Cellular and molecular determinants of all-trans retinoic acid sensitivity in breast cancer: Luminal phenotype and RARα expression

Forty-two cell lines recapitulating mammary carcinoma heterogeneity were profiled for all-trans retinoic acid (ATRA) sensitivity. Luminal and ER+ (estrogen-receptor-positive) cell lines are generally sensitive to ATRA, while refractoriness/low sensitivity is associated with a Basal phenotype and HER2 positivity. Indeed, only 2 Basal cell lines (MDA-MB157 and HCC-1599) are highly sensitive to the retinoid. Sensitivity of HCC-1599 cells is confirmed in xenotransplanted mice. Short-term tissue-slice cultures of surgical samples validate the cell-line results and support the concept that a high proportion of Luminal/ER+ carcinomas are ATRA sensitive, while triple-negative (Basal) and HER2-positive tumors tend to be retinoid resistant. Pathway-oriented analysis of the constitutive gene-expression profiles in the cell lines identifies RARα as the member of the retinoid pathway directly associated with a Luminal phenotype, estrogen positivity and ATRA sensitivity. RARα3 is the major transcript in ATRA-sensitive cells and tumors. Studies in selected cell lines with agonists/antagonists confirm that RARα is the principal mediator of ATRA responsiveness. RARα over-expression sensitizes retinoid-resistant MDA-MB453 cells to ATRA anti-proliferative action. Conversely, silencing of RARα in retinoid-sensitive SKBR3 cells abrogates ATRA responsiveness. All this is paralleled by similar effects on ATRA-dependent inhibition of cell motility, indicating that RARα may mediate also ATRA anti-metastatic effects. We define gene sets of predictive potential which are associated with ATRA sensitivity in breast cancer cell lines and validate them in short-term tissue cultures of Luminal/ER+ and triple-negative tumors. In these last models, we determine the perturbations in the transcriptomic profiles afforded by ATRA. The study provides fundamental information for the development of retinoid-based therapeutic strategies aimed at the stratified treatment of breast cancer subtypes.


Introduction
ATRA (all-trans retinoic acid) is used in the management of acute promyelocytic leukemia (Tallman et al, 1997;Lo-Coco et al, 2013), and the retinoid holds promise for the treatment of solid tumors like breast cancer (Garattini et al, 2014). The mechanisms underlying ATRA anti-tumor activity are unique, as the compound is endowed with anti-proliferative and cyto-differentiating activities, while it is only a weak cytotoxic agent (Garattini et al, 2007a,b). The retinoid pathway centers on ligand-dependent transcription factors belonging to the family of steroid nuclear receptors along with ERs/PRs (estrogen/progesterone receptors) and PPARs (peroxisome proliferator-activated receptors) (Chambon, 1996;Mark et al, 2009). Six retinoid receptors are known, that is RARa/b/c and RXRa/b/c. Each RAR and RXR isoform is encoded by a distinct gene which is transcribed into splicing variants (Garattini et al, 2014). The active receptors consist of RXR-RAR heterodimers or RXR-RXR homodimers. While the RXR-RXR homodimers are the target of 9-cis retinoic acid and synthetic rexinoids, which are also promising agents in the chemoprevention of mammary tumors (Wu et al, 2002;Kong et al, 2005;Kim et al, 2006;Abba et al, 2008;Uray & Brown, 2011), the RXR-RAR heterodimers are the classic mediators of ATRA activity. In the RXR-RAR complexes, RARs act as the ligand-binding moiety. ATRA is a pan-RAR agonist, binding RARa/b/c with the same affinity (Di Lorenzo et al, 1993;Gianni et al, 1996) and it also binds and activates PPARb/d (Shaw et al, 2003;Schug et al, 2007). In mammary tumors, ATRA-liganded RXR-RARs are purported to mediate growth inhibition, whereas RXR-PPARb/d induces proliferation (Noy, 2010). ATRA is transported to the nucleus by CRABP1 and CRABP2 (cytosolic retinoic acid-binding proteins-1/-2) as well as FABP5 (fatty acid-binding protein-5) (Schug et al, 2007). While CRABP2 delivers ATRA to RXR-RARs, FABP5 targets RXR-PPARb/d (Schug et al, 2007).
Breast cancer is a heterogeneous disease classified into subtypes according to the gene-expression profiles (Sorlie et al, 2003;Guedj et al, 2012). A rational use of ATRA and retinoids in breast cancer requires the definition of the sensitive subtypes (Garattini et al, 2014). Identification of the molecular determinants underlying retinoid sensitivity is another priority. The availability of representative cell lines characterized for the gene-expression profiles is a unique opportunity to establish the cellular/molecular determinants of retinoid sensitivity in mammary tumors. In this study, we tested the susceptibility of a large panel of breast cancer cell lines to ATRA, subsequently validating and extending the results in short-term tissue cultures of primary tumors. We identify RARa as the main retinoid receptor variant mediating the anti-tumor activity of the retinoid. In addition, we define gene sets, which are associated with ATRA sensitivity, and are of predictive potential. Finally, we determine the perturbations of the transcriptome afforded by ATRA in LuminalA/B and triple-negative (TN) tumors.

