Master regulator analysis of paragangliomas carrying SDHx, VHL, or MAML3 genetic alterations

RNA-seq gene expression profiles for 179 PPGL specimens (TCGA-PCPG) used to assemble an inferred transcriptional network were obtained via the firebrowse tool available through the Broad Institute. This dataset was described previously in a publication of the TCGA consortium [27].

Microarray data for eight normal adrenal medulla specimens and 44 PPGL tumors used to calculate gene expression signatures for SDH loss and VHL loss were obtained from NCBI GEO accession GSE39716. These data were described previously [41, 42]. Included in this dataset are 13 VHL-null pheochromocytomas, 18 SDHB-null pheochromocytomas and paragangliomas, and 13 SDHD-null pheochromocytomas and paragangliomas, including 8 head and neck tumors. For generation of SDH-loss and VHL-loss gene expression signatures used in the discovery phase of the project, the 18 SDH-null tumors or 13 VHL-null tumors were compared to the 8 normal adrenal medulla specimens.

Microarray gene expression profiles for 188 PPGL specimens comprising the COMETE cohort, used in the MR validation analysis, were obtained from ArrayExpress entry E-MTAB-733. This dataset has been described previously [43, 44]. This dataset includes the following number of tumors with specific known genetic defects: 1 SDHA, 17 SDHB, 2 SDHC, 3 SDHD, 27 VHL, 9 RET, 9 NF1, and 122 tumors of other or unknown genetic cause. For generation of the SDH-loss and VHL-loss gene expression signatures used in the validation phase of this project, the 23 SDH-loss tumors or 27 VHL-loss tumors were compared to the all other PPGL tumors in the cohort.

RNA-seq data used to derive the MEF-specific SDHC-loss transcriptomic signature were generated as described below and deposited in NCBI GEO under accession GSE114244. RNA-seq datasets used to assemble a MEF-specific inferred transcriptional network were obtained from ArrayExpress accessions E-GEOD-72275, E-GEOD-64489, E-GEOD-77351, E-GEOD-79095, E-MTAB-5089, E-GEOD-75631, E-GEOD-68902, E-GEOD-71209, E-MTAB-3875, E-GEOD-70816, E-GEOD-63794, E-GEOD-63756, and from NCBI GEO accession GSE103662.

Differential gene expression analysis of COMETE cohort and TCGA-PCPG cohort specimens was conducted using the R package limma, available through Bioconductor. SDH-loss and VHL-loss gene expression signatures were derived by differential gene expression analysis, comparing these specific PPGL tumor subtypes to normal adrenal medulla. Microarray datasets were pre-processed with the application of the robust multi-array average (RMA) algorithm, with downstream differential analysis including linear fitting and analysis of empirical Bayes statistics for differential expression. The MAML3 translocation gene expression signature used in the discovery phase of the project was derived from TCGA-PCPG data, comparing the 10 identified MAML3-mutant tumors to all other tumors (N = 169) in the cohort. Similarly, a differential gene expression signature for SDHC-loss in iMEF cells was calculated by performing differential expression analysis on RNA-seq count data, comparing SDHC-null cell lines to hemizygous control lines. The criteria for differential expression signatures in each human PPGL tumor analysis was an absolute log₂(fold-change) > 1.5 and adjusted p-value < 0.05. The criteria for differential expression for SDHC-loss iMEFs was an absolute log₂(fold-change) > 2.0 and adjusted p-value < 0.05.

Prior to transcriptional network inference (TNI), TCGA RNA-seq data were log₂-transformed. Transcriptional network inference and master regulator analysis (MRA) were conducted using the R package RTN, as previously described [34], re-implementing other previously-described algorithms [33, 40]. Following transcriptional network inference, data processing inequality filtering (DPI tolerance set to 0 or 0.05) was applied to remove the weakest among inferred interactions prior to master regulator analysis. Each differential expression signature was analyzed separately via MRA to infer putative transcriptional drivers of the observed signatures.

