Nuclear speckles regulate HIF-2α programs and correlate with patient survival in kidney cancer

SUMMARY Nuclear speckles are membrane-less bodies within the cell nucleus enriched in RNA biogenesis, processing, and export factors. In this study we investigated speckle phenotype variation in human cancer, finding a reproducible speckle signature, based on RNA expression of speckle-resident proteins, across >20 cancer types. Of these, clear cell renal cell carcinoma (ccRCC) exhibited a clear correlation between the presence of this speckle expression signature, imaging-based speckle phenotype, and clinical outcomes. ccRCC is typified by hyperactivation of the HIF-2α transcription factor, and we demonstrate here that HIF-2α drives physical association of a select subset of its target genes with nuclear speckles. Disruption of HIF-2α-driven speckle association via deletion of its speckle targeting motifs (STMs)—defined in this study—led to defective induction of speckle-associating HIF-2α target genes without impacting non-speckle-associating HIF-2α target genes. We further identify the RNA export complex, TREX, as being specifically altered in speckle signature, and knockdown of key TREX component, ALYREF, also compromises speckle-associated gene expression. By integrating tissue culture functional studies with tumor genomic and imaging analysis, we show that HIF-2α gene regulatory programs are impacted by specific manipulation of speckle phenotype and by abrogation of speckle targeting abilities of HIF-2α. These findings suggest that, in ccRCC, a key biological function of nuclear speckles is to modulate expression of a specific subset of HIF-2α-regulated target genes that, in turn, influence patient outcomes. We also identify STMs in other transcription factors, suggesting that DNA-speckle targeting may be a general mechanism of gene regulation.


INTRODUCTION
contributions to patient variation between cancer types (derived from PCA analysis; see Methods for detailed explanation), we observed remarkable correlations between cancer types (Fig 1Aiii, strong correlations are orange and red), indicating that speckle protein gene expression varies reproducibly in cancer.
Based on this consistent speckle protein gene expression variation across many cancer types, we generated a multi-cancer 117 gene "speckle signature" containing speckle protein genes that consistently contributed to patient variation (Fig 1Aiv; see Methods and Fig S1A for how speckle protein genes were selected). This included 40 "Signature I-high" speckle protein genes and 77 "Signature II-high" speckle protein genes that were consistently reciprocally expressed from one another, and that separated tumor samples into two groups based on their expression (Fig 1B). We assigned each patient a speckle signature score based on the collective expression of these 117 speckle protein genes (Fig 1B, speckle score to the left of each heatmap; see Methods) and used this quantitative measure for Kaplan-Meier analysis of disease outcomes. We assessed overall and disease-specific survival, separating patients by the top versus bottom quartiles or positive versus negative speckle scores (4 total analyses; Fig S1B). Of 24 cancers with consistent speckle protein gene contributions to patient variation (right grey bar in Fig 1Aiii), 21 showed no correlation between speckle signature and patient outcomes for any survival measurement, as shown by examples of melanoma (SKCM) and breast cancer (BRCA) (Fig 1C, left panels; Table S1). In head and neck cancer (HNSC), speckle Signature I was slightly correlated with poor overall survival in the quartile analysis (Table S1), while in ovarian cancer (OV), speckle Signature II slightly correlated with poor outcomes in the quartile analysis (Fig S1B, top  left).
Contrasting with other cancer types, speckle signature in clear cell renal cell carcinoma (ccRCC; called KIRC in TCGA data) showed remarkable correlation with patient outcomes (Fig 1C, right; Fig S1B bottom; Table S1). In ccRCC, patients expressing speckle Signature I had significantly worse outcomes compared to patients with speckle Signature II. Further, although speckle Signature I was more prevalent in later stage ccRCC (Fig S1C; bottom bars), speckle Signature I also occurred in early stage ccRCC (Fig S1C; top bars) and was predictive of survival in both early and late stages (Fig S1D). Hence, speckle signature can be an early-occurring predictor of patient outcomes in ccRCC. Of note, normal adjacent tissues of ccRCC, and most other cancer types, showed speckle Signature II expression patterns (Fig S1E), indicating that speckle Signature I reflects a dysfunction that emerges in tumors. Overall, these findings demonstrate that while the two speckle signatures were evident in 24 cancer types, they correlate strongly with patient outcomes specifically in ccRCC.

A putative speckle targeting motif (STM) is shared between HIF-2a and p53 and re-occurs among regulators of gene expression
We next sought to illuminate why patient survival in ccRCC was particularly correlated with speckle signature differences. One feature of ccRCC that distinguishes it from other cancer types is that ccRCC is typified by inactivating mutations in the VHL gene (reviewed in 24 ), whereas other cancers within the TCGA PanCan dataset have more heterogeneous etiologies. The pVHL protein encoded by VHL is a negative regulator of hypoxia-inducible transcription factors HIF-1a and HIF-2a, of which HIF-2a is proposed to be the oncogenic driver [24][25][26] . Given our previous demonstration of DNA-speckle targeting abilities by the p53 transcription factor, we hypothesized that HIF-2a may be a second example of a stress-responsive transcription factor that regulates association between specific genes and nuclear speckles.
HIF-2a and p53 belong to different transcription factor classes [27][28][29] . However, using EMBOSS Matcher 30 to perform local alignment, we discovered a 29 amino acid region as the best match between p53 and HIF-2a, with 55% sequence similarity (Fig S2A). This region comprised p53 amino acids 62-90, and encompasses the exact region required for DNA/target gene association with speckles that we mapped previously (p53 amino acids 62-77) 21 . The aligned sequence is characterized by prolines spaced every 5 amino acids (Fig 2A, labelled with *), with one proline preceded by a threonine (Fig 2A, green highlight). Examination of HIF-2a revealed a second region of spaced prolines with a serine/proline (SP) dinucleotide in the same relative position, and a similar sequence was present in GATA1 (Fig 2A), a transcription factor active during erythropoiesis, which also involves regulated DNA-speckle association 22,31 . Based on these findings, we termed this protein region the putative speckle targeting motif (STM). Pairwise Clustal Omega alignment between HIF-2a and HIF-1a showed that HIF-1a lacks the proline to the N-terminus of the TP dinucleotide found in the first HIF-2a STM (Fig S2B; Fig S2C, top red circle), and thus does not have prolines spaced every 5 amino acids, whereas the second HIF-2a STM is absent in HIF-1a (Fig S2B and Fig S2C). Hence, putative STMs are found within p53, HIF-2a, and GATA1, but are divergent between HIF-2a and HIF-1a.