Sensitivity of breast cancer cell lines to retinoids
To define ATRA sensitivity, we selected 42 cell lines representative of breast cancer heterogeneity (Supplementary Table S1) and characterized for ER, PR, and HER2 status as well as the Luminal or Basal phenotype according to PAM50 (Tibshirani et al, 2002;Parker et al, 2009) (Supplementary Fig S1). The concentration-dependent growth-inhibitory effects of ATRA (0.001-10 lM) at 3, 6, and 9 days were evaluated, as exemplified by the SKBR3, HCC-1954, and MDA-MB436 cell lines showing different ATRA sensitivity ( Fig 1A). The doubling time of each cell line and a number of other parameters associated with ATRA-dependent growth inhibition were determined (Supplementary Table S2). All these parameters are the basis for the calculation of the ATRA score, a new and robust index defining cell sensitivity to the growth-inhibitory action of ATRA (Supplementary Methods). The higher the ATRA score is, the higher is ATRA sensitivity. Development of this new index was necessary, since determination of standard IC 50 values for the definition of sensitivity to the anti-proliferative effect of ATRA was deemed inadequate for at least two reasons. The IC 50 is routinely and successfully used to assess cell sensitivity to cytotoxic compounds, while ATRA is predominantly a growth inhibitory and cyto-differentiating agent and it is largely devoid of a direct cytotoxic action (Garattini et al, 2007b(Garattini et al, , 2014. Given the slow kinetics of the response to the retinoid, we calculated the ATRA score between days 3 and 6. The ATRA score provides a continuous series of values across our panel of cell lines and identifies four separable groups (A-D, Fig 1B). The subsets with high and intermediate sensitivity (groups A and B) are enriched for cells with Luminal and ER + phenotypes. Indeed, 14/16 of the cell lines in combined groups A and B are Luminal and 11/16 are ER + . Interestingly, SKBR3 and AU565, representing a subgroup of HER2 + tumors which is predicted to be sensitive to ATRA due to co-amplification of the RARA and ERBB2 loci , are the only ER À /HER2 + cell lines present in group A. Similarly, HCC-1599 and MDA-MB157 are the only Basal cell lines in groups A and B, respectively. Group C clusters the cell lines characterized by low sensitivity to ATRA. In this group, the proportion of Luminal (6/14) and ER + (3/14) cell lines is reduced. Group D concentrates ATRA-resistant lines, the majority of which is Basal (10/12). Thus, the ATRA scores indicate that a Luminal phenotype and ER expression are major determinants of cell sensitivity to the anti-proliferative action of ATRA. In contrast, a Basal phenotype represents a negative factor. Indeed, the proportion of Basal cell lines increases as the ATRA score decreases if our panel is divided in tertiles (T1 = 2/14; T2 = 6/14; T3 = 12/14) ( Fig 1B).
Being one of the two Basal lines with a high ATRA score and one of the rare breast cancer lines transplantable in mice (Zhang et al, 2013), HCC-1599 represents a unique model to validate our ATRAsensitivity data in vivo. Thus, SCID mice bearing subcutaneous HCC-1599 xenografts were treated with ATRA (15 and 7.5 mg/kg) or vehicle on a daily basis for 3 weeks, and tumor growth was followed. A time-and dose-dependent reduction in the tumor volume is evident in mice treated with ATRA (Fig 2A). With the highest dose of ATRA, the effect is already significant after 17 days and is maintained for at least 10 days after treatment discontinuation. The total body weight of mice is not different in the experimental groups, demonstrating lack of ATRA-dependent toxicity (Supplementary Fig S2). The results were validated by MRI analyses performed at 24 days ( Fig 2B). Taken together, the results support the in vivo relevance of the cell-line studies.

Short-term cultures of mammary tumors: anti-proliferative responses to ATRA
To confirm the results obtained with the cell lines, we used shortterm cultures of mammary tumors (van der Kuip et al, 2006) derived from diagnostic Tru-cut procedures of 45 patients (Supplementary Table S3). To assess the anti-proliferative activity of ATRA, tissue slices were challenged with vehicle or the retinoid for 48 h, the maximal time interval maintaining tumor cell viability in basal culture conditions. The growth of tumor cells was evaluated with Ki67 ( Fig 3A and B), which is an established biomarker of cell division and it is routinely used in the clinics to assess the proliferation rate of breast cancer. Ki67 is rapidly down-regulated by a number of anti-proliferative agents in short-term tissue cultures of primary tumors (Alagesan et al, 2015). Rapid downregulation of the biomarker is of the outmost importance, given the relatively short exposure times to ATRA that our tissue culture model allows and the slow anti-proliferative effect exerted by the retinoid.
The cases used in our study are classified according to standard clinical criteria and consist of 17 Luminal-A, 14 Luminal-B, 7 HER2 + , and 7 TN (triple-negative) tumors. All the Luminal-A and Luminal-B tumors are characterized by > 70% ER + cells. ATRA reduces the proliferation of 11 Luminal-A and 10 Luminal-B tumors ( Fig 3B). Except for growth inhibition of the two cases characterized by co-amplification of the ERBB2 and RARA loci (patients 26 and 67) , the retinoid exerts no significant effect on HER2 + tumors. Only one of the TN or Basal cancers responds to ATRA. The data are consistent with the cell-line results and confirm that ATRA sensitivity is frequent in Luminal and ER + tumors.
Associations between the cellular phenotype and genes of the retinoid pathway Known members of the retinoid pathway are likely to be major mediators of ATRA anti-tumor activity. Given the respective associations with ATRA sensitivity and refractoriness observed in cell lines and primary tumors, we evaluated whether the Luminal and Basal A B Figure 1. Profiling of the breast cancer cell-line panel according to ATRA sensitivity. A panel of 42 breast cancer cell lines was challenged with increasing concentrations of ATRA (11 nM-10 lM) for 3, 6, and 9 days, and cell growth was determined.
A The graphs show the growth-inhibitory effect exerted by the indicated concentrations of ATRA in SKBR3, HCC-1954, and MDA-MB436 cells which are representative of lines characterized by a high, intermediate and low ATRA score, respectively. Each point is the mean AE SD of three replicate cultures. **Significantly lower than the corresponding vehicle-treated group (P-value < 0.01, Student's t-test). B Cell lines are ranked in ascending order according to the ATRA score. The plot distinguishes four separate groups of cell lines (D-A) with increasing ATRA scores, as indicated by the colored scale on the right. The cell lines are also grouped in tertiles, T1-T3, according to an ascending ATRA score. Each calculated value is representative of at least two independent experiments.
phenotype as well as ER and HER2 positivity influence the expression of retinoid receptors/binding proteins. Both the microarray and the RNA-seq data associated with our panel of cell lines indicate that the average levels of RARa, RXRa, and CRABP2 are significantly higher in Luminal than Basal cells (Fig 4), while FABP5 shows an opposite pattern. Thus, Luminal cells are predisposed to activate CRABP2/RARa upon ATRA challenge. In the context of Luminal cell lines, these mRNAs show the same expression profile in ER + relative to ER À cells. In the HER2 + cellular context (SKBR3, AU565, and UACC812 cell lines), the data confirm that RARA co-amplification results in high levels of RARa .
To evaluate whether the expression patterns of retinoid receptors/ binding proteins in cell lines recapitulate the situation in mammary tumors, we analyzed the TCGA RNA-seq dataset consisting of over 1,000 breast tumors classified into Luminal-A, Luminal-B, HER2-like, Normal-like and Basal according to PAM50. Consistent with the cell-line data, Basal tumors synthesize the smallest amounts of RARa, RXRa, and RXRc and the highest levels of FABP5 (Fig 5A and B). The analysis unmasks associations which are not evident in cell lines, that is, direct correlations between RARb/ RXRb/CRABP1/PPARb/d expression and the Basal phenotype. Given the poor responsiveness of Basal cell lines to the retinoid (Fig 1B), these results support the notion that FABP5 and PPARb/d are negative determinants of ATRA sensitivity (Balmer & Blomhoff, 2002;Kannan-Thulasiraman et al, 2010). As for RARb, its expression does not seem to be important for ATRA anti-tumor activity in breast cancer (Connolly et al, 2013). In conclusion, our data demonstrate that RARa is the only receptor with a high level of expression in the cellular phenotypes predicted to be responsive to ATRA.

