Revealing biomarkers associated with PARP inhibitors based on genetic interactions in cancer genome

Graphical abstract

Highlights

• •Candidate genomic biomarkers were revealed for PARPis from genetic interactions.
• •Gain-of-function mutation of EGFR induced resistance to PARP inhibitors.
• •Lung cancer may benefit from combination of PARP inhibitor and EGFR inhibitor.
• •Gene set of biomarkers for PARPis contributes to the prognosis of cancer patients.

Abstract

Poly (ADPribose) polymerase inhibitors (PARPis) are clinically approved drugs designed according to the concept of synthetic lethality (SL) interaction. It is crucial to expand the scale of patients who can benefit from PARPis, and overcome drug resistance associated with it. Genetic interactions (GIs) include SL and synthetic viability (SV) that participate in drug response in cancer cells. Based on the hypothesis that mutated genes with SL or SV interactions with PARP1/2/3 are potential sensitive or resistant PARPis biomarkers, respectively, we developed a novel computational method to identify them. We analyzed fitness variation of cell lines to identify PARP1/2/3-related GIs according to CRISPR/Cas9 and RNA interference functional screens. Potential resistant/sensitive mutated genes were identified using pharmacogenomic datasets. We identified 41 candidate resistant and 130 candidate sensitive PARPi-response related genes, and observed that EGFR with gain-of-function mutation induced PARPi resistance, and predicted a combination therapy with PARP inhibitor (veliparib) and EGFR inhibitor (erlotinib) for lung cancer. We also revealed that a resistant gene set (TNN, PLEC, and TRIP12) in lower grade glioma and a sensitive gene set (BRCA2, TOP3A, and ASCC3) in ovarian cancer, which were associated with prognosis. Thus, cancer genome-derived GIs provide new insights for identifying PARPi biomarkers and a new avenue for precision therapeutics.

1. Background

Poly (ADP-ribose) polymerases (PARPs) are DNA repair enzymes that play roles in DNA damage repair and replication fork protection. In cancer therapy, PARP inhibitors (PARPis) target PARP genes by reducing DNA repair function and increasing replication fork errors in cancer cells [1]. Currently, four PARPis (olaparib, rucaparib, niraparib, and talazoparib) have been approved by the U.S. Food and Drug Administration. Additionally, PARPis have been included in the clinical guidelines for treatment of ovarian, breast, pancreatic, and prostate cancers, and clinical trials of PARPis in other cancers are under way.

PARPis are associated with resistance in clinical therapeutics. A possible mechanism of PARPi resistance includes restoration of homologous recombination (HR) repair ability due to secondary mutations in BRCA1/2, and depletion of HR compensatory repair pathways such as the non-homologous end joining pathway [2], [3]. Patients with BRCA1/2 mutations are sensitive to PARPis, where PARPis lead to cancer cell death by utilizing a synthetic lethality (SL) interaction with BRCA1/2 deficiency. However, it is necessary to identify effective biomarkers to predict the response to PARPis in patients without BRCA1/2 mutations as well [4]. Thus, it is important to identify genomic biomarkers that overcome PARPi resistance and expand its application to promote precision medicine.

SL and synthetic viability (SV) interactions are known as genetic interactions (GIs). GIs refer to the effect of simultaneous changes in two genes that is different from the expected additive effect of an individual change [5]. SL interaction is a phenomenon when two gene defects that cause cell death or significant impairment of fitness, emerge simultaneously. SL interactions contribute to the identification of new drug targets or drug sensitivity biomarkers [6]. Meanwhile, SV interaction refers to the combination of gene effects that rescue the lethal effects of single gene alterations. Our previous work has demonstrated that SV interactions in the cancer genome could induce drug resistance [7], [8]. Our study aimed to detect PARPi response biomarkers using the GI network in cancer cells based on the hypothesis that genes with SL interactions with PARP1/2/3 may be potential sensitive biomarkers for PARPis, and genes with SV interactions with PARP1/2/3 may induce resistance to PARPis. However, the GI network for each cancer type is unavailable. Recent functional genomic technologies such as CRISPR/Cas9 assays and RNA interference (RNAi) have aided in detection of essential genes by gene knockdown and knockout, and have provided innovative tools to screen for cancer GIs [9]. For instance, CRISPR-based genetic screens have identified SL interactions and molecular functions of genes in acute myeloid leukemia cell lines [10]. Horn et al. have annotated metazoan genes based on GI profiles that were integrated by an RNAi-based experimental approach [11]. However, most studies focused on SL interactions and have ignored SV interactions as well as prediction values for determining drug response.

Thus, in this study, we proposed a novel computational method to identify candidate responsive resistant and sensitive biomarkers for PARPis by utilizing CRISPR/Cas9 and RNAi screens, and pan-cancer pharmacogenomic datasets.

