Chromosome 3p Deletion Leads to Extensive Genomic Alterations in Diverse Cancers and Confers Synthetic Lethality in Uveal Melanoma

Simple Summary

Isochromosomes are chromosomal aberrations consisting of two identical mirror-imaged arms and simultaneous deletion of the complementary arms. We show that SETD2 deleterious mutations or hemizygous deletion of chr3p containing the gene leads to isochromosomes in cancer patients. Dicentric isochromosomes, known to form a chromatin bridge between dividing cells, are prone to massive fragmentation, which can be modeled computationally. In uveal melanoma (UVM), chr3 deletion also includes MITF, but we show that MITF is a master regulator of key driver genes of UVM. We demonstrate how this accidental hemizygous deletion of MITF in UVM imposes a strong selection pressure on subsequent compensatory genome evolution, explains differences in patient survival, and provides therapeutic opportunities.

Abstract

Background: Chromosome 3p (chr3p) is frequently deleted in multiple cancers, indicating the presence of shared tumor suppressors. In aggressive uveal melanomas (UVM), this deletion often co-occurs with chr8q amplification (8q+), suggesting strong selection pressure during UVM evolution. Methods: To understand the pattern of genomic alterations mediated by chr3p deletion, we have developed an algorithm for detecting isochromosomes in 10,632 TCGA cancer patients. We further perform integrative genomics analysis to investigate how chr3p deletion could affect subsequent cancer genome evolution and synthetic lethality in UVM. Results: Analysis of genomic alterations in 33 different cancer types implicates the deletion or deleterious mutations of SET-domain-containing 2 (SETD2) at chr3p21 in significantly facilitating the formation of isochromosomes, thereby promoting genomic instability conducive to rapid cancer genome evolution. Fracturing of dicentric isochromosomes during cell division is pervasive and follows the dynamic fragmentation pattern of solids under impulse. In the most aggressive UVM subtype, chr3 deletion includes MITF, a master regulator of melanocyte survival and differentiation, and co-occurs with 8q+. We demonstrate that MITF is a master transcriptional regulator of GNAQ/GNA11 and associated synthetic-lethal genes in UVM. MITF maintains MAPK and calcium homeostasis in UVM, and its hemizygous deletion is thus accidental, likely creating an early crisis during oncogenesis. We further show that MITF, MYC, and GNAQ/GNA11 form coupled regulatory feedback loops in the melanocyte lineage, and MITF deletion in UVM creates acute dependency on MYC-mediated rescue via 8q+. The discovered feedback loops predict both overall and relapse-free patient survival within the most aggressive UVM subtype, explain sensitivity to therapeutic gene perturbations, and inform effective combinatorial therapies. Conclusions: SETD2 deletion potentiates isochromosome formation across diverse cancers. Combinatorial targeting of MITF together with a previously identified synthetic lethal gene may benefit UVM patients harboring both chr3 deletion and 8q+.

1. Introduction

Uveal melanoma (UVM), arising from the uveal tract melanocytes, is the most common primary intraocular cancer in adults, second to skin cutaneous melanoma (SKCM) in frequency among all melanoma cases in the US [1]. Co-occurring chr3 deletion and chr8q amplification (8q+) found in ~50% of UVM patients [2,3] comprise the most aggressive subtype with poor relapse-free survival and elevated metastasis risk [2,4]. Chr3 monosomy (M3) is thought to be an early oncogenic alteration in the evolution of aggressive UVM, likely preceding [5] the accompanying 8q+, and partly reflects the need to delete the tumor suppressor gene BAP1 [6], the second retained copy of which is often mutated to completely inactivate the gene [2,3,7]. Consistent with this putative ordering of events, recent single-cell sequencing [8] of a primary tumor with high metastatic potential showed the branching of ancestral M3 clones into large subclones subsequently acquiring 8q+. These phenomena thus suggest a strong selection pressure on cancer cells to amplify 8q in response to M3, but the source of this selection pressure and associated therapeutic opportunities remain unknown. In addition, over 90% of UVM patients harbor mutually exclusive mutations in the G-protein α-subunit genes GNAQ and GNA11. We show that hidden within these molecular stratifications resides hitherto-unrecognized information about the trajectory of cancer evolution and the selection pressure that forced the UVM cells to depend on a melanocyte-specific regulatory network.

The 8q+ event in M3 UVM often involves isochromosomes, consisting of two identical mirror-imaged arms and simultaneous deletion of the complementary arms (Figure 1A). Upon close examination of 10,632 patient datasets from 33 different The Cancer Genome Atlas (TCGA) cancers [3], we have found isochromosomes to be linked to chr3p (3p) deletion and contribute to chromosomal instability (CIN) across multiple cancers. CIN is a cellular state in which large-scale chromosomal rearrangements can emerge from mis-segregation of chromosomes during cell division [9]. CIN is thought to occur in bursts, especially at early stages of tumor evolution [10], and is a major source of rapid genome diversification on which natural selection can act during oncogenesis. Recent studies have identified multiple roles of SETD2 in protecting normal cells against CIN. First, SETD2 methylates α-tubulin in microtubules during mitosis, a step required for proper mitotic spindle formation, and the loss of this function leads to genomic instability [11,12]. Second, SETD2 is the sole enzyme that can catalyze the trimethylation of H3K36 (H3K36me3) [13], and the loss of H3K36me3 in SETD2-deficient cells promotes the formation of isochromosomes [14]; isochromosomes can function as a chromatin bridge between dividing cells and initiate CIN [14,15,16]. Finally, SETD2 loss of function causes nuclear morphological defects, promoting aneuploidy and CIN [17]. Even though the essential role of SETD2 in preventing isochromosomes has been demonstrated in cell lines [14], the impact of its aberration in cancer patients has not yet been demonstrated. Of note, SETD2 is located at 3p21 and thus deleted in M3 UVM. In UVM, isochromosomes are rarely observed on chr3, but frequently occur on chromosomes 1, 6, and 8 [18,19], leading to 1p-/1q+, 6p+/6q-, and 8p-/8q+ (Figure 1B); furthermore, 8q+ isochromosomes occur only in the context of M3 in TCGA-UVM, suggesting that SETD2 copy number loss (CNL) likely precedes and facilitates 8q+ in M3 patients. Extending this observation, our analysis of pan-TCGA cancers implicates SETD2 CNL in forming isochromosomes in diverse human cancers. Across cancers, isochromosomes occur most frequently on chr8, amplifying MYC as in UVM, and on chr17, deleting TP53. We further show that dicentric isochromosomes are prone to fragmentation, which can be accurately modeled using dynamic fracture analysis [20,21], and that such repeated fragmentation can cause ultra-high amplification of oncogenes.

