleterious BRCA1 or BRCA2 germline mutations as a gold standard for HRD positivity. We focused this analysis across breast, ovarian, pancreatic, and prostate cancer where BRCA1 or BRCA2 germline mutations are known to promote tumorigenesis. As indicated by the receiver-operator characteristic curve in Figure 2A, HRD score was highly accurate at recovering tumors with deleterious BRCA1 or BRCA2 germline mutations, demonstrated by an area under the curve (AUC) value of 0.83. Using this data, we determined an optimal HRD threshold score of 32, which identified numerous HRD-positive tumors across an array of different cancer types (Figure 2B). However, an analysis of the genetic events within these tumors indicated that only a small minority (9.7%) displayed alterations (mutation or methylation) in BRCA1/BRCA2. Therefore, we expanded the analysis of genetic events to include HR-associated genes from a larger, annotated list¹⁵ but could still only identify potential drivers for roughly 25% of HRD tumors (Figure 2C).

Figure 2.

Most HR deficiencies are of unknown etiology

(A) Receiver-operator characteristic (ROC) curve (blue) demonstrating the ability of the HRD score to predict germline mutations in BRCA1 or BRCA2 in the indicated tumor types. AUC is defined as the area under ROC curve. The dotted line represents the expectation due to random assignment. The red dot indicates the calculated optimal threshold value to separate HRD-positive (HRD⁺) and HR-competent tumors.

(B) Percentage of HRD-positive tumors by cancer type.

(C) Percentage of HRD-positive tumors caused by BRCA1, BRCA2, other canonical DDR genes, and those of unknown etiology.

(D) Correlation of gene set enrichment analysis performed on gene expression changes between DDR-driven HRD-positive tumors and HR-competent tumors (plotted on x axis) or HRD-positive tumors of unknown etiology and HR-competent tumors (plotted on y axis). Spearman correlation coefficient.

(E) Correlation of change in RPPA-derived protein levels between DDR-driven HRD-positive tumors and HR-competent tumors (plotted on x axis) or HRD-positive tumors of unknown etiology and HR-competent tumors (plotted on y axis). Spearman correlation coefficient.

The large fraction of HRD-positive tumors with causes that could not be attributed to known drivers of HRD may represent false positives or, alternatively, could indicate that the majority of drivers of HRD in patients with cancer are unknown. Based on the strong molecular alterations we observed to be associated with HRD (Figures 1C–1G), we hypothesized that if the HRD of unknown origin was due to true HRD, it would exhibit similar molecular changes to those caused by known drivers such as BRCA1/BRCA2. Analysis of GSEA scores from gene expression data were correlated for two groups: (1) DDR-driven HRD-positive tumors vs. HR-competent tumors, and (2) HRD tumors with unknown causes vs. HR-competent tumors. Both groups demonstrated remarkably concordant changes in gene expression (Figure 2D). Likewise, correlation analysis performed at the protein level using RPPA data also revealed a robust correlation between protein alterations in HRD-positive tumors with known DDR gene alterations and those of unknown origin (Figure 2E). These data indicate that a large fraction of HRD-positive tumors are driven by unknown causes.

Network-based discovery of undescribed drivers of HRD

Across all tumor types, approximately 75% of HRD-positive tumors had an intact complement of known HR-related genes, indicating that the majority of HRD drivers were of unknown etiology (Figure 3A). To define the causes of HRD in these tumors, we developed a network-based algorithm for identifying undescribed drivers of HRD (Figure 3B). We began with a list of verified inducers of HRD. Then we identified genetic events in the genes encoding these inducers in patient tumors. Tumors were considered to have a genetic event if they had either (1) mutations with high variant allele frequency (VAF) or (2) a methylation event that corresponded with downregulation of the methylated gene. VAF was used for assessment because if a mutation is driving HRD/tumorigenesis, then it should occur in the majority of tumor cells. Next, we assessed whether a genetic event was associated with an increased HRD score based on cancer type. Candidate genetic events that may drive HRD in an individual tumor were assigned based on the degree to which that event increased the HRD score, the number of tumors in which it occurred, and the VAF of the mutation in a given tumor. After assigning the genetic events that may cause HRD in an individual tumor, we hypothesized that proteins that interact with these drivers would also be more likely to cause HRD. Therefore, we used protein-protein interaction networks to expand the list of candidate HRD drivers. This prediction algorithm was iterated until convergence, when it could no longer identify additional putative HRD-driver candidates.

Figure 3.

Discovery of previously undescribed drivers of HRD

(A) Pie chart showing the fraction of HRD-positive (HRD⁺) tumors with known drivers of HRD and those with unknown drivers of HRD.

(B) Schematic of the network-driven pipeline used to discover new drivers of HRD.

(C) Resulting protein network modules of identified drivers of HRD in tumors from patients with cancer. The size of nodes represents the number of patients with HRD attributed to a given gene, corresponding to the scale shown in the purple spheres.

(D) Pie chart showing the fractions of previously identified HRD drivers (light red) and newly identified putative drivers.

(E) Distribution of HRD driver ontologies by cancer type.

(F) Percentage of HR defects driven by RBPs in tumors from male and female patients with cancer. Fisher’s exact test. See also Figures S2 and S3 and Table S1.

The final network of predicted HRD drivers (Figure 3C) could be grouped into three, large protein modules. As anticipated, the largest of these was a DNA damage module, followed by an RBP module and a smaller protein translation module. Only a small fraction of proteins (<3%) was not strongly associated with any one of these modules. In total, we identified HRD drivers for 626 of 1,296 patient tumors displaying HRD but lacking a discernable alteration in a previously defined HR pathway (unexplained HRD), with the largest fraction consisting of RBP genes (Figure 3D). Cross-referencing these RBP genes with prior proximity-CLIP (cross-linking and immunoprecipitation) analysis of RBPs indicated that all 16 candidates shared between the two studies were located in the nucleus.¹⁶ Further analysis of functional domains contained within putative drivers of HRD revealed a significant enrichment in BRCA1 C-terminal (BRCT), Sm, forkhead-associated (FHA), helicase ATP-binding, helicase C-terminal, and RNA recognition motif domains (Figure S2). Identified HRD causes and ontology annotations for each tumor are given in Table S1. Tumors that we failed to identify a putative driver for exhibited significantly lower (p = 5 × 10⁻¹¹) HRD scores, suggesting these samples may be enriched for false positives. The relative proportion of various HRD drivers varied across tumor types. Ovarian, bladder, and colorectal cancers showed the largest fractions driven by canonical DDR genes, whereas melanoma, LUAD, and pancreatic cancer showed the largest fractions driven by RBP genes (Figure 3E). Notably, tumors from men were more likely to have HRD driven by RBP mutations (Figure 3F). The majority of patients with BLCA in the TCGA cohort received cisplatin or similar chemotherapies and should respond favorably if the tumor is HRD. We found that all identified causes of HRD were associated with good prognosis in BLCA, indicating the identified HRD drivers likely contribute to HRD and thus chemosensitivity (Figure S3). When we analyzed the remaining HRD-positive tumors with unexplained HRD, we identified several miRNAs that could suppress the expression of HRD drivers that were upregulated in HRD-positive sarcoma tumors (Figure S4).

