Integrative copy number analysis of uveal melanoma reveals novel candidate genes involved in tumorigenesis including a tumor suppressor role for PHF10/ BAF45a

Abstract

Purpose:

Uveal melanoma (UM) is a primary malignancy of the eye with oncogenic mutations in GNAQ, GNA11, or CYSLTR2, and additional mutations in BAP1 (usually associated with LOH of Chr 3), SF3B1 or EIF1AX. There are other characteristic chromosomal alterations, but their significance is not clear.

Experimental Design:

To investigate genes driving chromosomal alterations we integrated copy number, transcriptome and mutation data from three cohorts and followed up key findings.

Results:

We observed significant enrichment of transcripts on Chromosomes 1p, 3, 6, 8 and 16q and identified seven shared focal copy number alterations (FCNAs) on Chr 1p36, 2q37, 3, 6q25, 6q27 and 8q24. Integrated analyses revealed clusters of genes in focal copy number regions whose expression was associated with metastasis and worse overall survival. This included genes from Chr 1p36, 3p21, and 8q24.3. At Chr 6q27 we identified two tumors with homozygous deletion of PHF10/BAF45a and one with a frameshift mutation with concomitant loss of the wild type allele. Down-regulation of PHF10 in UM cell lines and tumors altered a number of biological pathways including development and adhesion. These findings provide support for a role for PHF10 as a novel tumor suppressor at Chr 6q27.

Conclusions:

Integration of copy number, transcriptome and mutation data revealed novel candidate genes playing a role in UM pathogenesis and a potential tumor suppressor role for PHF10.

Introduction

Uveal melanoma (UM) is the most common primary intraocular malignant tumor diagnosed in approximately six cases per million per year. Approximately 40% of patients develop metastatic melanoma to the liver within ten years (1). Advances made in the local methods of treatment of primary uveal melanoma have not led to an improvement in survival and after metastasis there is a median survival time of less than six months (2). UMs display characteristic genomic signatures including recurrent chromosomal aberrations, gene mutations and gene expression profiles (GEP), some of which are predictive of metastatic risk (3,4). Mutations in GNAQ, GNA11, CYSLTR2 and PLCB4 that constitutively activate Gαq signaling are seen in almost all tumors in a mutually exclusive manner (5–8). Inactivating mutations in BAP1 at Chromosome 3p21 (9) are found in tumors with loss of heterozygosity for Chr 3 (LOH3) and a GEP predictive of metastatic risk (class 2 tumors) (10), and mutations in SF3B1 or EIF1AX are found in tumors with disomy 3 and a GEP associated with an intermediate and low likelihood of metastasis respectively (class 1 tumors) (4,11,12). Whole genome sequencing has also revealed potential mutations in other genes (13).

In addition to LOH3, UMs are characterized by chromosomal alterations that can include Chr 1p loss, 6p gain, 6q loss, 8p loss, 8q gain, and 16q loss (14–20). These DNA copy number aberrations (CNAs) are an important category of genetic alterations which can lead to development and progression of cancers by affecting gene dosage, sometimes amplifying genes conferring a proliferative or metastatic advantage or leading to loss of function of tumor suppressor genes. However, the significance of CNAs in most cancers, including in UM is poorly understood.

A global characterization of copy number, transcriptome, mutation, and methylation status in UM has been described (20). This confirmed much previous work and showed four distinct molecular subtypes of CNA, each associated with a varying degree of metastatic risk. Here, we describe copy number analysis of 182 tumor samples from three different cohorts followed by integration of transcriptomic, methylation and mutation data to look for genes driving copy number loss or gain, extending earlier studies investigating the relationship with chromosomal alterations and gene expression in UM with microarrays (21). We describe candidate genes on chromosomes 1p, 8q and 6q and provide evidence for a tumor suppressor role for PHF10 mapping to Chr 6q27. Functional studies provided insights into the consequences of loss of PHF10.

Materials and Methods

Data source and analysis

Data from 182 primary UM tumor samples were obtained from three different cohorts: Washington University (WU), Cleveland Clinic (CC) and TCGA. All of the samples were enucleated specimens obtained from adult patients after they had provided written informed consent. Studies were conducted in accordance with recognized ethical guidelines (e.g., Declaration of Helsinki, CIOMS, Belmont Report, U.S. Common Rule) and were approved by an institutional review board. The data analysis workflow and methods of analysis of SNP arrays (copy number), RNA sequencing and exome sequencing are described in Supplementary Figure 1.

PHF10 knock-down and RNA sequencing

Established cell lines derived from primary uveal melanomas (Mel202 (22), 92–1 (23) and Mel290 (24) (described in Supplementary Table 1) were subjected to PHF10 knockdown with siRNA (Santa-cruz, sc-95343) or a siRNA control that consisted of scrambled sequence/non-silencing siRNA (Santa-cruz, sc-37007) (SiCtrl). RNA seq was performed with routine methods. In the case of Mel202 and 92–1, RNA was sent to the Centre Nacional d’Anàlisi Genòmica (CNAG, Barcelona, Spain) for processing. In the case of Mel290 (25), RNA was sent to Genewiz (New Jersey, USA; www.genewiz.com). Raw counts were analyzed with DeSeq2 (26). RNA-Seq data from cell lines are available upon request.