Chr3 loss in UVM embodies another hidden driving force. The GNAQ/11 mutations induce constitutive activation of the G-proteins G_q and G₁₁, leading to the sustained production of diacylglycerol (DAG) and inositol 1,4,5 trisphosphate (IP3). Elevated DAG activates RAS-dependent MAPK pathway by binding and activating the guanine nucleotide exchange factor RASGRP3 [22], creating critical dependence of GNAQ/11-mutant cells on RASGRP3. Furthermore, recent high-throughput screening assays have identified synthetic-lethal genes that accompany the GNAQ/11 mutations [23]. In particular, constitutively synthesized IP3, if not degraded via the phosphatase INPP5A, can accumulate and increase intracellular calcium ions to a toxic level. Targeting RASGRP3 or INPP5A with small molecules in the context of GNAQ/11 mutations thus represents a promising strategy for treating a large fraction of UVM patients [23]. However, the UVM dependency on RASGRP3 and INPP5A in the background of M3 poses a perplexing paradox.

Namely, we show that both RASGRP3 and INPP5A, as well as several other critical UVM-vulnerability genes [23,24], are bound by and transcriptionally activated by MITF, the master regulator of melanocyte differentiation and survival [25]. Despite being a critical hub in the network of synthetic-lethal genes, however, MITF resides at 3p14 and thus gets co-deleted with SETD2/BAP1 in M3 UVM. Our discovered regulatory network thus predicts that MITF depletion reduces UVM-vulnerability genes and creates lethality in GNAQ/11-mutant UVM, as supported by independent screening experiments [24,26], implying that MITF hemizygous deletion is an accident that M3 UVM cells must resolve. We here show that MITF and MYC regulate each other’s transcription and form a positive feedback loop in melanocytes and that the M3-mediated crisis of MITF deletion poses a strong selection pressure on UVM to amplify MYC on 8q24 to restore the pre-M3 level of MITF. The positive feedback loop between MITF and MYC thus explains the coupling of M3 and 8q+ and resolves the paradox. This newly gained insight, in turn, uncovers an opportunity to develop combinatorial synthetic-lethal therapeutics that can benefit UVM patients.

Figure 1. Chr3 deletion precedes isochromosome formation in UVM and is linked to deep MITF amplification in SKCM. (A) Illustration of monocentric and dicentric isochromosome formation. (B) ASCAT3 allele-specific copy number data for TCGA-UVM patients with isochromosomes (black boxes). Isochromosomes are significantly enriched in the group with chr3 LOH (binomial test p-value = 2.9 × 10⁻⁴). Isochromosomes on chr8 occur only in the context of M3 (Fisher’s exact test p-value = 3.0 × 10⁻⁷). (C) Mean chr3 deletion profile in four TCGA cancers (142 BRCA, 338 KIRC, 70 SKCM, 17 KIRP patients). The KIRC (top right) plot shows a fitted Poisson break model (magenta line) initiated at the centromere to the location of BAP1. (D) Copy number alterations (CNA) showing loss (blue) and gain (red) in SKCM patients with either partial or full LOH on chr3. Green dashed line in the middle “chr3 CNA” panel indicates MITF. All focal MITF amplification loci in the context of chr3 LOH are heterozygous (HET). (B,D) are modified versions of UCSC Xena browser screenshots [27].

Figure 1. Chr3 deletion precedes isochromosome formation in UVM and is linked to deep MITF amplification in SKCM. (A) Illustration of monocentric and dicentric isochromosome formation. (B) ASCAT3 allele-specific copy number data for TCGA-UVM patients with isochromosomes (black boxes). Isochromosomes are significantly enriched in the group with chr3 LOH (binomial test p-value = 2.9 × 10⁻⁴). Isochromosomes on chr8 occur only in the context of M3 (Fisher’s exact test p-value = 3.0 × 10⁻⁷). (C) Mean chr3 deletion profile in four TCGA cancers (142 BRCA, 338 KIRC, 70 SKCM, 17 KIRP patients). The KIRC (top right) plot shows a fitted Poisson break model (magenta line) initiated at the centromere to the location of BAP1. (D) Copy number alterations (CNA) showing loss (blue) and gain (red) in SKCM patients with either partial or full LOH on chr3. Green dashed line in the middle “chr3 CNA” panel indicates MITF. All focal MITF amplification loci in the context of chr3 LOH are heterozygous (HET). (B,D) are modified versions of UCSC Xena browser screenshots [27].

2. Materials and Methods

2.1. The Human Reference Genome

Unless explicitly stated otherwise, all genomic coordinates in the manuscript are in hg38.

2.2. The Cancer Genome Atlas (TCGA) Data

The results shown here were based upon data generated by TCGA Research Network: https://www.cancer.gov/tcga (accessed on 17 June 2025). All TCGA data were accessed through the Genomic Data Commons (GDC) portal (v43.0) or through the UCSC XenaBrowser [27] (GDC v41.0).

2.3. UCSC Screenshots

The chromosome cytobands and the ChIP-seq/CUT&RUN tracks were obtained from visualizing the data on the UCSC Genome Browser [28].

2.4. Mean ASCAT3 Allele-Specific Copy Number Analysis Conditioned on Deletions or Amplifications

For a given chromosome arm, deletion-aware copy number alignment was performed using only those patients who have a chromosomal fragment with total ASCAT3 copy number <2 on the specified chromosome arm [29]. Amplification-aware alignment was similarly performed using only those patients who have a chromosomal fragment with total ASCAT3 copy number >2 on the specified chromosome arm.

2.5. Simulation of Poisson-Distributed chr3p Break Points

TCGA-KIRC patients tended to have chr3p contiguous partial-arm deletions including the telomere and BAP1, indicating that a single chromosomal break event generated the deletions. To simulate a null model of chr3p break between the centromere and BAP1 locus, we assumed the break point to be Poisson-distributed between 90 Mb and the BAP1 transcription start site (TSS). The mean ASCAT3 copy number of 338 KIRC patients with a deletion event on chr3p was 1.627 at the location 90 Mb, theoretically representing 212 patients who still retained two copies at this location and 126 M3 patients. We thus simulated 212 Poisson breaks in the indicated region and computed the mean copy number of the cohort consisting of these 212 simulated patients and 126 M3 patients.

2.6. Isochromosome Calling Algorithm

We called isochromosomes by calculating the modal copy number (the most frequently occurring copy number) of each chromosomal arm, using allele-specific ASCAT3 copy number data from TCGA. The locations of centromeres were determined from the “centromeres” table from UCSC Genome Browser Table Browser (hg38); p and q arms were demarcated by the midpoint of each centromere, which was determined using the smallest and largest genomic coordinates reported for the centromeric sequence start and end, respectively. Isochromosomes were then identified when all conditions were met as follows:

• One arm had loss of heterozygosity (LOH), which was enforced by requiring that the modal minor copy number be zero.
• The other arm was heterozygous, which was enforced by requiring that the modal minor copy number be greater than zero.
• The modal major copy number of the arm with LOH was at least two less than the total modal copy number of the heterozygous arm.
• If the modal major copy number of the LOH arm was closer to or equidistant to the modal minor copy number of the heterozygous arm compared to the modal major copy number of the heterozygous arm, then we reasoned that the major copy of the heterozygous arm formed an isochromosome and required its copy number to be at least two. Otherwise, we assumed that the minor copy of the heterozygous arm formed an isochromosome and required its copy number to be at least two.
• The allele-specific minor copy number of the telomeric end of the LOH arm had to be zero. This last condition prevented incorrectly calling non-isochromosome-forming mechanisms that retained the telomere while producing partial-arm LOH.