Functional validation of newly discovered drivers of HRD

We next sought to validate that the newly discovered putative drivers of HRD identified from patient tumors were functionally important in HR and not simply bystander events. To begin to validate the functional relevance of candidate genes, we began by utilizing the DR-GFP reporter assay. In this assay, expression of an eGFP variant can only be restored if it is accurately repaired by HR following cleavage with I-SceI.¹⁷ Pladienolide B, which inhibits the core spliceosome RBP SF3B1, significantly inhibited HR at concentrations as low as 1 nM in U2OS cells, suppressing HR comparable to inhibition of the Mre11-Rad50-Nbs1 complex critical for HR¹⁸ with 100 μM Mirin (Figure S5A). For more specific analysis, we utilized two independent siRNAs for 43 of the potential HRD mediators we identified, and we found that 95% induced HR defects (Figure 4A). The two genes that failed to induce HR defects, CEP72 and JMJD6, were both classified as “other.” The genes in this category were only weakly linked to any module, further indicating that our network-based approach increased the robustness of our ability to identify strong candidate drivers. To exclude the potential that reduced HR function is merely an artifact from arresting cell cycle, we analyzed the ability of cells to form irradiation (IR)-induced Rad51 foci specifically in cycling cells, as indicated by incorporation of the nucleotide analog EdU (5-ethynyl-2'-deoxyuridine). Suppression of candidate RBPs also hindered HR function as assessed by Rad51 foci formation, finding strongly concordant results with results from the DR-GFP assay (Figures S5B—S5D).

Figure 4.

Validation of newly discovered drivers of HR defects using flow cytometry-based DR-GFP reporter assay

(A) HR repair was measured using a DR-GFP assay and flow cytometry to detect GFP in cell lines following siRNA-mediated knockdown of each of the indicated genes. Relative HR repair is defined relative to the percentage of GFP-positive cells of untransfected and siCTRL-transfected controls. n = 3 per siRNA, two siRNAs per gene. Mean ± SEM.

(B) The relative amount of HR repair was measured as in (A). Cells were transfected with RFP-tagged wild-type (WT) or mutant protein-expressing constructs 24 h prior to transfection with an I-SceI expression plasmid. HR repair was measured in RFP⁺ cells and defined as mutant relative to WT overexpression. n = 3 per condition. Mean ± SEM. ANOVA with Dunnett’s post hoc test. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 1 × 10⁻⁴.

The two most discordant genes, MED9 and MRPS6, both exhibited a larger suppression of HR by the DR-GFP assay than Rad51 foci formation. Suppression of these two genes induced the first and third highest degree of cell-cycle inhibition, respectively, which likely resulted in exaggeration of the HR defect as quantified by the DR-GFP assay. However, overall, the percentage of EdU⁺ cells forming Rad51 foci was not associated with the percentage of cells positive for EdU (Figure S5E). A complete summary of the results of functional assays is given in Table S2. As depletion of RBP proteins may not be functionally equivalent to the effects of missense mutations, we engineered vectors expressing RBPs with mutations identified from HRD tumors. Cells were transiently transfected with RFP-tagged mutant proteins or wild-type controls, allowing for analysis of HR function in cells expressing the desired constructs. After we profiled 29 mutations derived from patient tumors across 10 genes, we found that 28/29 inhibited HR function (Figure 4B).

To further validate the role of RBPs in HR, we treated TNBC MDA-MB-231 cells with pladienolide B, and we found that it efficiently inhibited the formation of IR-induced Rad51 foci in cycling cells (Figures 5A and 5B). As PARP inhibitors are known to preferentially kill HRD cells, we hypothesized that the RNA spliceosome inhibitor pladienolide B would sensitize these cells to PARP inhibitors. Indeed, we found pladienolide B and the PARP inhibitor BMN-673 demonstrated synergy in two TNBC cell lines, MDA-MB-231 (Figure 5C) and BT-549 (Figure 5D). For more specific analysis, we repeated the IR-induced Rad51 foci assay following siRNA-mediated depletion of DDX3X and AQR, two of the most common mutant RBPs in TNBC. We found that both siDDX3X and siAQR inhibited foci formation across four TNBC cell lines (Figure 5E). We further validated that suppression of AQR likewise suppressed HR in ovarian cancer cells, where it was also predicted to be associated with HR function (Figure S5D). As HR-deficient tumors may be therapeutically targeted with PARP inhibitors, we next tested the effects of PARP inhibitors in MDA-MB-231 cells stably expressing shDDX3X, shSF3B3, or shBRCA2. We found that depletion of either of the RBPs DDX3X or SF3B3 increased cell sensitivity to the PARP inhibitors BMN-673 (Figure 5F) and AZD2281 (Figure 5G) as well as if not better than depletion of BRCA2. Long-term clonogenic assays performed in the presence of BMN-673 confirmed this increased sensitivity to PARP inhibition (Figure 5H). Together, these results indicate that the previously undescribed HRD drivers we identified are likely to be functionally relevant for HR repair across multiple cell lines and assays.

Figure 5.

Control of HR by RNA-binding proteins in breast cancer

(A) Images of cells treated with either 10 nM of the splicing inhibitor pladienolide B or a DMSO vehicle control, 24 h prior to irradiation (5 Gy). Images show Rad51 foci indicative of HR repair (green), as well as proliferating (S phase) cells with EdU (red), and DAPI nuclear stain (blue).

(B) Quantification of percentage of cycling (EdU⁺) cells that exhibit Rad51 foci. N = 3. Student’s t test. ∗p < 0.05, ∗∗p < 0.01. Mean ± SEM.