Migration, Invasion and Adhesion assays:

To assay for migration, cells were transfected in 6 cm dishes. After 48 hours of incubation they were trypsinized and subjected to an assay for chemotaxis as described elsewhere (27). Collagen invasion (ECM552 Milipore) and adhesion assays (ECM545 Milipore) were performed according to the manufacturer’s instructions.

Western blot analysis:

This was performed with standard approaches by querying a set of proteins identified following PHF10 knockdown in cell lines. Antibodies were: PHF10 (ab154637 Abcam), JAZF1 (ab80329 Abcam), SMAD2 (ab40855 Abcam) and PCGF3/5 (ab201510 Abcam).

Statistical analysis

Three replicates per cell line were used for PHF10 knock downs. Intergroup differences were assessed with a one-way ANOVA with Bonferroni’s post-hoc test. Results were expressed as mean and standard deviation or in boxplots. The K-M method and log-rank test were used to compare the survival plots and a univariate Cox proportion hazard model was used to compare the effects of high and low gene expression levels of select genes on overall survival. Time was computed for days. Survival was defined as the elapsed interval until date of last follow-up or date of death. An effect was considered significant at an FDR q-value less than 0.05.

Results

Broad copy number alterations in UM

Genomic DNA samples from 182 primary enucleated UMs (87 with matched normal DNA) from three different cohorts (Supplementary Table 2A) were profiled on three different high resolution SNP platforms encompassing >700K probes with a median inter-probe spacing of ~24 to ~18Kb. The analysis pipeline is described in Supplementary Figure 1 and revealed a high-degree of non-random CNAs across samples from all cohorts. Twelve broad events affected >10% of all samples. This included Chr 1q gain, 6p gain, 8p gain, 8q gain, 21q gain and Chr 1p loss, 3p loss, 3q loss, 6q loss, 8p loss, 9p loss and 16q loss (Supplementary Table 2B, Supplementary Figure 2A). Chr 8q gain and LOH3 were detected in more than 50% of all samples, consistent with previous studies (15,28). Ploidy estimates revealed an average of 2.4n in 103 tumors from the WU and CC cohorts (Supplementary Figure 2C) due to a minority of tumors being tetraploid which was more common in class 2 versus class 1 tumors consistent with BAP1 deficiency being the prognostic predictor for patients with polyploid tumors (29).

Unsupervised hierarchical clustering of all UM samples (N=182) revealed four distinct clusters similar to those described elsewhere (20) (Supplementary Figure 3): Group A and Group B correspond to those classified as class 1A and 1B tumors respectively and Group C and Group D correspond to those classified as class 2 tumors. Groups C and D were associated with monosomy 3 and were highly significantly associated with metastatic outcome (P value = 1.44e-08). The monosomy 3 tumors from both groups C & D showed gain of the entire 8q arm in nearly all samples as reported previously in TCGA cohort (20). Patients with monosomy 3, 8p loss, and 8q gain had lower overall survival (OS) as described elsewhere (30–35). When examined independently, chromosomal alterations associated with metastatic disease were loss of 3p (P value = 2.29 X 10⁻⁸), loss of 3q (P value = 2.95 X 10⁻⁷), loss of 6q (P value = 0.0246), gain of 8q (P value = 0.0001) and loss of 8p (P value = 6.29X 10⁻⁵). Chr 8p loss was more common in group C and Chr 8p gain was more common in group D. We identified 77 transcripts from Chr 8p that could explain this difference (Supplementary Figure 4).

Transcriptomic changes in regions with copy number alteration:

We integrated global copy number and transcriptomic data from TCGA and CC cohorts to identify transcripts driving the copy number alterations. We first performed a global correlation analysis between DNA copy number and transcript levels in the 80 tumor samples from TCGA cohort and the 57 tumor samples from the CC cohort. Results are shown in Supplementary Table 3. There were 922 transcripts whose expression levels were correlated with copy number changes at FDR P < 0.05 in both cohorts or within TCGA alone where data for the CC cohort was missing (an additional 1218 transcripts). Testing for enrichment of genes by computing the overlaps between our list and the Cytogenic set (C1) from MSigDB (36) revealed significant enrichment of genes/transcripts on Chrs 1p, 3, 6, 8 and 16q (Table 1A). Loss of transcripts from the cytoband in which BAP1 is located (3p21) was significant in both TCGA and CC cohorts and BAP1 showed strong correlation between copy number change and expression in both cohorts (Corr. Coeff. = 0.77, FDR P=8 X 10⁻¹⁵ and Corr. Coeff. = 0.53, FDR P = 0.001 in TCGA and CC respectively).

Table 1. A. Enrichment analysis between copy number and expression at the cytogenetic band level.

Results are presented for TCGA and CC cohorts where correlation coefficients were > 0.5. For each cytogenetic band the total number of genes analysed, and the frequency of the significantly correlated genes is shown, along with the P values and FDR P values. K is the number of genes in the set from MsigDB and k is the number of genes in the intersection of the query set with a set from MsigDB. The ratio of k/K in each cohort (CC and TCGA) is also provided.