Direct associations between RARa and ATRA sensitivity
After grouping all the cell lines in ascending tertiles (T3 to T1) according to the ATRA score, we looked for associations between retinoid receptors/binding proteins and ATRA sensitivity. According to both the microarray and RNA-seq results, the average amounts of RARa are significantly higher in T1 (ATRA sensitive) than T3 (ATRA resistant) cell lines ( Fig 6A). No difference in the amounts of PPARb/d ( Fig 6A), RARb, RARc, RXRa, RXRb, RXRc, CRABP2, and FABP5 (Supplementary Fig S3) is evident. RARa is also significantly overexpressed in the T1 group, if the analysis of the microarray data is limited to the 22 Luminal cell lines and the trend is confirmed by RNA-seq, although the results do not reach statistical significance. In the microarray data, T1 cells express larger amounts of PPARb/d ( Fig 6A) than the T3 counterparts, after restriction to the Basal subset (20 cell lines). Thus, RARa is likely to be a determinant of ATRA sensitivity in both the total and Luminal fraction of cell lines, while PPARb/d may represent a positive factor in Basal cell lines.
The complement of RAR splicing variants in cell lines and primary tumors: RARa3 as the major determinant of ATRA sensitivity RARa (RARa1-4), RARb (RARb1-2-5), and RARc (RARc1-5) splicing variants are known ( Supplementary Fig S4). Quantitative PCR was used to determine basal expression of these variants in our panel of cell lines. In the majority (40/42) of the cell lines ( Supplementary  Fig S5A), RARa3 is the most highly expressed RARa mRNA, being at least one order of magnitude more abundant than RARa2 and RARa1. Extremely low levels of RARa4 are generally observed. Only RARa3 and RARa4 show significant co-regulation across the panel ( Supplementary Fig S5B), consistent with transcriptional control by the same promoter. Luminal cells contain significantly larger amounts of RARa3, RARa1, and RARa4 mRNAs than the Basal counterparts (Table 1). Higher levels of the same transcripts are also associated with ER positivity, although statistical significance is not reached if analysis is restricted to the Luminal cell lines. If associations between RARa variants and ATRA sensitivity are searched for, significant over-expression of RARa3 in T1 relative to T3 cell lines is evident (Fig 6B). A similar, though less significant, association is A B Figure 2. ATRA-dependent anti-tumor activity in HCC-1599-derived xenografts in vivo. SCID mice were xenografted subcutaneously with 10 × 10 6 HCC-1599 cells on both sides. One week after transplantation 10 animals/experimental group were treated intraperitoneally with vehicle (DMSO) or two doses of ATRA (7.5 and 15.0 mg/kg) once/day, 5 days a week for a total of 24 days. At the end of this period, treatment was discontinued until sacrifice.
A The size of the tumors was determined with a caliper and the volume plotted.
Each point is the mean AE SE of 20 tumors. *Significantly lower than the corresponding vehicle-treated group (P-value < 0.05, Student's t-test). **Significantly lower than the corresponding vehicle-treated group (P-value < 0.01, Student's t-test).    Fig S6A). Despite regulation by distinct promoters, RARb1 and RARb2/5 are always co-expressed (Supplementary Fig S6B). Both RARb1 and RARb2/5 are over-expressed in Basal relative to Luminal cells. The same is true in ER À versus ER + cell lines, even if the analysis is restricted to the Luminal/ER À group (Supplementary Fig S6C). Regarding possible associations with ATRA sensitivity, no significant difference in the levels of RARb1 or RARb2/5 between the T1 and T3 cell lines is evident before or after stratification for the Luminal/Basal phenotype ( Fig 6B).
As for RARc, we focused our attention on RARc1-3, which are predicted to code for active transcription factors. The order of expression for the RARc forms is RARc1 >> RARc2 > RARc3 (Supplementary Fig S7A). Consistent with regulation by the same promoter, only the RARc2/RARc3 couple is characterized by coregulation across all the cell lines (Supplementary Fig S7B). As the RARb variants, RARc2 and RARc3 show a direct association with the Basal and ER À phenotypes ( Supplementary Fig S7C), while only RARc3 is significantly higher in HER2 À than HER2 + cells. In addition, RARc2 and RARc3 tend to be over-expressed in the T3 relative to the T1 group (Fig 6B), supporting the idea that they represent negative factors in terms of ATRA sensitivity (Bosch et al, 2012).
Taken together, the results point to RARa3 as the principal element of the retinoid pathway mediating the anti-proliferative responses of Luminal cells to ATRA.

RARa3 as a major player of ATRA sensitivity in primary tumors
The profiles of expression of the RAR splicing variants were defined in the primary tumors used to evaluate ATRA sensitivity (see Fig 3). In all the specimens considered, RARa3 and RARa2 are the most abundant RARa mRNAs and have a similar level of expression ( Fig 7A), which is different from what is observed in the cell lines. Across all the samples, RARb1 is more abundant than RARb2, although RARb1 levels are at least one order of magnitude lower than the RARa3/RARa2 counterparts. In the case of the RARc variants, RARc1 and RARc2 show intermediate levels of expression relative to RARa3/RARa2 and RARb1/RARb2. RARc2 in primary tumors is more abundant than expected from the cellline results, while RARc3 is by far the least abundant species. The expression of RARa3/RARa1, RARa3/RARa2, RARa3/RARa4, RARa2/RARa4, RARc1/RARc2, RARc1/RARc3, and RARc2/RARc3 across the tumor samples is highly correlated (Supplementary Fig S8).
As for possible associations between RAR splicing variants and tumor cell phenotype, in accordance with the cell-line data, the content of RARa3 is generally higher in Luminal, relative to TN cancers and HER2 + tumors with no co-amplification of the ERBB2 and RARA loci (Fig 7A). In the case of the RARb variants, expression is similar in the tissue samples and cell lines, as the average levels of RARb1 are significantly more abundant in TN (Basal) than The  7A). In terms of possible associations between RAR splicing variants and ATRA sensitivity, only the levels of RARa3 are significantly higher in sensitive than in refractory tumors ( Fig 7B).
In conclusion, the complement of RAR splicing variants in breast tumors and derived cell lines is not entirely superimposable. Nevertheless, the profiles of RAR splicing variants in our cohort of mammary tumors support a major role of RARa3 in the anti-tumor responses to ATRA, which is in line with the conclusions drawn in cell lines.