HIF-2a maintains DNA-speckle contacts of its target genes in ccRCC cells
While hundreds of gene regulators contain STMs, it is unclear whether the presence of an STM predicts DNAspeckle targeting. To test speckle targeting functions of HIF-2a, we utilized 786-O ccRCC cells, a patientderived cell line that has elevated endogenous HIF-2a 32 , and inhibited the interaction between HIF-2a and its obligate DNA-binding heterodimer ARNT/HIF-1b with the PT2399 small molecule 33,34 . ChIP-seq of HIF-2a with the inhibitor showed loss of HIF-2a genomic binding at 3 hours after PT2399 compared to vehicle treatment (Fig 2C). We used this time point to assess whether HIF-2a DNA binding is essential for speckle association, employing two genomic methods to measure DNA-speckle contacts. The first, SON TSA-seq, is a proximitylabeling approach that utilizes an antibody against the SON speckle marker protein to capture DNA that resides near speckles 14,19 . The second, SON Cut&Run, is a pA-MNase method that captures speckleassociated DNA by digesting chromatin fragments that surround SON 35 . We found high correlation between genomic SON signal from TSA-seq and Cut&Run (Fig S2E). SON TSA-seq and SON Cut&Run detected significant PT2399-dependent decreases in SON (padj < 0.01), and decreasing regions significantly overlapped between the two methods. However, each method detected decreases that were not detected by the other (Fig S2F). In comparing them, SON TSA-seq more effectively identified alterations at lower SON regions, while SON Cut&Run more readily detected changes at higher SON regions (Fig S2G). Distinct types of regions detected as differential between the two methods was not surprising given that they use distinct chemistries to identify speckle associated chromatin. In summary, we consider that both methods are valid, capturing different types of decreasing regions, we thus combined them in the analyses below.
As a third method to evaluate DNA-speckle contacts upon HIF-2a inhibition, we employed immunoDNA-FISH with immunofluorescence against SON to label speckles and DNA-FISH against gene loci to label DNA. We designed probe sets against regions that decreased in SON TSA-seq or Cut&Run signal upon PT2399 treatment (for example, Fig 2D, DDIT4 orange box), as well as regions that did not change (for example, Fig 2D, DDX21 blue box). We treated 786-O cells with DMSO control or PT2399, and found that HIF-2a inhibition resulted in the DDIT4 HIF-2a target gene localizing farther from speckles, while the DDX21 control gene did not change speckle association (Fig 2E -example images; Fig 2F -quantification). Likewise, FBXW5, APOL1 and ELFN2 HIF-2a target genes that significantly decreased SON in TSA-seq and Cut&Run showed reduced speckle association by immunoDNA-FISH, as did MYEOV (Fig 2F), which had decreased SON signal by Cut&Run, but not by TSA-seq. A second control gene, GTF3C4, which is within the same larger speckle domain as FBXW5, did not decrease SON signal in either genomic method, and did not show changes in speckle association by immunoDNA-FISH (Fig 2F). Comparing the immunoDNA-FISH-measured proportion of foci at speckles to the amount of SON signal for SON TSA-seq (Fig 2G, left) and SON Cut&Run (Fig 2G,  right) for each gene, we observed similar correlations between imaging and genomic speckle-association measurements, supporting the contention that each method reflects the amount of DNA-speckle association. In total, these three orthogonal approaches demonstrate that inhibition of HIF-2a DNA binding lowers HIF-2a target gene speckle association, indicating a new mechanistic function of HIF-2a for maintaining DNA-speckle contacts in ccRCC cells.