2. Results

2.1. Identification of gene mutations as candidate biomarkers for PARPis

We used CRISPR/Cas9 and RNAi screens to identify PARP1/2/3-related GIs (Fig. 1A). If PARP1/2/3 was knocked down and gene A mutations caused a significant selective enhancement (or reduction) in cancer cell viability, then gene A mutations may have an SV (or SL) relationship with PARP1/2/3 inhibition. We identified mutated genes that significantly affected PARPi response (IC50 or area under the curve [AUC]) in specific cancer cell types (Fig. 1A; P < 0.01; one-sided Wilcoxon rank-sum test). Mutated genes, have SL interactions with PARP1/2/3 knockdown, were associated with PARPi sensitivity in cancer cells and predicted as candidate sensitive biomarkers of PARPis. Whereas, mutated genes that had an SV effect with PARP1/2/3 knockdown, and were associated with PARPi resistance in cancer cells, were predicted as resistant candidate biomarkers of PARPis (Fig. 1A; P < 0.01; one-sided Wilcoxon rank-sum test). In total, 41 resistant and 130 sensitive genes were identified as candidate biomarkers for PARPis (Fig. 2; Table S1).

Fig. 1.

Workflow of the study. (A) Identification of potential biomarkers for PARP inhibitors (PARPis) using cancer genome data. (B) Mechanism analysis for biomarkers. (C) Verification of biomarkers for PARPis in The Cancer Genome Atlas (TCGA). (D) Prediction of drug combination with PARPis.

Fig. 2.

Candidate genomic biomarkers for PARP inhibitors (PARPis) in cancer. Each diamond-marked module shows mutated genes as biomarkers for PARPis predicted in 11 cancer tissues. Circle nodes indicate PARPi resistance-related mutated genes and square nodes indicate PARPi sensitivity-related mutated genes. Node color corresponds to -logP, where P values were calculated using Wilcoxon rank-sum test. The red part indicates the significance that was tested from pharmacogenomic datasets of PARPis, and the green part depicts the significance that was tested using CRISPR/Cas9 and RNA interference functional screens. The border color of the nodes in red indicates different drugs: pink is olaparib, yellow is veliparib, green is rucaparib, orange is talazoparib, and purple is niraparib. Dashed line of the node border represents the P value from IC50, and the solid line indicates the P value from area under the curve (AUC). Border color of the node in the green part shows the significance detected in DepMap_CRISPR (green), DepMap_shRNA (red), and ProjectScore_CRISPR (blue). Edge colors denote the five PARPis. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Patients with BRCA1/2 mutations are sensitive to PARPis in cancer cells. Our results showed that cancer cells with BRCA2 mutations were sensitive to olaparib and niraparib in large intestine tissues (Fig. 3A, B; P < 0.01; one-sided Wilcoxon rank-sum test). In the Achilles shRNA dataset, after PARP1 knockdown, large intestine cancer cells with BRCA2 mutation showed poorer proliferation than cancer cells with wild type BRCA2 (Fig. 3C; P < 0.01; one-sided Wilcoxon rank-sum test). Besides, according to olaparib response data in Genomics of Drug Sensitivity in Cancer (GDSC) and PARP2 knockdown data, mutation of BRCA1 was confirmed as a sensitive biomarker for PARPis (Fig. S1; P < 0.05; one-sided Wilcoxon rank-sum test). Moreover, mutations in some cancer genes, such as RB1 (Fig. 3D-I; P < 0.01; one-sided Wilcoxon rank-sum test) and MACF1 (Fig. S2A-F; P < 0.01; one-sided Wilcoxon rank-sum test), were predicted as candidate sensitive biomarkers for PARPis, and RB1 and TOP3A mutations were predicted as candidate sensitive biomarkers for PARPis based on both AUC and IC50 values (Fig. 3D-P; P < 0.01; one-sided Wilcoxon rank-sum test). In GDSC database, TTN and TRIP12 mutations were predicted as resistant biomarkers for PARPis based on both AUC and IC50 values (Fig. 3Q-V; P < 0.01; one-sided Wilcoxon rank-sum test). Importantly, RB1 and NRDC mutations were predicted as candidate sensitive biomarkers in both the GDSC database and Cancer Therapeutics Response Portal (CTRP; Fig. 3D-I for RB1; Fig. S2G-S for NRDC; P < 0.01; one-sided Wilcoxon rank-sum test).

Fig. 3.

Candidate biomarkers for PARP inhibitors (PARPis) in cancer. (A-C) Large intestine cell lines with BRCA2 mutation are sensitive to PARPis in Genomics of Drug Sensitivity in Cancer (GDSC) data, and have worse viability when PARP1 is knocked down in DepMap_shRNA data. (D-J) Lung cell lines with RB1 mutation are sensitive to PARPis in GDSC and Cancer Therapeutics Response Portal (CTRP), and have worse viability when PARP3 is knocked down in DepMap_shRNA data. (K-P) Large intestine cell lines with TOP3A mutation are sensitive to PARPis in GDSC, and have worse viability when PARP2 is knocked down in ProjectScore_CRISPR data. (Q-S) Central nervous system cell lines with TTN mutation are resistant to PARPis in GDSC, and have better viability when PARP3 is knocked down in DepMap_CRISPR data. (T-V) Central nervous system cell lines with TRIP12 mutation are resistant to PARPis in GDSC, and have better viability when PARP3 is knocked down in DepMap_CRISPR data. Biomarkers identified in GDSC1 or GDSC2 are labeled as “GDSC”.