(C and D) Viability of TNBC cells following a 5-day incubation with the splicing inhibitor pladienolide B (concentration indicated on top x axis), PARP inhibitor BMN673 (concentration indicated on bottom x axis), or a combination thereof, relative to a DMSO vehicle control for MDA-MB-231 cells (C) and BT-549 cells (D). Mean ± SEM. C.I., Chou-Talalay combination index. N = 2.

(E) Quantification of IR-induced Rad51 foci in cycling (EdU⁺) TNBC cell lines following siRNA-mediated depletion of the indicated RBPs. N = 2 per cell line. Mean ± SD.

(F) Viability of individual MDA-MB-231 cell lines with stable single knockdowns of RBPs DDX3X and SF3B3, as well as HR protein BRCA2, following a 5-day treatment with BMN673. Mean ± SD.

(G) As in (F) but following a 5-day treatment with AZD2281. Mean ± SD.

(H) Clonogenic assay with MDA-MB-231 cells with stable knockdown of RBPs DDX3X and SF3B3, following a 2-week treatment with BMN673. N = 2, and points represent independent biological replicates. See also Figure S5. Mean ± SD.

Induction of HR defects may occur by modulation of DDR genes as well as independent pathways

Next, we sought to understand how the newly identified drivers of HRD might influence HR repair. We hypothesized that loss of RBPs could interfere with DDR gene expression, by either decreasing mRNA stability or by interfering with splicing, either of which could result in hindered protein function. To test whether RBPs were modulating DDR genes, we performed multiple experimental and computational analyses. First was RNA-seq analysis following siRNA-mediated depletion of 17 RBPs in three cell lines to identify differential DDR gene expression relative to either siCTRL or siBRCA1/siBRCA2. Next, TCGA patient tumors were analyzed to detect decreased DDR gene expression relative to HR-competent tumors or tumors with HRD caused by DDR genes. Finally, the same comparisons were made using TCGA alternative splicing analysis rather than gene expression levels. All comparisons were made to both siCTRL/HR-competent samples and siDDR/HRD caused by DDR, as we had found that HRD itself can cause transcriptional rewiring (Figures 1C and 2D). The integration of these results is shown in Figure 6A, with specific comparisons shown in Figure S6. We identified DDR genes that were either suppressed or alternatively spliced for 55% (26 out of 47) of the RBPs that are candidates for affecting HR. The largest influence on gene expression was seen for members of the mediator complex (75% of genes), followed by core spliceosome members (61% of genes), with less influence seen by other RBPs (36% of genes). We validated 10 proteins identified to be modulated at the gene expression level by three RBPs at the protein level by western blot, and we found all candidates showed lower levels of protein as expected (Figures S6G–S6I). Assessment of RNAs interacting with the RBP PRPF8 by enhanced crosslinking and immunoprecipitation followed by high-throughput sequencing (eCLIP-seq) detected binding of all four identified DDR target genes (Figure S6J).

Figure 6.

RNA-binding proteins control HR repair by multiple mechanisms

(A) RBPs that can modulate DDR gene expression were identified through three approaches: (1) in vitro RNA-seq screening following siRNA-mediated knockdown of selected (N = 17, Table S2) pipeline-identified RBPs compared to either control siRNA (siCTRL) or siBRCA1 and siBRCA2 (siDDR) in isogenic cell lines; (2) identification of downregulated DDR genes expression levels in TCGA tumors with HRD driven by RBP loss compared to HR-competent tumors or tumors with HRD driven by DDR loss; and (3) identification of alternatively spliced DDR genes in TCGA patient tumors with HRD driven by RBP loss compared to HR-competent tumors or tumors with HRD driven by DDR loss. The network nodes indicate RBPs (green) identified to modulate DDR genes (red), with edges indicating how the modulation was identified. Genes listed across the bottom were not identified as being modulated by RBPs. The key for the network diagram, with solid arrows representing decreased expression and dotted arrows representing altered splicing, is on the left.

(B) IR-induced foci formation showing merged images of co-staining of RFP-SNRPE (red) and γH2AX (green). Total and chromatin-bound proteins prior to irradiation and chromatin-bound proteins at 1, 3, and 6 h following irradiation (5 Gy). Scale bar represents 10 μm.

(C) Magnified image of boxed cell indicated in (B) showing single-channel SNRPE (top) and γH2AX (middle) and the merged image (bottom). Arrows mark foci with co-localization of RFP-SNRPE and γH2AX.

(D) Quantification of fluorescence signal due to chromatin-bound RFP-SNRPE in SNRPE⁺ cells prior to and at 1, 3, and 6 h following irradiation. Intensity is reported relative to the median total intensity in cells not subjected to extraction. Mean ± SD.

(E) Quantification of the fluorescence signal due to γH2AX prior to and at 1, 3, and 6 h following irradiation in RFP-SNRPE⁺ and RFP-SNRPE^– cells. See also Figures S6 and S7. Mean ± SD.

For those RBPs that did not appear to be directly modulating expression of DDR genes, we hypothesized that they might act directly at sites of DNA damage. To evaluate whether RBPs might act at sites of damage, we tested whether the RBP SNRPE formed foci in response to IR-induced damage. Although at baseline SNRPE is largely nuclear, it is not tightly chromatin bound and most can be extracted (Figure 6B). Within 1 h after cells were irradiated, SNRPE became increasingly chromatin bound and co-localized with the DNA DSB marker γH2AX (Figures 6C and 6D). Furthermore, cells overexpressing SNRPE demonstrated quicker DSB repair as quantified by quicker loss of γH2AX foci (Figure 6E).

Based on the ability of RBPs to modulate expression of DDR genes, we hypothesized that a similar paradigm may apply to mutations in genes regulating translation by altering protein levels of DDR proteins. However, our analysis of TCGA RPPA data for DDR proteins revealed no significant relationships between mutations in translation genes, such as E4F1, and decreased DDR protein expression (Figures S7A and S7B). Nonetheless, as observed for SNRPE, we were able to detect IR-induced E4F1 foci (Figure S7C), suggesting that E4F1 might have a functional role at sites of DNA damage.

Putative HR drivers generalize to independent cohorts and are associated with cancer risk

To evaluate whether the undescribed HR drivers we uncovered in the TCGA data might be more broadly applicable to patient tumors in general, we analyzed additional patient cohorts. Interrogation of the International Cancer Genome Consortium (ICGC) breast cancer patient cohort validated that mutations from all identified ontologies were significantly associated with higher HRD scores (Figure 7A). Furthermore, RBP mutants accounted for a similar fraction of HRD tumors as that observed in the TCGA cohort (Figures 7B and 3E). The association of HRD score with candidate drivers was maintained when analyzing TNBC tumor alone (Figure 7C).