Cytogenetic band	# Genes in Gene Set (K)	TCGA k/K	TCGA FDR P value	CC k/K	CC FDR q value
Chr 16q22	168	0.3155	2.29 X 10−31	ns	ns
Chr 1p34	195	0.3487	1.59 X 10−43	ns	ns
Chr 1p35	116	0.3276	3.39 X 10−23	ns	ns
Chr 1p36	504	0.244	5.51 X 10−59	0.0496	6.97 X 10−10
Chr 3p21	271	0.3542	8.61 X 10−62	0.1218	2.08 X 10−24
Chr 3p22	ns	ns	ns	0.1047	3.73 X 10−6
Chr 3p25	101	0.3564	2.29 X 10−23	0.1584	1.58 X 10−13
Chr 3q21	130	0.2615	2.36 X 10−17	0.1308	4.22 X 10−13
Chr 6p21	544	0.2059	6.75 X 10−46	0.0551	4.52 X 10−13
Chr 8p21	ns	ns	ns	0.0952	2.03 X 10−6
Chr 8p22	ns	ns	ns	0.186	2.37 X 10−7
Chr 8q13	ns	ns	ns	0.125	2.65 X 10−5
Chr 8q22	102	0.3137	6.55 X 10−19		6.97 X 10−10
Chr 8q24	214	0.3551	3.57 X 10−49	0.1075	1.09 X 10−15
B. High confidence overlapping GISTIC focal peaks in the of uveal melanoma cohorts, genes in minimal overlapping regions, and overlap with a Pan Cancer study.
Copy Number event	GISTIC focal peak	Overlapping cohorts	Minimum Overlapping region (MOR)	Genes in MOR	Genes reported in Pan Cancer study (64)
Amplification	8q22.1	WU & TCGA	chr8:94721058-94749641	RBM12B*, FAM92A1
Amplification	8q24.3	WU & CC	chr8:144992,452-144993,067	PLEC	PARP10, CYC1
Deletion	1p36.11	CC & TCGA	chr1:26387424-26491354	PDIK1L, TRIM63, FAM110D	SFN (lies nearby)
Deletion	2q37.3	CC & TCGA	chr2:241835544-241943252	SNED1, LOC200772, CROCC2, C2orf54	ING5 (lies nearby)
Deletion	3p11.1-q11.1	CC & WU	Chr3:89516653- 93735022	Centromere, EPHA3, PROS1, U3, ARL13B, STX19, U6
Deletion	6q25.2	CC & TCGA	chr6:154470218-154726987	IPCEF1, CNKSR3, OPRM1
Deletion	6q27	WU & CC	chr6:169652570-170152990	THBS2, WDR27, C6orf120, PHF10, TCTE3, C6orf7

Focal copy number alterations (FCNAs) in UM:

The significance of most large-scale chromosomal changes in cancer is not known. Some are thought to reflect altered dosage of a set of genes (for example, loss of genes for lipid biosynthesis on Chr 8p in breast cancer (37)). However, loss or gain of some chromosomes or chromosomal arms can reflect loss or gain of an underlying tumor suppressor or oncogene respectively, as with LOH3 and BAP1 (10). To search for genes driving regions of gain or loss in UM we looked for regions of common focal amplification or deletion and then asked which of the genes whose expression was correlated with CNA, resided in these regions. We then asked if any of the genes in regions with focal alterations also harbored deleterious mutations in any UMs since this could be further evidence that the correct gene had been identified.

Copy number data from the 182 UM primary cancer specimens were analyzed with GISTIC (38). By employing a q value threshold of 0.1, 246 independent regions of significantly recurrent somatic FCNAs were identified. These included 130 amplifications and 116 deletions. On average, after common germline variants were removed, there were 21 focal alterations per sample. Among these 246 regions, the CNA boundaries for amplification events revealed a median size of 46.81 kb (0.6–7614 kb) and a median size of 259 kb (24–17328 kb) for deletions. For each of these significant FCNAs, a “peak” region lying within a 99% confidence window was then identified that was most likely to contain the putative locus involved in tumorigenesis. These peaks each contained a median of 2 genes (range 0–176) and included microRNAs and other non-coding RNAs). Thirteen regions contained more than 25 genes each and the remaining 233 regions encompassed a total of 677 potential target genes (Supplementary Table 4, Figure 1). In addition, there were 217 copy number and gene expression correlated transcripts present within one or more reported GISTIC peaks across the three cohorts (Supplementary table 3).

Figure 1.

Results of GISTIC analysis of copy number data from all three cohorts (TCGA, WU and CC). TCGA data included matched tumor/normal tissue for all samples and germline CNVs could be filtered out in all cases. In the case of the WU and CC cohorts where matched normal DNA was not always available, the germline CNAs were filtered by adopting a platform independent compiled list of genomic variants from the Database of Genomic Variants (DGV) and HapMap normals. Focal copy number alterations (FCNAs) occurring at a significantly higher frequency in UMs compared to the average background rate are shown. Overlapping peaks shared between any two cohorts are indicated. Red indicates gains and blue indicates losses.

Overlapping focal changes in independent UM cohorts:

Since focal alterations can be spurious and unrelated to tumorigenesis, peaks from different cohorts that overlap provide further evidence that a biologically relevant interval has been identified. Globally we identified seven regions with focal alterations that overlapped in two of the three cohorts (Table 2). Six of these regions mapped to chromosomal regions with copy number change already implicated in UM: Chr 3, Chr 1p36, Chr 6q, and Chr 8q. These are discussed below.