RARa protein and ATRA sensitivity in breast cancer cell lines
Given the observed relevance of the RARa3 transcript in our models, the basal levels of the corresponding RARa protein were determined in breast cancer cell lines with a specific antibody (Fig 8A and B). The average levels of RARa are significantly higher in T1 versus T3 ATRA score groups (Fig 8C), and the same trend is observed if the analysis is restricted to Luminal cells. Larger amounts of RARa are also observed in Luminal versus Basal and ER + versus ER À cell lines. As for possible correlations with the RARa mRNA variants across the cell lines, the highest R 2 values were calculated for the RARa protein and the RARa3/RARa4 mRNAs ( Fig 8D). This indicates that the protein is encoded by either the RARa3 or the RARa4 transcript. Given the low relative expression levels of RARa4, we favor RARa3.

Effects of RAR agonists/antagonists in breast cancer cells
The functional role of RARa in the anti-proliferative action of ATRA was evaluated in Luminal and Basal cell lines with different ATRA scores and RAR-variant expression profiles (Fig 9) with a pharmacological approach, using the validated (Supplementary Fig S9) AM580 RARa agonist (Gianni et al, 1996), the UVI2003 RARb agonist (Alvarez et al, 2014), and the BMS961 RARAc agonist . The cell lines were challenged with increasing concentrations of ATRA, AM580, UVI2003, and BMS961 for 3 (data not shown) and 6 days prior to evaluation of cell growth. In the ATRA-sensitive Luminal lines, AM580 is the only agonist which inhibits growth in a dose-dependent manner. In ER + /HER2 À HCC-1428 cells, AM580 is more effective than ATRA, while the opposite is true in the ER À /HER2 À EVSAT counterpart. In the remaining Luminal lines, AM580 and ATRA show similar efficacy. AM580 is also the sole agonist inhibiting the growth of the retinoid-sensitive Basal cell lines, HCC-1599, MDA-MB157, and HCC-1954. In these cellular contexts, no significant difference in the anti-proliferative activity of AM580 and ATRA is noticeable. AM580, UVI2003, BMS961, and ATRA are equally ineffective in retinoid-resistant HCC-38 cells.
To corroborate the results obtained with the RAR agonists, we evaluated the effects of the RARa antagonist, ER50891 (Kikuchi et al, 2001;Somenzi et al, 2007), and the RARb/c antagonist, CD2665 (Szondy et al, 1997), on ATRA-dependent growth inhibition of HCC-1428 and SKBR3 cells, which are characterized by very high ATRA scores. To obtain maximal blockade of the two RARs without off-target effects, cells were treated with 100 nM ATRA and 3 lM of ER50891 or CD2665, as this concentration of the antagonists blocks the trans-activating potential of ATRA in a RARa-and RARb/c-specific fashion, respectively (Supplementary Fig S10). In HCC-1428 and SKBR3 cells, only ER50891 blocks the anti-proliferative action of ATRA (Supplementary Fig S11).

RARa and ATRA sensitivity: over-expression and knock-down studies
To obtain direct proof that the RARa3 protein is mediating the action of ATRA, we over-expressed it in retinoid-resistant, HER2 + / ER À , and Luminal MDA-MB453 cells. Two RARa-over-expressing (RARA-C5 and RARA-C7), two vector-transfected control (Vect-C1 and Vect-C2) clones, and the parental MDA-MB453 cells (WT) were used in comparative experiments. WT, Vect-C1, and Vect-C2 express barely detectable levels of the RARa protein, while large amounts of the product are synthesized by RARA-C5 and RARA-C7 cells (Fig 10A). RARA-C5 and RARA-C7 express a transcriptionally active RARa form, as indicated by ATRA-dependent activation of the luciferase-based retinoid reporter, DR5-RARE-Luc. Over-expression of RARa does not exert major effects on the basal growth rate of the MDA-MB-453 clones (Fig 10B). Upon treatment with increasing concentrations of ATRA for 3, 6, and 9 days, Vect-C1 and Vect-C2 and WT cells are equally unresponsive to retinoid-dependent growth inhibition (Fig 10C). In contrast, RARA-C5/RARA-C7 proliferation is inhibited dose-and time-dependently by ATRA. Thus, stable overexpression of RARa renders MDA-MB-453 cells sensitive to the retinoid with an~4-fold increase in the calculated ATRA score at 9 days.
In a mirror series of experiments, we knocked down RARa in the retinoid-sensitive HER2 + /ER À and Luminal SKBR3 cells by stable transfection of a RARa1/3-targeting shRNA. The two shRNAtransfected RARA-sh18 and RARA-sh19 clones express < 10% of the RARa protein levels in the parental (data not shown) and void vector-transfected Vect-C6 or Vect-C8 cells (Fig 10D). Transfection of DR5-RARE-Luc in RARA-sh18 and RARA-sh19 cells demonstrates inhibition of ATRA-dependent transcriptional activity relative to the Vect-C6 or Vect-C8 counterparts (Fig 10D). While the shRNA constructs and void vectors do not alter the basal growth rate of SKBR3 cells (Fig 10E), RARa knockdown attenuates the anti-proliferative action of ATRA ( Fig 10F). Attenuation is observed at concentrations of ATRA between 0.001 and 0.1 lM and tends to be lost at the two highest concentrations considered, where ATRA exerts off-target effects.
We evaluated whether modulation of RARa has any effect on ATRA-dependent expression of four direct retinoid target genes. ATRA-dependent induction of CYP26A1, CYP26B1, BTG2, and RARRES3 is not observed in Vect-C2 cells (Supplementary Fig  S12A). In contrast, ATRA induces the expression of the first three transcripts in RARA-C5 cells. Similar differential effects are observed if the levels of the two retinoid-dependent epithelial differentiation markers, b-catenin and SMAD3 , are measured in Vect-C1, Vect-C2, RARA-C5, and RARA-C7 cells before and after treatment with ATRA ( Supplementary Fig S12B). Conversely, ATRA-dependent induction of the CYP26A1, CYP26B1, BTG2, and RARRES3 mRNAs as well as the b-catenin and SMAD3 proteins observed in Vect-C8 cells is blocked in RARA-Sh18 cells. Thus, RARa is the predominant mediator not only of the anti-proliferative, but also of the transcriptional effects afforded by ATRA in the two breast cancer cells. To determine whether RARa mediates other ATRA-dependent responses of relevance for the anti-tumor activity of the retinoid, we measured single-cell random motility, as the process is a determinant of invasive/metastatic behavior and it is inhibited by ATRA in breast cancer cells (Terao et al, 2011). As expected from the results in parental lines (data not shown), MDA-MB453-derived Vect-C1 cells are unresponsive to the anti-motility action of ATRA (Fig 10G), while SKBR3-derived Vect-C6 cells respond with a significant reduction in random motility. RARa over-expression sensitizes RARA-C5 and RARA-C7 cells to ATRA, while RARa knockdown induces ATRA resistance in RARA-sh18 and RARA-sh19 cells. These data support a key role of RARa in mediating the anti-metastatic activity of ATRA.