Prior to TNI on MEF RNA-seq data to generate a MEF-specific transcriptional network, raw FASTQ datasets were downloaded from public repositories and aligned to the mm9 genome using the HISAT2 fast read aligner. SAM files yielded from the alignment step were converted to BAM format using Samtools [45], and FPKM gene expression values calculated using R package systemPipeR [46]. Following these steps, FPKM values were log₂-transformed prior to TNI, as described above for the human PPGL tumor data. Following TNI, MRA was performed using the SDHC-loss MEF gene expression signature to infer transcriptional drivers of the observed signature.

Known motif position-weight matrices for inferred master regulator DNA binding sequences were downloaded from HOMER (http://​homer.​ucsd.​edu/​homer/​) or JASPAR (http://​jaspar.​genereg.​net/​) motif databases. The unfiltered regulon for a subset of human PPGL master regulators was searched for nearest motif match to the transcription start site (TSS) within a 50-kbp window. This search was then repeated 200 times for a scrambled version of the same motif. For each motif search, the fraction of motif matches localizing within 2.5 kbp of the TSS was calculated, and the distribution of scrambled motif fraction compared to the observed fraction for the original motif via calculation of an empiric p-value.

Cell culture methods and protocol for RNA-seq gene expression analysis for stable SDHC-loss and hemizygous control iMEF cell lines were as described previously [47]. Following high-throughput sequencing, RNA-seq data was deposited in NCBI GEO under accession GSE114244.

Stable SDHC-loss and hemizygous control cell lines were plated into 96-well plates at a density of approximately 20,000 cells per well in 100 μL of DMEM medium containing 10% FBS, penicillin (100 units/mL), streptomycin (100 μg/mL), 1 mM pyruvate, 1X MEM NEAA, and 10 mM HEPES buffer (pH 7.2–7.5) and cultured at 21% O₂ and 5% CO₂. The next day, medium was aspirated and cells were washed with 100 μL PBS, fixed with 100 μL 3.7% formaldehyde in 1XPBS for 20 min at room temperature, permeabilized with 100 μL 0.1% Triton X-100 in PBS for 15 min at room temperature, and blocked for 15 min at room temperature in PBS containing 10% FBS. Solutions of primary antibody were prepared by dilution of anti-SOX11 (Abcam cat# ab134107) or anti-ZFP423 (Abcam cat# ab94451) primary antibodies 1:100 into PBS containing 10% FBS. Primary antibody solutions were added to plates and incubated for 1 h at room temperature with gentle agitation on an orbital shaker. Wells were then washed 3X with 100 μL PBS, and then incubated for 30 min at room temperature with goat anti-rabbit IgG Alexa Fluor 594 (Molecular Probes cat# R37117) diluted 2 drops per mL into PBS containing 10% FBS. Cells were then washed 3X with 100 μL PBS, and then counterstained with DAPI (5 μg/mL in PBS) for 5 min prior to imaging on a Zeiss LSM 780 confocal microscope using a 10X objective. For each visual field, an autofocus routine was implemented to capture the plane with the highest DAPI staining intensity. Automated image analysis was then employed to quantify compartment-specific immunostaining patterns, as previously described [48].

Stable SDHC-loss and hemizygous control cell lines were plated into 96-well plates at a density of approximately 10,000 cells per well in 100 μL medium as described above. All-trans retinoic acid (ATRA; Sigma Aldrich cat# R2625) solutions in DMSO were then added to the plated cells to a final concentration of between 78 nM and 5 μM, with final DMSO concentration of 1% of final medium volume. Plates were tapped gently to mix, and then cultured at 21% O₂ and 5% CO₂ for 4 d. On the fourth day, the medium and retinoic acid solutions were replaced, at which point phenol red was omitted from the medium, and cells allowed to grow for an additional 2 d under the same conditions. On the 6th day, 10 μL Alamer Blue (Thermo Fisher) cell viability reagent was added to plates and incubated for 6 h prior to taking absorbance measurements at 570 and 600 nm. Cell viability was calculated from Alamar Blue reduction as described previously [47].