Table 2:

Copy number driven gene clusters significantly associated with overall survival and metastasis in UM

Focal loci	# genes	Genes associated with survival
8q24	13	COL22A1, ADCK5, ARC, C8orf33, CHRAC1, COMMD5, CYC1, DENND3, GPR20, MAFA, PTK2, PTP4A3, ZNF7
1p36	11	CNKSR1, EXTL1, NIPAL3, PAQR7, RCAN3, RHCE, UBXN11, WDTC1, HSPG2, PLA2G2C, IFFO2
3p21	9	BAP1, DNAH1, ITIH4, NISCH, SFMBT1, DUSP7, RPL29, PTPN23, SCAP
2q37	9	LIMS2, AGFG1, MFF, EIF4E2, GBX2, COPS8, HDAC4, HES6, ILKAP, KLHL30, PASK, SCLY
8q22	6	CDH17, FLJ46284, GEM, KIAA1429, PDP1, TMEM67
11p15	4	LRRC56, PNPLA2, RASSF7, SIRT3
6q23	4	RPS12, SLC2A12, HEBP2, NHSL1
6q24	4	CCDC28A, REPS1, EPM2A, SASH1
16q24	3	ZFPM1, MVD, ZNF778
2p22	3	DPY30, SLC30A6, SPAST
3p25	3	CAND2, RPL32, VGLL4
2q36	2	AGFG1, MFF
8q23	2	OXR1, ZFPM2
11q13	1	PC
14q21	1	LRR1
17p12	1	PMP22
19p13	1	PCSK4
1q25	1	TOR3A
20p13	1	SPEF1
20q13	1	ZNF512B
2p23	1	MEMO1
2q14	1	LIMS2
3q11	1	PROS1
3q28	1	CCDC50
4q24	1	CISD2
5p15	1	CCT5
6p21	1	FKBP5
6p22	1	RNF144B
6p25	1	SERPINB9
6q22	1	C6orf58
6q25	1	MTHFD1L
7q36	1	EZH2
8p22	1	MTUS1
8q11	1	ATP6V1H
8q13	1	SLCO5A1
8q21	1	RUNX1T1

An overlapping region of deletion at Chr 1p36.11 (chr1:26387424–26491354) deleted a segment of TRIM63, the entire PDIK1L gene and a segment of FAM110D. Of these genes, TRIM63 showed the strongest correlation between deletion and expression in TCGA cohort (correlation coeff. = 0.64, FDR P = 3.15X 10⁹. Supplementary Table 3). However, loss of expression of 1p is historically associated with poorer survival and some transcripts outside this region were strongly correlated with this (see below).

Other regions of overlapping deletion were: Chr 2q37 (chr2:241835544–241943252) detected in 8% of tumors. This region harbors SNED1, LOC200772, CROCC2 and C2orf54 but they did not exhibit correlation with between expression and copy number change. Chr 2q27 was also deleted in a Pan Cancer study (38) (ING5) and harbors HTR2B whose upregulation discriminates class 2 from class 1 tumors (39). We also identified an overlapping deletion in CC and WU cohorts around the Chr 3 centromere (chr3:89516653–93735022); and an overlapping deletion on Chr 6q25 (chr6:154470218–154726987) which harbors CNKSR3 (40).

At Chr 6q27 we identified an overlapping deletion of a region harboring WDR27, c6orf120, PHF10 and TCTE3 (chr6:169652569–170152990). In the case of WDR27 and PHF10, expression was correlated with deletion in both TCGA and CC cohorts (Supplementary Table 3). Further inspection of this region revealed that UM patient MM016 (WU cohort) harbored a homozygous deletion of this region that resulted from a complex deletion of one Chr 6q homolog and a second deletion that spanned 54kb at Chr6:170099399–170152990. This tumor had been diagnosed in a young woman of 24 years and it had very quickly metastasized. It had a differentiated histology (spindle), was very large (24 × 22 mm), and was disomic for Chr 3. It had the following chromosomal changes: +1q, +6p, −6q, +8, +11p, −11q, +22. MM016 did not harbor a mutation in any previously identified prognostic driver (BAP1, EIF1AX, SF3B1) although it did harbor an oncogenic mutation in GNAQ and had upregulated PRAME (not shown) consistent with its metastatic features and its classification as a class IB tumor. Homozygous deletions are rare and can point to the locale of tumor suppressors. Further inspection revealed a second homozygous deletion spanning this region (Chr 6:170073603–170123579 (hg19)) in a tumor of a 91-year-old female patient from the CC cohort (GSM1082739). The overlapping regions of homozygous deletion in both of these tumors disrupted PHF10, C6orf120 and WDR27 (Figure 2). Analysis of exome data also revealed a somatic frameshift mutation in PHF10 in a class 2 tumor MM133 (c.678delT, p.F226fs) (Supplementary Table 5) that was validated with Sanger sequencing (Supplementary Figure 5). Read counts revealed loss of the wild type allele consistent with a potential role for PHF10 as a tumor suppressor. Tumor MM133 arose in a patient of 54 years of age. It was medium-large (19 × 15 mm) of undifferentiated histology (epithelioid), no metastasis at enucleation and lacked a detectable BAP1 mutation. We also identified a deleterious missense mutation (p.D453E, rs761295711) in PHF10 in TCGA tumor A8KM (Supplementary Table 5).