Identification of gene sets associated with ATRA sensitivity
Besides RARa, other gene products are likely to play a role in the anti-tumor action of ATRA. Thus, we looked for genes whose basal levels of expression are correlated to the ATRA score across our panel of cell lines using the microarray/RNA-seq databases and a regressed Random Forest approach (Supplementary Fig S13). The goal was the generation of ranked lists of genes associated with ATRA sensitivity based on the variable importance scores (Supplementary Table S4).
We generated two distinct expression heat-maps of the top 100 RNA-seq (Fig 11, left) and microarray (Fig 11, right) genes associated with ATRA sensitivity in the two databases. Fourteen of the genes are common to the microarray and RNA-seq gene sets. This is a high proportion considering the large difference in the quantifiable gene products between the two datasets (RNA-seq = 57,789; microarray = 15,543). Cluster analysis of both the microarray and RNA-seq data allows a clear separation of the lines belonging to the T1 and T3 groups identified by the ATRA score. Although our gene sets may contain elements specific to Luminal or Basal cell lines, as indicated by the presence of three PAM50 genes (estrogen receptor 1, ESR1; progesterone receptor, PGR; CXXC finger protein 5, CXXC5), it must be noticed that they do not simply stratify the cell lines according to the Luminal or Basal phenotype. To validate the expression results with an independent assay, we performed quantitative real-time PCR on 14 selected genes. The PCR, microarray, and RNA-seq results are concordant (Supplementary Fig S14).
By far the highest ranking retinoid nuclear receptor in the list generated from the microarray data is RARa, standing within the first 3.6% of the ranked mRNAs. RARa is also highly ranked in the RNA-seq list (top 16.5%). Some of the 100 top-ranking genes present in both lists may be of interest for the anti-tumor action of ATRA. For instance, CXXC5 is a retinoid-inducible gene (Knappskog et al, 2011;Astori et al, 2013), and it mediates the proliferative responses of IGFs and HER2 in breast cancer (Montero et al, 2011(Montero et al, , 2013. The ATRA-regulated SYT7 (Synaptotagmin-VII) (Ekici et al, 2008) and VAMP3 (vesicle-associated-membrane protein-3) proteins control the homeostasis of micro-vesicles, which, in turn, regulate mammary tumorigenesis (Wright et al, 2009). A splicing variant of CPE (carboxypeptidase-E) stimulates growth, and it is a biomarker for mammary tumor metastatic spread (Lee et al, 2011). In IRAK1 (interleukin-1-receptor-associated kinase-1) knockout macrophages, RARa expression is higher than in the parental counterparts (Maitra et al, 2009).
The two identified gene sets may be useful for the stratification of patients who are likely to benefit from retinoid-based therapeutic approaches. We evaluated whether primary Luminal tumors, which are generally responsive to the anti-proliferative action of ATRA, are enriched for elements present in both gene sets (Fig 11, bottom), comparing the 11 Luminal and 5 TN tumors used to assess the genomic effects of ATRA (see Fig 12). The enrichment (single sample Gene Set Enrichment Analysis, ssGSEA) of genes whose basal levels of expression is higher in ATRA-sensitive (T1 group) than in ATRArefractory (T3 group) cell lines is significantly higher in Luminal relative to TN tumors. In contrast, the enrichment of genes whose basal levels of expression is lower in the T1 than in the T3 group tends to be lower in Luminal than TN tumors. The two identified gene sets are the basis for the generation of an optimized gene signature predictive of ATRA sensitivity.

Transcriptional responses to ATRA in short-term cultures of mammary tumors
The results obtained in the cell lines and the short-term cultures of primary tumors support the concept that Luminal and ER + phenotypes are positive determinants, while Basal and ER À phenotypes are negative determinants of sensitivity to the anti-proliferative action of ATRA. To evaluate the transcriptional effects of the retinoid, we performed microarray gene-expression studies in 16 of the 45 primary-tumor samples profiled with the Ki67 biomarker (see Fig 3). The cohort analyzed consisted of 11 Luminal and 5 TN cases. Microarray data were validated by quantitative PCR on a selected number of tumors and genes (CYP26A1, CYP26B1, RARRES3, and BTG2) (Supplementary Fig S15).   Table S5). Cluster analysis of the regulated transcripts results in a clear separation between Luminal/ER + and TN tumors on the basis of the genomic responses to ATRA. It is interesting to notice that cluster analysis groups together the three Ki67unresponsive Luminal/ER + cases (Patients No. 41,55 and 61). A total of 1,702 genes are significantly regulated by ATRA (P < 0.005, paired t-test) in either TN or Luminal/ER + tumors. Approximately 20% of the genes (up-regulated genes = 198; down-regulated genes = 134) are similarly modulated by ATRA in both the Luminal/ER + and TN cases (Fig 12B).
We focused our attention on the subset of genes identified as retinoid targets in various cell types (Balmer & Blomhoff, 2002, 2005Topletz et al, 2015). Among these 402 genes, 34 are significantly up-or down-regulated by ATRA (P < 0.001, paired t-test) in tissue slices derived from either Luminal/ER + or TN tumors ( Fig 12C). Five of these genes are differentially regulated in Luminal/ER + and TN tumors. The up-regulation of RARRES3 (retinoic acid receptor responder 3), TGM2 (transglutaminase 2), S100A8 (S100 calcium binding protein A8), and CYP26A1 (cytochrome P-450 26A1) is significantly higher in Luminal/ER + tumors. In contrast, THBD (thrombomodulin) is significantly up-regulated only in TN tumors. RARRES3 up-regulation in Luminal/ER + may play a role in the anti-motility and anti-metastatic effects of ATRA (Nwankwo, 2002;Terao et al, 2011), as the factor has been shown to suppress metastases to the lung in breast cancer (Errico, 2014;Morales et al, 2014). In contrast, increased induction of CYP26A1 may be detrimental for the anti-tumor action of ATRA, as the enzyme metabolizes and inactivates the retinoid (Thatcher et al, 2010;Topletz et al, 2012).
To define the biochemical pathways regulated by ATRA and potentially involved in the anti-tumor action of the retinoid, we performed gene-network enrichment analysis of the microarray data, focusing on Luminal/ER + tumors. Among the top 10 processes enriched (Supplementary Table S6), the ER nuclear signal transduction pathway is of interest for its role in Luminal breast cancer growth. For instance, down-regulation of IRS1 and one of the regulatory subunit of PI3K by ATRA are likely to block the proliferation of Luminal/ER + tumors caused by the growth factors IGF1 and EGF. Out of the 28 types of interactions between couples of proteins belonging to the ER pathway (Supplementary Table S7), 17 are consistent with an inhibition of the ER pathway by ATRA. Taken together, the data demonstrate an anti-estrogenic action of the retinoid in primary tumors, which is in line with what was reported in breast cancer cell lines (Hua et al, 2009).
We performed an interactome analysis (Fig 12D) looking for gene products significantly over-connected in the network modulated by ATRA in Luminal/ER + tumors. We focused on two groups of genes relevant for the molecular mechanisms underlying the anti-tumor action of ATRA, that is, transcription factors and kinases (Supplementary Table S8). As for transcription factors, the list of over-connected genes contains RARA, RARG, and RXRA. The presence of ESR1 and ESR2 among the top-ranked transcription factors is in line with the process enrichment analysis described above. Finally, the inclusion of STAT1, STAT5B, and STAT3 is of relevance given the cross talk between retinoid receptors and this group of transcription factors in acute myeloid leukemias (Gianni et al, 1997). As for the kinases, PI3K and AKT stand out, as ATRA has been shown to inhibit the two corresponding signal transduction pathways in certain breast cancer cell lines .