2.7. Analysis of the Association Between SETD2/PSIP1 LOH and Isochromosome Formation

We analyzed the effect of SETD2/PSIP1 LOH on isochromosome formation as follows. Isochromosome occurrences were called in each of the 33 TCGA cancers (10,632 total patients) using the previously outlined procedures for 18 of the 24 human chromosomes; the acrocentric chromosomes 13, 14, 15, 21, 22, and Y were removed.

We determined the SETD2 or PSIP1 LOH status for each patient based on the gene’s ASCAT3 allele-specific minor copy number being zero. The possible number of isochromosome calls ranged from 0 to 17 for each patient after the exclusion of acrocentric chromosomes and chr3 or chr9 during the analysis using SETD2 or PSIP1, respectively, and was further reduced to 16 when analyzing the effect of their double-LOH status. No patients carried both SETD2 and PSIP1 LOH in LAML, PRAD, and THCA cancers, which were excluded from the analyses involving SETD2/PSIP1 double-LOH status. The Wald statistic

$$T={\displaystyle \frac{{\widehat{p}}_1-{\widehat{p}}_2}{\sqrt{p\left(1-p\right)\left(\frac{1}{N_1}+\frac{1}{N_2}\right)}}},$$

was calculated where $N_1$ and $N_2$ are the number of patients with or without gene LOH, respectively. We calculated the probability of a chromosome being an isochromosome as

$${\widehat{p}}_1={\displaystyle \frac{\mathrm{n}\mathrm{u}\mathrm{m}\mathrm{b}\mathrm{e}\mathrm{r} \mathrm{o}\mathrm{f} \mathrm{i}\mathrm{s}\mathrm{o}\mathrm{c}\mathrm{h}\mathrm{r}\mathrm{o}\mathrm{m}\mathrm{o}\mathrm{s}\mathrm{o}\mathrm{m}\mathrm{e} \mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}\mathrm{s} \mathrm{i}\mathrm{n} \mathrm{L}\mathrm{O}\mathrm{H} \mathrm{p}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{s}}{17\times N_1}}$$

for patients with SETD2/PSIP1 LOH, and similarly for patients with heterozygous SETD2/PSIP1 (substituting $N_1$ for $N_2$ and ${\widehat{p}}_1$ for ${\widehat{p}}_2$). The probability $\widehat{p}$ is given by

$$\widehat{p}={\displaystyle \frac{\mathrm{n}\mathrm{u}\mathrm{m}\mathrm{b}\mathrm{e}\mathrm{r} \mathrm{o}\mathrm{f} \mathrm{i}\mathrm{s}\mathrm{o}\mathrm{c}\mathrm{h}\mathrm{r}\mathrm{o}\mathrm{m}\mathrm{o}\mathrm{s}\mathrm{o}\mathrm{m}\mathrm{e} \mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}\mathrm{s} \mathrm{i}\mathrm{n} \mathrm{L}\mathrm{O}\mathrm{H} \mathrm{o}\mathrm{r} \mathrm{h}\mathrm{e}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{o}\mathrm{z}\mathrm{y}\mathrm{g}\mathrm{o}\mathrm{u}\mathrm{s} \mathrm{p}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{s}}{17\times (N_1+N_2)}}.$$

The Wald statistic $T$ was evaluated against a standard normal distribution to calculate the p-values, along with the odds ratio

$$\mathrm{o}\mathrm{d}\mathrm{d}\mathrm{s} \mathrm{r}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}={\displaystyle \frac{{\widehat{p}}_1/(1-{\widehat{p}}_1)}{{\widehat{p}}_2/(1-{\widehat{p}}_2)}}.$$

To prevent Type I errors, the false discovery rate (FDR) was calculated using the Benjamini–Hochberg procedure from SciPy’s (v1.15.1) false_discovery_control and was thresholded at ${10}^{-2}$ FDR for calling significance.

When computing the Wald statistic comparing the isochromosome occurrence rate in patients with or without LOH at a given location on chr3 or chr9, the locations were sampled every 100 kb along chr3 and chr9. For display, we averaged the −log₁₀ p-values using a running window of 20 bins.

2.8. Isochromosome Occurrence Probability Calculation

Isochromosomes were called for each of the 33 TCGA cancers for 18 chromosomes (excluding chr13, 14, 15, 21, 22, and Y), using the isochromosome-calling method outlined in Section 2.6. For each cancer type, the number of patients with a specific isochromosome was divided by the total number of patients to calculate the occurrence probability. In the final column (“all”), we summed the number of occurrences of each isochromosome across all cancers and divided the sum by the total number of patients (10,632).

2.9. Analysis of Chromosomal Fragmentation as a Function of the Distance of Isochromosome Fusion Site from Centromere

We analyzed the fragmentation pattern of the isochromosome arm subjected to LOH as follows. The centromere midpoint on each chromosome was determined by first concatenating all hg38 centromeric sequences annotated in the UCSC Genome Browser and then taking the center of the merged region. The p and q arms of each chromosome were then defined to be the left and right side of the corresponding centromere midpoint. Defining heterozygous fragments as the chromosomal fragments in the ASCAT3 allele-specific copy number data with a strictly positive minor copy number, we counted all heterozygous fragments entirely contained in the isochromosome LOH arm. To account for the centromere region itself, we added a count of 1 to the calculated number of heterozygous fragments. The fusion site of isochromosome was determined to be the location of the outer boundary of the farthest heterozygous fragment on the LOH arm from the centromere midpoint. Combining all 11,658 isochromosomes, we used R’s (v4.4.3) linear model (lm; default parameters) to regress the number of heterozygous fragments on the LOH arm against the fusion site distance from the centromere. The regression plot (including the fitted line, the error bars, and the 95% confidence interval) was made with ggplot2, using stat_summary_bin for the error bars and geom_smooth for the fitted line and 95% confidence interval.

2.10. Analysis of Chromosomal Fragmentation as a Function of Dicentric Isochromosome Status

To compare the number of heterozygous fragments occurring on isochromosomes vs. non-isochromosomes, we called isochromosomes according to the methods outlined in Section 2.6 using ASCAT3 allele-specific copy number. Equal copy number fragments immediately flanking the centromere were identified as a single fragment to compensate for missing or incomplete copy number data. The boundaries of the centromere were defined using the smallest and largest genomic coordinates for centromere endpoints listed in the “centromeres” table from UCSC Genome Browser. The number of heterozygous fragments were counted in dicentric isochromosomes and non-isochromosomes. Isochromosomes were considered monocentric and excluded from the analysis if they had LOH on an entire arm (divided at the centromere midpoint); otherwise, they were considered dicentric. In cases where the copy number fragment did not span the entire called centromere region, one heterozygous fragment count was subtracted if the copy number of the fragments immediately before and after the centromere region were equal, as this would indicate that the fragments were an artifact of missing copy number data at the centromere. We then tested for the difference between the number of heterozygous fragments in chromosomes with dicentric isochromosome vs. non-isochromosome status using a Mann–Whitney U-test (mannwhitneyu from SciPy v1.15.1; default parameters). The ratio between the mean number of heterozygous fragments with or without isochromosome formation was reported in addition to p-value in the results Section 3.5.