The triple-negative breast cancer patients who treated with veliparib plus DNA-damaging agents (carboplatin and paclitaxel) were used to validate proposed biomarkers [12]. Mutations of PPP1R12B was predicted as a sensitive biomarker of PARPi in breast cancer (BRCA). BRCA with PPP1R12B mutations showed lower expression than BRCA with wild type PPP1R12B in The Cancer Genome Atlas (TCGA) (Fig. S3A; P = 9.0E-02; one-sided Wilcoxon rank-sum test). The expression of PPP1R12B in pathological complete response (pCR) patients were significantly lower than those in residual disease (RD) patients (Fig. S3B; P = 5.0E-02; one-sided Wilcoxon rank-sum test).

In addition, we obtained drug response data (IC50 and AUC) of four PARPis (PARP_0108, PARP_9482, PARP_9495 and TANK_1366) from GDSC for further validation. IC50 and AUC values in cell lines with mutations of resistant genes (e.g. FREM1 and PCDH12) were higher than cell lines with wild type resistant genes in at least three PARPis (Fig. S4A, B; P < 0.05; one-sided Wilcoxon rank-sum test). IC50 and AUC values in cell lines with mutations of some sensitive genes (e.g. MACF1, TOP3A, and NOTCH1) were lower than cell lines with wild type resistant genes in three PARPis (Fig. S4C-E; P < 0.05; one-sided Wilcoxon rank-sum test).

2.2. Disrupting HR to mediate response to PARPis in cancer cells

PARP inhibition causes SL in cancers with defects in HR. However, HR restoration leads to resistance to PARPis even in cancers with BRCA1/2 mutations or hypermethylation [13]. Thus, we investigated whether the genomic biomarkers predicted by GI could disrupt HR to mediate the response of PARPis in cancer cells (Fig. 1B). We found that HR proteins were significantly differentially expressed in cancer samples with mutations of biomarker genes than in cancer samples with wild type biomarker genes (Fig. 4; P < 0.05; one-sided Wilcoxon rank-sum test). According to HR protein expression analysis, resistant genes affected the expression of HR proteins such as RAD51 and BRCA2 (Fig. 4A, B). Expression of RAD50 was significantly downregulated by 44.0% (62/141) sensitive genes in TCGA data and 44.3% (43/97) sensitive genes in Cancer Cell Line Encyclopedia (CCLE) data (Fig. 4C, D). RAD50 deletion is a candidate marker of response to PARPis in BRCA wild type ovarian cancer [14]. Furthermore, expression of ATM protein was significantly downregulated by 37.6% (53/141) sensitive genes in TCGA data and 42.3% (41/97) sensitive genes in CCLE data (Fig. 4C, D). Further, ATM mutant cell line significantly reduces cancer load and increases survival of animals after olaparib treatment in vivo, and ATM depletion sensitizes breast cancer cells to PARP inhibition [15], [16].

Fig. 4.

Differential expression of homologous recombination (HR) proteins. (A) Resistant genes disrupt the expression of HR proteins in The Cancer Genome Atlas (TCGA). (B) Resistant genes disrupt the expression of HR proteins in Cancer Cell Line Encyclopedia (CCLE). (C) Sensitive genes disrupt the expression of HR proteins in TCGA. (D) Sensitive genes disrupt the expression of HR proteins in CCLE. Rows of the matrix indicate resistant/sensitive genes. Red entries represent upregulation of the HR proteins, and green entries represent downregulation of the HR proteins in the heatmap. Heatmap shows -logP values, where P values were calculated by one-sided Wilcoxon rank-sum test. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Homologous recombination deficiency (HRD) genomic scarring was scored by sum of telomeric allelic imbalance (TAI) score, the large-scale state transitions (LST) score and the loss of heterozygosity (LOH) score [17]. We investigated the effect of candidate biomarker mutation on HRD by testing the HRD score between groups of mutated and wild type candidate biomarkers. HRD score was significantly lower in the mutation group than wild type of proposed resistant genes in TCGA, such as DYNC1H1, BCL9L, and MFHAS1 (Fig. S5A-C; P = 3.1E-09 for DYNC1H1, P = 4.6E-05 for BCL9L, P = 7.1E-04 for MFHAS1; one-sided Wilcoxon rank-sum test). HRD score was significantly higher in the mutation group than wild type of proposed sensitive genes, such as PCDH10, GHR, and TRIP11 in LUAD (Fig. S5D-F; P = 1.2E-05 for PCDH10, P = 1.5E-05 for GHR, P = 4.9E-04 for TRIP11; one-sided Wilcoxon rank-sum test). The above results confirmed that mutated genes that were predicted as candidate biomarkers might affect the therapeutic effect of PARPis by disrupting the HR mechanism.