Figure 7.

External validation of the involvement of RNA-binding proteins in HR repair

(A) HRD scores were determined for an independent cohort of tumors from patients with breast cancer from ICGC. Tumors were classified based on their candidate drivers into either known DDR (pink), newly identified DDR (red), RBP (green), or other (purple) genes. Kruskal-Wallis with Dunn’s post hoc comparing each group to WT.

(B) Pie chart showing the relative proportions of proposed causes of HRD defects in patient tumors from the ICGC cohort.

(C) HRD scores as in (A), but only showing tumors from TNBC patients. Kruskal-Wallis with Dunn’s post hoc.

(D) Newly identified drivers of HR defects are enriched for genes associated with cancer risk from genome-wide association studies. The graph shows the distribution of number of genes associated with cancer risk using a random set of genes of equal size. The arrow indicates the actual observed number of genes associated with cancer risk. Empirical p value.

(E) Enrichment in genes associated with cancer risk as described in (D), but for individual gene ontologies.

(F) Network showing genes identified from the GWASs associated with cancer risk as well as the cancer types where mutations in those genes were identified as causing an HR defect. DDR genes (red), RBP genes (green), translation genes (blue), and other (purple). ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, ∗∗∗∗p < 1 × 10⁻⁴.

As loss of function in genes associated with HR is associated with increased propensity for cancer development, the genes we identified as candidate drivers of HRD should be enriched in genes associated with cancer risk in genome-wide association studies (GWASs). Therefore, we assembled all genes associated with cancer risk from GWAS DB³⁰ and looked for the HRD drivers identified through our pipeline. Overall, we found that our identified HRD drivers were significantly enriched for genes associated with cancer risk (Figure 7D). Analysis of individual gene ontologies indicated that this enrichment was largely driven by DDR and RBP genes (Figure 7E). Mutations in genes associated with cancer risk were detected across numerous cancer types (Figure 7F). These mutations, when occurring in the germline, could provide vital information for genetic counseling to improve cancer screening/prevention.

Discussion

Our analysis of genomic scars indicative of HRD across tumors in the TCGA indicated that only about 25% of HRD could be attributed to alterations in known drivers of HRD. The remaining 75% of HR-deficient tumors had no identifiable defects in known DDR genes. However, these tumors displayed gene and protein expression changes consistent with HRD caused by aberrations in DDR genes known to cause HRD, including activation of cell-cycle checkpoints and suppression of senescence/apoptosis pathways. Suppression of senescence/apoptosis pathways may be critical for HRD tumor cells to circumvent tumor-suppressive DDR checkpoints,¹⁹ thus enabling genomically unstable cells to continue proliferating and acquire additional mutations. This transcriptional rewiring may explain how tumor cells continue to proliferate in absence of BRCA1/BRCA2, whereas depletion of these genes in non-malignant cells often reduces cellular fitness. Using a networks-based approach, we more than doubled (from 462 to 1,088) the number of tumors in TCGA with an attributable driver of HRD. Among the newly discovered drivers, we found a particular enrichment for aberrations in genes encoding RBPs, which represented over half of newly identified drivers of HRD.

The potential role of RBPs in controlling HR is consistent with limited, previously published reports. For example, genome-wide siRNA depletion screens designed to evaluate HR function identified RBPs, namely RBMX²⁰ and CDC73,²¹ as potential drivers of HRD. An orthogonal genome-wide screen for PARP inhibitor sensitivity using CRISPR-mediated deletion also recovered genes consistent with our results, including SF3B3 and SF3B5.²² However, in vitro screening approaches do not necessarily correspond to bona fide drivers of HRD observed in patient tumors. Genes identified as critical for HR through loss of function screens might be essential genes, meaning loss is incompatible with cell viability, or they may simply not be mutated at appreciable frequencies in human populations. Additionally, the deletion/depletion of genes may not reflect the phenotypes observed when those same genes are mutated. The disparity between in vitro data and observations in patient tumors is best highlighted by the two aforementioned siRNA screens that identified 6,137²⁰ and 10,050²¹ genes that reduced HR function more than the average reduction observed following loss of canonical HR/BRCA-ness genes.¹⁵ In contrast, our study indicated that only 1.58% of the 6,137 genes and 1.04% of the 10,050 genes may be relevant in tumors from patients with cancer.

The genomic scar HRD score calculation and corresponding analyses performed in our study are subject to several limitations. The primary limitations are centered around the accuracy of the HRD score itself, as well as the chosen threshold for HRD-positive and -negative tumors. Although the combination of three different measures of genomic instability offers an improved signal over any single metric, it may still not capture all HRD tumors.¹¹ Further, we assumed a constant threshold value for HRD positivity across all tumor types, but the validity of this assumption is unclear. At the molecular level, we observed consistent changes between HRD driven by canonical drivers and those with originally unidentified etiology, suggesting that HRD score reflects loss of HR function in both contexts and that the HRD-positive tumors of unidentified etiology are not generally false positives. However, at the single-tumor level, the HRD score is still subject to false positives/negatives. For example, at the optimal threshold identified for bifurcation into HRD-positive and -negative tumors, roughly 20% of BRCA1/BRCA2 germline mutations were classified as existing in HR-competent tumors. The observed false-negative germline BRCA1/BRCA2 mutations could represent actual false negatives or may represent non-deleterious variants of BRCA1/BRCA2. By enforcing occurrence of genetic events in multiple tumors, we were largely able to avoid the effects of sporadic false positives, as evidenced by the over 90% experimental validation rate.