SH-SY5Y neuroblastoma cells were plated into 96-well plates at a density of approximately 15,000 cells per well in 100 μL of DMEM/F12 medium containing 10% FBS, penicillin (100 units/mL), and streptomycin (100 μg/mL) and grown overnight to allow for cell attachment. The next day, diethyl malonate (Sigma Aldrich cat# W237507) stock solutions prepared in ethanol at 500 mM concentration were added to the plate to 5 mM final concentration, and the cells returned to the incubator overnight. Equal volume of ethanol as vehicle control was added to untreated cells. The next day, fresh stock solutions of all-trans retinoic acid (Sigma Aldrich cat# R2625) were prepared in DMSO at 1.2 mM concentration and added to cells to final concentration of 12 μM. Equal volume of DMSO was added to vehicle-treated cells. Cells were returned to the incubator for 2 days to allow for initiation of retinoic acid-induced neuronal differentiation. Cells were then trypsinized (30 μL trypsin), diluted into 500 μL media, and transferred to a poly-D-lysine-coated glass coverslip. Diethyl malonate, retinoic acid, and appropriate vehicle control volumes were then added to previous concentrations and the cells returned to the incubator overnight to allow for cell attachment. The next day, media was aspirated and cells washed with 1 mL PBS, fixed with 3.7% formaldehyde in PBS for 10 min at room temperature, and permeabilized with 0.1% Triton X-100 in PBS for 15 min at room temperature. Cells were the stained with DAPI (Roche cat# 10236276001) and ActinRed 555 ReadyProbes reagent (Invitrogen cat# R37112). Cells were then covered in PBS and imaged by confocal microscopy. Cell morphology of actin staining was then analysed using the MorphologicalSkeleton and MeasureObjectSkeleton subroutines in CellProfiler.

To infer transcription factors responsible for gene dysregulation in PPGL subtypes, signatures of differentially-expressed genes were first derived to seed the MR inference algorithm (MARINa) [40]. We derived these signatures [absolute log₂(fold-change) > 1.5 and P-Value < 0.05] for SDH-loss, VHL-loss, and MAML3 translocation PPGL subtypes, as described in Methods (Fig. 1a). In total, the identified signatures consist of 230 (SDH-loss), 249 (VHL-loss), and 532 (MAML3 translocation) differentially-expressed genes (Fig. 1b; Additional file 1: Dataset S1, Additional file 2: Dataset S2, and Additional file 3: Dataset S3). A pattern of general gene expression down-regulation was evident for the SDH-loss signature, with 68% of genes being down-regulated. Approximately equal numbers of genes were up- and down-regulated in each of the VHL-loss and MAML3 translocation signatures. Interestingly, the gene sets differentially expressed in the SDH-loss and VHL-loss are highly overlapping (empiric p-value: 5E-136), compared to the expected overlap for randomly-selected gene sets of this same size (empiric p-value 4E-7; Additional file 12: Figure S1). This high degree of overlap between SDH-loss and VHL-loss signatures is consistent with the hypothesis that both tumor molecular subtypes activate pseudohypoxic signalling [49]. We also examined the degree of overlap between the MAML3 translocation differential gene expression signature for PPGL and that previously observed in MAML3 translocation-positive neuroblastoma tumors [50] (Fig. 1d), and identify a significant similarity (empiric p-value: 3E-35; Additional file 13: Figure S2). This suggests that MAML3 translocation has similar downstream impacts upon gene expression in these highly similar tumor types.

We then assembled a PPGL-specific inferred transcriptional network using 179 RNA-seq experiments generated by the TCGA Research Consortium (http://​cancergenome.​nih.​gov/​). This dataset has been described previously [27]. As described in Methods, we used these data to assemble inferred transcriptional networks modeling gene regulation in PPGL tumors, using an implementation of the ARACNE transcriptional network inference (TNI) algorithm available in the R package RTN [34]. We subsequently validated the structure of the inferred transcriptional network by searching the promoter sequences (50 kbp window centered on TSS) of the inferred transcription factor (TF) regulons (i.e. sets of target genes) for known TF DNA-binding domain motifs, extracting the nearest pattern match for each TSS, and comparing to the same searches performed iteratively for scrambled versions of the same motif. We then assessed the fraction of nearest pattern matches for each search localizing to within 2.5 kbp of the TSS. In each of the examined cases, we find that the fraction of motif pattern matches in this window is higher for the original motif than for most scrambled versions (Additional file 14: Figure S3), suggesting that the inferred transcriptional network successfully models real gene-regulatory interactions.