Figure 2. Region of homozygous deletion at Chr 6q27 in the CC and WU cohorts.

The region of overlap that harbors PHF10, WDR27 and c6orf120 is shown. The location of the frameshift mutation identified in MM133 (a class 2 tumor) is also shown. The CC sample with the PHF10 HD was GSM1082739. The WU cohort sample with a PHF10 HD was MM016, a class 2 tumor in a female who was diagnosed with UM in her 20s, did not harbor a detectable mutation in BAP1, SF3B1 or EIF1AX and who died from metastasis.

There were two regions of copy number gain on Chr 8q. A region at Chr 8q22.1 was detected in WU and TCGA cohorts (chr8: 94721058–94749641) and harbors RBM12B. The second region mapped to Chr 8q24.3 (chr8:144992452–144993067) but only harboured a segment of PLEC whose expression was only weakly correlated with copy number gain in TCGA (Supplementary Table 3). However, a number of nearby transcripts from Chr 8q24.3 exhibited stronger correlation between copy number gain and expression and were also correlated with overall survival (discussed further below)

Identification of additional candidate genes driven by integrating copy number, exome sequencing and CpG methylation:

Methylated CpG sites correlated with gene expression in both TCGA and CC cohorts are shown in Supplementary Table 6. There were 1984 transcripts with a significant correlation with hypo- or hyper methylated CpG sites. Expression of 206 of these transcripts was also correlated with chromosomal copy number levels. These were primarily from Chr 3p21, 6p21, 3q21, 1p36, 8p21 and 8q24 (Supplementary Figure 6). Genes in deleted regions where there was also hypermethylation included EXTL1, RUNX3 and WASF2. These also resided in a region of focal deletion at Chr 1p36.11 (chr1:18079813–27712917) (Supplementary Table 4). At several regions of copy number gain there were clusters of upregulated genes exhibiting hypomethylation including on Chr 6p21.1 (KLHDC3, SLC29A1, BYSL, POLR1C, PPP2R5D and XPO5) and Chr 8q24.3 (PTK2, PTP4A3, CYHR1, PPP1R16A, RECQL4 and ZNF517). Upregulation of ENPP2, mapping to 8q24.3 and previously reported (as a gene expressed at high levels in class 2 tumors (21) appears to have been driven exclusively by methylation (Supplementary Table 6). MTUS1, mapping to Chr 8p22 is one of the 12 prognostic transcripts (41) differentiating class 1 from the aggressive class 2 tumors. Its expression was correlated with both copy number and methylation in TCGA, however it was not significant in the CC cohort (Supplementary Table 3, Supplementary Table 6). Expression of MTUS1 was also significantly associated with metastasis and overall survival (see below).

Mutation analysis:

We looked for protein damaging mutations in genes affected by copy number alteration (Supplementary Table 7). In addition to Chr 3p21 which harbors BAP1 and the handful of deleterious alterations in PHF10 at Chr 6q27 there was very little enrichment of mutations in any other single gene, and very few genes had the hallmarks of a tumor suppressor. On Chr 1p36, two tumors harbored mutations in MTOR, and at Chr 1p34.1, one TCGA tumor (AA9A) harboured a mutation in KDM4A (exon13:c.C1942T:p.P648S) with loss of the wild type allele. Exon 32 of PTK2 mapping to Chr 8q24.3 was somatically mutated in MM127 (c.G3091T:p.A1031S), a class 2 tumor with an activating mutation in GNAQ, no identified mutation in BAP1, SF3B1 or EIF1AX and no apparent copy number change (Supplementary Table 2B). CYC1 on Chr 8q24.3 was mutated in TCGA tumor A883 (p.D209delinsDYY). On Chr 16q23.1, MON1 homolog B (MON1B) showed the highest correlation with expression in TCGA cohort (Supplementary Table 3) and was mutated in TCGA tumor A87Y (c.C1390T:p.R464X). A87Y was a primary tumor from an enucleated specimen that had metastasized. The same mutation has been described in an aggressive cutaneous neuroendocrine tumor (Merkel cell) where Merkel cell polyomavirus has been implicated in a subset of tumors (42). In rare cases neuroendocrine tumors such as those of Merkel cells can metastasize to the uvea (43). Hence, it is possible that the UM in A87Y represents a metastasis from a different primary tumor or that in some instances loss of function of MON1B is associated with the development of UM.

Survival analysis:

We asked how genes from regions exhibiting significantly high correlation with expression and CNAs influenced patient overall survival and compared the results to chromosomal alterations alone. We performed a genome-wide univariate Cox proportional hazard model and compared the effects of high and low gene expression in TCGA data. 6341 transcripts were significantly associated with overall survival at q-value < 0.05 (Supplementary Table 8). Next, we selected the most significant transcripts from regions of copy number alteration and asked if they were also associated with metastasis (Supplementary Table 3) which resulted in 1071 transcripts. Pathway enrichment analysis revealed significant enrichment for metabolism of RNA, immune system and mRNA stability. The majority of these 1072 transcripts were from genes encoded by Chr 3p (165 genes), 3q (108 genes), 8q (152 genes), 6p (80 genes) and 1p (73 genes). Out of these 1072 genes, 424 genes had a hazard ratio < 1 (lower expression associated with good prognosis) and 647 genes had hazard ratio > 1 (lower gene expression associated with poor prognosis).