2.11. Mott Distribution Fitted to the Cumulative Distribution of Chromosomal Fragment Lengths

Dynamic fragmentation of solids under transient force can be modeled using either stochastic or energy-based physics methods [20]. These theories predict that the cumulative fragment-size distribution decays exponentially in material-dependent power of the fragment size, and this distribution is called a Mott distribution. The heterozygous fragments on the isochromosome LOH arm were determined as described in the Methods Section 2.9. The intermediate regions with LOH between adjacent heterozygous fragments were discarded, because their fragmentation pattern was lost during cancer evolution. We thus assumed that the observed heterozygous fragments were sampled uniformly from the set of all chromosomal fragments upon fracturing. To demonstrate that the heterozygous fragments occurring on the LOH arm of isochromosomes followed a distribution of dynamic fragmentation predicted by the theory of solids under impulse [20], we fitted the cumulative number of fragments of length larger than $L$ to a Mott distribution:

$$N\left(L\right)=N_0{e}^{-\mu L^{\beta }}.$$

Specifically, the logarithm of the Mott distribution was fitted using R’s (v4.4.3) nonlinear least squares function (nls; default parameters) to numerically determine the parameters $\mu$, $\beta$, and $log N_0$. The mean fragment size of this distribution was calculated to be ${\mathsf{\mu }}^{-1/\mathsf{\beta }}\mathsf{\Gamma }\left(1+1/\mathsf{\beta }\right)$. To construct a canonical representative of the fragmentation pattern across multiple cancer types, we aggregated heterozygous fragmentation data from chr1, 2, and 3 (large, metacentric chromosomes) for all 33 cancer types from TCGA. The aggregated data was then fitted using a Mott distribution (red) and compared to the best-fit Poisson distribution (blue). Because different chromosomes have different arm lengths, we also fitted Mott distributions for individual chromosomes pooled across all cancer types in TCGA. The LOH status of the acrocentric isochromosomes 13, 14, 15, 21, and 22 could not be reliably determined and were, in addition to chrY, excluded from the analysis.

2.12. Gene Network Construction and Betweenness Centrality

The undirected co-dependency network of synthetic-lethal genes was constructed with edge weights equal to the Pearson correlation $r$ between genes’ CRISPR gene effects (DepMap Public 25Q2+Score, Chronos). The network was restricted to MITF and the 29 synthetic-lethal targets identified by Elbatsch et al. [23], with an edge between two genes included only if either gene appeared in the other’s top 100 co-dependency list. The resulting graph contained three nontrivial connected components; the remaining vertices, all with degree zero, were not shown. Since only positive edge weights were present, the Pearson correlation edge weights were transformed into edge weight penalties by the formula penalty $=1-r$, making the analysis of vertex importance amenable to standard techniques based on shortest paths. We then employed the Floyd–Warshall algorithm to compute the shortest paths between all gene pairs and calculate the betweenness centrality score for each vertex in the connected component of MITF. The betweenness centrality of node $v$ was computed by

$$g\left(v\right)=\sum _{s\ne v, t\ne v}{\displaystyle \frac{{\sigma }_{st}\left(v\right)}{{\sigma }_{st}}}$$

where ${\sigma }_{st}\left(v\right)$ denotes the number of shortest paths between vertices $s$ and $t$ which pass through vertex $v$, and ${\sigma }_{st}$ denotes the total number of shortest paths between vertices $s$ and $t$. These centrality scores were then subjected to the normalization

$$\mathrm{normalized}(g(v))={\displaystyle \frac{g\left(v\right)-\min \left(g\right)}{\max \left(g\right)-\min \left(g\right)}}$$

such that the vertex with largest betweenness centrality score was one and the smallest was zero.

2.13. Summary of MITF or MYC KD/KO/Mutation or OE Experiments

The curation of multiple over-expression (OE) and knock-down (KD)/knock-out (KO)/mutation data from Gene Expression Omnibus (GEO) for both MITF and MYC followed several standard conventions: if multiple transcript IDs mapped to the same gene name, the data corresponding to the most statistically significant transcript was used. In the case of GSE163646, which contained data from both OE and KO experiments, ENST00000526292.1 was the most statistically significant transcript for MTMR9 during MITF OE but was absent in the MITF KO experiment where instead ENST00000221086 was most significant. To maintain consistency in which isoform’s data was displayed, the MTMR9 log₂-FC value for GSE163646 MITF KO was removed and instead shown in gray; data for ENST00000526292.1 was shown in GSE163646 MITF OE.

The experiments utilized for the construction of heatmaps are listed below by GEO accession number, followed by additional experiments used to corroborate the double-positive feedback loop between MITF and MYC. Unless otherwise stated, all analyses were performed using GEO’s GEO2R webtool with DESeq2 (v1.49.3) [30] for differential expression analysis; p-value adjustment was performed within DESeq2 using the Benjamini–Hochberg procedure. GEO Sample accession numbers are provided for each replicate in the groups (case/control) analyzed.

