Cross-platform pathway-based analysis identifies markers of response to the PARP inhibitor olaparib

Twenty-two breast cancer cell lines previously profiled for RNA transcript levels were tested for response to nine concentrations of olaparib (see Table 1). These cells mirror many of the transcriptional and genomic characteristics of primary breast tumors and have been used to model responses to a large number of experimental and approved therapeutic compounds [28, 30]. The concentration of olaparib needed to reduce survival to 50 % (SF50) was used as a quantitative measure of sensitivity and ranged from 0.44 nM to 32 µM. The SF50 was not reached for five cell lines at the maximum treatment concentration of 50 µM olaparib. Olaparib response obtained with the growth inhibition assay was not influenced by growth rate assessed as doubling time (Spearman correlation coefficient −0.036, p value 0.874). Figure 2 shows the waterfall plot of SF50 with cell lines ordered from most resistant at the left to most sensitive at the right. Cell lines were divided into a group of 15 resistant and 7 sensitive cell lines, based on an SF50 threshold of 1 µM. Drug response was not significantly associated with breast cancer subtype (p value luminal vs. basal 0.136; Fig. 3), and did not differ between ERBB2-amplified and non-ERBB2-amplified cell lines (p value 1), with transcriptional subtypes assigned to cell lines as previously reported [28]. Four of the seven sensitive cell lines (57 %) were triple negative, compared to 5 of 15 (33 %) resistant cell lines (p value 0.376). Table 1 summarizes characteristics for the 22 cell lines, with SF50, doubling time, transcriptional ER, PR, and ERBB2 status, and the molecular data available for each of them.

We selected candidate molecular features that might be developed as biomarkers for prediction of response to olaparib as those features involved in DNA repair activities that were associated with quantitative response to olaparib in the cell line panel. Molecular features included pretreatment RNA transcript levels, mutation status, copy number variation, and promoter methylation status. Specific genes tested involved aspects of DNA repair listed by Wang and Weaver [31]; ER, PR and ERBB2 due to the importance of PARP inhibition for triple negative breast cancer [17, 19]; and PARP family members PARP1, PARP2, VPARP, TNKS, and TNKS2. This approach is based on observations that in vitro models showing high sensitivity to PARP inhibitors often have BRCA and PTEN deficiencies [7, 8], copy number variations involving BRCA1 and PARP1 [32], and/or hypermethylation of the promoter regions of genes BRCA1 and FANCF [20]. Molecular features showing statistically significant associations with SF50 values are summarized in Supplementary Table 1 and illustrated in Fig. 4.

The transcription levels of MRE11A, NBS1, TNKS, TNKS2, XPA, and XRCC5 were significantly lower (p < 0.05; fold-change >2) in the sensitive compared to the resistant cell lines for at least one expression platform (U133A, exon array and RNA-seq), whilst transcription levels for BRCA1, ERCC4, FANCD2, and PR tended to be lower in sensitive lines (p < 0.1). We refer to Supplementary Table 1a for the list of significant associations per platform. PR protein levels measured using reverse phase protein lysate arrays [33] were also significantly reduced in the sensitive cell lines (p < 0.05). Transcript levels for CHEK2 and MK2 were significantly higher in the sensitive compared to the resistant lines (p < 0.05), with a similar trend for PARP2 and XRCC3 (p < 0.1). Although PARP1 has been shown to be overexpressed in 58 % of invasive breast cancer samples [34] and upregulated at protein level in 82 % of BRCA1-associated breast cancer samples [35], there is no consensus on its importance as a biomarker of response to PARP inhibitors [36, 37]. In our cell line panel, expression of PARP1 mRNA levels were not significantly higher in the sensitive lines compared to the resistant lines (median p value 0.277) (Supplementary Table 1a).

The BRCA1-mutated cell lines MDAMB436 and SUM149PT had a trend to be more sensitive to olaparib compared to the wild-type cell lines (p value 0.091) (Supplementary Table 1b). Likewise, cells with reduced BRCA1 copy number were significantly more sensitive to olaparib than cells with normal copy number at this locus (p value 0.012) (Supplementary Table 1c). PTEN loss of function, which was defined as mutation and/or lack of expression, was not significantly associated with olaparib SF50 response (p value 0.145), even though previous studies from our group suggested that PTEN deficiency can cause olaparib sensitivity [8, 9]. Lack of association in the cell line panel could be ascribed to the small sample size and/or to the possibility that the univariate associations do not take into account important multigene effects. Since BRCA1 mutations have been associated with reduced PTEN expression [38], we tested for association of either BRCA1 mutation or PTEN deficiency with olaparib sensitivity. We found that cell lines with a deficiency in either gene tended to be more sensitive to olaparib than cell lines with functional BRCA1 and PTEN (p value 0.052) (Supplementary Table 1b). No association was found between TP53 mutation status and drug response (p value 0.376).