Discussion
Exploitation of the clinical potential of ATRA requires definition of the sensitive mammary tumor subtypes and the molecular determinants of this sensitivity. In this study, we examined the responsiveness of a large panel of breast cancer cell lines, recapitulating the heterogeneity of the disease, to the anti-proliferative action of ATRA. A Luminal phenotype and ER expression are identified as major determinants of ATRA sensitivity. In contrast, a Basal phenotype, which is characteristic of TN tumors, is associated with ATRA refractoriness. The observations made in cell lines reflect the situation delineated in primary tumors using short-term tissue-slice cultures. We propose that ATRA should be used in a neo-adjuvant or adjuvant setting for the treatment and chemoprevention of Figure 10. Over-expression of RARa1/3 in ATRA-resistant MDA-MB453 cells and silencing of RARa1/3 in ATRA-sensitive SKBR3 cells. A RARa1/3 plasmid construct and the corresponding control void vector were stably transfected into ATRA-resistant MDA-MB453 cells. Two cell clones over-expressing RARa (RARA-C5 and RARA-C7) and two appropriate control clones (Vect-C1 and Vect-C2) were isolated. A RARa3 shRNA plasmid construct and the corresponding void vector were stably transfected into ATRA-sensitive SKBR3 cells. After selection, two cell clones silenced for RARa3 (RARA-sh18 and RARA-sh19) and two appropriate control clones (Vect-C6 and Vect-C8) were isolated.
A, D The indicated clones and the parental cell line (WT) were transiently transfected with the RARE-DR5-Luc retinoid reporter construct and the level of luciferase activity was measured 24 h after treatment with vehicle (DMSO) and ATRA (100 nM), as illustrated in the upper bar graph. Each value is the mean AE SD of three replicate cultures. The levels of the RARa protein measured in the indicated clones by Western blot analysis is shown under the bar graph. To demonstrate that similar levels of total proteins were loaded in each lane, the b-actin band signal obtained after re-blotting of the gel is shown. FI = Fluorescence intensity. B, E The panels illustrate the growth curves of the indicated MDA-MB453 and SKBR3 clones and the SKBR3 parental cell lines (WT) measured with the sulforhodamine assay. C The bar graphs illustrate the effect of increasing concentrations of ATRA on the growth of the indicated MDA-MB453 clones and the parental cell line. The cell lines were challenged with vehicle (DMSO) or ATRA for 3, 6, and 9 days prior to the sulforhodamine assay. OD = optical density at 540 nm. Each value is the mean AE SD of five replicate culture wells. F The graphs illustrate the effect of increasing concentrations of ATRA on the growth of the indicated SKBR3 clones and the parental cell line. The cell lines were challenged with vehicle (DMSO) or ATRA for 3 and 6 days prior to the sulforhodamine assay. Each value is the mean AE SD of five replicate culture wells. G The panel illustrates the effects exerted by ATRA (0.1 lM) on random cell motility of the indicated MDA-MB453 and SKBR3 clones. The results are representative of two independent experiments. Each value is the mean AE SE of the motility of at least 60 cells.
Data information: * and **, significantly different from the corresponding vehicle-treated control (*P-value < 0.05, **P-value < 0.01, Student's t-test). Luminal ER + tumors. In ER + breast cancer, ATRA may represent a rational addition to anti-estrogens particularly in conditions of induced resistance to these agents (Belosay et al, 2006;Johansson et al, 2013). Despite their importance, the Luminal/Basal phenotypes and ER positivity/negativity are not sufficient determinants of ATRA sensitivity or resistance. In fact, there is a minority of the Luminal or ER + cell lines and tumors which are refractory to the retinoid and a few Basal cell lines and tumors responding to ATRA. This indicates that factors other than the cell origin control the responsiveness of breast cancer cells to this anti-tumor agent.
The results obtained in short-term tissue-slice cultures demonstrate that ATRA exerts major quantitative effects on the transcriptomes not only of Luminal/ER + tumors, but also of TN cancers. Thus, our transcriptomic data support the concept that the general refractoriness of TN tumors and the corresponding Basal cell lines to the anti-proliferative action of ATRA is not associated with a similar resistance to the transcriptional effects of the retinoid. In contrast, it is likely that ATRA sensitivity of Luminal/ER + relative to TN/Basal tumor cells is the consequence of different transcription programs activated by the retinoid in the two cell types. The different complement of RAR isoforms and splicing variants present in Luminal/ER + relative to TN/Basal tumor cells may be at the basis of these differential responses to ATRA.
The biological activity of ATRA is deemed to be mediated by the RXR/RAR and RXR/PPARb/d transcription factors via the distinct cytosolic binding proteins, CRABP2 and FABP5 (Shaw et al, 2003;Schug et al, 2007;Kannan-Thulasiraman et al, 2010). The correlative results obtained in cell lines and breast tumors indicate that expression of the RARa3 mRNA and the corresponding protein is directly associated with ATRA sensitivity, the Luminal phenotype, and ER positivity. The role of RARa in the anti-proliferative responses triggered by ATRA is supported by functional studies involving specific pharmacologic RAR agonists/ antagonists performed in selected Luminal and Basal cell lines. RARa overexpression and knockdown experiments provide direct evidence for the involvement of the receptor not only in ATRA-dependent growth inhibition, but also in other aspects of ATRA anti-tumor activity. In addition, RARa is a biomarker of ATRA sensitivity and the major target for retinoids in breast cancer. This suggests that specific RARa agonists should be developed for the management of the disease to avoid side effects and toxicity associated with the clinical use of a pan-RAR agonist like ATRA (Garattini et al, 2007b(Garattini et al, , 2014. Although RARa is an important mediator of ATRA anti-tumor activity, it is unlikely to represent the only determinant of sensitivity. The search for other genes outside the retinoid pathway performed in this study resulted in the definition of two gene sets whose basal expression levels are associated with retinoid sensitivity/resistance in our panel of cell lines. These gene sets are relevant from both a basic and an applied perspective. At the basic level, the two gene sets provide information on previously unrecognized genes and gene networks which may control/influence the sensitivity of cancer cells to ATRA. Our gene sets show a significant overlap (M. Bolis, unpublished observations) with the gene-expression signatures determined for PI3K-inhibitors (Daemen et al, 2013). This suggests that part of the anti-proliferative action of ATRA may involve inhibition of the PI3K pathway, which is often turned on in breast cancer cells. The contention is supported by the presence of PI3K among the over-connected kinases in the network of gene products modulated by ATRA in primary tumors challenged with ATRA ex vivo. With respect to this, a major link may be represented by PREX1 (RAC-exchanger-factor-1), an important determinant of the sensitivity of breast cancer cells to PI3K inhibitors (Ebi et al, 2013). Interestingly, combinations of PI3K inhibitors and ATRA show additive or synergistic growth effects in selected breast cancer cell lines (MT, unpublished results). At the applied level, the two gene sets contain possible pharmacological targets for the design of therapeutic combinations based on ATRA or derived retinoids. In addition, these gene sets have the potential to be optimized in view of their use as diagnostic tools for the selection of breast cancer patients who may benefit from retinoidbased treatments.
In conclusion, this work is a first step toward a rational use of ATRA and derived retinoids in breast cancer. The data obtained with both the cell lines and the short-term tissue cultures indicate that~70% of Luminal breast cancers are likely to be responsive to ATRA. As ATRA is characterized by low toxicity as well as mild side effects, our data strongly suggest that the compound is of potential interest in the adjuvant therapy of the majority of Luminal breast cancer with particular reference to ER + tumors. Indeed, the results obtained represent the rationale for an independent clinical trial (AZ and EG, personal communication), which we will conduct in post-menopausal patients suffering from ER + breast cancer aimed at evaluating the efficacy of ATRA addition to aromatase inhibitors.