HIF-2a regulates speckle association of a functionally-distinct subset of target genes
Our findings indicate that, like p53, HIF-2a has a speckle targeting function that drives association between specific genes and nuclear speckles. Previously, we showed that p53 regulates speckle association of 20-30% of target genes; these speckle-associating target genes experience boosted induction upon p53 activation and belong to distinct functional categories compared to non-associating p53 target genes 21 . To determine if similar principles were true for HIF-2a-mediated DNA-speckle association, we performed RNA-seq at 3, 6, 9, and 12 hours after PT2399 treatment in 786-O cells (Fig S2H). We found that genes exhibiting decreased speckle association also showed a general decrease in gene expression at 3 hours, which became more pronounced at later time points (Fig 2H, 2007 genes, compare to log2 fold change of 0 red line). We then evaluated how many HIF-2a-responsive genes, as defined by reduced RNA expression at any time point upon PT2399 treatment, showed HIF-2a-regulated DNA-speckle association, as measured by SON TSA-seq or Cut&Run (Fig 2I). This analysis showed significant overlap between HIF-2a target genes and genes with decreased SON signal upon PT2399 treatment, revealing that ~25% of HIF-2a target genes experience PT2399-mediated decreases in speckle association. Thus, like p53, HIF-2a regulates speckle association of a subset of its target genes.
To evaluate whether speckle-associating and non-speckle-associating HIF-2a target genes differ in their level of expression, we examined RNA-seq data of these two gene groups, with 250 HIF-2a target genes that had decreased SON in TSA-seq or Cut&Run with PT2399 (padj < 0.01) defined as "associating", and 522 HIF-2a target genes that did not have decreased SON in TSA-seq or Cut&Run (padj > 0.1) defined as "nonassociating". We found that speckle-associating HIF-2a target genes tended to have higher RNA expression at baseline without HIF-2a inhibition (Fig 2J, left; DMSO treated 786-O cells), and showed slower reduction in gene expression upon PT2399 treatment compared to non-speckle-associating targets (Fig 2J, right). Hence, HIF-2a target genes that experience HIF-2a-mediated speckle association are more highly expressed and show distinct responses to PT2399 HIF-2a inhibition.
We next assessed whether speckle-associating versus non-associating HIF-2a target genes belonged to distinct functional categories using Gene Ontology (GO) analysis. We found that speckle-associating HIF-2a target genes were specifically enriched in terms related to inflammation (e.g. IkB/NFkB signaling and cell response to IL4), protein transport, metabolism, and cell division (Fig 2K, left), while non-speckle-associating HIF-2a target genes were specifically enriched in terms related to cell differentiation and tissue development, importantly including many terms relating to angiogenesis (Fig 2K, right). These findings indicate that nuclear speckles act upon a specific subset of HIF-2a target genes.
The speckle-targeting-defective HIF2As-∆STMs mutant provides a specific tool to evaluate the gene expression consequences of disrupting DNA-speckle association. We employed single molecule RNA-FISH (smRNA-FISH) of DDIT4 in the above MCF7 cells (with doxycycline-induced HIF2As-wtSTM or HIF2As-∆STM) using intron and exon probe sets which enable transcription site exon probe intensity to be measured concurrently with transcription site distance to the nearest speckle (Fig 3D, diagram on left). Transcription site intensities increased upon application of HIF2As-wtSTM, whereas they failed to increase in HIF2As-∆STMs (Fig 3E). We compared transcription site exon intensity to speckle distances, and observed a characteristic Lshaped distribution (also observed in our p53 results in 21 ) with RNA building up at speckle-adjacent sites in cells with HIF2As-wtSTM, but not in cells without induction of HIF-2a or with HIF2As-∆STMs (Fig 3F). We observed the same trends with DDIT4 intronic probes (Fig S3B-C), suggesting that HIF-2a STMs are essential for production of nascent unspliced DDIT4 RNAs within transcription sites. We quantified the number of exonic spots to estimate total DDIT4 RNA molecules per cell, revealing that HIF2As-∆STMs greatly lowered induction of DDIT4 compared to HIF2As-wtSTM (Fig 3G). These results indicate that intact speckle-targeting of HIF-2a is required for expression of speckle-associating HIF-2a target gene, DDIT4.
We then evaluated the genome-wide consequences of HIF-2a STM deletions via RNA-seq, testing whether disrupting speckle targeting had specific consequences based on speckle association. We found that HIF2As-∆STMs was specifically defective at inducing speckle-associated HIF-2a target genes (Fig 3H; yellow bin 5 -most speckle associated genes), while non-speckle-associated HIF-2a target genes were similarly induced by HIF2As-wtSTM and HIF2As-∆STMs (Fig 3H; grey bin 1 -least speckle associated genes). Similarly, by comparing induction of HIF-2a target genes that did or did not have speckle association regulated by PT2399 (from Fig 2), we found that HIF2As-∆STMs was defective at inducing genes that had speckle association regulated by HIF-2a (Fig 3I; ∆STMs blue "non-associating" versus ∆STMs blue "associating"), whereas HIF2As-wtSTM showed comparable induction of HIF-2a target genes that did or did not have HIF-2adependent changes in speckle association (Fig 3I; wtSTM orange "Non-associating" versus wtSTM orange "Associating"). These results demonstrate that deletion of HIF-2a STMs shifts the balance of HIF-2a target gene induction, with specific consequences for speckle-associated target genes.
In aggregate, we identified a motif for speckle association (the STM) that reoccurs among many key regulators of gene expression, such as HIF-2a. Our results show (1) that HIF-2a works via its STMs to accomplish DNA-speckle association, and (2) that disruption of HIF-2a speckle targeting ability compromises induction of the speckle-associating subset of its target genes. These findings together support a model whereby nuclear speckles and DNA-speckle targeting represent a distinct layer of gene regulation that shifts the balance of genes that are preferentially expressed. We hypothesize that this layer of gene regulation can be manipulated by altering the transcription factor speckle targeting abilities, as with HIF-2a STM deletions above, or by alterations in the speckles themselves, as with the cancer speckle signature (see Fig 1), explored further below.
Speckle Signature I tumors are enriched in oxidative phosphorylation and ribosome pathways, with higher RNA expression of speckle-associating megabase-sized gene neighborhoods The speckle signature, while present in many cancers, was particularly predictive of survival in ccRCC (see Fig  1). Given our findings that HIF-2a regulates DNA-speckle association, we hypothesized that HIF-2a combines with speckle phenotypes, resulting in poor ccRCC outcomes. To first broadly understand the consequences of cancer speckle signature, we shifted our attention to deeper analysis of gene expression differences between speckle signature patient groups in TCGA data. We divided TCGA samples into Signature I and Signature II groups using the top and bottom 25% of sample speckle scores (from Fig 1; Signature I -top 25%, Signature II -bottom 25%), calculated gene expression fold changes, and used Gene Set Enrichment Analysis (GSEA) to identify which biological pathways were altered between the two patient groups. We found striking enrichment of "Oxidative phosphorylation" and "Ribosome" in the Signature I group among all cancer types, including ccRCC (KIRC) ( . Hence, across many cancer types, speckle Signature I correlates with increased oxidative phosphorylation and ribosomal pathways, supporting that speckle Signature I tumors, which reflect the aberrant speckle signature acquired in cancer, may exist in a "hyper-productive" state with enhanced metabolic and protein production capacity. We also performed positional gene set analysis, which consists of gene sets corresponding to human chromosome cytogenetic bands, delineating megabase-sized neighborhoods of heterochromatin or euchromatin on the linear genome 39 . Our analysis revealed astonishing expression biases of certain chromosome cytogenetic bands. For example, genes in the CHR1P31 band were highly enriched in the speckle Signature II patient group, while genes in CHR19Q13 were highly enriched in the speckle Signature I patient group (Fig S4A, right-most plots). We plotted cytogenetic band gene set enrichment statistics on entire chromosomes side-by-side with SON TSA-seq (from Fig 2), finding that Signature I-biased bands corresponded to regions of the genome with higher SON signal (Fig 4B and Fig S4C; e.g. CHR6P21, CHR9Q34, all of Chr22). Reciprocally, Signature II-biased bands corresponded to regions of the genome with lower SON signal (Fig 4B and Fig S4C; e.g. Q arm of Chr6, all of Chr18). These data suggest that speckle signature affects large groups of genes depending on speckle association.
We quantified the per-gene relationship between speckle association and expression in the speckle signature patient groups. We separated all expressed genes into deciles depending on their speckle association and graphed the expression fold change of the speckle Signature I versus II patient groups for each gene (Fig 4C). In agreement with our cytogenetic band enrichment results, this analysis revealed that speckle associated genes were biased towards higher expression in the Signature I patient group (Fig 4C; yellow bin 10, compare to log2 fold change of 0, red line), and the genes that were most speckle-distal were biased towards higher expression in the Signature II patient group (Fig 4C; grey bin 1). These data reveal a propensity of speckle associated genes to be more highly expressed in speckle Signature I samples, suggesting that Signature I speckle components are positive regulators of gene-activating speckle functions (explored further in Fig 5).