We focussed on 92 genes present within focal regions of alterations reported from our GISTIC analysis and compared the top three cluster of genes present at 8q24, 1p36 and 3p21 with chromosomal alterations alone (Table 2). In the case of Chr 3, downregulation of BAP1 expression (HR=11.39, CI:3.36–38.54, P value=9.0e-05) was more significant than monosomy 3 (HR=14.5, CI:3.37–62.44, P value =3.3e-04) or the presence of BAP1 mutations (HR=6.58, CI:2.4–17.89, p-value=2.18e-04) (Figure 3A). In addition, there were eight other transcripts within this cluster which were highly correlated with each other (Figure 3A), two of which have been previously reported as downregulated in class 2 versus class 1 UM (RPL29 and SCAP) (39). On Chr 1p36 we identified a cluster of genes (Table 2) highly correlated with poor outcome and associated with metastasis although 1p loss had not been significant in TCGA cohort (HR=1.62, CI:0.66–3.98, P value = 0.29) (Figure 3B, Supplementary Table 3). A cluster of transcripts at chromosome 8q24.3, encoded by 13 genes were strongly correlated with metastasis and poor overall survival (Table 2, Figure 3B). Compared to chromosomal 8q gain (HR = 6.10, CI:1.42–26.34, P value = 0.01) the HR were lower for the cluster of genes on 8q24 (Supplementary Table 3), however the confidence interval are wide and more accurate estimates will require larger sample sizes. MTUS1 on chromosome 8p was more strongly correlated with metastasis than Chr 8p gain (Supplementary Table 3, Table 2, Figure 3B),

Figure 3.

Kaplan-Meier curves for overall survival (OS) in TCGA cohort (n = 80) based on broad chromosomal changes versus candidate gene expression. In the case of BAP1, mutation status is included. A. Chromosome 3 loss versus BAP1 expression and mutation and a correlation plot showing the cluster of transcripts from chr 3p21 which are highly correlated with BAP1 expression. B. Examples for Chr 1p, 8q24.3 and 8p transcripts versus broad Chr 1p loss, 8q gain and 8p loss.

Transcriptional profiling of PHF10 knockdown in uveal melanoma cell lines

We used small interfering RNAs (siRNAs) to investigate the consequences of PHF10 knockdown in the established UM cell lines Mel202, 92.1 and Mel290 (Figure 3). We first profiled transcriptome-wide changes with RNA-Seq (Supplementary Tables 9, 10 and 11). All lines were wild type for PHF10 coding sequence, although Mel202 and 92.1 had lost one copy of Chr 6q. There were 363 differentially expressed transcripts with a fold change of 2 (adjusted P value < 0.05) shared by at least two of the three cell lines (Figure 4A). Significant gene ontology terms for these shared transcripts (Supplementary Table 12) included cell adhesion (GO:0007155), developmental process (GO:0032502), cell development (GO:0048468), regulation of multicellular organismal process (GO:0051239) and extracellular matrix organization (GO:0030198) (Figure 4B).

Figure 4. Consequence of PHF10 knockdown in three uveal melanoma cell lines (Mel202, 92.1 and Mel290).

A. GO analysis of differentially expressed transcripts. For RNA sequencing, three biological replicates were prepared for each condition. Transcripts with a fold change >2 and adjusted p-value < 0.5 following PHF10 kd were selected, and those shared by at least two cell lines were used for pathway analysis with g:Profiler (https://biit.cs.ut.ee/gprofiler/gost). GO redundant terms were removed with REVIGO (65); B. Venn diagram showing the number of differentially expressed transcripts unique to and shared by each cell line; C. Western blot analysis of some transcripts encoding chromatin remodelers that were differentially expressed in PHF10 mutant tumors. 30μg of whole cell lysate was loaded. Knockdown was confirmed with qRT-PCR, RNA-Seq and Knockdown was confirmed with Western blotting and an antibody to PHF10 at a concentration of 1:1000; D, E, F: Results of adhesion assays: ~10⁵ cells (PHF10kd or siCtrl for 24h) were seeded on top of a collagen insert in serum-free media and medium containing 10% FBS and extracellular matrix proteins were added to the lower chamber. Cells were left to invade for 48h and then lysed and stained with CyQuant GR dye and fluorescence was measured at 480/520nm. The extracellular matrix proteins tested here were Col I, Col II and Col IV corresponding to Collagen 1, 2 and 4 respectively, FN (Fibronectin), LN (laminin), TN (tenascin), VN (vitronectin). Neg corresponds to the negative control. Three biological replicates were analysed and results are shown with error bars represent standard deviation of the mean * P-value <0.05. G: Result of migration assays performed as described elsewhere (27).