• GSE273527 [31] utilized choroidal melanocytes cultured from eyeball of human donors. The supplementary materials therein provided differentially expressed genes (DEGs) between RNA-seq of lentiviral transduction-generated mutant GNAQ^Q209L vs. normal choroidal melanocytes (12915_2025_2118_MOESM6_ESM.xlsx, sheet name ‘20-RNAseq-NCMvsMutCM DEG’).
• GSE163646 [32]. MITF OE was accomplished via FLAG-tagged MITF in the stable doxycycline inducible overexpression A375P cell line. Differential expression analysis was performed between MITF OE (GSM4983430, GSM4983431, GSM4983432) and empty vector FLAG control (GSM4983427, GSM4983428, GSM4983429). Additionally, differential expression analysis was performed for CRISPR-Cas9-generated MITF KO in the human hypotetraploid SK-MEL-28 melanoma cell line (GSM4983417, GSM4983418, GSM4983419, GSM4983420) relative to control (GSM4983413, GSM4983414, GSM4983415, GSM4983416). Sleuth [33] (v0.30.2) was used for both MITF OE and KO analyses.
• GSE167496. MITF OE in A2058 human melanoma cells was performed in three replicates using lentiviral transduction for MITF OE (GSM5106134, GSM5106135, GSM5106136), and DEGs were computed relative to control (GSM5106131, GSM5106132, GSM5106133).
• GSE188469 [34]. Differential gene analysis was performed between Vemurafenib-treated replicates (GSM5683486, GSM5683487) relative to a control treated with dimethyl sulfoxide (DMSO) (GSM5683470, GSM5683471) in the metastatic melanoma cell line Ma-Mel-63a harboring the BRAF^V600E mutation. Both groups were treated for 24 h.
• GSE50686 [35] treated the COLO829 cell line with the BRAF-inhibitor PLX4032 (a.k.a. Vemurafenib), leading to accumulation of MITF. DEGs were computed between Vemurafenib-treated replicates (GSM1225723, GSM1225724) relative to control (GSM1225721, GSM1225721).
• GSE190071 [36]. RNA-seq was performed 40 h post application of the BRAF-inhibitor Vemurafenib (GSM5750555, GSM5750556, GSM5750557, GSM5750558) in the melanoma cell line A375. Three replicates (GSM5750548, GSM5750549, GSM5750550) were used as control.
• GSE209769 [37]. Replicates of the melanoma A375 cell line were treated with 10 $\mathsf{\mu }\mathrm{M}$ of Vemurafenib for 3 days and then sequenced (GSM6395047, GSM6395048, GSM6395049). Differential expression analysis was then performed between the Vemurafenib-treated replicates and control (GSM6395044, GSM6395045, GSM6395046).
• GSE252781 [38]. The M397 melanoma cell line was sequenced after 12, 24, 36, 48, and 72 h of BRAF inhibitor (Vemurafenib) application, as well as 12, 24, 48, and 72 h and 6 days after the removal of Vemurafenib. Controls using DMSO were sequenced at day 0 and 72 and 216 h after the DMSO treatment. Differential gene expression was performed between replicates treated with Vemurafenib for 12, 24, 48, and 72 h (respectively: GSM8007258, GSM8007259, GSM8007261, GSM8007262) relative control replicates sequenced at day 0 and after 72 h of DMSO treatment (respectively: GSM8007257, GSM8345559). Data for the replicate treated with Vemurafenib for 36 h, the replicate 6 days post Vemurafenib removal, and the control replicate sequenced at 216 h post DMSO treatment were all unavailable for analysis using GEO2R.
• GSE115845 [39]. We aligned the raw sequencing reads to the hg38 reference genome using the package HISAT2 [40] (v2.2.1) and assembled transcripts using the package StringTie [41] (v2.1.1); both used default parameters. We called differentially expressed genes between MITF KD samples (GSM3191783, GSM3191784) and control (GSM3191785, GSM3191786) using DESeq2 [30] (v1.20.0) with default parameters.
• GSE200674 [42] utilized melanocytes cultured from induced pluripotent stem cells (iPSCs) derived from a Waardenburg Syndrome patient carrying a heterozygous mutation in MITF. Differential expression analysis was performed between mutant MITF (GSM6041871, GSM6041872, GSM6041873) and wild-type (WT) MITF (GSM6041874, GSM6041875, GSM6041876) iPSC-derived melanocyte replicates.
• GSE255546 [43] used BRAF^V600E-mutant and WT melanocytes isolated from neonatal foreskin. Differential expression analysis between mutant (GSM8074510, GSM8074511, GSM8074512) and WT (GSM8074504, GSM8074505, GSM8074506) was performed.
• GSE255256 [44] used Doxycycline-inducible shMITF KD in MeWo and SK-MEL-28 melanoma cell lines. For MeWo, differential expression analysis was performed between replicates treated with 1 $\mathsf{\mu }\mathrm{g}/\mathrm{m}\mathrm{L}$ Doxycycline (GSM8067677, GSM8067678, GSM8067679) compared to a group without treatment (GSM8067674, GSM8067675, GSM8067676). Likewise, three SK-MEL-28 replicates were treated with Doxycycline (GSM8067683, GSM8067684, GSM8067685) and compared to control (GSM8067680, GSM8067681, GSM8067682).
• GSE50686 [35] used shMITF in COLO829; differentially expressed genes were computed between shMITF replicates (GSM1225717, GSM1225718) and scrambled controls (GSM1225721, GSM1225722). Differential expression analysis was performed using the R package limma (v3.54.0) [45], and the microarray platform used was GPL6244.
• GSE16249 analyzed siMITF KD in two metastatic melanoma cell lines, NZM012 and NZM015. We merged siMITF replicates (GSM409128, GSM409132) in these two cell lines to increase the statistical power of the resulting differential gene expression relative to control (GSM409126, GSM409127, GSM409130, GSM409131). The media control and negative control groups were further merged. The R package limma [45] (v3.54.0) was used for differential expression analysis. The experiment was done using microarray platform GPL570.
• GSE227840 [46] used the MYC inhibitor Omomyc to modulate MYC expression in the BRAF-mutant A375 cell line. Differentially expressed genes were provided in the associated supplementary materials [46] (‘Supplementary_Tables.xlsx’, sheet name ‘Table S1A A375 Omomyc in vitro’).
• GSE164726 [47] performed MYC KD using siRNA in A375. Differentially expressed genes for MYC KD sorted by log₂-FC (FDR < 0.05) were provided in the associated supplementary materials [47].

2.14. Construction of Signal Track Figures

MITF CUT&RUN data from Dilshat et al. [32] was acquired from GEO. Called peaks were determined from the associated supplementary materials [32] (‘elife-63093-supp2.xlsx’, sheet name ‘CutandRunMITFChIPseq-all’). All ChIP-seq and CUT&RUN track figures were made using UCSC Genome Browser [28].

2.15. Analysis of BRAF^V600E-Mutant Melanocytes

We used the ENSG00000136997 Ensembl gene ID for MYC when identifying genes in the GEO data GSE255546. The gene-level counts were normalized by the total counts in each sample and then multiplied by 1000 for visualization.

2.16. Analysis of Gene Expression Levels as a Function of MYC or MITF Copy Number

For the functional analyses of MYC and MITF, we chose to use the ABSOLUTE copy number data, instead of the ASCAT3 allele-specific copy number data, because the previous integrative functional study that reported the TCGA-UVM data used ABSOLUTE [3]. We nevertheless checked that our results hold true even when ASCAT3 were used. All boxplots used either TCGA-UVM TPM expression data or ABSOLUTE copy number. SciPy’s (v1.16.0) ranksums function with default parameters was used to compute p-values in the boxplots grouped by MITF copy number. Of the 80 TCGA-UVM patients, 37 had MITF copy = 1, 42 had MITF copy = 2, and 1 patient had MITF copy = 3. The t-test p-values were computed using the ttest_ind function in SciPy v1.16.0 assuming unequal variances. In the comparison of MITF-regulated pigment gene expressions between M3 and D3 groups, the ranksums function was again used to compute p-values, but these were further subjected to genome-wide p-value adjustment using R’s p.adjust function with FDR method.

2.17. Analysis of MITF Correlation with MYC:MAX Targets

The Genomic Regions Enrichment of Annotations Tool [48] (GREAT, v4.0.4) was used to determine MYC:MAX targets using MAX ChIP-seq data in human primary melanocytes [39]. A gene was assigned to be a MYC:MAX target if a MAX ChIP-seq peak resided within 500bp upstream of its TSS. An identical procedure using MITF ChIP-seq in human primary melanocytes [35] was performed for the determination of MITF targets. Genes on chr3 were classified as either being a MYC:MAX target gene or a background (non-target) gene. MITF target genes on chr3 were excluded from this analysis due to the possibility of shared binding sites between MYC:MAX and MITF, both of which could recognize E-box elements. The Spearman correlations between MITF expression and MYC:MAX target gene expressions (TPM) were calculated using TCGA-UVM patients with the MITF copy = 1 (37 patients) or 2 (42 patients). Genes which were not expressed (TPM = 0.0) in patients with MITF copy = 1 or 2 were excluded from their respective analyses. The statistical significance between the MYC:MAX target distribution and background distribution was computed using the Kolmogorov–Smirnov test implemented in SciPy stats (v1.16.0; ks_2samp; default parameters).