2.3. PARPi response proposed biomarkers have functional interactions with PARP1/2/3

We suspected that PARPi-related resistant or sensitive genes might interact with PARP1/2/3 to affect the pharmacodynamic effect of PARPis through functional regulation. Thus, we constructed protein–protein interaction (PPI) networks of proposed resistant or sensitive genes and PARP1/2/3. Most drug response genes directly or indirectly interacted with PARP1/2/3, especially cancer-related genes (e.g., EGFR, ABL1, and RB1) and DNA damage repair (DDR) genes (e.g., BRCA2, ASCC3, TOP3A, and SMARCA4). EGFR was a hub gene in the PARPi-resistant PPI network, and controlled the state change of PARP1. Candidate resistant TJP1 and TTN genes, neighbors of EGFR, indirectly interacted with PARP1 (Fig. 5A). TJP1 plays a role in regulating cell adhesion and matrix remodeling, and loss of TJP1 causes MCF10A cells to become resistant to PARPis, including olaparib, veliparib, talazoparib, and rucaparib [18], [19]. Fig. 5B shows the proposed sensitive genes of PARPis and PARP1/2/3 interaction network, within which the SL effects observed between hub genes (RB1, BRCA2, KRAS, etc.) and PARP1/2/3 have been reported in previous studies. For instance, mutations in KRAS have SL interactions with PARP1/2 knockout in colon adenocarcinoma [20]. Large intestine cancer cell lines with KRAS mutations were sensitive to PARPis in GDSC (AUC; P = 5.8E-03 for niraparib; one-side Wilcoxon rank-sum test) and had SL interactions with PARP3 in ProjectScore_CRISPR data (P = 2.5E-04; one-sided Wilcoxon rank-sum test). Moreover, we found that RB1 mutations could mediate the sensitivity to PARPis in lung cancer cell lines (Fig. 3D-J), which is consistent with previous findings that RB1 has SL interactions with PARP1 in multiple cancers, including lung adenocarcinoma [9]. Large intestine cell lines with cancer gene STAT3 mutation were sensitive to PARPis in GDSC (IC50; P = 7.3E-03 for veliparib; one-sided Wilcoxon rank-sum test) and had SL interaction with PARP1 in DepMap_shRNA (P = 4.3E-03, one-side Wilcoxon rank-sum test), and STAT3 had indirect interaction with PARP1.

Fig. 5.

Protein-protein interaction networks (PPI) between PARP1/2/3 and resistant or sensitive genes of PARP inhibitors. (A) PPI network of resistant genes and PARP1/2/3. (B) PPI network of sensitive genes and PARP1/2/3. Red nodes represent PARP1/2/3, orange nodes indicate genes that have direct interactions with PARP1/2/3, and gray nodes are genes that have indirect interactions with PARP1/2/3. Color of the edges denotes seven types of interaction relationships. Red squares represent cancer genes, and green squares represent genes with DNA damage repair function. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

2.4. Functional analysis of candidate PARPis responsive biomarkers

PARPi-related resistant or sensitive genes had the same Gene Ontology (GO) term annotations as that of PARP1/2/3 (Fig. S6), which suggests that the biomarkers have SL or SV effects with PARPis by disrupting specific molecular functions and biological processes. To explore the biological functions of candidate PARPi response biomarkers, we performed enrichment analysis using molecular function and biological process terms from the GO database.

Resistant genes of PARPis were significantly enriched in molecular functions associated with microtubule and dynein activities, such as “ATP-dependent microtubule motor activity, minus-end-directed” (Fig. 6A; P = 9.28E-06, hypergeometric test), “dynein intermediate chain binding” (Fig. 6A; P = 4.08E-05, hypergeometric test), and “dynein light intermediate chain binding” (Fig. 6A; P = 4.08E-05, hypergeometric test). Paclitaxel can stabilize cellular microtubules and block chromosome segregation, and has shown success in clinical treatment of breast and ovarian cancers. Therefore, PARPi resistance can be overcome by combining it with paclitaxel. Sensitive genes of PARPis were significantly enriched in “calcium ion binding” (Fig. 6B; P = 3.97E-06; hypergeometric test), “alpha-1, 4-glucosidase activity” (Fig. 6B; P = 1.61E-04; hypergeometric test), and “maltose alpha-glucosidase activity” (Fig. 6B; P = 1.61E-04; hypergeometric test). Calcium (Ca²⁺) signaling is involved in regulating important biological processes such as apoptosis [21]. Thus, these results indicate that mutations in sensitive genes induce an increase in free Ca²⁺, which could promote cancer cell death.

Fig. 6.