Signature I and Signature II-biased HIF-2a target genes reflect distinct functional categories
We next focused on HIF-2a target genes within the speckle signature analysis to elucidate consequences of speckle alterations in ccRCC. Compared to normal tissues, HIF-2a target genes were more highly expressed in both speckle Signature I and Signature II patient groups (Fig 4D; N -normal adjacent tissue samples; Ispeckle Signature I samples; II -speckle Signature II samples). However, certain target genes were more highly expressed in the Signature I patient group (e.g. PHPT1 in Fig 4D, left), while others were more highly expressed in the Signature II patient group (e.g. ASAP1 in Fig 4D, right). These expression biases corresponded to differences in speckle association, shown at representative genes (Fig 4D, top SON TSA-seq tracks), and for all HIF-2a target genes (Fig 4E). In total, 471 of 986 HIF-2a target genes were more highly expressed in the Signature I patient group, and 216 were more highly expressed in the Signature II patient group (Fig 4F). Using Gene Ontology analysis, we found that Signature I-biased HIF-2a target genes were enriched in cell cycle, metabolism, and inflammatory pathways, and Signature II-biased HIF-2a target genes were enriched in endothelium development (angiogenesis) (Fig 4G). Importantly, these functional terms were similar to the terms we found as differential between HIF-2a target genes that did or did not have HIF-2a regulated speckle association, with shared pathways between Signature I-biased genes and speckleassociating genes as well as between Signature II-biased genes and non-speckle-associating target genes. (compare Fig 2K with Fig 4G). These findings indicate that HIF-2a functional programs in ccRCC correlate with speckle signature. We propose that these speckle-relevant HIF-2a target gene expression biases may underly the speckle-based differences in patient outcomes in ccRCC.

Loss of HIF-2a DNA-speckle targeting phenocopies HIF-2a target gene expression biases of the betteroutcome speckle Signature II patient group
Our data thus far indicates that HIF-2a regulates expression of speckle-associating target genes by mediating their speckle association. Meanwhile, our correlational evidence supports that speckles regulate expression of speckle-associating HIF-2a target genes by changing the gene activation functions of nuclear speckles. We hypothesized that HIF-2a target genes differentially expressed between the two speckle signatures were specifically and similarly dependent on the HIF-2a speckle targeting motif, STM. We thus evaluated HIF-2a target gene expression biases between speckle signatures in the context of HIF-2a target genes that were differentially regulated between HIF2As-wtSTM and HIF2As-∆STMs. Indeed, nearly all the HIF-2a target genes that were preferentially expressed in HIF2As-wtSTM were more highly expressed in the speckle Signature I patient group (Fig 4H; orange box, above 0 means higher in Signature I patient group). In contrast, the HIF-2a target genes that were preferentially expressed in HIF2As-∆STMs were more highly expressed in the speckle Signature II patient group (Fig 4H; blue box, below 0 means higher in Signature II patient group); meanwhile, HIF-2a target genes that did not have a wtSTM/∆STMs preference showed a mixture of I and IIbiased expression patterns (Fig 4H, grey). We also separated HIF-2a target genes into quintiles based on speckle Signature I or II gene expression biases and graphed fold change of HIF2As-wtSTM versus HIF2As-∆STMs (Fig 4I). This analysis demonstrated that the most Signature I-biased HIF-2a target genes had the highest expression in the HIF2As-wtSTM condition (Fig 4I; yellow bin 5, above 0 means higher in HIF2As-wtSTM), and the most Signature II-biased HIF-2a target genes had the highest expression in the HIF2As-∆STMs condition (Fig 4I; purple bin 1; below 0 means higher in HIF2As-∆STM). Hence, there was a strong concordance between gene expression differences between speckle signature patient groups and changes observed upon disrupting HIF-2a speckle targeting abilities in tissue culture. Speckle Signature I corresponds to boosted expression of speckle-associated HIF-2a targets and mimics the effects of HIF-2a that is speckletargeting-competent (wtSTM), whereas speckle Signature II corresponds to boosted expression of nonspeckle-associating HIF-2a targets and mimics the effects of HIF-2a that is speckle-targeting-defective (∆STMs) (see models in Fig 3J and Fig 4J). Thus, our results show striking parallels of HIF-2a target gene expression upon differential speckles in human tumors and differential HIF-2a DNA-speckle targeting in tissue culture, supporting a model whereby HIF-2a target gene expression can be similarly regulated by changes in speckles or DNA-speckle association.

