HDAC6 activity is a non-oncogene addiction hub for inflammatory breast cancers

Loss-of-function screening using genetic tools [12, 39, 40] represent a powerful strategy to interrogate gene function at the genome-wide level. We [9, 12] and others [40, 41] have developed RNAi-based genetic approaches to perform high-throughput (HTP) screens in mammalian systems. Using this technology, we performed genome-wide pooled RNAi screens in 13 breast cancer cell lines (2 IBC and 11 non-IBC lines, including 4 luminal, 4 basal-B, 3 basal-A) and 2 non-transformed mammary epithelial lines using a lentiviral library of shRNA-miRs [9] containing approximately 58,000 different shRNAs targeting approximately 18,500 human genes (Fig. 1a). These models were selected because they recapitulate the genetics and drug sensitivity of the main molecular subtypes of human breast cancer [42]. The screens were performed as we have previously described [12, 13] (see also description in “Methods”). The resulting dataset contained data points from 90 independent cell populations.

As a first step in our studies we performed QC studies in our screens. Screens were highly reproducible between biological replicates with correlation between 0.8 and 0.97 for all cell lines (Figure S1a in Additional file 2). Next, we looked for essential genes across multiple cell lines. For this, housekeeping and highly conserved genes are commonly found depleted in shRNA screens, independent of cell type [12, 15, 40, 43]. We thus used these genes as a first metric of screen quality. As previously reported, genes significantly depleted (p <0.05 in > =3 screens, 2,555 genes) were significantly enriched in housekeeping functions involving the ribosome, proteasome, spliceosome, DNA replication, protein metabolism and mRNA processing (Figure S1b in Additional file 2). Notably, there was highly significant overlap (p <7.2 × 10^−18; Fisher’s exact test) between general essential genes identified by our study and those previously reported [15] (Figure S1c in Additional file 2).

Next, we determined whether essential genes emerging from these screens could classify breast cancer cell lines consistently with functional genomics studies, as we [12] and others [15] have previously shown. As expected, unsupervised hierarchical cluster analysis divided the cell lines into two major groups enriched in luminal and basal subtypes due to subtype-specific sensitivities (Fig. 1b). Interestingly, the IBC cell lines appeared as an independent sub-cluster within the basal-enriched cluster subtype. This suggests that IBC cells present a highly specific profile of essential genes that is not recapitulated by other breast cancer subtypes.

Finally, to achieve an overall profile of IBC vs. non-IBC dependencies, we selected shRNAs significantly and globally depleted in IBC lines vs. non-IBC (p <0.05 and log2FC or log2FC <-1). Additionally, to prevent selection of genes that were essential in non-transformed cells we required that selected shRNAs were not significantly depleted (p <0.05 and log2FC <-1) in the two non-transformed lines. This yielded 71 candidate genes (Table S1 in Additional file 3). We show the top 20 as a heatmap, in order of global IBC-specific depletion significance (Fig. 1c).

Next, we investigated whether significantly depleted shRNAs specific to IBC cells cluster within specific functional categories. To create a thorough portrait of functionally enriched IBC pathways, we used both DAVID [28] and GSEA [29] as complementary approaches in order to perform functional enrichment analysis. DAVID analysis, using the 71 candidate genes selectively depleted in IBC vs. non IBC cells, yielded a set of Gene Ontology (GO) biological processes that were directly and specifically related to one of the candidate genes in the list (i.e., HDAC6) (Fig. 1d). Thus, HDAC6 was the only one of the 71 candidate genes that consistently emerged as part of the top 15 statistically enriched biological processes identified by DAVID. Interestingly, GSEA analysis, including all screened shRNAs ranked by their depletion in IBC vs. non-IBC cells, yielded biological processes that were also specifically related to HDAC6 (Fig. 1d) and HDAC6 was part of 1/3 of the top 15 statistically enriched processes. Thus, both functional enrichment analysis tools provided a comprehensive and intriguing portrait of the role of HDAC6 in IBC survival. Critically, to achieve maximum translational relevance, we paid special attention to candidate targets for which there were clinically relevant pharmacological inhibitors. In this aspect, HDAC6 [18, 20, 44] was also especially interesting, as it represents a druggable target with highly selective inhibitors [21, 45] already available in the clinics, including Ricolinostat [21], which is currently being evaluated in multiple clinical trials (Myeloma NCT01997840, NCT01323751 and NCT02189343 and Lymphoma NCT02091063) as an anticancer drug. Taken together, all of the above provide a strong rationale to select HDAC6 as a primary candidate to validate our screen and further investigate its role in IBC cell survival.