2.18. Analysis of ZBTB17 Expression and Deletion in the Contexts of MITF Deletion and MYC Amplification

The relationship between ZBTB17, MITF, and MYC was explored using both TCGA-UVM and TCGA-SKCM expression data by computing the Spearman correlation between ZBTB17 and MITF expression (TPM). Using ABSOLUTE LiftOver copy number, we found ZBTB17 was preferentially co-deleted with MITF in TCGA-UVM patients using a one-sided binomial test to compute the statistical significance of ZBTB17 deletion probability when MITF copy = 1 (37 patients) relative to MITF copy $\ge 2$ (43 patients). Similarly, the statistical significance of co-occurring ZBTB17 deletion and MYC amplification was assessed using a one-sided binomial test to compare ZBTB17 deletion probability between MYC copy $=2$ (20 patients) and MYC copy $\ge 3$ (60 patients). SciPy’s (v1.16.0) binomtest was used in both cases.

2.19. Bivariate Regression Analysis of MITF Expression Using MYC and HUWE1

Utilizing gene expression (TPM) data for MITF, MYC, and HUWE1 in the cohort of 80 TCGA-UVM patients, we performed bivariate regression analysis (Python, statsmodels (v0.14.5); OLS; default parameters) to predict the expression of MITF given MYC and HUWE1 as predictors. Both response (MITF) and predictor (MYC, HUWE1) variables were z-score standardized prior to fitting but transformed back afterwards for visualization. Linear regression was done without a y-intercept, because the data were first mean centered.

2.20. Bimodality Analysis for MITF, MYC, and GNAQ/11

We stratified the TCGA-UVM data and identified 34 patients with MITF copy number = 1 and GNAQ/11 mutations. We used the ABSOLUTE copy number data, and the GNAQ/11 mutation data were retrieved through UCSC Xena Browser [27]. We calculated the bimodality index [49] for MITF, MYC, and GNAQ/11 for these patients. The bimodality index shows the degree to which a 1D distribution can be described as a two-component mixture model and has a recommended value of 1.1 as a cutoff (higher values are deemed bimodal). The R (v4.4.3) implementation, BimodalIndex (v1.1.11), was used to calculate the index and confirm the bimodal distributions of MITF, MYC and GNAQ/11.

2.21. Estimating the Frequency of Uveal Melanoma Patients with M3 and 8q+

We integrated the copy number data from TCGA-UVM (ABSOLUTE LiftOver) and the information from Johansson et al. [2] to estimate the frequency of the M3/8q+ UVM subgroup. Of the 80 TCGA-UVM patients, 37 (46.3%) carried M3, of which 34/37 (91.9%) carried 8q+. M3 and 8q+ status were determined using modal copy number. The Johannson et al. UVM patient data reported 103 patients, 59 of which carried M3 (57.3%) and 46/59 (78.0%) had “gain” or “high gain” in chr8q ploidy. Combined, there were 96/183 (52.5%) of patients with M3, and 80/96 (83.3%) of those patients carried 8q+.

2.22. Survival Analysis

Kaplan–Meier curves and log-rank tests were computed with Python’s lifelines [50] package (v0.30.0). Overall survival data were retrieved from TCGA-UVM. Relapse-free survival from TCGA-UVM was calculated by taking the time to first occurrence of either death or recurrence, which does not include instances of a second primary tumor. MITF and MYC ABSOLUTE copy number were used to determine chr3 and chr8q ploidy for TCGA-UVM data. Data from Johansson et al. labeled as category 3 or 4 (LOH in chr3 with WT copy number chr 8 or 8q+, respectively) were also included. Patients with status ‘Unknown’ were removed from the analysis.

2.23. DepMap CRISPR Screenings

Genome-wide CRISPR KO screening data for UVM cell lines were downloaded from the Cancer Dependency Map (DepMap) [26] (DepMap Public 25Q2+Score, Chronos). We calculated chr3 arithmetic mean expression from expression data available on “DepMap (Expression Public 25Q2),” utilizing all known genes on chr3 from the hg38 GENCODE V48 tracks (retrieved through UCSC Genome Browser Table Browser). WM3772F had no validated chr3 karyotyping data; we thus used chr3 expression analysis as a proxy for chr3 status.

3. Results

3.1. The Pattern of 3p Deletion Reflects Shared Selection Pressure to Delete SETD2 in Multiple Cancers

Chr3, in particular the p arm, is often lost in diverse cancers, including breast cancers [51] (~30%), renal carcinomas [52,53] (~60%), pancreatic adenocarcinoma [54,55] (PAAD) (~20%), and UVM [3] (~50%). In UVM, entire chr3 is lost, making the search for potential tumor suppressor genes difficult (Figure S1A). We thus sought to learn from the pattern of 3p deletion in TCGA breast cancer (BRCA), clear-cell renal cell carcinoma (KIRC), papillary renal cell carcinoma (KIRP), skin cutaneous melanoma (SKCM), and PAAD data. Aligning the copy number profiles of only those patients who had a deletion event on 3p revealed similarities and key differences (Figure 1C and Figure S1B; Section 2.4). Notably, the 5.2 Mb region containing SETD2 and BAP1 was preferentially deleted in all five cancers; but, MITF deletion was prevalent only in BRCA and KIRP, neutral in KIRC (as simulated by Poisson breaks), under negative selection in SKCM (Figure 1C), and mostly neutral in PAAD, but with one patient who had deep MITF amplification in the background of SETD2/BAP1 deletion (Figure S1B). These results strongly supported that in addition to the known tumor suppressor BAP1, SETD2 inactivation also was targeted by multiple cancer types.

3.2. MITF Deletion Is Likely Accidental in Melanocyte-Lineage Cancers

Given the essential role of proper MITF dosage in melanocyte-lineage survival [56,57], we sought to understand the pattern of MITF copy number alterations in the context of partial chr3p deletion in SKCM. MITF is amplified in ~10% of primary and ~20% of metastatic SKCM, leading to the lineage-addiction hypothesis proposing that certain cancers depend on lineage-specific pathways that are regulated by critical differentiation factors and promote cancer initiation/progression [58]. In TCGA-SKCM data, however, 17 (3.7%) and 30 (6.5%) out of 461 patients had a single copy of MITF and SETD2, respectively; 28 patients who had BAP1 deletion also had SETD2 deletion (Figure 1C); and, 6 of the 30 patients with SETD2 copy = 1 had focal amplification of MITF (Figure 1C,D). This pattern indicated that BAP1/SETD2 deletion facilitated skin cutaneous melanomagenesis, but the reduced deletion frequency and focal amplification of MITF in the background of BAP1/SETD2 deletions also suggested that MITF deletion in the melanocyte lineage might have been an undesirable accidental consequence of the proximity between the three genes. Closer investigation revealed that the BAP1/SETD2 locus actually harbored higher frequency of cryptic deletions: 39 patients (8.5%) had local CNL LOH of SETD2 and 33 patients (7.2%) had pan-chr3 LOH, sometimes interspersed by many heterozygous fragments (Figure 1D). Importantly, 20 (28.8%) of these 72 patients had focal MITF amplifications, all in the heterozygous MITF-locus context (Figure 1D); furthermore, 42 patients had duplicated the intact chr3, likely at a later stage of cancer progression after the initial BAP1/SETD2 deletions had already facilitated oncogenesis. Co-occurring partial deletion of 3p and focal amplification of MITF could be explained by break-fusion-bridge [59] cycles of dicentric chromosome 3, while focal MITF amplification in the background of pan-chr3 LOH possessed signatures of extensive fragmentation in chromothripsis [59,60] (Figure 1D). Together, these results strongly supported that codeletion of MITF with BAP1/SETD2 was not under positive selection in SKCM and that, on the contrary, a large fraction of MITF amplification in SKCM was associated with BAP1/SETD2 deletion.