The TREX complex is elevated in speckle Signature I patients and TREX complex member ALYREF is required for expression of speckle-associated genes
Our results link speckle Signature I to higher expression of speckle associated genes (Fig 4B-C), leading us to investigate whether variation in the speckle composition of Signature I compared to Signature II might explain gene expression differences. To reveal the protein complexes and functional categories of proteins that reside within speckles and predominate within the speckle signatures, we utilized speckle resident protein annotations from the Human Protein Atlas and physical interactions from STRING-DB to perform network analysis of speckle resident proteins and their interactors. This analysis revealed hubs of interconnected proteins that represented distinct protein complexes, providing a comprehensive view of complexes residing within speckles (Fig 5A; speckle-resident proteins are solid colors, non-speckle-resident interactors are transparent, only connected nodes are displayed). Using STRING-DB to identify the key functional annotations, we classified seven major categories of speckle-resident proteins: RNA metabolism, DNA repair, transcription, chromatin organization, splicing (subdivided into U2 pre-catalytic and U2 catalytic complexes), and the transcriptional export (TREX) complex (Fig 5A). Examining RNA expression differences between Signature I and Signature II speckle patient groups, we found that 8 of 10 speckle-resident TREX complex members were more highly expressed in the Signature I group, while the other functional groups were approximately evenly split between the two signatures (Fig 5B; yellow box). These results highlight TREX as a potentially key differential protein complex that is more highly expressed in the speckle Signature I patient group.
Elevated TREX in speckle Signature I patients led us to hypothesize that the TREX complex may contribute to the gene-activating functions of nuclear speckles. TREX plays a central role in gene expression, accompanying mRNAs from transcription to nuclear export (reviewed in 40 ). Within TREX, the ALYREF protein is a central component that is loaded onto mRNAs co-transcriptionally, promoting pre-mRNA processing and export, and we find that it was elevated in Signature I tumor samples (Fig 5C). Depletion of ALYREF alters gene expression and reduces Poll II gene occupancy 41 , supporting an additional role for ALYREF in transcription efficiency. To investigate a correlation between requirement of ALYREF for speckle-associated expression, we used siRNA to knockdown ALYREF in 786-O ccRCC cells (Fig 5D). We found that the most speckle associated genes (bin 10, yellow) were the most decreased in expression upon ALYREF knockdown, while the least associated genes were the most increased in expression (bin 1, grey) (Fig 5E, ALYREF siRNA  #5; Fig S5A, ALYREF siRNA #8). We also found higher levels of speckle association of genes significantly decreasing in expression upon ALYREF knockdown compared to unchanging genes and significantly higher genes (Fig 5F). This trend existed genome-wide (Fig 5F), and was true of HIF-2a target genes (Fig 5G).
These results indicate that ALYREF is required for expression of speckle-associated genes.
Because ALYREF is more highly expressed in the speckle Signature I patient group (Fig 5C), we hypothesized that knockdown of ALYREF may exhibit similar gene expression biases to the speckle Signature. Consistent with this, we found that genes decreasing with ALYREF knockdown were more highly expressed in the speckle Signature I patient group, while genes increasing with ALYREF knockdown were more highly expressed in the speckle Signature II patient group (Fig 5H). This was also observed for HIF-2a-responsive genes (Fig 5I). Hence, ALYREF knockdown in 786-O ccRCC cells phenocopies gene expression differences of ccRCC speckle signature patient groups.
To further evaluate the relationship between ALYREF and speckle-associated gene expression, we performed smRNA-FISH using DDIT4 HIF-2a target gene probes (as in Fig 3). As a reference point for disrupting DDIT4 expression and speckle association, we compared ALYREF knockdown to HIF-2a knockdown in 786-O cells (Fig S5B shows confirmation of HIF-2a knockdown). We found that ALYREF knockdown partially compromised RNA levels within DDIT4 transcription sites (Fig 5J, compare NTC and ALYkd panels; quantified in Fig 5K), with lower total DDIT4 RNA levels (Fig 5L). ALYREF knockdown did not affect DDIT4 speckle association (Fig 5M). Hence, ALYREF knockdown compromised expression of speckle associated genes without influencing speckle association status (Fig 5N). These results suggest that ALYREF contributes to the ability of nuclear speckles to boost gene expression but is not required for HIF-2a-mediated association of genes with speckles.

Nuclear positioning of the SON speckle marker predicts patient outcomes in ccRCC
Our findings with ALYREF knockdown recapitulated some of the speckle signature group differences in ccRCC tumor samples. We also noted changes in nuclear speckle morphology, including fewer, but larger, speckles with ALYREF knockdown and a slight reduction in the amount of SON signal at the nuclear periphery (Fig  S5C-G). These findings prompted us to evaluate whether ccRCC tumor samples show variation of imagingbased speckle phenotypes. To this end, we obtained a tumor microarray that contained 86 ccRCC formalinfixed paraffin-embedded (FFPE) tumor sections with matched adjacent tissue and associated patient survival data. We performed SON immunofluorescence and scanned the entirety of each tissue section at 20x magnification (Fig 6A, left). Using CellProfiler, we identified nuclei based on DAPI staining and calculated pernucleus measurements of SON texture, intensity, and radial distribution. For radial distribution, the nucleus was divided into four concentric bins and the proportion of immunofluorescence signal residing in each bin was calculated (Fig 6B, right schematic). Based on the median sample value of each measurement, we separated ccRCC tumors into the top and bottom half and performed Kaplan-Meier analysis for each measurement (Fig  6A, right schematic). Plotting Kaplan-Meier statistical measures, we found that SON radial distribution correlated most significantly with survival (Fig 6B, yellow points). Specifically, the most significant predictors of survival were fraction of SON in the most central nucleus bin followed by fraction of SON in the most peripheral nucleus bin (Fig 6B, top two labelled yellow points). These differences were clearly apparent in tumor images: high central SON tumors showed speckles in the center of nuclei (Fig 6C; C11 and I15 tumors), and high peripheral SON tumors showed speckles spread throughout the nuclei (Fig 6C; C9 and E17 tumors). As these are reciprocal measurements, samples with low central and high peripheral SON reflect the same patient group. Tumors with low central (high peripheral) SON were associated with worse patient outcomes compared to tumors with high central (low peripheral) SON (Fig 6D-E). These results reveal that imaging measurements of speckle phenotypes predict outcomes in ccRCC.
We compared ccRCC samples to matched normal adjacent tissues and separated ccRCC based on tumor grade. This showed that normal adjacent tissues have higher central SON (lower peripheral SON) compared to ccRCC, and SON becomes less central (more peripheral) in later grade ccRCC (Fig 6F-G). We evaluated whether altered radial distribution of SON in higher grade tumors strongly influenced survival correlations, performing Kaplan-Meier analysis on the subset of Grade1 and 2 tumors, which had similar SON nuclear distributions. Despite excluding higher grade ccRCC, radial positioning of SON was predictive of ccRCC outcomes (Fig 6H), indicating that nuclear speckle phenotypes predict outcomes even in early-stage ccRCC, providing potential diagnostic relevance.
We next directly assessed both RNA-and imaging-based measurements of speckle phenotype in the same cohort of samples, hypothesizing that samples with lower central SON would correspond to Signature I speckle protein gene expression (see Fig 1). We obtained clinical ccRCC tumor and tumor-adjacent samples and divided them for RNA-seq and FFPE SON immunofluorescence, including three tumor-adjacent primary tubule renal epithelium samples, three primary human ccRCC tumors, and four mouse xenograft-derived ccRCC tumors derived from the same individual. We calculated speckle protein gene expression scores via RNA-seq as previously (Fig 1). As predicted, primary renal tubule epithelial samples (normal adjacent) displayed Signature II RNA speckle scores and high central SON by imaging (Fig 6I; blue points; N -normal primary renal tubule epithelium). Two primary ccRCC tumors also showed Signature II speckle scores with high central SON (Fig 6I; yellow points; T -primary tumor). In contrast, the remaining primary ccRCC tumor and all four patient-derived xenograft samples showed the opposite Signature I speckle scores, with corresponding low central SON, as predicted (Fig 6I; see yellow primary tumor point and red xenograft points (Tx) on upper left portion of graph). These direct comparisons strongly indicate that speckle Signature I manifests as more spread out and less central speckles associated with poor ccRCC survival (e.g. Fig 6C,  bottom), while speckle Signature II manifests as more central larger speckles associated with better ccRCC survival (e. g., Fig 6C, top). Therefore, these data link RNA-seq and imaging-based speckle phenotypes, demonstrating that they may be used interchangeably to predict survival in ccRCC and suggesting that RNAseq estimations of speckle phenotypes reflect bona fide differences in speckle morphologies, adding to potential therapeutic relevance.