Chemicals plasmids and cell lines
The following compounds were used: ATRA (Sigma-Aldrich, https://www.sigmaaldrich.com), AM580 (Tocris, http://www. Figure 11. Gene sets associated with ATRA sensitivity. Using the microarray and RNA-seq data associated with the breast cancer cell lines, two ATRA score-associated gene lists ranked for their variable importance were generated. Upper Panels: The gene-expression results of the first 100-ranking genes in the RNA-seq (left) and microarray (right) datasets were used to perform a cluster analysis of the breast cancer cell lines according to the gene-expression profiles. Data are expressed using a log 2 scale of the expression signal intensity after normalization of the data across the different cell lines. The genes marked in red are present in both the microarray and the RNA-seq gene sets. The cell lines marked in red are those belonging to the ATRA score T1 group and are sensitive to ATRA, while the ones marked in blue belong to the T3 group and are refractory to the retinoid. The left dark blue lines indicate the genes with higher levels of constitutive expression in the ATRA-sensitive cell lines, while the light blue lines indicate the genes with higher levels of basal expression in the ATRA-refractory cell lines. Lower Panels: The box plots show the enrichment score (single sample Gene Set Enrichment Analysis, ssGSEA) of the microarray (left) and RNA-seq (right) gene sets in the TN (patients 9,22,23,31,50) and Lum (patients 13,18,27,36,41,44,55,60,61,62,64)   Tissue slices corresponding to the indicated patients were treated with vehicle (DMSO) or ATRA (0.1 lM) for 48 h. Whole-genome gene expression studies were performed on the extracted total RNA using a microarray platform.
A The heat-map shows the genes significantly up-or down-regulated by ATRA (P < 0.005, paired t-test) in either Luminal-A and -B (Lum) or triple-negative (TN) tumors, and the results are expressed as the log 2 ratio observed between the ATRA and vehicle-treated samples. B A Venn diagram of the genes up-or down-regulated by ATRA in TN and Lum tumors is shown. The number of genes commonly or selectively regulated in TN and Lum tumors is indicated. C The heat-map shows the regulation patterns of the retinoid-dependent genes significantly modulated by ATRA (P < 0.001) in either TN or Lum tumors. The symbols highlighted in red represent the five genes differentially and significantly regulated by ATRA in Lum versus TN tumors. D The panel shows the estrogen-receptor (ESR1) pathway, which is significantly enriched for genes regulated by ATRA in Lum tumors. The green arrows indicate up-regulatory or stimulating interactions, while the red arrows indicate down-regulatory or inhibitory interactions. The gray arrows indicate unknown types of interactions. The red and blue dots above the protein symbols indicate the effect of ATRA in Lum tumors (red = up-regulation; blue = down-regulation).  Table S1. The plasmid constructs used for RARa3 overexpression in MDA-MB453 cells and knockdown in SKBR3 cells are described below.

Plasmid construction
To generate the RARa plasmid used for the over-expression in MDA-MB453 cells, 5 0 FLAG-tagged RARa1/3 cDNA was introduced into pcDNA3, using the NdeI and XhoI sites in the multiple cloning region downstream of the pCMV promoter. To obtain the RARa silencing construct, a custom-synthesized double-stranded DNA coding for a RARa-targeting shRNA (5 0 -GATCCGCGGGCACCTCAATGGGTACTT CCTGTCAGATACCCATTGAGGTGCCCGCTTTTTG-3 0 , the underlined sequence corresponds to nt 629-646 of the NM_001145301.2 sequence, Sigma-Aldrich) was introduced into the pGreenPuro plasmid (System Biosciences Inc., http://www.systembio.com), using the EcoRI and BamHI sites in the multiple cloning region downstream of the H1 gene promoter. The RARa over-expressing MDA-MB453 clones and the RARA-silenced SKBR3 clones were selected, after transfection with Fugene HD (Promega, www.promega.com), in the presence of 400 lg/ml G418 and 1 lg/ml puromycin (Sigma-Aldrich), respectively.