We used a breast cancer cell line panel comprised luminal, basal, and claudin-low cell lines to develop a multi-transcript predictor of sensitivity to olaparib according to the REMARK recommendations [39]. We limited the predictor to transcript levels to facilitate clinical application. We considered all breast cancer subtypes for the development of the predictor based on a study of RAD51 focus formation in cells responding to a PARP inhibitor. That study showed that 30–40 % of triple negative breast cancers appeared not to have defective HR and therefore might not benefit from a PARP inhibitor whilst ~20 % of non-triple negative breast cancers appeared to have defective HR and therefore might respond to a PARP inhibitor [40]. Thus, we reasoned that a predictor developed using the complete cell line panel might be applicable to the full spectrum of breast cancer covered by the cell line panel. As shown in Fig. 1, the molecular features tested as candidate biomarkers were limited to genes involved in DNA repair pathways BER, NER, MMR, HR/FA, NHEJ, and DDR as defined by Wang and Weaver [31] and in the Kyoto Encyclopedia of Genes and Genomes (KEGG) database release 55.1 [41]. This led to the selection of 118 genes (see Supplementary Table 2) that were tested for association between transcript levels and response to olaparib. These transcript levels were measured using three different mRNA analysis platforms (Affymetrix U133A arrays, Affymetrix exon arrays and Illumina RNA-seq).

We identified the most important transcripts by applying LR with forward feature selection (fivefold CV) 100 times. Markers significantly associated with olaparib response in over half of the iterations are shown in Table 2. These were further reduced to seven gene transcripts that were significantly associated with olaparib response in at least 2 out of 3 mRNA analysis platforms. Five transcript levels (candidate resistance markers BRCA1, MRE11A, NBS1, TDG, and XPA) were inversely associated with predicted probability of response and two transcript levels (candidate sensitivity markers CHEK2 and MK2) were positively associated with predicted probability of response (see Table 3). BRCA1 is involved in DSB repair via RAD51-mediated HR [42, 43]. CHEK2 is a kinase with signal transduction function in cell-cycle regulation and checkpoint responses [2], and is involved in the major parallel DDR pathway ATM-CHEK2 [31]. CHEK2 has also been reported as an intermediate-level breast cancer risk gene, regardless of family history [44, 45]. Besides the standard DDR pathways, the cell-cycle checkpoint pathway p38MAPK/MK2 is additionally activated in TP53 mutant cells [46]. MK2 activity is critical for prolonged checkpoint maintenance through a process of posttranscriptional regulation of gene expression [47]. MRE11A and NBS1 are part of the MRN complex, a multifaceted molecular machine for DSB recognition [48]. Finally, TDG is part of the BER pathway, whilst XPA encodes a zinc finger protein that is part of the NER complex.

We combined information on the seven-transcript levels to form a predictive signature using a weighted voting algorithm (Supplementary Material and [28]). This algorithm assigns a weight and decision boundary to each of the seven genes, based on their expression distribution for the class of sensitive versus resistant cell lines (see Table 3). For this signature to work on external samples, the transcript levels were normalized to the geometric mean of seven control genes, followed by median normalization across the cell lines (see Supplementary Material). The larger the weight for a gene transcript level, the more influence this gene has on predicted probability of response. Positive weights were assigned for sensitivity markers and negative weights were assigned for resistance markers.

We analyzed expression profiles measured for breast cancer patients not treated with PARP inhibitors to understand which patients would have a likelihood of response to olaparib according to our seven-transcript predictor. We used seven U133A and one U133 plus two data sets on 1,846 primary breast tumors with or without metastasis, heterogeneous in treatment and ER/PR/LN status. Our seven-transcript response algorithm predicted that 8–21 % of patients in the eight data sets would be responsive to olaparib (Table 4), using threshold 0.0372 obtained from the cell lines to distinguish sensitive from resistant (see Supplementary Material). The fraction predicted to respond was inversely related to the fraction of ER-positive patients in each data set (Pearson correlation coefficient −0.614, p value 0.1). We also tested the seven-transcript predictor in Agilent mRNA transcript profiles measured for 536 breast invasive carcinoma samples collected by The Cancer Genome Atlas (TCGA) [49]. This required that an Agilent-specific threshold distinguishing sensitive from resistant be established. We accomplished this using a set of Affymetrix and Agilent mRNA transcript profiles measured for 80 I-SPY 1 samples [50, 51]. The Agilent threshold was set so that the fraction of I-SPY 1 samples in the Agilent data set predicted to be sensitive was the same as that predicted to be sensitive using the Affymetrix data (see Supplementary Materials). The fraction of samples predicted to be sensitive in the TCGA data set was 12 % (Table 4). We assessed the transcriptional subtypes of the patient populations predicted to respond to olaparib in 464 samples from GSE25066 and in 528 TCGA tumor samples after exclusion of the normal-like samples. The tumors predicted to respond were enriched in samples classified as basal-like compared to samples classified as luminal A, luminal B or HER2 (p value 0.002 and 2.6 × 10⁻²⁸ for GSE25066 and TCGA, respectively; Table 5).