DISCUSSION
This study spotlights nuclear speckles as a distinct layer of gene regulation that skews HIF-2a functional programs and is correlated with poor patient outcomes in ccRCC. Our data indicate DNA-speckle targeting by transcription factors is a generalizable mechanism employed by transcription factors p53 and HIF-2a, via their speckle targeting motifs (STMs) defined herein. Identification of putative STMs that reoccur among gene regulatory proteins (Fig 2) suggests that speckle targeting may be a broadly utilized gene-regulatory mechanism. Based on mutagenesis of HIF-2a STMs (Fig 3), analysis of speckle variation in human samples (Fig 4), and manipulation of speckles in tissue culture (Fig 5), we reveal that nuclear speckles function by shifting the types of target genes that are preferentially induced by a transcription factor. We propose that this alteration of HIF-2a functional programs by speckles in ccRCC underlies speckle-based differences in patient survival.
While 23 other cancer types displayed similar speckle variation to ccRCC, none showed such dramatic survival differences (Fig 1). We speculate that the reason other cancer types do not show speckle-based survival differences may be due to more heterologous etiologies of other cancer types compared to ccRCC, which involves hyperactivation of HIF-2a in nearly all clinical cases 24 . Based on our findings that nuclear speckles skew HIF-2a functional programs, we hypothesize that speckle signature may be consequential for cancer subsets depending on specific transcription factor or gene expression dependencies. However, further studies are required to elucidate which specific cancer contexts may combine with nuclear speckle phenotypes to predict patient outcomes. Across most cancer types, the tumor adjacent normal tissue showed a speckle Signature II gene expression pattern (Fig S1E), suggesting that speckle Signature I represents an aberrant speckle state that is acquired in some tumors. The speckle Signature I patient group showed remarkable gene expression enrichment of ribosome and oxidative phosphorylation pathways in each cancer type (Fig 4A and  Fig S4B), suggesting that speckle Signature I could reflect a hyper-productive cell state with enhanced metabolism and protein production capabilities. Given that some existing cancer therapies target protein synthesis pathways (e.g. mTOR inhibitors 42 ), we propose that future studies should investigate whether speckle signature predicts patient responses to particular therapeutic agents, hypothesizing that, across many cancer types, the strong enrichment of ribosome protein gene expression in speckle Signature I tumors renders them more sensitive to mTOR inhibition or other therapies targeting metabolism and protein production.
Each cancer type also showed remarkable biases in chromosome position-based gene set enrichment, where entire groups of genes within a cytogenetically-defined chromosomal band were heavily skewed towards higher expression in the speckle Signature I patient group at speckle-associated chromosome bands, while non-speckle-associated chromosome bands were skewed towards higher expression in the speckle Signature II patient group (Fig 4B and Fig S4A,C). These findings provide compelling evidence for the longstanding model that nuclear speckles represent gene regulatory neighborhoods in higher eukaryotes [43][44][45] . Together, the existence of the speckle signature in many cancer types, its strong correlation with certain functional pathways, and the extreme cytogenetic chromosome segment biases observed in our study all support the model that the speckle signature is a widespread phenomenon that coordinates functional pathways by influencing entire gene neighborhoods based on speckle association status.
We propose that nuclear speckles exist in two states (model in Fig 7). In one state (Fig 7, left panel), defined by Signature I speckle protein gene expression patterns, speckles are distributed throughout the nucleus (Fig 6), members of the TREX complex are more highly expressed (Fig 5), and expression of speckleassociated genes is elevated (Fig 4) (see model in Fig 7). In the other state, defined by Signature II speckle protein gene expression patterns that are more similar to normal tissues, speckles are located more centrally within the nucleus (Fig 6), members of the TREX complex exhibit lower expression (Fig 5), and non-speckleassociated gene expression is elevated (Fig 4). Reduction of the TREX complex member, ALYREF, suppressed expression of speckle associated genes, and shifted cells towards a more Signature II-like gene expression phenotype (Fig 5). Despite the observed decrease in speckle-associated gene expression with ALYREF knockdown, speckle association of the DDIT4 gene was not perturbed (Fig 5), indicating that this disruption of speckle associated gene expression was due to altered speckle functions rather than an inability of genes to be speckle localized. This is a critical finding because it indicates that speckle phenotypes and DNA-speckle association are separate gene regulatory strategies used by cells.
While the speckle signature and ALYREF knockdown represent changes that broadly impact speckles, our experiments pinpointing the speckle targeting portions of HIF-2a provide a specific test for how speckles impact gene expression of associating DNA. Remarkably, the effects on HIF-2a target gene expression were paralleled by disruptions of HIF-2a speckle targeting abilities and differential ccRCC patient speckle signature (Fig 4H-I), supporting that disruption of DNA-speckle contacts or changes in speckles can shift the types of genes that are preferentially activated by a transcription factor. Speckle-associating and speckle Signature Ibiased HIF-2a target genes were enriched in metabolism, inflammation, and cell cycle Gene Ontology categories, whereas non-speckle-associating and speckle Signature II-biased HIF-2a target genes were enriched in angiogenesis-related pathways (Fig 2K and Fig 4G). This functional shift of HIF-2a gene expression programs likely explain the striking survival differences based on speckle signature in ccRCC. However, further research is required to fully understand how these functional pathways link to different patient outcomes. Importantly, in ccRCC, several systemic therapies target HIF-2a and its downstream pathways 46 . Because our data suggests differential roles for HIF-2a depending on speckle signature, we speculate that speckles may be useful predictors for which therapies will be effective in particular ccRCC tumors.
Our identification of putative STMs within hundreds of gene regulatory proteins provide a platform for future investigations into the extent that DNA-speckle targeting is utilized by different factors. STMs occur in both direct DNA-binding transcription factors, as well as in many co-activators that engage with the genome by binding histones or DNA-binding proteins. Among these were several co-activators involved in enhancer functions, such as Histone K27 acetyltransferases, members of the Mediator complex, and Histone K4 methyltransferases. Given previous reports that enhancers and super-enhancers are highly enriched among speckle associated chromatin 14,15 , one hypothesis is that these STM-containing proteins work together to both mediate DNA-speckle association and promote deposition of activating histone modifications, creating a potent gene-activating environment where activating histone modifications are coordinated in nuclear space with the high concentration of RNA binding and processing factors found in nuclear speckles. Further research will be required in this area to disentangle enzymatic from speckle targeting functions of these co-activators, as well as to gain a full understanding of how histones, transcription, RNA processing, RNA stability, and RNA nuclear export may be coordinated at speckles to boost gene expression.
At present, the mechanisms used by STMs to accomplish their functions are not known. In one model, STMs increase DNA affinity with nuclear speckles via interacting with a particular speckle-resident protein.
Alternatively, STMs may assist in trafficking DNA from other nuclear locations to speckles. Although a handful of speckle resident proteins contain STMs, we envision that speckle residency and speckle targeting are distinct molecular functions accomplished by different types of protein sequences. Other studies have identified protein components required for speckle residency, including a TP region of SF3B and the phosphorylated SR tracks of SR speckle resident proteins [47][48][49] . We note that these speckle residency protein domains do not overlap with the speckle targeting motifs identified in this study. While further efforts will be required to disentangle speckle targeting from speckle residency, we propose that STMs within speckle resident proteins may assist their speckle targeting, while other molecular features dictate speckle residency. Overall, our identification of STMs within both HIF-2a and p53 provides a molecular handle for future detailed inquiry of the molecular mechanisms underlying STMs and how they drive DNA-speckle association.
Overall, our findings illuminate nuclear speckles as underexplored regulators of gene expression that are consequential for human disease. We anticipate that our study of HIF-2a and ccRCC represents just one example, and that as nuclear speckles and DNA-speckle association by transcription factors become more broadly studied, additional examples of critical speckle functions in human health and disease will become evident. Figure 1. A recurring speckle signature predicts patient outcomes in ccRCC.