We then applied the MR inference algorithm (MARINa; also implemented in R package RTN), using the inferred transcriptional network, to infer TFs whose target genes are significantly enriched for the input SDH-loss, VHL-loss, or MAML3-translocation gene expression signatures. First, we performed MRA on ARACNE-inferred transcriptional networks stringently trimmed to remove all but the strongest inferred regulatory connections (DPI tolerance = 0). The results of these analyses, presented in Additional file 4: Dataset S4, Additional file 5: Dataset S5, and Additional file 6: Dataset S6, nominate several dozen high confidence TFs whose target genes are perturbed in these specific PPGL molecular subtypes, with each TF typically controlling between 2 and 10% of differentially-expressed genes. This TF subset nominated by MRA with adjusted p-value < 0.05 are shown in Fig. 2a, with the degree of statistical significance from the MRA indicated on the y-axis, and the fraction of the input gene expression signature attributed to each MR indicated on the x-axis.

Surprisingly, hypoxia-inducible factors were not among the nominated high confidence SDH-loss MRs, although HIF2α (encoded by EPAS1) was nominated as a MR for VHL-loss tumors, controlling ~ 5% of the observed differentially-expressed genes in that tumor subtype. We therefore repeated the MRA procedure using ARACNE-inferred networks trimmed with slightly less stringent criteria (DPI tolerance = 0.05), so as to enhance the sensitivity of detection of master regulators. These analyses, presented in Additional file 7: Dataset S7, Additional file 8: Dataset S8, and Additional file 9: Dataset S9, reveal evidence of some enhanced EPAS1-related transcriptional effects in SDH-loss tumors, although EPAS1 is by no means among the top MRs inferred to cause observed patterns of tumorigenic transcriptional perturbation. Importantly, this suggests that HIF-related transcriptional dysregulation is detectible, but that it inadequately accounts for the majority of transcriptional perturbations observed in these “pseudohypoxic” tumor subtypes, and that dysregulation of other transcription factors may play a much more important role in driving oncogenic transcriptomic patterns than previously appreciated. Strikingly, for MAML3 translocation-positive tumors, we find that a single TF, IRX4, is predicted to account for > 20% of the observed transcriptional perturbation. This is particularly intriguing because the reported MAML3 translocations involve chr18~chr4 fusion (TCF4~MAML3 gene fusion) or chr17~chr4 fusion (UBTF-MAML3 gene fusion), with the IRX4 locus on chromosome 5 being unaffected in either case [27]. It is intriguing and currently unknown how MAML3 translocation is connected to IRX4-mediated transcriptional dysregulation.

Next, we examined the structure of the inferred PPGL transcriptional network to determine whether the inferred subtype-specific MRs specifically co-regulate common sets of target genes. We therefore calculated in a pairwise fashion the degree to which the regulon (i.e. set of target genes) of a given TF overlaps with the regulon of other known TFs (Additional file 10: Dataset S10). This pairwise matrix of regulon overlap values was then ordered using an unbiased hierarchical clustering algorithm to identify groups of TFs with highly overlapping sets of target genes. This data, presented in Fig. 2b, shows that TFs in PPGL tumors exert their activities on two distinctly different and highly co-regulated subnetworks (labelled #1 and #2, respectively), with an identifiable, but somewhat less co-regulated third subnetwork (labelled #3). Intriguingly, SDH-loss and VHL-loss MRs exert their activities almost exclusively on subnetworks #1 and #3, while sparing subnetwork #2. This asymmetry in the transcriptional subnetworks impacted by PPGL MRs suggested to us the possibility that transcriptional subnetwork #2 is essential for cell survival.