Short-term tissue slice cultures
Tissue cultures of primary breast tumors were performed as described (van der Kuip et al, 2006). Briefly, tissue slices (thickness, 200 lm) deriving from surgical specimens of 45 breast cancer patients who underwent a diagnostic Tru-cut procedure were challenged with vehicle (DMSO) or ATRA (0.1 lM) for 48 h in Mammary Epithelial Cell Growth Medium (Lonza, Allendale, NJ). At the end of the treatment, samples were fixed, paraffin-included, and dissected into 5-lm slices, which were subjected to immunohistochemical staining with an antibody targeting the Ki67 proliferation-associated marker. The percentage of Ki67-positive tumor cells in the various samples was assessed in a quantitative manner by automatic image analysis, and the results are illustrated. Scoring of Ki67 was blinded as to treatment. Each value represents the mean AE SE of at least five separate fields for each experimental sample. The fresh primary tumor samples used for the short-term tissue slice cultures aimed at assessing ATRA sensitivity were supplied by Fondazione S. Maugeri, Pavia. All the procedures were approved by the internal ethical committee of the Fondazione S. Maugeri, and an informed consent for the donation of the sample was obtained from patients.

ATRA score
Cell lines were exposed to increasing concentrations of ATRA (0.001-10.0 lM) for 3, 6, and 9 day, and cell growth was determined with sulforhodamine assays (Skehan et al, 1990;Voigt, 2005;Vichai & Kirtikara, 2006). A detailed description of the ATRA scores and associated mathematical equations and models is available in Supplementary Methods.

Xenotransplants of HCC-1599 cells
HCC-1599 cells (1 × 10 7 / animal) were injected subcutaneously on both flanks of female 6-week-old SCID mice weighing~18 g (Harlan Laboratories, http://www.harlan.com). All the experiments were performed following approval of the internal Ethical Committee on Animal Experimentation and were conducted in compliance with the Italian legislation. Tumor volume was determined with a caliper and by magnetic resonance imaging (MRI, Supplementary Methods).

PCR and Western blot analyses
Real-time PCR was performed using Taqman assays (Terao et al, 2011). Amplimers and Taqman probes (Life Technologies Italia, The paper explained Problem All-trans retinoic acid (ATRA) is the primary vitamin A metabolite and a promising agent in the treatment/chemoprevention of breast cancer. Breast cancer is very heterogeneous, being a collection of different diseases. A rational use of ATRA in breast cancer requires the definition of the sensitive subtypes and the molecular determinants underlying retinoid sensitivity.

Results
We examined ATRA sensitivity in a large panel of breast cancer cell lines, recapitulating the heterogeneity of the disease after development of a new parameter (ATRA score) defining ATRA sensitivity in a quantitative manner. Luminal and ER + (estrogen-receptor-positive) cell lines are generally sensitive to ATRA. In contrast, refractoriness or low sensitivity is associated with a Basal phenotype and HER2 positivity. The associations between cellular phenotype and ATRA sensitivity are confirmed using short-term tissue-slice cultures of primary tumors, which also result in the definition of the modifications in the transcriptome afforded by ATRA in Luminal and Basal tumors. Using a retinoid-pathway-oriented approach, RARa is identified as the only member of the retinoid pathway directly associated with the Luminal phenotype, estrogen positivity, and ATRA sensitivity. Studies in selected Luminal and Basal cell lines with RAR-specific agonists/antagonists confirm that RARa is the principal mediator of ATRA responsiveness. In addition, RARa over-expression sensitizes ATRA-resistant cells to the retinoid. In contrast, RARa silencing in ATRA-sensitive cells abrogates the activity of the retinoid. All this is paralleled by similar effects on ATRA-dependent inhibition of cell motility, indicating that RARa mediates also ATRA anti-metastatic effects. Whole-genome gene-expression data allow the definition of two overlapping gene sets expressed in basal conditions and characterized by predictive potential and associated with ATRA sensitivity in breast cancer cell lines and tumors.

Impact
From a basic point of view, we provide information as to the breast cancer cellular phenotypes and the RAR isoform regulating ATRA sensitivity. In addition, we identify a first list of genes of functional relevance for the anti-tumor activity of ATRA and derived retinoids in breast cancer. From an applied perspective, the study provides fundamental information for the development of retinoid-based therapeutic strategies aimed at the stratified treatment of breast cancer. Finally, the data suggest that therapeutic strategies based on the use of RARa-specific retinoids may overcome the toxicity problems associated with the clinical use of a pan-RAR agonist like ATRA.
Gene-expression studies in short-term tissue cultures of primary tumors Tissue slices were incubated with vehicle (DMSO) or ATRA (0.1 lM) for 48 h. Total RNA was extracted with the miRNeasy Mini kit (QIAGEN), labeled with the Lowinput Quick Amp labeling kit (Cy3 mono color, Agilent), and hybridized to whole-genome gene expression microarrays (Agilent). Fluorescent signals were determined and quantified with an Agilent microarray laser scanner. The microarray raw data and experimental protocols were deposited in the Arrayexpress database (accession No. E-MTAB-3313).
Bioinformatic analysis of the gene-expression microarray and RNA-seq data Gene expression data for the cell lines were derived from the Affymetrix GeneChip Human Genome U133 Plus 2.0 arrays (provided by CCLE, http://www.broadinstitute.org/ccle). RAW sequencing data (Illumina-paired end reads) were derived from two distinct datasets. The first dataset is publicly available in the CCLE project and .BAM files where downloaded through the cgdownload utility from the Cancer Genomics Hub (CGHub/ UCSC, https://cghub.ucsc.edu). Sequencing data (.FASTQ files) for those cell lines that were not part of this first set were downloaded from a second GenBank dataset under the accession GSE48216 (GenBank). The heat-maps were generated using the algorithms available in T-Mev (http://www.tm4.org). Further details on the bioinformatic analyses performed on all the geneexpression data are available in the appropriate sections of the Supplementary Information.
Supplementary information for this article is available online: http://embomolmed.embopress.org