Gene Ontology (GO) enrichment terms for candidate resistant and sensitive genes of PARP inhibitors. (A) Resistant and (B) sensitive gene enrichment with GO term “molecular function.” (C) Resistant and (D) sensitive gene enrichment with GO term “biological process.” Red and green circles indicate resistant and sensitive gene enrichment, respectively. Graph exhibits the top 10 enrichment results with false discovery rate (FDR) < 0.1. Left part of the circles represent resistant or sensitive genes, and right part of the circles represent the enrichment terms. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Resistant genes of PARPis were significantly enriched in biological processes of tyrosine phosphorylation and synaptic transmission, such as “peptidyl-tyrosine phosphorylation” (Fig. 6C; P = 1.02E-05; hypergeometric test), “positive regulation of synaptic transmission, glutamatergic” (Fig. 6C; P = 1.96E-05; hypergeometric test), and “positive regulation of excitatory postsynaptic potential” (Fig. 6C; P = 2.87E-05; hypergeometric test). Tyrosine phosphorylation governs many cancer hallmarks, such as cell proliferation [22], which could induce resistance to PARPis. Sensitive genes of PARPis were significantly enriched in multiple biological processes such as “homophilic cell adhesion via plasma membrane adhesion molecules” (Fig. 6D; P = 3.69E-06; hypergeometric test), “aortic valve morphogenesis” (Fig. 6D; P = 4.91E-05; hypergeometric test), and “sodium ion transmembrane transport” (Fig. 6D; P = 1.51E-04; hypergeometric test). Adhesive interactions play a critical role in metastatic cancer dissemination [23].

2.5. Mutations in PARPi response-related gene sets to indicate prognosis

As the mutation frequency of individual PARPi response-related genes was quite low, we merged the PARPi resistant or sensitive genes in one specific cancer tissue or biological process to form a gene set. For candidate resistant genes, we focused on three genes (TNN, PLEC, and TRIP12) in the central nervous system cell lines, which were detected in lower grade glioma (LGG) and glioblastoma (GBM) from TCGA (Fig. 3Q-V and 7A-C). In LGG patients with PARP copy number deletions (as a simulation of PARP inhibition), those with mutations in the resistant gene set showed poorer prognosis than those without mutations in this gene set (Fig. 7D, E; P = 0.27; log-rank test with overall survival [OS]; P = 9.4E-03, log-rank test with Progress Free Survival [PFS]). At the transcriptome level, in cancer patients with PARP1/2/3 downregulation (also a simulation of PARP inhibition), cholangiocarcinoma (CHOL) and kidney renal papillary cell carcinoma (KIRP) patients with mutations in this candidate resistant gene set had worse PFS than those without mutations in this candidate resistant gene set (Fig. S7A, B; P = 1.9E-04, log-rank test with PFS, CHOL; P = 9.4E-03, log-rank test with PFS, KIRP).

Fig. 7.

PARP inhibitors response-related gene sets to indicate prognosis. (A)-(C) Central nervous system cell lines with PLEC mutation were resistant to olaparib in Genomics of Drug Sensitivity in Cancer (GDSC; LN_IC50, area under the curve [AUC]), and had better viability when PARP3 were knocked down. (D)-(E) Kaplan-Meier overall survival (OS) and progression-free survival (PFS) analysis of lower grade glioma (LGG) patients is represented in two groups: TTN, PLEC, and TRIP12 mutations, and others. (F)-(G) Lung cell lines with ASCC3 mutation were sensitive to veliparib in Cancer Genome Project (CGP; IC50), and had worse viability when PARP2 were knocked down. (H)-(I) Kaplan-Meier OS and PFS analysis of ovarian cancer patients with BRCA2, ASCC3, and TOP3A mutations, and others. P values were calculated based on log-rank test.

For candidate sensitive genes, we focused on three DDR genes (BRCA2, TOP3A, and ASCC3), which interacted with PARP1/2/3 (Fig. 3A-C, 3K-P, and 7F, G). Ovarian cancer (OV) patients with PARP copy number deletions and mutations in the DDR gene set showed better survival than those without mutations in this gene set (Fig. 7H, I; P = 0.011, log-rank test with OS; P = 0.023, log-rank test with PFS). In OV patients with PARP1/2/3 downregulation, patients with mutations in the DDR gene set showed better prognosis than those without mutations in this gene set (Fig. S7C, D; P = 1.4E-03, log-rank test with OS; P = 0.011, log-rank test with PFS). Prognostic analysis demonstrated the reliability of our predicted candidate biomarkers for PARPis.

2.6. Prediction of drug combination for PARPis

Frequently acquired resistance to PARPis has spurred efforts to combine PARPis with other agents, such as chemotherapeutic agents, cell cycle checkpoint inhibitors, antiangiogenic agents, and PI3K inhibitors [1]. Inhibition of resistance-related activated mutations may provide clues to develop a combination strategy for PARPis (Fig. 1D). Cancer cell lines with EGFR mutations tended to be resistant to olaparib according to the pharmacogenomic data in CTRP (Fig. 8A; P = 6.5E-03, one-sided Wilcoxon rank-sum test), and had SV interaction with PARP2 knockdown (Fig. 8B; P = 9.9E-03, one-sided Wilcoxon rank-sum test). Cancer patients or cell lines with EGFR mutations showed significantly higher expression of EGFR at the gene and protein levels than those without EGFR mutations (Fig. 8C-G; P < 0.05, one-sided Wilcoxon rank-sum test). EGFR is one of the targets of erlotinib. According to the drug combination sensitivity score (CSS) for PARPis from DrugComb data portal, the CSS of the EGFR inhibitor erlotinib and PARPi ABT-888 (veliparib) in human lung cancer cell line MSTO were ranked one and two (Fig. 8H).