Although there can be considerable differences in cell lines versus primary tumors, a number of transcripts altered by PHF10 knockdown exhibited the same trend in the PHF10 mutant tumors MM16 and A8KB (TCGA) (Supplementary Table 13). These included downregulated transcripts IL6ST, JAM3, JAZF1, PCGF5, SMAD2 and TRIM22 and up-regulated transcripts PHRF1, POLR3D and TCF3. Many of these were transcription factors/chromatin remodelers and western analysis of the PHF10 knockdowns with antibodies to some of the encoded peptides confirmed altered protein expression of JAZF1, PCGF5 and SMAD2 in the cell lines (Figure 4C).

We also asked if there were transcripts that were also upregulated or downregulated in correlation with PHF10 changes in the larger cohort of UMs. We categorized all 80 TCGA samples into those with low and high PHF10 gene expression based on the survival data using maximally selected rank statistics. We then compared these two groups (PHF10 high and low) and identified 10272 significant differentially expressed transcripts (FDR P value <0.05) (Supplementary Table 14A). There were 4388 transcripts shared between TCGA PHF10 low samples and the UM PHF10 knockdown cell lines that included IL6ST, JAM3, PCGF5 and PHRF1. Employing a more stringent filtering criteria for the TCGA data (FC>2) there were 222 transcripts enriched in pathways including development (GO:2000026, regulation of angiogenesis (GO:0045765), and focal adhesion (KEGG:04510) (Supplementary Table 14).

We used migration, invasion and adhesion assays to contrast cells treated with either a PHF10 suppressing siRNA or a non-silencing control. No change in invasion was observed with PHF10 knockdown (Supplementary Figure 8). However, PHF10 suppression trended toward less adhesion to most ECM proteins in both Mel202 and 92.1 cell lines, but conversely showed no effect in Mel290 (Figure 4D, E, F). Knockdown of PHF10 resulted in significantly reduced migration in a chemotaxis assay in Mel202 and 92.1, but not in Mel290 (Figure 4G).

Discussion

Here we describe large scale and focal CNAs in UM identified by interrogating 182 primary tumors from three different datasets and nominate candidate genes that may contribute to tumorigenesis following integration of gene expression, methylation and mutation data. Clustering of broad copy number changes revealed four tumor groups characterized primarily by alterations involving 1p loss, 1p gain, monosomy 3, 6p gain, 6q loss, 8p gain/loss 8q gain and 16q loss. This is similar to what was recently described for TCGA (20) and previously identified with gene expression profiling alone (44). Although, there were no significant differences in the ploidy estimates between the class 1 and class 2 subgroups, class 2 tumors tended to exhibit higher ploidy states (tetraploidy), consistent with BAP1 deficiency being the prognostic predictor for patients with polyploid tumors (29)

Integration of copy number with expression showed enrichment of genes in cytogenetic bands 1p36, 6p21, 8q22, 8q24 and 3p21 (the locale of BAP1), consistent with comparative genomic hybridization analysis (18). To search for candidate genes that might be driving copy number alterations, we looked for focal alterations and identified seven regions overlapping events in two independent cohorts. Deletion events localized to Chrs 1p36.11, 2q37, 3cen, 6q25 and 6q27 and amplification events mapped to 8q22.1 and 8q24.3.

Loss of Chr 3 was correlated with down-regulation of expression of BAP1 (10). From our studies of clonality in UM, BAP1 mutations were found to be occasionally present in the subclones suggesting loss of one Chr 3 homolog precedes mutations in BAP1(33). Thus, it is unlikely that haplo-insufficienty of BAP1 is driving LOH3. However, we identified an overlapping region of deletion at 3p11.1-q11.1 that contains PROS1 and ARL13B as well as pericentromeric sequence. This has not been described before in the context of UM and its significance is not clear and will require further investigation.

The region of 8q gain in UM has been narrowed to 8q23-q24 in a number of studies (16–19) and we identified focal amplification of a segment of PLEC at chr. 8q24.3. However, there were a number of candidates outside this region with higher correlation coefficients and significant P values (Supplementary Table 3). 8q gain is associated with poor prognosis and expression of a number of transcripts from Chr 8q24.3 exhibited high correlation with copy number gain. Their over-expression was more significantly associated with metastasis and worse overall survival. We also showed that in addition to Chr 8q24.3 gain, upregulation of PTK2 and PTP4A3 could be a consequence of de-methylation. These findings are consistent with earlier studies of PTP4A3 which also showed it to be a strong predictor of metastasis where its over-expression increased cell migration and invasiveness (45). These results are also consistent with earlier array CGH studies where the importance of chromosome 8q gain on metastasis was demonstrated, along with losses of 3, 8p, and 16q (46).

PTK2/FAK is an established driver of some tumors and is a target for amplification at 8q in primary hepatocellular carcinoma (HCC) (47) and breast cancer (48). The number of copies of Chr 8q24.3 in the region harboring PTK2 ranged from 3 to 9 in the CC and WU cohorts and given its association with poor overall survival, these data are consistent with earlier data, including a clear association between metastatic potential of UM and 8q ploidy of five copies or more(13). Tumor MM127 harbored an A1031S alteration in PTK2 and no other detectable abnormalities. This alteration alters a highly conserved “DAKNL” motif in the C-terminal end of PTK2 that is conserved from humans to flies. It lies in a paxillin binding subdomain of the FAT domain which is required for localizing PTK2 to focal adhesions in response to integrin stimulation (49,50). We hypothesize that this p.A1031S alteration strengthens cellular adhesion mediated by FAK. Inhibition of FAK has been implicated in improving local control in HPV-negative head and neck squamous cell carcinoma (HNSCC) (51) and in castrate resistant prostate cancer (52) and might also be a therapeutic target in UM. This is consistent with recent studies showing that oncogenic mutations found in GNAQ in UM activates YAP through FAK and that inhibition of FAK reduces UM growth (53). Chr 8q24 also harbors MYC which is amplified in about 30% of UMs (54), but its expression was not significantly upregulated as a consequence of 8q24 amplification.