Our genome-wide lentiviral shRNA library contains two shRNAs against HDAC6. Thus, in order to individually validate HDAC6 as a screen candidate, we first tested the silencing efficiency of these shRNAs. Lentiviral-mediated individual transduction of both shRNAs in the IBC cell line SUM149 strongly reduced the protein expression of HDAC6 (Fig. 2a). Next, these two shRNAs were used to individually silence the expression of HDAC6 in a series of cell lines consisting of two non-IBC cell lines (MDA-MB-231 and MDA-MB-436) randomly selected from the cell line series used in the shRNA screens and three IBC cell lines. The IBC cell lines consisted of the two lines used in our screen (SUM149 and SUM190) and a new and independent line, MDA-MD-IBC-3 [46].

We utilized the Annexin V/7-AAD assay to study the induction of apoptosis in each cell line via shRNA knockdown in comparison with control cells transduced with a scrambled shRNA. This assay is able to monitor both early and late events in apoptosis by flow cytometry. These studies revealed that although HDAC6 silencing efficiencies were comparable between IBC and non-IBC models (Fig. 2a), significant apoptotic response was only observed in the IBC lines with minor effects in the non-IBC cells (Fig. 2b and c). As expected, growth curve studies showed reduction of cell proliferation in the IBC cell lines expressing shRNAs targeting HDAC6 compared to controls expressing a scrambled shRNA (Fig. 2d).

To translate our discovery to preclinical animal models, we decided to evaluate the impact of two of the most potent and specific HDAC6 inhibitors previously described, Tubastatin A [45] and Ricolinostat [21], in the viability of IBC cells. HDAC6 is well known to be responsible for the deacetylation of α-tubulin [44] and accumulation of Ac-α-tubulin is commonly used to evaluate the efficacy of HDAC6 inhibition [18, 20, 21, 44, 45]. Thus, we first compared accumulation of Ac-α-tubulin in SUM149 cells when equal doses of Tubastatin A and Ricolinostat were used. Our results showed that Ricolinostat is a more potent inhibitor of HDAC6 in vitro (Figure S2a in Additional file 4) and in vivo (Figure S2b in Additional file 4).

Next, we evaluated the anticancer activity of Ricolinostat in IBC and non-IBC breast cancer models. For these studies we used three IBC and four non-IBC models [42]. Dose titration curves in cell culture showed that Ricolinostat inhibited the growth of IBC cells more efficiently than non-IBC cells (Fig. 3a). As expected, selective inhibition of cell growth in IBC lines was associated with induction of apoptosis (Fig. 3b). Finally, we performed in vivo preclinical efficacy studies. We used three IBC and two of the non-IBC xenograft models (one luminal and one basal) mentioned above. The IBC cell models included both lines used in our screen (SUM149 and SUM190) and a unique IBC human-patient-derived xenograft (PDX) model (Mary-X) that faithfully recapitulates the dermal lymphatic invasion phenotype characteristic of human IBC [47, 48]. Animals were dosed with 50 mg/kg/day of Ricolinostat, which was previously shown to result in plasma exposure levels consistent with those observed clinically [21]. As observed in vitro, treatment with Ricolinostat in vivo significantly reduced the growth of IBC models without affecting the non-IBC cells (Fig. 3c). For comparison of the anticancer response we performed parallel treatments with a commonly used chemotherapeutic drug (paclitaxel) (Fig. 3c). Interestingly, in IBC xenograft models Ricolinostat reduced tumor growth at least as much as was observed with paclitaxel treatment.