This hypothesis was indeed borne out by more detailed functional analysis of putative transcription-regulatory interactions. We considered the subset of genes inferred to be jointly controlled by at least two transcription factors within a given subnetwork and assessed enrichment for specific gene ontologies. Strikingly, this revealed that each of the three identified subnetworks is highly enriched for specific transcriptional programs (Additional file 14: Figure S3G-I). Subnetwork #1 was found to be highly enriched for genes involved in plasma membrane organization, extracellular matrix (ECM) organization, and regulation of cell migration (Additional file 14: Figure S3G). Subnetwork #3 was found to be highly enriched for genes expressed in the endoplasmic reticulum and various metabolic processes (Additional file 14: Figure S3I). In contrast, subnetwork #2 was found to be highly enriched for nuclear proteins, splicing factors, and proteins involved in ubiquitin-dependent catabolic processes (Additional file 14: Figure S3H). These data support the notion that subnetwork #2 is likely important for cell survival, and additionally, that modulation of subnetwork #1 may be a mechanism by which tumor cells acquire an invasive phenotype. Rigorous testing of these specific hypotheses will be a topic of interest for future follow-up.

We then assessed the degree to which the cumulative effects of the top 20 MRs in each tumor subtype can account for the observed PPGL molecular subtype gene expression signatures. This analysis, presented in Fig. 2c, shows that for each tumor subtype, the top 20 MRs are predicted to account for > 60% of the input tumor subtype-specific gene expression signatures. This suggests that the majority of transcriptomic perturbation for these tumor subtypes may be explicable in terms of the cumulative effects of multiple dysregulated MRs.

Next, we assessed the degree to which the MRs nominated for SDH-loss, VHL-loss and MAML3 translocation-positive PPGL tumors are similar. We therefore assessed the overlap between master regulator lists for these tumor subtypes. This analysis revealed significant overlap between SDH-loss and VHL-loss tumors (Fig. 2d), with a degree of overlap is larger than would be expected by chance considering the size of the specific nominated MR lists (empiric p-value: 9E-8; Additional file 15: Figure S4). This suggests an aspect of similarity in transcription factor dysregulation in SDH-loss and VHL-loss PPGL tumors.

Beyond nomination of MRs that collectively explain patterns of dysregulated gene expression in specific PPGL molecular subtypes, we sought to validate the nominated MRs using data from an independent large cohort of PPGL specimens (N = 188). This dataset has been described previously [43, 44]. Using known information regarding SDHx and VHL mutation status of these tumor specimens, we calculated the average log₂(fold-change) in gene expression for tumors of each of these molecular subtypes relative to other PPGL tumors. Then, using known regulon structures of the PPGL inferred transcriptional network, we calculated the average log₂(fold-change) for all genes under regulatory control of each of the SDH-loss and VHL-loss MRs nominated from MRA performed on the ARACNE network trimmed with DPI tolerance of 0.05. We performed this same calculation using the input gene expression signature used for MR discovery, and then examined whether there was correlation between the regulon log₂(fold-change) values for discovery and validation data sets. Strikingly, we observe strong agreement between the two analyses (Fig. 3a, b). The strength of statistical correlation is especially remarkable considering the fact that the discovery signature involves a tumor-normal comparison while the discovery signature involves a tumor-tumor comparison. The validation analysis therefore supports the claim that the inferred dysregulated MR activities are biologically-reproducible phenomena that are consistent across independent patient cohorts.

We pursued further analysis of the discovery and validation cohort data, asking whether these altered TF activities are sufficient to drive hierarchical clustering of PPGL specimens that accurately separates specimens according to molecular subtype. For this analysis, we leveraged the VIPER algorithm, which is capable of robustly converting individual transcriptomic datasets into TF activity profiles [51]. When this algorithm is applied to discovery cohort specimens including SDHB-null, SDHD-null, and VHL-null PPGL tumors from various anatomical locations, as well as normal adrenal medulla specimens, the hierarchical clustering algorithm generally separates normal adrenal samples, VHL-null pheochromocytomas, and SDHD-null head and neck paragangliomas away from the remaining SDHB-null and SDHD-null PPGL tumors (Fig. 3c). This suggests that significant differences in transcription factor activity profiles exist between SDHD-null head and neck PPGL tumors and SDH-null tumors arising from other anatomical locations, although SDH-loss tumors are generally more self-similar than PPGL tumors arising from other genetic defects (Fig. 3d). These patterns were further reproduced by t-SNE clustering (Additional file 16: Figure S5), similarly underscoring the unique TF activity profile of head and neck PPGLs compared to tumors arising from SDH subunit defects in abdomen or thorax, but with little other difference that can be attributed to specific SDH subunit defects.