Fig. 8.

Drug combination of PARP inhibitors (PARPis) and EGFR inhibitor. (A) Comparing area under the curve (AUC) from Cancer Therapeutics Response Portal (CTRP) and (B) fitness score (DepMap_CRISPR) in cell lines with and without EGFR mutation. (C)-(D) Boxplots show the distribution of EGFR protein expression (The Cancer Genome Atlas [TCGA] and Cell Line Encyclopedia [CCLE]) in cell lines with mutation of resistant gene EGFR and those with wild type expression (P < 0.05). (E)-(G) Boxplots show the distribution of EGFR expression (TCGA, Catalog of Somatic Mutations in Cancer [COSMIC], and CCLE) in cell lines with resistant gene EGFR mutation compared with wild type expression (P < 0.05). P values were calculated by one-sided Wilcoxon rank-sum test, and μ is the mean value. (H) Prediction of drug combination with PARPis in DrugComb Portal Database. Color of the dots represents five synergy prediction models (PLISS, HSA, LOEWE, ZIP, and CSS). The top one and two CSS are combined TBT-888 (veliparib) with erlotinib in MSTO.

Fig. S8 shows the drug-target and PPI networks provided by DrugComb. Many targets of erlotinib interacted with veliparib or PARP1/2, indicating the potential cross talks between erlotinib and veliparib. Cancer xenografts show that combination treatment with erlotinib and olaparib has a markedly enhanced anticancer effect in human ovarian cancer cell line A2780 [24].

In LGG patients with PARP1/2/3 copy number deletions or downregulation of expression, patients with TTN mutations showed worse prognosis than those with wild type TTN (Fig. S9A, P = 8.4E-06, log-rank test with OS; Fig. S9B, P = 1.7E-05, log-rank test with PFS). The mutation frequency of TTN was 48.6% (17/35) and 43.2% (19/44) in Catalog of Somatic Mutations in Cancer (COSMIC) and CCLE central nervous system cell lines, respectively, and 11.4% (58/508) and 26.1% (6/23) in LGG patients from TCGA and cBioportal, respectively. In cancers, TTN stimulates angiogenesis by elongation, migration, and sprouting of endothelial cells [25]. According to a phase III PAOLA-1/ENGOT-ov25 trial, the combination of olaparib and bevacizumab extended PFS in patients with advanced-stage ovarian cancer, regardless of the mutation status of BRCA1/2. Hence, we conjectured that LGG patients with TTN mutations might benefit from the treatment of combining PARPis and antiangiogenic drugs, such as bevacizumab.

3. Discussion

Our study proposed a novel computational methodology to identify candidate responsive biomarkers for PARPis based on the GI network of PARP1/2/3 according to large-scale CRISPR/Cas9 and RNAi screens, and PARPi pharmacological screening data. In total, we identified 41 resistant genes and 130 sensitive genes related to PARPis. Most PARPi response-related genes were associated with HR deregulation, and interacted with PARP1/2/3 in the PPI network. Functional analysis indicated that tyrosine phosphorylation imbalance might mediate the response of PARPis in cancer cells. Furthermore, in patients with copy number deletion or downregulation of PARP 1/2/3, LGG patients with mutations in the candidate resistant gene set (TNN, PLEC, and TRIP12) showed poor prognosis, and OV patients with mutations in the sensitive gene set (BRCA2, TOP3A, and ASCC3) had better prognosis. In addition, we predicted a combination therapy of EGFR inhibitor (erlotinib) and PARPi (veliparib) in lung cancer patients, and that LGG patients with TTN mutation can be treated with a combination of PARPis and antiangiogenic drugs (bevacizumab). Thus, our findings provide an avenue for expanding the scale of beneficiary cancer patients for whom PARPi can be considered, and provide novel treatment options to target resistance mechanisms of PARPis.

Among the 41 resistant and 130 sensitive genes, PRRC2A, USP34, and DNAH10 overlapped, which was not contrary to our hypothesis because mutations of these three genes mediate the response to PARPis in different cancer types. In different cancers, the gene mutation frequency is quite different. For example, IDH1 was mutated in 77.3% (394/510) LGG samples, but showed < 12% mutation in other cancer types in TCGA. Moreover, the basic drug sensitivity of cell lines is tissue-specific. Therefore, we used our computational methodology on samples of specific cancer types and cell lines from specific tissues. However, statistical power was limited by the small sample sizes of specific cell line tissues. Thus, we did not use FDR to filter the biomarkers. Notably, PARPi biomarkers revealed by our method were validated in different pharmacological datasets via the HR mechanism and function correlation.

Preclinical and clinical research on combination strategies encompassing PARPis are ongoing to expand the scope of benefit in patients and overcome resistance to PARPis. Relevant clinical studies have suggested the use of a combination of PARPis and immune checkpoint inhibitors with anticancer activity in many cancer types. Combination of PARPis with bevacizumab or chemotherapy has significantly improved PFS in patients with OV [26]. In our study, we predicted combination strategies for PARPis by directly targeting PARPi resistant biomarkers such as EGFR and TTN. We estimated the functional effects of somatic mutations according to transcriptome and proteome expression data, and identified EGFR mutations as gain-of-function mutations. However, further detailed and in-depth analyses, such as structure analysis and animal model experiments, are warranted in future.