To further demonstrate that inhibition of HDAC6 compromised the growth of IBC cells we performed cell culture growth studies where SUM149 and MDA-MB-231 were treated with second generation HDAC6 inhibitors that are more selective for HDAC6 than Ricolinostat for off-target inhibition of class-I HDACs. These studies showed that despite efficient inhibition of HDAC6 in both cells lines (as demonstrated by accumulation of acetylated α-tubulin) all these selective HDAC6 inhibitors efficiently reduced the growth of SUM-149 but had a minimal impact on MDA-MB-231 viability (Fig. 3d).

Next, we aimed to investigate the dependency of HDAC6 in IBCs. We hypothesized that differential expression and/or activity of HDAC6 between IBC and non-IBC cells could mediate IBC cell sensitivity to HDAC6 inhibition. We studied a series of primary breast cancers (63 IBC and 134 non-IBC) representing the largest IBC data series with matched expression and copy number variant (CNV) data from untreated tumors [49]. The HDAC6 locus is located in the chromosome-X at the p11.23 region. This region is rarely amplified in breast cancer, and we found no differences in the mRNA expression level of HDAC6 between IBC and non-IBC samples (Fig. 4d and data not shown). Thus, differential expression of HDAC6 cannot be linked to the different response observed after HDAC6 inhibition in IBC and non-IBC. However, protein activity can be affected by factors such as post-translational modifications, which do not change protein or mRNA levels. We [36, 50, 51] and others [52] have developed methods to infer protein activity in primary cancer samples by reconstructing regulatory networks using mRNA expression profiles. Thus, we used the gene expression profile signatures in over 900 breast cancer samples available in the TCGA BRCA dataset to reconstruct the genome-wide regulatory networks of breast cancer cells, using the ARACNe [30, 36] algorithm. These methods identified a regulon consisting of 162 transcripts as a set of transcriptional targets whose expression is affected by HDAC6 activity (Fig. 4a). GO term enrichment analysis (DAVID) confirmed that this list was enriched in genes involved in canonical HDAC6 functions, such as response to unfolded protein-induced stress [18‐20] (Fig. 4a). Interestingly, when we analyzed lung (TCGA LUAD)-specific and colorectal cancer (TCGA COAD-READ)-specific HDAC6 regulons, generated by ARACNe analysis of the corresponding TCGA datasets, we obtained a list of 147 and 138 genes, respectively, for which thge overlap with the breast cancer regulon was highly significant (Fig. 4b). This suggests that the transcriptional footprint of the HDAC6 regulon is highly conserved among epithelial cancer cells. Finally we integrated the expression of all transcripts in the HDAC6 regulon in a single score, termed the HDAC6 score (see “Methods”).

To demonstrate that the HDAC6 score is an indicator of the HDAC6 activity, SUM149 cells were treated for 3, 6 and 12 hours with 2.5 uM of Ricolinostat and the HDAC6 score for treated samples was compared to controls. This study revealed that inhibition of HDAC6 significantly attenuated the HDAC6 score (Fig. 4c and Figure S3a in Additional file 5).

Finally, we evaluated the HDAC6 score in our series of 63 IBC and 134 non-IBC primary specimens. Importantly, IBCs had a significantly higher HDAC6 score than non-IBCs (Fig. 4d). To further study whether the HDAC6 score was influenced by the different composition in molecular subtypes between IBCs and non-IBCs [53] we evaluated the HDAC6 score after stratifying the tumor series based on their hormone receptor (HR) status and their intrinsic molecular subtype [54]. Our results revealed that the HDAC6 score was significantly higher in IBCs compared with non-IBC independently of those molecular characteristics (Figure S3b in Additional file 5). Furthermore, multivariate analysis taking into account these molecular classifications demonstrated that there is no significant difference between the multi-variable model, considering PAM50, ER–PR or both, and the single model with IBC only. These data show that inflammatory vs. non-inflammatory is the main feature that impacts on the HDAC6 score (see table in Additional file 1). Overall these data revealed correlation between IBC disease and the HDAC6 score, which suggests a rationale for IBC dependency on HDAC6.