FIGURE LEGENDS
A) Schematic showing generation of multi-cancer speckle signature. Proteins residing within speckles were identified (i), their expression evaluated (ii), contributions to patient variation compared (iii; Pearson's correlations), and consistent speckle protein gene contributors to patient variation were identified (iv). See also Fig S1A.  B) Heatmaps showing z-scores of speckle signature speckle protein gene RNA expression in melanoma (SKCM), breast cancer (BRCA), and renal cell carcinoma (KIRC). Bar on left represents speckle scores. Bar above represents Signature I-(blue) or Signature II-(pink) high speckle protein genes. C) Kaplan Meier plots separating cancer cohorts by the top and bottom 25% of speckle scores. See also Fig S1B and Table S1. ** -p < 0.01; *** -p < 0.001; **** -p < 0.0001, ns -not significant For full list of STM-containing proteins and STM sequences, see Table S2 and Table S3.    Fig S5A) split into speckle association deciles. F) SON TSA-seq signal of genes that decrease upon ALYREF knockdown (orange, "down"), are unchanged (grey, "ns"), and increase (blue, "up"). G) Same as F for HIF-2a target genes H) Signature I to Signature II expression ratio in the KIRC TCGA cohort for genes that decrease upon ALYREF knockdown (orange, "down"), are unchanged (grey, "ns"), and increase (blue, "up"). Significance calculated using Wilcoxon tests. * -p < 0.05; ** -p < 0.01; *** -p < 0.001; **** -p < 0.0001, ns -not significant SON TSA-seq data in E, F, and G is from Fig 2 786-O DMSO condition.  Our data supports that speckles exist in two states in cancer, with differential induction of speckle associated genes, speckle size and nuclear distribution, and speckle content. In ccRCC, speckle state correlates with patient survival and differential expression of HIF-2a functional pathways. Our findings indicate that speckle associated gene expression can be manipulated by changes in either speckles or transcription factor-driven DNA-speckle targeting. Figure S1. Supplement to Figure 1. A) Schematic showing how speckle protein genes with high contributions to patient variation were selected based on Principal Component 1 (PC1) weights. B) Kaplan Meier disease-specific (DSS) or overall survival (OS) analysis for ovarian cancer (OV, top) and ccRCC (KIRC, bottom) splitting individuals based on the top/bottom quartiles of speckle score (left) or on positive versus negative speckle score (right). C) Tumor AJCC stage frequency of ccRCC (TCGA KIRC cohort) tumors based on speckle scores split into quartiles. First quartile are the most Signature II tumors; fourth quartile are the most Signature I tumors. D) Kaplan Meier plots for early (left) and late (right) stage ccRCC (TCGA KIRC cohort). Patients were split into Signature I (positive speckle score) and Signature II (negative speckle score). E) Speckle scores of tumor adjacent normal tissues. Speckle scores are calculated independently for each tissue/cancer type.

SUPPLEMENTAL FIGURE LEGENDS
See Table S1 for survival statistics for each cancer and Github (https://github.com/katealexander/speckleSignature.git) for detailed instructions, scripts used, and additional files generated relating to speckle signature calculations and survival analysis.

Speckle signature and TCGA survival analysis
We took 446 protein genes annotated as "Enhanced", "Supported", or "Approved" for subcellular localization within nuclear speckles in The Human Protein Atlas and extracted their upper-quartile normalized RNA expression from the 30 PanCan TCGA projects that had greater than 50 samples. We performed Principal Component Analysis on these 446 speckle protein genes. In doing so, each speckle protein gene was assigned a weight (called rotation in the analysis) that was used in the analysis to separate tumor sample along the first Principal Component (PC1). The absolute value of a speckle protein gene PC1 weight thus estimates the contribution of each speckle protein gene to patient variation and the PC1 weight sign, positive or negative, reflects genes that have opposite expression patterns to one another (for example, in Fig 1B, "Sig-I high" and "Sig-II high" have, by definition, opposite PC1 weight signs and heatmaps show opposite expression patterns between the two groups). To compare speckle protein gene expression contributions to patient variation between cancer types, we took the pairwise Pearson's correlation coefficients of the speckle protein PC1 weights (an example single pairwise comparison is shown in Fig S1A, right scatterplot; the correlation coefficients of each cancer-to-cancer pairwise comparison are summarized in Fig 1Aiii heatmap).
To obtain a set of speckle protein genes that consistently contributed to patient variation in many cancer types, we flipped the rotation signs so that the speckle protein gene, SON, was always assigned a negative weight. We extracted the speckle protein genes that had consistently signed rotations across 22 cancer types (the 22 cancer types in Fig 1Aiii at the bottom of the heatmap that showed highly similar speckle protein gene PC1 weights to one another) and used these 117 speckle protein genes to calculate speckle scores (40 Signature Ihigh speckle protein genes; 77 Signature II-high speckle protein genes). Speckle scores were calculated by taking the z-scores of speckle protein gene expression, calculated per cancer, and applying the following formula: sum((z-score Sig I speckle protein gene)*1/(number Sig I speckle protein genes)) + sum((z-score Sig II speckle protein gene)*-1/(number Sig II speckle protein genes). In this manner a speckle score was assigned to sample so that it would be strongly positive for tumors with the strongest Signature I expression pattern and strongly negative for tumors with the strongest Signature II expression pattern. Speckle score was then used to separate samples into groups for Kaplan Meier and gene expression analysis between the two groups. For detailed instructions, scripts, and output files for the RNA-based speckle signature, see the speckle signature GitHub page: https://github.com/katealexander/speckleSignature.git.