Cancer patients with HR defect can benefit from the use of PARPis, thereby providing new opportunities for cancer treatment modalities. We proved our hypothesis that resistant or sensitive genes could disrupt HR to mediate the response to PARPis in cancer cells. HR analysis results showed that downregulated expression of RAD50 and ATM protein expression were associated with candidate sensitive gene mutations. Moreover, we found that RAD51 expression was significantly upregulated by 51.1% (72/141) sensitive genes in TCGA. However, previous studies have indicated that upregulation of RAD51 contributes to drug resistance in cancer cells [27]. RAD51 has also been identified as a tumor suppressor, and the upregulation of RAD51 may mediate cancer cells to be sensitive to PARPis, which warrants further experiments [28].

In summary, our study systematically predicted candidate resistant and sensitive biomarkers for PARPis, pan-cancer, by constructing a cancer genome GI network. Consequently, this could pave the way for novel treatment options to target resistance mechanisms or acquired vulnerabilities of target drugs, and guide precise combination therapies.

4. Conclusion

PARPis have emerged as a beneficial therapeutic option for ovarian, breast, pancreatic, and prostate cancers. However, expanding the scale of patients who can benefit from PARPis and overcoming drug resistance are critical for furthering precision medicine. Our study systematically identified candidate responsive biomarkers (resistant and sensitive biomarkers within specific cancer tissues) for PARPis, by constructing GIs in the cancer genome. Gene mutations may mediate resistance or sensitivity to PARPis by disrupting the HR mechanism. Furthermore, by analyzing the transcriptome and proteome data, we found that EGFR with gain-of-function mutation induced resistance to PARPis, and predicted the combination therapy of PARPi (veliparib) and EGFR inhibitor (erlotinib) for lung cancer. Therefore, this study paves the way for precise combination therapies with PARPis for cancer.

5. Material and methods

5.1. Identification of PARP-related GIs

We used two publicly available CRISPR-Cas9 screening datasets. DepMap Portal (https://depmap.org/portal/) (version, 19Q1), which consists of 558 cell lines, and parental cell lines from CCLE. DepMap Portal (version, 19Q1) uses Avana CRISPR-Cas9 genome-scale knockout library and CERES algorithm to estimate gene dependency, wherein a lower CERES score indicates more essential genes [29]. Gene mutation information of cell lines was downloaded from the DepMap Portal (version, 19Q1). Another CRISPR-Cas9 fitness screen was obtained from DepMap Project Score (https://score.depmap.sanger.ac.uk/). We used the CRISPR depletion log-transformed fold change values (fitness scores) to represent the ability required for cell growth or viability, wherein a lower score indicated that a gene was more essential [30]. Gene mutation profiles of 324 cell lines were downloaded from COSMIC (https://cancer.sanger.ac.uk/cosmic).

RNAi screens were downloaded from the DepMap Portal, Achilles 2.20.2, which includes genome-scale RNAi-based loss-of-function screens of 501 cell lines [31]. We used gene knockdown viability effects (gene dependency scores) for further analysis. Here, higher shRNA scores indicated enhancement of cell viability and a gene with lower dependency score was considered more essential. However, there was no shRNA depletion scores for PARP2 in this dataset.

For somatic mutation data processing, we excluded mutations in the 5′-flanks, introns, intergenic regions, 5′-UTR, and 3′-UTR, and excluded silent mutations [32]. For cell lines with knockdown of PARP1/2/3, a one-sided Wilcoxon rank-sum test was used to test whether the fitness scores were significantly different in cell lines with and without mutation of gene A. If the fitness scores of gene A mutation samples were significantly higher or lower than that of the wild type samples (P < 0.01), gene A was predicted as a candidate partner gene with an SV or SL interaction with PARP1/2/3 (Fig. 1A).

5.2. Identification of PARPi candidate responsive biomarkers

We downloaded pharmacological screening data from GDSC (https://www.cancerrxgene.org/), CTRP, and Cancer Genome Project (CGP) (Table S2) [33]. We obtained LN_IC50 (natural log of the fitted half-maximal inhibitory concentration) values and AUC (area under the drug inhibition curve) values of five PARPis (olaparib, veliparib, talazoparib, rucaparib, and niraparib) from GDSC, two PARPis (olaparib, veliparib) from CTRP, and IC50 values of two PARPis (olaparib, veliparib) from CGP. The cell line information of the GDSC and CGP datasets was referenced from COSMIC, and cell line information of CTRP was available from CCLE.

For more than three cell lines in a specific tissue, we tested the difference in drug response values using a one-sided Wilcoxon rank-sum test between cell lines with mutated and wild type gene A. If the IC50/AUC values of cell lines with gene A mutation samples were significantly higher or lower than that of the wild type samples (P < 0.01), we predicted that mutations in gene A might induce resistance or sensitivity to PARPis in cancer cell lines.