Integrated analyses also revealed that a cluster of genes on chr1p36 is associated with metastasis and worse overall survival compared to broad chromosomal changes.

At Chr 6q27 we obtained evidence for PHF10/BAF45a as a novel tumor suppressor for UM on the basis of three tumors with definitive loss of function mutations: two tumors with overlapping homozygous deletion (HD) and one with a frameshift mutation and concomitant loss of the wild type allele (MM133). GEP had classified one MM016 as a class 1B tumor and MM133 as a class 2 tumor although they lacked SF3B1 and BAP1 mutations respectively. However, in addition to the PHF10 frameshift mutation, MM133 also harbored a mutation in SF3B1 leading to a R625C alteration and was disomic for Chr 3 despite being classified as a class 2 tumor (Supplementary Table 2). We also identified a mis-sense mutation in TCGA tumor A8KM (p.D453E, rs761295711). This mutation was very rare (frequency in Exac is 8.3 X 10⁻⁶) and although tumorigenicity of mis-sense alterations is hard to predict, this alteration was predicted to be damaging. Tumor A8KM had also lost one copy of Chr 3 (Supplementary Table 1) and PRAME was upregulated but it lacked identified mutations in BAP1, SF3B1 or EIF1AX (Supplementary Table 2). Hence, the presence of alterations in PHF10 and BAP1 were mutually exclusive in the small number of tumors sampled. Loss of 6q where PHF10 resides was not correlated with worse overall survival in TCGA, nor were levels of PHF10.

PHF10 (AKA BAF45a) is a component of the PBAF complex (a SWI/SNF-like complex) that is involved in chromatin remodeling(55) which in turn can affect transcriptional regulation. During PHF10 silencing, other components of the PBAF complex (SMARCE1, SMARCC1 and SMARCA4) (Supplementary Figure 7) were upregulated, potentially as a compensation mechanism. RNA sequencing data were available for MM016 and A8KM and we identified a number of chromatin remodelers with altered expression that exhibited the same trend in cell lines with PHF10 knockdown. We confirmed alteration at the protein level of some of these including PCGF5, SMAD2 and JAZF1 (55). JAZF1 is required for ciliated cell differentiation in vitro (56) and is involved in transcriptional repression. It can be fused to other genes in endometrial stromal tumors where it disrupts the polycomb repressive complex 2 (PRC2), abolishes histone methyl transferase activity and activates chromatin/genes normally repressed by PRC2 (57). PCGF5 encodes a component of PRC1. It is required for the differentiation of mouse embryonic stem cells (mESCs) towards a neural cell fate where it functions both as a repressor for the SMAD2/TGF-beta signaling pathway and a facilitator for neural differentiation. Its loss impairs the reduction of H2AK119ub1 and H3K27me3 around neural specific genes, keeping them repressed.

PHF10 loss also led to downregulation of adhesion and cell migration in Mel202 and 92.1 cell lines. This was not observed in Mel290, but this line lacks mutations in the known drivers BAP1, SF3B1, EIF1AX so is not a “typical” UM cell line. Moreover, this line is very migratory and not very adhesive at base-line. Some transcripts such as ITGA7, ITGA10, JAM3, LAMC1, EPHB3, HES1 and PIK3CB that are involved in adhesion and migration were downregulated after PHF10 kd. However some such as EPHA3, PPP1R9B, ADAM15, POSTN and COL4A1 were upregulated, highlighting the complex relationship between metastasis, migration and adhesion. Further studies will be required to understand the role PHF10 plays in these processes.

Mutations in PHF10 that drive tumor development have not been described before, although mutations in other PBAF components such as ARID2, PBRM1, SMARCA4 (BRG), SMARCB1 (BAF47) are implicated in other cancer types (58,59). PHF10 lies in a region of Chr 6q which harbors an unidentified tumor suppressor for a variety of epithelial cancers (60–62) so it should also be strongly considered as a candidate for this elusive gene. (63)

Translational Relevance.

Here we describe integration of copy number, transcriptomic, epigenomic and mutational data in uveal melanoma across three independent datasets. Although loss of one copy of Chromosome 3 is usually accompanied by loss of function mutations in BAP1, the significance of other chromosomal changes such as loss of Chr 1p, 6q and 8p, and gain of Chr 6p and 8q are not clear. We describe candidate genes on altered chromosomes affecting uveal melanoma pathogenesis and patient survival. We also provide evidence for a tumor suppressor role for PHF10 on Chr 6q27. Its loss affects early pathways in cell development such as transcriptional regulation as well as adhesion and migration. This knowledge will be important as we progress towards a more comprehensive molecular diagnosis of uveal melanoma and determination of prognosis. Besides providing important insights in the development of uveal melanoma these studies provide novel therapeutic targets for this cancer.