We therefore pursued a more detailed analysis of differences in TF activity profiles that exist between SDHD-null head and neck PPGL tumors and SDHD-null tumors from abdomen and thorax. This analysis, presented in Additional file 17: Figure S6, reveals several specific differences in transcription factor activity that correlate with anatomical location. Among the most impressive differences in TF activity, head and neck tumors display lower activities of FEV, TFAP2B and GATA2 and higher activities of TFEC, LEF1, and MAFB compared to SDHD-null tumors of other anatomical locations. The basis for this is unclear, but this finding clearly suggests that tumor anatomic location and underlying genotype synergize to drive observed molecular signatures at the transcriptional level. Whether any of these differences is responsible for the generally more benign character of head and neck neoplasms remains to be seen.

We then specifically considered EPAS1, a transcription factor of much interest to the field, testing the hypothesis that tumors originating from different SDH subunit mutations have variable EPAS1 activity. This analysis, presented in Fig. 3e, revealed similarly elevated levels of EPAS1 activity in SDHB-null, SDHC-null, and SDHD-null PPGL tumors compared to those attributable to RET of NF1 gain-of-function mutations. Additionally, EPAS1 activities for all SDHx-loss tumors were roughly similar to those observed in VHL-loss tumors. These data suggest that EPAS1 activity is similarly elevated for all PPGL tumors exhibiting VHL loss or SDH complex bi-allelic loss of function, regardless of the particular SDH subunit involved.

We next tested the hypothesis that EPAS1 activity is significantly different in SDHB-null metastatic tumors vs. those annotated as “non-malignant”. This analysis, presented in Fig. 3f, did not detect any significant differences in EPAS1 activity between metastatic and “non-malignant” tumors. Indeed, an expanded analysis considering the activities of all transcription factors inferred via the VIPER algorithm did not identify a single TF that is differentially active between metastatic SDHB-null tumors and those annotated as “non-malignant”. This suggests either a lack of statistical power in the analysis, or else the lack of a true biological distinction between the involved samples. Also, as a subset of the specific analysis of SDH MR activities in SDHD-null tumors of the head and neck vs. those arising from other anatomical locations, we did not note any statistical difference in EPAS1 activity between tumors arising from these locations (Fig. 3g). Collectively, these data suggest that differences in EPAS1 activity are not appreciably different between SDH-null tumors of variable malignant character or those arising from disparate anatomic locations.

Beyond analysis of SDH-loss MRs in human PPGL tumors, we next asked whether there is any evidence for conservation of SDH-loss MR responses across species and cell types. To answer this question, we generated a transcriptomic signature of SDHC-loss immortalized mouse embryonic fibroblasts (iMEFs). This model of SDHC-loss was used rather than the more common SDHB and SDHD subunit defects seen in human PPGL tumors specifically because the SDHC gene trapped allele used to derive the line was readily available from the Wellcome Trust Sanger Institute Gene Trap resource. Examination of transcriptional patterns in this model revealed general transcriptional up-regulation (Fig. 4a) that is different from that observed in human tumors. The reason for this is unclear. We next performed MR analysis on a MEF-specific inferred transcriptional network, as described in Methods. In total, this approach nominated 73 MR TFs whose target genes are perturbed upon SDHC-loss (Additional file 11: Dataset S11). Notably, EPAS1 was nominated as the top-ranking SDHC-loss MR gene, in contrast to the SDH-loss human PPGL tumor analysis (Fig. 4b). The reason for this difference is unclear, but it seems likely that cultured cell lines provide homogeneous populations, whereas tumors have significant cellular heterogeneity. In any case, the observation that the top SDHC-loss MEF MR is EPAS1 (HIF2α) lends considerable support to the pseudohypoxia hypothesis of succinate-mediated poisoning of prolyl hydroxylases catalysing HIF factor hydroxylation. This being the case, we emphasize that EPAS1 dysregulation is still inferred to account for only ~ 5% of dysregulated genes in SDHC-loss MEFs, indicating that alternative explanations must be sought to account for the majority of the observed dysregulated gene expression signature. These explanations likely include dysregulation of other TFs, as observed in our analysis.