5.3. Genome data from TCGA

Somatic mutation and copy number alteration (CNA) data of 32 cancer types from TCGA consortium were downloaded from Genomic Data Commons (GDC, https://portal.gdc.cancer.gov/). CNAs were determined by Genomic Identification of Significant Targets in Cancer (GISTIC2) [34]. As per Shi et al., the log2 ratio cutoff values <-0.25 were defined as deletion of PARP1/2/3 [35]. To further optimize CNAs, a one-sided t-test was used to select PARP1/2/3, whose copy numbers had a positive correlation with gene expression.

5.4. Survival analysis

We downloaded the OS and PFS data from TCGA. For PARP1/2/3 deficiency (deletion or downregulation), log-rank test was used to assess the difference in the survival time between patients with and without mutations of PARPi response-related genes. Kaplan-Meier plots were used to present the results (Fig. 1C). We ranked samples according to gene expression values of PARP1, PARP2, and PARP3. Lower quartile expression values were defined as the downregulation of PARP1/2/3.

5.5. Expression analysis

Protein expression data of CCLE cell lines were downloaded from DepMap (version 18Q4), consisting of 213 proteins across 899 cell lines. Protein expression data of 31 cancer tissues were downloaded from TCGA. All protein expression data were measured using reverse phase protein arrays. Gene expression values (transcripts per million reads) of CCLE cell lines were downloaded from the DepMap Portal (version 19Q1). Gene expression of COSMIC cell lines was log scale robust multi-array from microarray analysis. Fragments per kilobase per million mapped fragments values were used to evaluate the gene expression of TCGA samples.

We grouped samples according to the status (mutated or wild type) of PARPi response-related genes. One-sided Wilcoxon rank-sum test was used to identify the significant differential expression of HR proteins (P < 0.05; Fig. 1B).

To investigate the activities (gain-of-function or loss-of-function) of gene A mutation, we grouped the gene or protein expression data into two groups (gene A mutation vs. gene A wild type). Gene A mutation was considered a gain-of-function mutation if the gene or protein expression in mutated samples was significantly higher than that in the wild type samples, which was tested using the one-sided Wilcoxon rank-sum test (P < 0.05).

5.6. Functional analysis of PPI network

We collected 241 DNA repair-related genes from the Kyoto Encyclopedia of Genes and Genomes database, and studies by Liu et al. and Kang et al. [36], [37], [38]. There were 88 homology-dependent recombination genes. Cancer-related genes were downloaded from the Cancer Gene Census of COSMIC to explain how dysfunction of these genes drives cancer.

GO terms were downloaded from http://geneontology.org/, and a hypergeometric distribution model was used to test whether the GO terms (molecular function, biological process) were significantly enriched in the PARPis response-related genes. P values were corrected using the Benjamini-Hochberg correction for multiple tests, and the top 10 results with a false discovery rate (FDR) < 0.1 were reported (Fig. 1B).

The PPI network of PARPi response-related genes was constructed, and Cytoscape software was used to visualize the networks (https://cytoscape.org/). PPI data were assembled from the Pathway Commons database (Fig. 1B) [39].

5.7. Drug combination prediction

Drug combination prediction data were downloaded from the DrugComb Portal Database [40]. DrugComb provides drug CSS, which were interpreted as a normalized average inhibition of the drug combination response, and four other synergy prediction indexes, including bliss independence (BLISS), highest single agent (HSA), Loewe additivity (LOEWE), and zero interaction potency (ZIP). DrugComb recommends that only a drug combination that achieves a higher synergy score in all the models as well as a higher CSS should be a reliable drug combination (Fig. 1D).

We extracted the synergy prediction indexes greater than zero of PARPis, and sorted them by CSS, which involved olaparib, veliparib/AT-888, and rucaparib with 46 combination drugs from 46 cell lines. The drug-target profiles and PPI networks provided by DrugComb were obtained from PubChem (https://pubchem.ncbi.nlm.nih.gov/), ChEMBL (https://www.ebi.ac.uk/chembl/), and STITCH (http://stitch.embl.de/) databases.

Conflicts of interests

The authors declare that they have no competing interests.

CRediT authorship contribution statement

Qi Dong: Conceptualization, Methodology, Validation, Formal analysis, Investigation, Writing–original draft, Visualization. Mingyue Liu: Conceptualization, Methodology, Validation, Formal analysis, Investigation, Writing - original draft, Visualization. Bo Chen: Conceptualization, Methodology, Validation, Formal analysis, Investigation, Writing - original draft, Visualization. Zhangxiang Zhao: Data curation, Visualization. Tingting Chen: Data curation, Visualization. Chengyu Wang: Data curation, Visualization. Shuping Zhuang: Data curation, Visualization. Yawei Li: Methodology, Investigation. Yuquan Wang: Methodology, Investigation. Liqiang Ai: Methodology, Investigation. Yaoyao Liu: Methodology, Investigation. Haihai Liang: Writing - original draft. Lishuang Qi: Writing - original draft. Yunyan Gu: Supervision, Project administration, Writing - original draft.

Appendix A. Supplementary data

The following are the Supplementary data to this article: