Tamoxifen Genome-Wide RNAi Screen.

To identify determinants of tamoxifen response, we designed a high-throughput RNAi screen (SI Appendix, Fig. S1A). In brief, the screening procedure involved (i) infecting a 4-hydroxytamoxifen (4OHT, an active metabolite of tamoxifen) sensitive, estrogen receptor positive (ER⁺), breast tumor cell line (MCF7) with a library encompassing 56,670 shRNA-coding lentiviral constructs designed to target 16,487 human protein-coding genes (OpenBiosystems GIPZ shRNAmir human genome-wide library), (ii) after allowing the viral constructs to integrate into the tumor cell genome, (iii) exposing cells to 4OHT for 21 d, (iv) recovering shRNA target sequences from the genomic DNA of the surviving cell fraction by PCR amplification, and then (v) estimating the frequency of each shRNA sequence in surviving cells (and thus the effect of each shRNA upon 4OHT response) by sequencing PCR amplification products by massively parallel sequencing.

Before MCF7 cell infection, we divided the genome-wide shRNA library into six pools (each pool comprising 9,600 shRNAs) and infected cells with each pool separately. Cells were infected at multiplicity of infection of 0.7, and on average 2,000 cells per shRNA construct were infected (SI Appendix, Fig. S2). Seventy-two hours after infection, cells were selected in puromycin for 2 d to remove the nontransduced fraction and then divided into two samples: one subsequently exposed to 4OHT solubilized in ethanol, the other exposed to ethanol alone. Five days after initial infection, drug exposure was initiated and sustained for 21 d to model chronic tamoxifen exposure used clinically. A final 4OHT concentration of 500 nM was used, which caused 40% inhibition of cell survival over the time course of the screen (surviving fraction 60%, SF₆₀). To identify shRNA constructs that modulated the response to 4OHT, we estimated shRNA frequency in 4OHT and vehicle-treated cultures at the end of this 21-d period using massively parallel sequencing.

After PCR products were sequenced (SI Appendix, Table S1), each short read was matched to the corresponding RNAi target in the reference library. We then used the frequency of each individual short-read sequence to estimate the frequency of shRNAs in each surviving population. To identify shRNAs that conferred either 4OHT resistance or sensitivity effects, we compared the representation of shRNAs in the 4OHT-treated samples with those in vehicle-treated samples. First, data from each of three biological replicate screens were normalized to account for experimental intrascreen variation (variation between different viral pools). In doing this, we also corrected biases associated with differential starting representations of shRNAs. Second, we applied an interreplica screen rank normalization and used these values to ascribe each shRNA a drug effect (DE) score that represented the magnitude of 4OHT resistance (positive DE) or sensitivity (negative DE) (SI Appendix, Fig. S1B). Here overrepresentation of an shRNA in a 4OHT-treated sample indicated a resistance-causing effect, whereas underrepresentation indicated a sensitization effect.

Identification of Screen Hits.

To define 4OHT-modulating effects from the screen data, we used three parallel strategies previously used in RNA interference screens. In the first instance we selected significant effects according to the variance of the entire dataset. We calculated the median absolute deviation (MAD) to estimate the variance of the normalized data (16) and defined resistance-causing hits as those shRNAs that gave DE scores >2 × MAD (Z score >2), a threshold approximately equal to 2 SDs from the median. Sensitization effects were defined as shRNAs that returned DE scores of Z < −2. In addition to this approach we also used RNAi Gene Enrichment Ranking (RIGER) (17), as implemented in the Broad Institute's GENE-E software package. In brief, RIGER calculated the weighted sum of the two top-ranked shRNAs for each gene on the basis of the log fold change between three biological replicates for each condition, and provided a normalized enrichment score per gene. Finally, we also used RNAi Set Analysis (RSA), a modification of Gene Set Analysis (http://www-stat.stanford.edu/∼tibs/GSA/) that uses maximum–mean statistics to identify significantly enriched or depleted shRNA sets.

In total, Z score threshold identified 1,049 resistance-causing genes and 1,126 sensitization genes, RIGER generated a list of 504 resistance-causing and 498 sensitization genes with a P value of <0.05, and RSA generated a list of 443 resistance-causing genes and 477 sensitization genes with a false discovery rate approaching zero. Given the limitations of each method, we took a pragmatic approach and considered a subset of the genes identified by all three methods for further examination. This intersection approach identified 121 candidate genes mediating sensitization and 131 candidate genes mediating resistance to tamoxifen (Fig. 1A and SI Appendix, Table S2).

We examined screen performance using positive and negative controls. As expected, nonsilencing control shRNAs did not modulate the cellular response to 4OHT, whereas four shRNAs targeting ESR1, the gene encoding ERα, caused sensitivity to 4OHT (Fig. 1B). In vitro functional studies as well as clinical data suggest that reduced PTEN activity causes resistance to tamoxifen (17). Five shRNAs targeting PTEN caused 4OHT resistance in the screen (Fig. 1B), and three PTEN shRNAs chosen for further investigation caused 4OHT resistance as assessed using a GFP competition assay (SI Appendix, Fig. S3 A and C). Similarly, an siRNA pool targeting PTEN was able to recapitulate the resistance phenotype in MCF7 cells treated with either 4OHT or a pure antiestrogen, ICI182,780, using a previously validated 96-well arrayed method (15) (SI Appendix, Fig. S3 B and C). As the PTEN effect identified in the genome-wide screen was confirmed using two orthogonal assays, we reasoned that the screen itself and the analysis methods used were able to identify true endocrine therapy-modulating effects.

Validation of Screen Effects.

Despite their utility, RNAi screens are prone to artifacts such as off-target and screen format-specific effects (18). To overcome these issues, we selected genes identified in the genome-wide shRNA screen and examined them using an independent mode of RNA interference (siRNA) and a 96-well arrayed method. The target sequences for siRNAs used in this validation step did not overlap with the target sequences of shRNAs present in the initial screen (SI Appendix, Table S3), and this allowed us to minimize the impact of off-target effects.

Using this assay system and a single concentration of 4OHT, siRNA duplexes recapitulated 23 tamoxifen-modulating effects (Fig. 2A). Focusing on this 23-gene set, we established 4OHT dose–response relationships for each gene (SI Appendix, Table S4). In addition to PTEN, we validated 11 genes whose RNAi targeting caused 4OHT resistance (BAP1, CLPP, GPRC5D, NAE1, NF1, NIPBL, NSD1, RAD21, RARG, SMC3, and UBA3) (Fig. 2B) and 11 genes that when targeted caused sensitivity (ESR1, C10orf72, C15orf55/NUT, EDF1, ING5, KRAS, NOC3L, PPP1R15B, RRAS2, TMPRSS2, and TPM4) (Fig. 2C and SI Appendix, Table S5).

Network analysis using STRING (19) identified a number of well-established molecular associations among these validated genes (Fig. 3A). These included the identification of both components of the UBA3/NAE1 heterodimer, which mediates the conjugation of NEDD8, an ubiquitin-like protein, to substrates (20). siRNAs targeting either UBA3 or NAE1 (SI Appendix, Fig. S4), in addition to causing resistance to 4OHT, also caused resistance to ICI182,780 (21). The resistant phenotype was also reproduced by siRNA silencing of NEDD8 itself (Fig. 3B), thus validating our screen observations. Interestingly, members of this neddylation pathway, and the UBA3/NAE1 complex in particular, have previously been reported to mediate the degradation of ER (22).

We also demonstrated that a number of cohesion-associated/chromatin remodeling proteins modulated response to both 4OHT and ICI182,780. The cohesin protein complex is involved in maintaining sister chromatid proximity in dividing cells and has recently been implicated in mediating transcriptional insulation and control via interactions with CTCF (23). siRNA targeting RAD21 (24), SMC3, and NIPBL (25), caused 4OHT and ICI182,780 resistance (Figs. 2B and 3B). The precise mechanism whereby cohesin components affect tamoxifen sensitivity is not yet clear, but it seems possible that imbalance of these proteins alters the transcriptional profile of ER⁺ tumor cells, enabling continued growth in the face of endocrine therapy (23).

Finally, a number of proteins involved in RAS signaling were also found to modulate the response to 4OHT. Both shRNA and siRNAs targeting NF1 (neurofibromin), a GTPase tumor suppressor protein, caused 4OHT resistance (Figs. 2B and 3B). NF1 mediates the inactivation of classic RAS proteins (e.g., KRAS and HRAS) but also nonclassic RAS proteins, including RRAS2 (26). siRNAs targeting KRAS and RRAS2 both caused 4OHT sensitivity, and KRAS silencing also caused moderate ICI182,780 sensitivity (Figs. 2C and 3B), as did the downstream mediator of RAS signaling, c-RAF (RAF1) (Fig. 3B). To verify that silencing of NF1 affected RAS signaling in the face of 4OHT treatment, we measured levels of active RAS (GTP-RAS) and phosphorylated ERK in MCF7 cells treated with 4OHT, and these were indeed elevated with NF1 silencing (Fig. 4A). Furthermore, silencing of c-Raf in MCF7 cells with stable NF1 silencing alleviated tamoxifen resistance (Fig. 4B), supporting the hypothesis that these effects were, in part at least, via RAS/RAF signaling.