We hypothesized that although human PPGL tumors and mouse SDH-loss iMEFs are obviously very different, fundamental similarities might point to drivers of PPGL tumorigenesis. We therefore examined the overlap between the inferred SDHC-loss iMEF MRs and MRs inferred via unbiased MR analyses in human SDH-loss PPGL tumors. This analysis revealed five potentially conserved MRs, a number that is higher than would be predicted by random chance, suggesting that perturbation of these MRs represents a conserved biological response to SDH loss (Fig. 4c). We then assessed whether the expression changes for the inferred regulons of each of these MRs in SDH-loss human PPGL tumors and in SDHC-loss MEFs display similar patterns of regulatory perturbation. This analysis, presented in Fig. 4d, reveals that 4 of the 5 conserved SDH-loss MRs also show conserved patterns of regulon gene expression perturbation. These 4 MRs (ZNF423/ZFP423, SOX9, ETV5, and SOX11) thus represent high-confidence conserved SDH-loss MRs. Since SOX11 and ZFP423 exhibit the most dramatic patterns of altered regulation, these were chosen for follow-up analysis.

We examined the patterns of differential gene regulation for the inferred regulons of ZFP423 and SOX11, assessing the degree to which expression of these gene subsets is sufficient to drive unbiased hierarchical clustering of iMEF experimental (SDHC-loss) and hemizygous control specimens (Fig. 4e, f). The results of these analyses show that the regulon for each of these MRs is indeed sufficient to segregate experimental and control cell lines, suggesting that the regulatory activities of these factors are indeed characteristically perturbed in SDHC-loss cell lines. Interestingly, despite down-regulation being the predominant pattern of gene dysregulation for the ZFP423 regulon in SDHC-loss cells, a subset of genes displays consistent up-regulation relative to the control cell lines. This suggested to us the possibility that ZFP423 may not be acting simply as a transcriptional co-activator.

We therefore assessed patterns of gene expression regulatory synergy between ZFP423 and other known TFs to determine whether co-regulation may modulate the transcriptional effects of ZFP423. This analysis involved assessing in a pairwise fashion whether the portion of the ZFP423 regulon overlapping with another TF is differentially-expressed relative to the full ZFP423 regulon, and to the full regulon of the other TF. This analysis, presented in Fig. 4g, revealed three distinct modes of transcriptional modulation by ZFP423, suggesting complicated TF function. For a subset of TFs (HOXA13, SOX18, NKX3–2, HOXC6, EVX2, POU1F1, HOXC8, HOXB7, HOXC4, EBF1, HOXC5, and HOXC9), ZFP423 displays synergistic transcriptional repression. Encouragingly, EBF1 has previously been described as a ZNF423 interaction partner, with synergistic repressive effect, suggesting that the other synergistic repressive partners inferred by this method may also be genuine [52]. A second subset of synergistic partners (FOS, HMX2, HOXA11, ARNTL, RXRG, ZIC2, POU3F2, TAL1, EHF, POU3F1, KLF15, NR1H3, NFE2L3, AR, and BCL6B) was inferred for which ZFP423 binding attenuates the general transcription-activating activity of these TFs. A third subset of synergistic partners (ZFP12B, ARID5B, MXD1, BCL6, SPDEF, VAX1, TFDP1, and ZFP410) was inferred with synergistic activation of gene expression in the co-regulated gene subset. These synergistic activating interactions likely account for the small subset of genes in the ZFP423 regulon that display up-regulation upon SDHC-loss. The complex pattern of gene up- and down-regulation upon SDHC-loss reflects the known complex role of ZFP423 in gene regulation.

Similar analysis of synergistic interactions for SOX11 was simpler, with generally only synergistic transcription-activating interactions detected. Putative synergistic activating partners include ALX3, DBP, FGF9, DBX2, EPAS1, BRCA1, and FLI1 (Fig. 4h). Immunostaining for SOX11 levels in SDHC-loss and control MEFs revealed an increased SOX11 signal in the SDHC-loss context (Fig. 5a, b). Taken together with the general pattern of regulon up-regulation upon SDHC-loss and the sparse synergistic up-regulatory interactions with other factors, this is consistent with SOX11 having a role as a simple transcriptional activator. The reasons why the cellular concentration of SOX11 is increased in SDHC-loss MEFs remain unclear, however.