We thus partitioned the MITF amplification events into four groups: shallow amplification (copy = 3, 4) with or without SETD2 LOH, and deep amplification (copy $\ge$ 5) with or without SETD2 LOH, following the published criteria [61]. For shallow amplification, copy number alteration (CNA) in the presence of SETD2 LOH (8%) exhibited a propensity to selectively amplify MITF; in the absence of SETD2 LOH (92%); however, there was no evidence of specifically amplifying MITF over other genes on chr3 (Figure S1C). Deep amplifications showed a clear signature of positive selection on focal amplification of MITF, with 19 (34.5%) out of the 55 deep amplifications occurring in the background context of SETD2 LOH and showing ~2-fold higher amplification than when chr3p was heterozygous (Figure 1D and Figure S1C; Table S1A). Deep MITF amplifications in SKCM were thus significantly associated with SETD2 LOH (Fisher’s exact test p-value =4.7 × 10⁻⁶, odds ratio = 6.3), while shallow MITF amplifications did not exhibit a trace of selection specifically for MITF itself, unless 3p was partially deleted. These results thus revealed that partial 3p deletion crucially contributed to amplifying the oncogene MITF in SKCM and supported that MITF hemizygous deletion in UVM was an accidental byproduct of entire chr3 loss.

3.3. Deleterious Mutations of SETD2 Are Significantly Associated with Isochromosome Formation in Papillary Renal-Cell Carcinoma

Frequent SETD2 mutations were previously reported in the TCGA-KIRP cohort [53]. Given the uncertainty in determining the functional consequence of missense mutations, we focused on unambiguous deleterious mutations. We identified nine KIRP patients harboring a stop-codon gain or a frame-shift mutation in SETD2 at a minimum allele-frequency of 0.5. Out of these nine, five patients had $3\pm 2$ isochromosomes (Table S1B). We found a significant association between the high-frequency deleterious SETD2 mutation status and the presence of isochromosomes in 269 KIRP patients with intersecting data (Fisher’s exact test p-value = 3.0 × 10⁻⁴, odds ratio = 16.4). Other TCGA cancers, including KIRC, did not have enough patients with high-frequency deleterious SETD2 mutations to test for association. In UVM, no patient had a deleterious SETD2 mutation meeting the specified criteria.

3.4. Deletion of SETD2 or PSIP1 Is Strongly Associated with Isochromosome Formation in the Majority of TCGA Cancer Types

In addition to or in lieu of deleterious point mutations, cancer cells often suppress a gene’s function by deleting one copy of the gene. In particular, we hypothesized that although SETD2 is seldom mutated in UVM, its reduced dose in M3 UVM patients might have facilitated isochromosome formation. Importantly, a critical gene deletion event, e.g., transiently activating CIN, might later get compensated by amplifying the second copy to avoid mitotic catastrophe [62]. We thus examined not only copy number changes but also the LOH status of SETD2 in 33 TCGA cancer types, spanning 10,632 patients (Section 2.7). In addition to SETD2 on chr3, the only known writer of H3K36me3, we also examined PSIP1 on chr9, the only known reader of H3K36me3 involved in homologous recombination and RAD51 recruitment to double-strand breaks (DSBs) [63]. The LOH status of SETD2 was significantly associated with isochromosome formation in non-acrocentric chromosomes in 10 (30%) out of 33 TCGA cancers (Wald-test FDR threshold of ${10}^{-2}$, Figure 2A; Table S1C,D; Section 2.7); chr3 was excluded in this analysis to avoid correlations arising from over-counting isochromosomes that themselves led to SETD2 LOH. Likewise, PSIP1 LOH was significantly associated with isochromosome occurrence, excluding chr9, in 17 (52%) out of 33 TCGA cancers (Wald-test FDR threshold of ${10}^{-2}$, Figure 2A), sharing all 10 cancers with the SETD2 LOH result. In UVM, 36 out of 77 patients (47%) with ASCAT3 data had one copy of SETD2, and five additional patients had SETD2 LOH; three patients had PSIP1 deletions, and two additional patients had PSIP1 LOH. SETD2 LOH and PSIP1 LOH synergized, and the double-LOH status of SETD2/PSIP1 was significantly associated with isochromosome occurrence compared to the complement set and the double-heterozygous SETD2/PSIP1 cohort, with the most pronounced effect seen in KIRP (Figure 2A, Table S1D). Isochromosomes occurred most frequently on chr8 (99% 8q+ including MYC, 1% 8q-), followed by chr3 (96% 3p-including SETD2/BAP1, 4% 3p+) and chr17 (98% 17p- including TP53, 2% 17p+) (Figure 2B). Isochromosomes thus led to critical oncogene amplifications and tumor suppressor deletions across cancer types.

Figure 2. SETD2 and PSIP1 deletions are associated with isochromosome formation and chromosome fragmentation across multiple cancers. (A) Benjamini–Hochberg-corrected Wald-test p-values comparing isochromosome formation rates in patients with vs. without LOH in either SETD2 (top left) or PSIP1 (top right), and in patients with double-LOH SETD2/PSIP1 vs. the complement (bottom left) or double-heterozygous SETD2/PSIP1 (bottom right). Red indicates passing the statistical significance threshold (adjusted p-value < 0.01) for each of the 33 TCGA cancers. (B) Isochromosome rates as probability of occurrence per patient for each cancer type; e.g., ~50% of LUSC patients have isochromosome 5. (C) Wald-test p-value for observing higher isochromosome occurrence probability in patients with LOH at each chromosome location compared to patients without LOH at the same location in TCGA-KIRP. Chr3 (left) and chr9 (right) locations were evenly sampled at 100kb intervals, and the transformed p-values were smoothed using a running window of 20 for visualization. (D) Illustration of how a dicentric isochromosome can fragment during cell division. (E) Examples of increased fragmentation as a function of the distance from centromere (yellow dotted line) to fusion point (black dot) on chr1 isochromosomes of TCGA-BRCA patients. Gray regions indicate unavailable data. (F) Number of heterozygous fragments on the LOH arm as a function of the distance of fusion site from centromere on 11,658 isochromosomes in 33 TCGA cancers. The shaded region indicates the 95% confidence interval of fragment numbers predicted by univariate linear regression against the distance, including an intercept. (G) Fragment-size distribution of 2188 isochromosomes detected on metacentric chr1,2,3 in 33 TCGA cancers. The patient data (black) agree well with a dynamic fragmentation model (red) and deviates from the Poisson model using the same estimated mean (blue).