The findings presented in this study warrant follow up studies to further elucidate the mechanisms underlying how the newly identified drivers of HRD control HR repair. Previous reports have documented the role of RBPs in controlling expression of DDR factors; for example, RBMX has been shown to be required for BRCA2 expression.²⁰ We likewise detected that 26 and 47 RBPs analyzed may suppress expression of canonical HR genes. Consistent with our observation that ILF2 can modulate multiple DDR genes, ILF2 overexpression facilitates expression of DDR genes leading to resistance to DNA-damaging agents in 1q21-amplified multiple myeloma.²³ Following DNA damage, the spliceosome has also been shown to undergo rapid mobilization resulting in alternative splicing that may facilitate repair of DNA lesions.²⁴ Although analysis of Rad51 foci formation indicates RBPs largely induce HR defects at or before Rad51 loading, future studies to elucidate precisely which step(s) of HR putative drivers of HR defects are responsible for are warranted. Alternatively, RBPs may not interact directly with HR proteins but interfere with the generation of RNA species required for HR repair. For instance, recent work indicated that DICER and DROSHA RNA products are involved in activation of the DDR, and that DDR foci can be abrogated by RNase A-mediated RNA degradation.²⁵ Alternatively, evidence exists that endogenous RNA transcripts may serve as templates for HR repair, and loss of RBPs may have a primary effect on these template RNA molecules.²⁶ Furthermore, it remains unknown if RBP-deficient HR defective tumors utilize Rad52 for repair, as has been documented following loss of canonical HR repair genes such as BRCA1, BRCA2, PALB2, and RAD51C.²⁷ Further mechanistic insight into how RBPs and other newly identified HRD drivers modulate HR repair will advance our understanding of the diverse mechanisms used to promote genomic stability in human cells.

PARP inhibitors are gaining FDA approval in a variety of contexts. However, they are most commonly used for treating breast and gynecological cancers, and most on-going PARP inhibitor clinical trials are performed within this context. Across TCGA, we found that men tended to have higher HRD scores, strongly observed in the context of chromophobe kidney cancer, acute myeloid leukemia, pancreatic cancer, head and neck squamous cell carcinoma, and esophageal carcinoma. It is possible that because these cancers are enriched for HRD driven by previously undocumented RBPs that are likely to drive HRD, men with these specific HRD-positive cancers may constitute an understudied population. Loss of RBPs may promote sensitivity to PARPi directly via induction of HRD, through secondary mechanisms such as induction of R-loops or a combination thereof. Use of the HRD score and/or screening for mutations in the HRD driver genes we identified may provide biomarkers that could be used to stratify patients based on their predicted response to PARP inhibitors and expand the number of cancers that might benefit from such treatment. Alternatively, pharmacological induction of HRD by inhibiting RBPs may offer an approach to sensitize tumor cells to PARP inhibitors.