Interestingly, immunostaining for ZFP423 in SDHC-loss and control MEFs did not reveal dramatic differences in cellular abundance of this MR, suggesting that bulk increase or decrease in cellular ZFP423 protein is not a feature of SDHC-loss (Fig. 6a, b). We did, however, detect generally increased nuclear localization of ZFP423 in SDHC-loss context (Fig. 6c). Understanding the basis for this altered nuclear localization of ZFP423 upon SDHC-loss will require future study.

ZNF423 has previously been described as a transcriptional cofactor of RARα/RXRα that is critically required for retinoic acid-induced differentiation of neuroblastoma [53]. An aspect of this prior work was the demonstration that ZNF423 physically associates with RARα/RXRα in the region of the RARβ promoter, potentially driving transcriptional activation through the RARα/RXRα complex. We therefore examined our RNA-seq gene expression data from SDHC-loss and control MEF lines to determine whether there is any evidence for differential expression among any of the known retinoic acid receptors (RAR) or retinoid X receptors (RXR). Intriguingly, we find that Rarβ gene expression is significantly up-regulated in SDHC-loss cells (Fig. 6d). If ZFP423 is a transcriptional cofactor required for activation of Rarβ gene expression, increased nuclear localization of ZFP423 upon SDHC-loss and subsequent increased occupancy of the Rarβ gene locus would potentially account for the observed transcriptional activation of Rarβ expression.

The observation that SDHC-loss cells display increased nuclear localization of ZFP423 and concomitant up-regulation of Rarβ gene expression suggested to us the possibility of differential cellular response to retinoic acid. We tested this as described in Methods. After 6 d of exposure to retinoic acid over a range of concentrations, we assayed cell viability as a function of genotype. Strikingly, we observe robust induction of cell death in both control cell lines at relatively low concentrations of retinoic acid, but little effect on viability for SDHC-loss lines (Fig. 6e). This suggests that SDHC-loss status fundamentally modulates the cellular response to retinoic acid, one of the most powerful known endogenous morphogens. How perturbation of ZFP423 and RARβ results in decreased retinoic acid-induced cell death is unclear. Previous work has suggested that BCL2 overexpression is predictive of a differentiation response rather over induction of apoptosis in cultured tumor cell lines [54]. Intriguingly, we see that BCL2 expression is generally up-regulated in SDHC-loss MEF lines relative to controls, although the comparison is not statistically significant (Fig. 6f). Our results suggest that SDHC loss may fundamentally modulate cells away from a death response to retinoic acid.

Since retinoic acid responses differ between cell types, with the apoptotic responses on MEFs not necessarily predicting the retinoic acid responses of neuronal cells, we therefore decided to test the effects of SDH complex inhibition on the retinoic acid-induced differentiation of SH-SY5Y neuroblastoma cells. The differentiation responses of SH-SY5Y to retinoic acid have been previously well-characterized, with resultant phenotypic changes in neurite outgrowth that are readily apparent through morphological image analysis that has been broadly used to probe the pathways involved in neuronal differentiation [55‐58]. We therefore studied the effects of SDH complex inhibitor malonate (formulated as cell-permeable diethyl ester) upon retinoic acid-induced differentiation in the SH-SY5Y neuroblastoma cell line, assessing for changes in neurite outgrowth as a function of exposure to diethyl malonate (DEM) and/or retinoic acid. Intriguingly, we observe that the cellular differentiation response to 12 μM all-trans retinoic acid for 72 h is blunted when cells are pre- and co-treated with 5 mM diethyl malonate (DEM), evident through relatively attenuated neurite outgrowth (Fig. 7). This result aligns with the date from the MEF studies demonstrating that SDH loss attenuates the normal cellular response to retinoic acid, suggesting that attenuated retinoic acid response may be a generalized property of cells lacking SDH complex activity. How these phenomena relate to human PPGL tumors remains to be seen.