Speckle targeting motif
To identify speckle targeting motifs across the proteome, we first used Motif Search to download all the motifs in the human proteome that followed this rule: x(30)-[TSED]-P-x(30), meaning any 30 amino acids followed by a threonine, serine, glutamine, or asparagine, followed by a proline, followed by any 30 amino acids. From these motifs, we applied the following rules to identify speckle targeting motifs: 1) no more than 4 prolines in a row, 2) at least 3 spaced prolines of 7 possibilities spaced every 5 amino acids from the central [TSED]P (Python indexes 11,16,21,26,36,41,46), 3) at least 5 of TSED on the C-terminal side of the central [TSED]P, 4) at least 7 of AMVFLIG on the N-terminal side of the central [TSED]P, 5) fewer than 16 total RHK amino acids (summarized in Fig S2D). We note that these rules are for the purpose of identifying new putative speckle targeting motifs and may be modified as additional information about the required biochemistries of speckle targeting becomes available in future studies. Detailed instructions can be found on the speckle targeting motif GitHub page: https://github.com/katealexander/speckleTargetingMotif.git. Lists of the identified STM-containing proteins (Table S2), and their STM sequences (

SON TSA-seq and SON Cut&Run analysis
SON TSA-seq or SON Cut&Run reads were aligned to the human reference genome assembly GRC37/hg19 using Bowtie2 allowing for a maximum fragment size of 1000 base pairs 52 . PCR duplicates were removed using Picard, and SON signal was quantified over sliding windows of 50 or 100kb slid by 1/10 th of the window size using DiffBind with input (TSA-seq) or IgG (Cut&Run) subtraction 53 . Differential 50kb and 100kb windows identified from DiffBind were merged using BEDTools 54 , resulting in differential domains. Genes within differential domains were then extracted using Python. Genes within decreasing domains that had an adjusted p-value of less than 0.01 in either SON TSA-seq or SON Cut&Run were considered to have speckle association regulated by HIF-2a; genes were considered unchanging if they were in domains with an adjusted p-value of greater than 0.1 by both methods. For more detailed analysis instructions and associated scripts, see Github: https://github.com/katealexander/TSAseq-Alexander2020/tree/master/genomicBins_DiffBind.

immunoDNA-FISH
Oligopaint DNA-FISH probes for genes were designed across a 50kb region centered on the transcription start site as described 55 . ImmunoDNA-FISH was performed first with the standard immunofluorescence protocol, followed by 10 minutes in 4% paraformaldehyde to fix the secondary antibody prior to DNA-FISH. Specifically, cells were fixed in 4% paraformaldehyde for 10 minutes, permeabilized using 0.1% Triton X-100 in PBS for 10 minutes, incubated with 1:200 SON antibody diluted in PBS overnight at 4°C, and incubated in 1:200 antirabbit A488 secondary for 1 hour at room temperature, with 3x5 minute washes in PBS after fixation and antibody steps. After fixation and PBS washes, DNA-FISH was performed as described 56 .
immunoRNA-FISH RNA-FISH probes were designed using Stellaris probe design software, inputting DDIT4 exons (for exonic probe set) and introns (for intronic probe set). Immunofluorescence was performed as in immunoDNA-FISH, with the exception that primary and secondary antibodies were used at a 1:1000 dilution. After immunofluorescence, cells were incubated overnight in 70% ethanol at 4°C. RNA-FISH was then performed as previously described 57 . DDIT4 exonic probes were labelled with Quasar 570, DDIT4 intronic probes were labelled with CAL Fluor Red 610, and a mouse anti-rabbit A488-labelled secondary was used.

RNA-seq
For RNA-seq of 786-O cells, cells were treated with DMSO for 12 hours or with 2µM of PT2399 for 3, 6, 9, or 12 hours. Cells were lysed in TRIzol (15596018, Thermo Fisher Scientific) and snap frozen. RNA was then isolated using chloroform extraction, followed by QIAGEN RNeasy Mini Kit isolation (#74106), including DNA digestion with DNase. Poly(A) + RNA was then isolated using double selection with poly-dT bead (E7490, NEB), and RNA-seq libraries were prepared using a NEB-Next Ultra II Directional Library Prep Kit for Illumina (E7760, NEB). Library sizes were determined on a Bioanalyzer, and concentrations determined using NEBNext Lib Quant Kit (E7630, NEB). Libraries were sequenced on an Illumina NextSeq 550, using pairedend sequencing of 42 bases per read. For RNA-seq of ccRCC and normal adjacent tissues, the above protocol was applied to flash frozen tissue samples with the added step of manually homogenizing the tissue during the TRIzol step prior to chloroform extraction.

Definitions of HIF-2a target genes
Because exact HIF-2a target genes are known to be context dependent 58 , our working definition of what constitutes a HIF-2a target gene varied depending on the analysis/cell type. For 786-O cells (Fig 2H-K), 874 genes decreasing with PT2399 at an adjusted p-value of less than 0.1 at 3, 6, 9, or 12 hours were used. For MCF7 cells (Fig 3H-I and Fig 4H-I), 1169 genes increasing with dox-inducible HIF2As-wtSTM or HIF2As-∆STMs at an adjusted p-value of less than 0.1 were used. For analysis of TCGA tumor samples (Fig 4D-G), we combined the 786-O HIF-2a targets, MCF7 HIF-2a targets, and a list of published HIF-2a targets 59 and applied the additional filter of having increased expression in either tumor patient group compared to tissueadjacent normal samples (p-value < 0.05, log2 fold change > 0.3), for 986 genes.