The previously undescribed drivers of HRD identified here were enriched for loci associated with cancer risk in prior GWASs. Risk-reduction surgery for BRCA-related breast cancer can decrease the risk of developing cancer from 85%–100%, but these surgeries are not without potential risk and should be focused on high-risk individuals.¹⁰ Integration of GWASs with mechanistic understanding can increase the confidence in the relevance of GWAS-uncovered candidate cancer risk loci. In turn, this could advance cancer prevention programs by improving the genetic information available to genetic counselors and patients.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
γH2AX, clone JBW301	Millipore Sigma	Cat# 05-636; RRID:AB_309864
SNRPB	GeneTex	Cat# GTX101883; RRID:AB_1951963
NBS1	GeneTex	Cat# GTX70224; RRID:AB_372445
ERCC1, clone EPR7277	abcam	Cat# ab129267; RRID:AB_11157618
Rad51, clone EPR4030(3))	abcam	Cat# ab133534; RRID:AB_2722613
PRPF8	Bethyl	Cat# A303-921A; RRID:AB_2620270
SF3B3	Bethyl	Cat# A302-508A; RRID:AB_1966103
FANCA, Clone D1L2Z	Cell Signaling	Cat# 14657; RRID:AB_2798558
FANCD2, Clone D5L5X	Cell Signaling	Cat# 16323; RRID:AB_2798761
Actin, Clone 8H10D10	Cell Signaling	Cat# 3700; RRID:AB_2242334
anti-Mouse HRP	Cell Signaling	Cat# 7076; RRID:AB_330924
anti-Rabbit HRP	Cell Signaling	Cat# 7074; RRID:AB_2099233
anti-Rabbit Alexa Fluor 488	ThermoFisher	Cat# A32731; RRID:AB_2633280
anti-Mouse Alexa Fluor 488	ThermoFisher	Cat# A-11029; RRID:AB_2534088
Streptavidin, Alexa Fluor 647	ThermoFisher	Cat# S21374; RRID:AB_2336066
Bacterial and virus strains
DH10B	ThermoFisher	Cat# 18297010
Chemicals, peptides, and recombinant proteins
KOD HotStart Polymerase	Sigma	Cat# 71086
Pladienolide B	Tocris	Cat# 60709
Mirin	Sigma	Cat# M9948
BMN673 (Talazoparib)	Selleck	Cat# S7048
AZD2281 (Olaparib)	Selleck	Cat# S1060
EdU	CarboSynth	Cat# NE08701
Azide-PEG3-Biotin	Sigma	Cat# 762024
Critical commercial assays
PrestoBlue	ThermoFisher	Cat# A-13262
Lipofectamine 3000	ThermoFisher	Cat# L3000015
Deposited data
RNAseq data following siRNA transfection	This paper	GEO: GSE153396
TCGA Pan-Cancer Atlas Data	TCGA Consortium	https://portal.gdc.cancer.gov
ICGC Data	ICGC	https://dcc.icgc.org
MicroRNA Targets	miRDB	http://mirdb.org
GWAS data	GWASdb v2	http://jjwanglab.org/gwasdb
PRPF8 eCLIP Data	ENCODE	ENCODE: ENCFF582YLB, ENCFF160MVU
Experimental models: Cell lines
BT-549	ATCC	Cat# HTB-122; RRID:CVCL_1092
MDA-MB-231	ATCC	Cat# HTB-26; RRID:CVCL_0062
HCC1806	ATCC	Cat# CRL-2335; RRID:CVCL_1258
HCC38	ATCC	Cat# CRL-2314; RRID:CVCL_1267
U2OS	ATCC	Cat# HTB-96; RRID:CVCL_A4CF
Oligonucleotides
Tag1-M13F: 5′-GGCAGACGTGCCTCACTC CCAGTCACGACGTTGTAAAACG-3′	Sigma	Custom synthesized
Tag2-M13R: 5′-CTGAGCTTGACGCATTGC TAGTGTCTCAAAATCTCTGATGTTAC-3′	Sigma	Custom synthesized
Tag1: 5′-GGCAGACGTGCCTCACTACT-3′	Sigma	Custom synthesized
Tag2: 5′-CTGAGCTTGACGCATTGCTA-3′	Sigma	Custom synthesized
siRNAs	Sigma	See Table S3
Recombinant DNA
pGIPZ shDDX3X-1	Dharmacon	Cat# V3LHS_644473
pGIPZ shDDX3X-2	Dharmacon	Cat# V2LHS_202531
pGIPZ shSF3B3-1	Dharmacon	Cat# V3LHS_644840
pGIPZ shSF3B3-2	Dharmacon	Cat# V2LHS_43924
pGIPZ shBRCA2-1	Dharmacon	Cat# V2LHS_89238
pGIPZ shBRCA2-2	Dharmacon	Cat# V2LHS_89237
pGIPZ shControl	Dharmacon	Cat# RHS5346
pCBASceI	Addgene	Cat# 26477
pDRGFP	Addgene	Cat# 26475
pCAG-EGFP	Addgene	Cat# 89684
pcDNA3.1-ccdB-TagRFP SNRPE WT	This Paper	N/A
pcDNA3.1-ccdB-TagRFP SNRPE c.208T>A	This Paper	N/A
pcDNA3.1-ccdB-TagRFP PRPF3 WT	This Paper	N/A
pcDNA3.1-ccdB-TagRFP PRPF3 c.227C>G	This Paper	N/A
pcDNA3.1-ccdB-TagRFP RBM15 WT	This Paper	N/A
pcDNA3.1-ccdB-TagRFP RBM15 c.1186C>T	This Paper	N/A
pcDNA3.1-ccdB-TagRFP RBM15 c.2009C>T	This Paper	N/A
pcDNA3.1-ccdB-TagRFP HNRNPC WT	This Paper	N/A
pcDNA3.1-ccdB-TagRFP HNRNPC c.211G>A	This Paper	N/A
pcDNA3.1-ccdB-TagRFP HNRNPC c.341C>T	This Paper	N/A
pcDNA3.1-ccdB-TagRFP LUC7L2 WT	This Paper	N/A
pcDNA3.1-ccdB-TagRFP LUC7L2 c.1205C>G	This Paper	N/A
pcDNA3.1-ccdB-TagRFP RBMX WT	This Paper	N/A
pcDNA3.1-ccdB-TagRFP RBMX c.462A>T	This Paper	N/A
pcDNA3.1-ccdB-TagRFP RBMX c.743A>G	This Paper	N/A
pcDNA3.1-ccdB-TagRFP RBMX c.760G>C	This Paper	N/A
pcDNA3.1-ccdB-TagRFP PRPF31 WT	This Paper	N/A
pcDNA3.1-ccdB-TagRFP PRPF31 c.18G>A	This Paper	N/A
pcDNA3.1-ccdB-TagRFP PRPF31 c.456T>A	This Paper	N/A
pcDNA3.1-ccdB-TagRFP SF3B5 WT	This Paper	N/A
pcDNA3.1-ccdB-TagRFP SF3B5 c.169G>T	This Paper	N/A
pcDNA3.1-ccdB-TagRFP SF3B5 c.57C>T	This Paper	N/A
pcDNA3.1-ccdB-TagRFP SF3B5 c.7G>A	This Paper	N/A
pcDNA3.1-ccdB-TagRFP CDC73 WT	This Paper	N/A
pcDNA3.1-ccdB-TagRFP CDC73 c.10G>A	This Paper	N/A
pcDNA3.1-ccdB-TagRFP CDC73 c.349G>A	This Paper	N/A
pcDNA3.1-ccdB-TagRFP CDC73 c.729G>T	This Paper	N/A
pcDNA3.1-ccdB-TagRFP SNRPB WT	This Paper	N/A
pcDNA3.1-ccdB-TagRFP SNRPB c.18C>T	This Paper	N/A
pcDNA3.1-ccdB-TagRFP SNRPB c.247G>T	This Paper	N/A
pcDNA3.1-ccdB-TagRFP SNRPB c.249G>C	This Paper	N/A
pcDNA3.1-ccdB-TagRFP SNRPB c.35A>G	This Paper	N/A
pcDNA3.1-ccdB-TagRFP SNRPB c.387G>T	This Paper	N/A
pcDNA3.1-ccdB-TagRFP SNRPB c.395G>T	This Paper	N/A
pcDNA3.1-ccdB-TagRFP SNRPB c.535A>G	This Paper	N/A
pcDNA3.1-ccdB-TagRFP SNRPB c.691C>G	This Paper	N/A
pcDNA3.1-ccdB-TagRFP WBP11 WT	This Paper	N/A
pcDNA3.1-ccdB-TagRFP WBP11 c.1016G>T	This Paper	N/A
pcDNA3.1-ccdB-TagRFP WBP11 c.1123A>G	This Paper	N/A
pcDNA3.1-ccdB-TagRFP WBP11 c.1216C>T	This Paper	N/A
pcDNA3.1-ccdB-TagRFP WBP11 c.387G>A	This Paper	N/A
pcDNA3.1-ccdB-TagRFP WBP11 c.900A>G	This Paper	N/A
Software and algorithms
MATLAB 2020a	Mathworks	https://www.mathworks.com
FlowJo v10.6.1 GraphPad	FlowJo, LLC	https://www.flowjo.com
Graphad Prism 9	Graphpad	https://www.graphpad.com
NIS-Elements Advanced Research	Nikon	https://www.microscope.healthcare.nikon.com
Cytoscape v3.5.1	Cytoscape	https://cytoscape.org
kallisto	Bray et al.28	https://pachterlab.github.io/kallisto

Resource availability

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Nidhi Sahni (nsahni@mdanderson.org).

Materials availability

All plasmids generated in this study are available from the lead contact Dr. Nidhi Sahni with a completed material transfer agreement.

Experimental model and subject participant details

Clinical cohorts

Two primary datasets were utilized within this study. First, data from The Cancer Genome Atlas pan-cancer atlas was used to identify new putative drivers of HR defects. These results were then validated in breast cancer samples from the International Cancer Genome Consortium. Informed consent was obtained at the time of the original studies. No a priori sample sizes were determined. Biological gender and ethnicity were analyzed as potential co-variates (see Figures 3F and S1).

Cell lines

Cell lines BT-549, MDA-MB-231, HCC1806, HCC38, and U2OS were acquired from ATCC. Cell lines were cultured in recommended media at 37°C in a humidified incubator with 5% CO₂. Cell identity was validated by STR testing and routinely tested for mycoplasma infection using the MD Anderson Cancer Center cell line core facility.