Clinical Significance.

Having identified a compendium of genes that modulated the response to endocrine therapy, we assessed whether they were associated with clinical response to endocrine therapy. To do this we interrogated publicly available microarray gene expression profiles from tumors of patients subsequently treated with adjuvant tamoxifen. Specifically, we used five patient datasets for this analysis: Stockholm GSE1456, n = 87 (STOCK in Fig. 5) (27); Oxford GSE6532, n = 109 (OXFT) (28); Karolinska Institute GSE3494, n = 72 (KIT) (29); Guy's Hospital GSE9195, n = 77 (GUYS77) (30); and Guy's Hospital GSE6532, n = 87 (GUYS87) (28). In each of these datasets, biopsies were taken from ER⁺ breast tumors before tamoxifen treatment. We obtained the normalized gene expression levels for each of the 23 functionally validated genes and compared these with a surrogate marker of tamoxifen response, the time to distant relapse. To quantify associations between gene expression and the likelihood of relapse/tamoxifen resistance, we calculated Cox proportional log hazard ratios (HRs) (31) for each gene as the relative risk of distant relapse in terms of gene expression. This analysis suggested that reduced expression of NF1, a gene whose silencing caused tamoxifen resistance in the functional screen, was associated with a statistically significant higher risk to distant relapse when considering all datasets (P < 0.05; Fig. 5A). Reduced NF1 expression did not, however, correlate with well-established prognostic factors (SI Appendix, Table S6), suggesting that it was a relatively independent marker of response to tamoxifen. A similar analysis of the validated sensitivity-causing effects indicated that reduced expression of EDF1 was associated with a statistically significant lower risk to distant relapse in a combined series of independent cohorts (P < 0.05; SI Appendix, Fig. S5).

Although interrogating the relationship between individual gene expression and clinical outcome parameters in tamoxifen-treated patients can provide some insight into the clinical relevance of the genetics of tamoxifen resistance, it is also possible that a constellation of individually modest gene effects combine to generate the final resistant phenotype. Given this, we tested whether aggregate measures of expression from groups or modules of genes (metagenes) also correlated with time to distant relapse in patients treated with tamoxifen. Using the same patient datasets and established methodology (32), we defined two functional metagenes for this analysis: one populated by genes shown in this or our previous study (15, 33) to cause resistance to tamoxifen, and a second similarly curated gene set defined by sensitivity-causing effects that also correlated with outcome data. In total, these modules comprised eight genes functionally predicting resistance [BAP1, RARG, PTEN, SMC3, NF1, NIPBL, UBA3, and CDK10 (15)] (Fig. 5B) and six genes functionally predicting sensitivity [EDF1, KRAS, RAF1, TMPRSS2, TPM4, and PDPK1 (33)] (Fig. 5C). This analysis suggested that both “resistance” and “sensitivity” metagenes correlated with clinical outcome (P < 0.05) and gave moderately greater predictions of outcome than the use of individual genes such as NF1.

Pooled shRNA Screen Method.

As a screening library, we used the OpenBiosystems GIPZ human genome-wide shRNA library. Lentiviral DNA was generated according to the manufacturer's instructions and then pooled to generate six pools, each containing 9,600 different shRNA constructs. Virus was packaged in 293T cells and MCF7 cells transduced with a final representation of ≈1,000 cells per shRNA (20,000,000 per shRNA pool). Transduction efficiency was estimated by GFP detection using FACS 72 h after infection. Cells were then cultured in 1 mg/L puromycin for 2 d to enrich for transduced cells and then continuously cultured in media containing 500 nM 4OHT or 0.5% (vol/vol) ethanol for 21 d. Media with drug was replenished every 2 to 3 d. MCF7 cells were cultured in phenol red-free RPMI-1640 media supplemented with 10% (vol/vol) charcoal-dextran stripped FBS. Massively parallel sequencing for shRNA retrieval and screen data analysis are described in SI Appendix, SI Methods and Materials.

Clinical Analysis.

This was performed using tumor samples from five independent datasets of ER⁺ patients subject to tamoxifen adjuvant therapy. Raw Affymetrix U133A microarray expression files were downloaded from the Gene Expression Omnibus repository (accession codes per study indicated in main text) and processed using the RMA function from the Bioconductor Affy package for R (37). Log–rank analyses and Kaplan-Meier plots were produced using the “survdiff” and “plot.survfit” S-plus (TIBCO Software) functions, respectively. Normalized gene expression values were partitioned into quartiles. Cox proportional hazards regression was carried out using the “coxph” S-plus function. Here we used gene expression values as the sole continuous predictor variable. Random effects metaanalysis was carried out using the β values estimated by this regression.