Method details

Calculation of HRD score

As described,³² we determined the genomic scar HRD scores from SNP arrays following the previous published 3 components of HRD score: loss of heterozygosity (LOH),³³ large scale transitions (LST),³⁴ and telomere allelic imbalance (TAI or NtAI),³⁵ and used the sum of all 3 scores as the final HRD score.^11,36,32

Association of molecular data with HRD score

To analyze RNA, miRNA, and proteins that were associated with high HRD scores we performed regression with a generalized linear mixed effects model, taking the tumor type as a random effect. For gene expression changes from RNAseq data, the resulting regression coefficients were used for gene set enrichment analysis as described.¹³ For protein data from RPPA, regression coefficients and corresponding p values were reported directly with false discovery rate determined with Benjamini–Hochberg procedure. For miRNA data, regression coefficients and corresponding p values were reported directly with the FDR determined using Storey’s method. To define those pathways that were preferentially suppressed by miRNAs, we retrieved predicted miRNA gene targets from MicroRNA Target Prediction Database (miRDB). For each significant microRNA, the regression coefficient was added to all candidate target genes, yielding a positive score for genes predicted to be preferentially targeted by microRNAs in HRD tumors. Resulting microRNA scores were utilized for GSEA, and the top 5 significant pathways were reported.

Threshold score to separate HRD positive vs. negative

To define a gold standard by which to define a threshold to identify HRD positive tumors, we selected breast, ovarian, prostate, and pancreatic tumors that had germline mutations in BRCA1 or BRCA2. For determining the optimal threshold value, we performed stochastic sub-sampling using 50% of all patients for 1000 iterations. For each iteration, we determined a modified Youden’s J statistic defined as J’ = 2∗true positive rate + true negative rate – 2, with increased weighting on the true positive rate because the definition of true positives was more robust than that of true negatives. The modal threshold value was determined to be 31, which was used to classify tumors as HRD positive or HRD negative.

Identification of previously undescribed candidate HRD drivers

To identify candidate HRD drivers, we initially used canonical DDR genes¹⁵ as well as highly validated candidates from two previous studies.^20,21 Genetic events were defined as either mutations or methylation events. Mutation events were constrained to variant allele frequencies (VAF) of at least 0.1. Methylation events were only considered if in the given tumor type methylation significantly correlated with gene expression (r ≤ −0.25), and the specific tumor exhibited both methylation and down-regulation (1 standard deviation) of the candidate gene. To assign genes to individual tumors, a scoring metric of 2∗(HRD score, upper tertile) + (HRD score, lower 20^th percentile) – 0.5∗(Coefficient of variation) was used to rank genes, with the highest scoring gene being assigned as the cause for a given tumor. To identify candidate genes for the next iteration, we first computed an interaction score using a merged protein interaction list from both the BioPlex affinity purification-mass spectrometry network³⁷ as well as a literature curated network.³⁸ For all genes, we determined the number of interactions with genes identified to be candidate drivers of HR defects, and then z-transformed these values. This z-normalized network score was averaged with a z-normalized score for induction of HR defects from two siRNA screens.^20,21 All genes that score above the 25^th percentile of candidate genes identified in the prior iteration were added for the next iteration. This process was completed until convergence, that is no additional candidate genes were identified.

DR-GFP HR reporter assay

The DR-GFP reporter assay was performed in U2OS cells per previous publications.¹⁷ Transfection with siRNA or the RFP-tagged protein of interest was performed 12 h prior to transfection with I-SceI (Addgene #26477) or GFP control (Addgene #89684). Alternatively, cells were treated with Pladienolide B or Mirin prior to I-SceI/GFP transfection. All transfections were performed with Lipofectamine 3000 per the manufacturer’s instructions. Two days after I-SceI/GFP transfection, samples were analyzed by flow cytometry. After gating for singlets and cell size, GFP positivity was gated based on SSC vs. GFP intensity. All values were normalized to the average of siCTRL and non-transfected controls for siRNA experiments, and RFP-tagged wild-type protein overexpression for analysis of RFP-tagged mutant proteins.

RFP-tagged mutant constructs

To generate point mutations, we implemented a modified high-throughput site-directed mutagenesis pipeline described previously.^39,40 Briefly, we used the corresponding wild-type reference ORFs from their Entry clones in human ORFeome as template for a 3-step PCR experiment. For a given mutation, PCR cloning consisted of two “primary PCRs” to generate gene fragments, and one “fusion PCR” to obtain the mutated ORF. For the primary PCRs, two universal primers, Tag1-M13F and Tag2-M13R (key resources table), and two ORF-specific internal forward and reverse primers were employed. The two universal primers allowed the preservation of the attL sites on both ends of the ORF. The mutation-specific primers (namely MutF and MutR), encompassing the desired single nucleotide change, were designed to have an overlapping region of ∼40 base pairs. The two ORF fragments flanking the mutation of a gene were amplified using the primer pair Tag1-M13F and MutR, and the primer pair Tag2-M13R and MutF, respectively. For the fusion PCR, the two primary PCR fragments were fused together using the primer pair Tag1 and Tag2 (key resources table) to generate the single amino acid change mutation allele. The final product was a full-length ORF harboring the desired mutation. All wild-type and mutant allele clones were transferred by Gateway recombination into a mammalian expression vector containing a C-terminal RFP tag. For subsequent sequence confirmation, the inserts were PCR amplified with KOD HotStart Polymerase and verified by Sanger sequencing.

IR induced foci formation

Cells were grown on glass coverslips, irradiated with either 5 Gy or 10 Gy, and incubated as specified. To analyze chromatin bound fractions, the soluble fraction was extracted prior to fixation, as described.⁴¹ For analysis of RFP-tagged protein foci formation, cells were transfected 48 h prior to irradiation using Lipofectamine 3000 per manufacturer’s instructions prior to fixation and detection of phosphorylated histone variant H2AX using indirect immunofluorescence with anti-γH2AX (clone JBW301, Millipore Sigma). For IR-induced Rad51 foci, cycling cells were pulse labeled with 10 μM EdU prior to irradiation. EdU was labeled with CLICK chemistry as described.⁴² Nuclei and Rad51 foci were segmented, and EdU positivity was determined from integrated intensity in each nuclei compared to a no EdU stained control. Rad51 positivity was analyzed only in EdU positive cells, and defined relative to control cells without irradiation. Threshold values for positivity were not modified between samples. For siRNA experiments, cells were transfected with siRNA 48 h prior to irradiation. For Pladienolide B experiments, cells were treated with Pladienolide B (Tocris) 24 h prior to irradiation. Cells were imaged by fluorescence microscopy (Eclipse TE2000E, Nikon), capturing all images for a given replicate simultaneously to assure no variances in light intensity. All quantification was performed in MATLAB R2016a.

PARP sensitivity assays

For short term assays, cells were plated in 96 well plates before treatment with specified concentrations of BMN673 (Selleck), AZD2281 (Selleck), Pladienolide B, or vehicle control. Cells were incubated for 5 days, and viability was assessed using PrestoBlue (Invitrogen) relative to vehicle controls (DMSO for BMN673 and AZD2281, ethanol for Pladienolide B). Synergy was assessed using the Chou-Talalay combination index.⁴³ For clonogenic assays, cells were plated in 12 well plates and incubated with drugs for two weeks before fixation and staining with crystal violet. Plates were scanned, and the crystal violet was extracted with 10% acetic acid. Absorbance of solubilized crystal violet was measured at 590 nm using a plate reader, and viability was normalized to a solvent-treated control. For stable shRNA cells, Dharmacon pGIPZ Lenti shRNA vectors were used including a pGIPZ non-silencing control (RHS5346), shDDX3X (V3LHS_644473, V2LHS_202531), shSF3B3 (V3LHS_644840, V2LHS_43924), and shBRCA2 (V2LHS_89238, V2LHS_89237).

RNAseq with depletion of RNA binding proteins

Cells (BT-549, MDA-MB-231, or U2OS) were transfected with the desired siRNAs (Table S3) using Lipofectamine 3000 48 h prior to RNA isolation with a Qiagen RNeasy Kit. RNA quality was confirmed using the Agilent TapeStation RNA reagents according to the manufacturer’s protocol. Libraries were prepped using the Lexogen QuantSeq 3′ mRNA-Seq Library Prep Kit FWD for Illumina with 6nt unique dual indexing using. 100ng of RNA was used as input material for an automated protocol adapted for the PerkinElmer SciClone NGS Workstation. Libraries were analyzed for quality using the Agilent TapeStation High Sensitivity DNA kit and for quantity using Qubit dsDNA HS assay, in 384-well format, using 20μL reactions in triplicate (19μL working reagent + 1μL sample or standard) with an 11-point standard curve from 0 to 10 ng/μL. Plates were shaken in the plate reader for 5 s, then measured for fluorescence (excitation: 480nm, emission: 530nm). Sample concentrations were determined using the standard curve. Libraries made from each RNA sample were then pooled at 25 nM each, denatured with 1M NaOH added to a 0.2M final concentration (5 min at room temperature), and quenched with 200mM Tris HCl (pH 7). 1% PhyX spike-in (Illumina) was added then pooled, denatured libraries were run on an Illumina NovaSeq with a NovaSeq 6000 SP Reagent Kit (100 cycles) using 51bp reads, 6bp index reads, and paired-end single read parameters. Resulting reads were quantified using kallisto.²⁸

Identification of RNA binding proteins that may modulate DDR genes

To test whether RBPs were modulating DDR genes, we performed multiple analyses. First, RNAseq analysis following siRNA-mediated depletion of 17 RBPs in 3 cell lines was used to identify differential DDR gene expression relative to either siCTRL, or siBRCA1/siBRCA2 by paired t-test. Next, TCGA tumors were analyzed to detect decreased DDR gene expression relative to HR competent tumors or tumors with HRD caused by DDR genes using a generalized linear mixed effects model, taking the tumor type as a random effect. Finally, the same comparisons in TCGA tumors was made using alternative splicing instead of gene expression levels. We considered gene expression changes that caused decreased gene expression with an FDR of less than 10%, or events that were detected in both patient tumors and cell lines with a nominal p value of at least 0.05. For alternative splicing, increased or decreased alternative splicing at 10% FDR was considered an event. The resulting network was visualized in Cytoscape v3.5.1. Interactions between PRPF8 and putative target genes were assessed using as previously described.⁴⁴ Raw data may be using identifiers ENCFF582YLB and ENCFF160MVU.

Western blot. Cells were lysed in urea lysis buffer (8M urea, 1% β-mercaptoethanol, 50 mM Tris pH 7.5), separated by SDS-polyacrylamide gel electrophoresis, and then transferred onto nitrocellulose membranes. Membranes were blocked in 5% BSA, and then probed with desired primary antibody followed by horseradish peroxidase-conjugated secondary antibodies for detection. Primary antibodies were SNRPB (GeneTex, GTX101883), NBS1 (GeneTex GTX70224, Clone 1D7), ERCC1 (abcam, ab129267, Clone EPR7277), Rad51 (abcam, ab133534, clone EPR4030(3)), PRPF8 (Bethyl, A303-921A), SF3B3 (Bethyl, A302-508A), FANCA (Cell Signaling, 14657, Clone D1L2Z), FANCD2 (Cell Signaling, 16323, Clone D5L5X), and Actin (Cell Signaling, 3700, Clone 8H10D10). Secondary antibodies were both acquired from Cell Signaling, anti-Mouse HRP #7076 and anti-Rabbit HRP #7074.

Quantification and statistical analysis

Specific statistical tests are discussed within corresponding sections. In general, pan-cancer associations with HRD score were determined using a generalized linear mixed effects model, taking the tumor type as a random effect. Multiple comparisons were corrected for using either Storey’s method (large number of variables) or the Benjamini–Hochberg procedure (smaller number of variables). Comparisons of normally distributed data were made using either t tests (2 groups) or one-way ANOVA (3+ groups) with appropriate post-hoc analysis. Comparisons of non-normally distributed data were made using rank-sum test (2 groups) or Kruskal-Wallis (3+ groups) with appropriate post-hoc test. Correlations were assessed using Spearman rank correlation coefficients. Categorical variables were compared using Fisher’s exact test.

Published: October 30, 2023

Data and code availability

TCGA RNAseq gene expression, alternative splicing, methylation, copy number, RPPA, and clinical data were downloaded from the GDC data commons (https://portal.gdc.cancer.gov). TCGA CPTAC data were acquired from the manuscript’s supplemental information.²⁹ International Cancer Genome Consortium (ICGC) data were downloaded from the ICGC data portal (https://dcc.icgc.org). GWAS data were downloaded from GWASdb v2 (http://jjwanglab.org/gwasdb).³⁰ MicroRNA targets were downloaded from miRDB (http://mirdb.org).³¹ Table S1 contains all HRD tumors, with identified HRD drivers and gene annotations. RNAseq data generated in this manuscript are available at NCBI GEO: GSE153396. Any analysis code is available from the authors upon request without restriction. Any additional information required to reanalyze the data reported in this work paper is available from the lead contact upon request.