As segmental deletions and amplifications tend to have long-range correlations, we further examined the statistical association of LOH status at specific locations on entire chr3 or chr9 with isochromosome occurrence on the remaining chromosomes. We performed the analysis in seven TCGA cancer types that showed the highest levels of association between SETD2/PSIP1 LOH and isochromosome formation (Figure 2C and Figure S2A,B). In six out of the seven cancers, the 3p arm containing SETD2 showed much higher association than the 3q arm with isochromosome formation on the remaining non-acrocentric chromosomes; KIRC was an exception to this pattern, because we removed chr3 isochromosomes which were most prevalent in that cancer. In five out of the seven cancer types, the 9p arm containing PSIP1 clearly showed much higher association with isochromosome formation than the 9q arm; in head-and-neck squamous cell carcinoma (HNSC), even though the p arm was more significant than the q arm, the LOH status on the q arm itself showed very high association with isochromosomes, reflecting the fact that 185 HNSC patients (36%) had pan-chr9 LOH. BRCA and UCEC showed elevated association on the q arm, suggesting that there might be another gene on chr9q contributing to isochromosomes.

3.5. Dicentric Isochromosomes Show Massive Fragmentation Between Centromeres

Isochromosomes can be either monocentric, if fused at the centromere, or dicentric, if fused at the boundary of partial arm-deletion [14] (Figure 1A). As a dicentric chromosome can form a bridge between dividing cells and fragment [59] (Figure 2D), we counted the number of heterozygous copy-number fragments on the arm experiencing LOH. The number grew linearly as a function of the distance between fusion site and centromere, yielding 1 new fragment every ~12Mb on average (Section 2.9; Figure 2E,F and Figure S2C). Dicentric isochromosomes were also significantly more fragmented than non-isochromosomes (Mann–Whitney U-statistic p-value $<5\times {10}^{-324}$, 1.9-fold enrichment; Section 2.10). Furthermore, the size distributions of heterozygous fragments between centromeres on dicentric isochromosomes showed excellent fitting to dynamic fragmentation models of solids under transient force [20] (Figure 2G and Figure S2D; Section 2.11). The fitted exponent $\beta$ being less than 1 indicated heterogeneity in fracture propensity across each chromosome and deviation from homogeneous Poisson fragmentation (Figure 2G). Even though the analyses of fragment numbers and sizes are limited in describing the details of dynamic chromosomal fracture and provide only descriptive statistics that serve as indirect proxies for fragmentation events, these results supported that isochromosomes served as a major source of complex genome alterations, yielding numerous fragmented and amplified segments, reminiscent of chromothripsis [59,60,64]. Some of these amplified fragments contained oncogenes related to cancer progression and poor patient survival. For example, ultra-deep amplifications of the driver genes CCND1 [65] and GAB2 [66] in TCGA-BRCA arose from the fragmentation of chr11q- isochromosomes (Figure S2E,F), similar to the amplification mechanism of MITF in SKCM (Figure 1C,D and Figure S1C).

3.6. MITF Is a Central Hub in UVM Cancer Co-Dependency Network

A recent CRISPR-screening study by Elbatsh et al. [23] reported 29 UVM synthetic-lethal genes, the therapeutic depletion of which significantly impeded proliferation in the presence of GNAQ/11 mutations. To understand the relation among these hit genes, we examined their top 100 co-dependency genes from DepMap [26] database and found that MITF was the top shared gene. Examining the screening data showed that the effect of MITF KO correlated strongly with the reported hits, but perhaps did not meet the threshold criteria employed in that study (Figure S3A). Of note, another independent screening study found MITF to be among the top 10 UVM-specific essential genes [24]. We thus included MITF in the hit list and constructed a co-dependency graph for the 30 genes (Figure 3A); the highest betweenness centrality score of MITF confirmed the visible role of MITF as a central hub in the network (Figure 3B). This central role of MITF was consistent with the strongly selective effect of its dosage perturbation on melanocyte-lineage cancers in DepMap screening experiments (Figure S3B). The effect of MITF KO was most acute in UVM, even more pronounced than in SKCM. These results thus highlighted the prominent role of MITF in UVM biology and strongly supported that MITF hemizygous deletion in the M3 subtype was accidental, as was the case in SKCM (Figure 1C,D, Figures S1C and S3B). Even though these analyses based on correlation structure highlighted the potential role of MITF, confirming causal relations would require functional validations, some of which we provide in the subsequent sections using publicly available genomic and synthetic perturbation data.

Figure 3. MITF is a master regulator of the network of synthetic-lethal genes in UVM. (A) A network of UVM synthetic-lethal genes and MITF constructed from DepMap’s top 100 co-dependency genes and corresponding correlation scores [26]. Genes with degree zero are not shown. (B) The relative betweenness centrality scores for the 8 genes in the first connected component of the co-dependency network in (A). (C) Heatmap of log₂ fold-change (FC) of expression levels in MITF overexpression (top block) and MITF KD/KO/mutation (bottom block) experiments in melanocyte-lineage cells. The figure shows differential regulation of the vulnerability gene CDS2 [24], 29 synthetic-lethal genes from Elbatsh et al. [23], and GNAQ/11. The pigmentation gene MLANA, a known transcriptional target of MITF, is included as a positive control. Black bordering boxes indicate statistically significant differential expression (adjusted p-value < 0.05). Missing data are shown in gray. MITF binding in promoter and intron regions, marked by magenta and teal gene names, respectively, support direct regulation of the indicated genes by MITF. (D) MITF CUT&RUN data in the melanoma cell line SK-MEL-28 [32] showing binding in GNAQ introns. (E) Same as (D), but for GNA11 promoter. (F) Scatter plot of MITF and GNAQ expression levels in TCGA-UVM. Pearson correlation coefficient r is calculated by pooling the M3 and D3 patients together. (G) Proposed regulatory network of MITF and GNAQ/11 in UVM.

3.7. MITF Is a Direct Transcriptional Regulator of Synthetic-Lethal Genes and Forms a Double-Positive Regulatory Feedback Loop with Mutated GNAQ/11

To further elucidate the role of MITF in melanocytes, we examined published experimental data modulating MITF expression level and found that MITF transcriptionally regulated many of Elbatsh et al.’s 29 synthetic-lethal genes [23], as well as cytidine diphosphate diacylglycerol synthase 2 (CDS2) identified by Chan et al. [24] (Figure 3C). In particular, RASGRP3, the downstream effector of GNAQ/11 mutations and the key mediator of the MAPK pathway in UVM [22], was consistently elevated upon MITF overexpression; conversely, MITF suppression consistently reduced RASGRP3 expression (Figure 3C). Furthermore, MITF cleavage under targets and release using nucleosome (CUT&RUN) data [32] in melanocyte-lineage cells showed direct binding of MITF in the RASGRP3 promoter (Figure S3C). The top synthetic-lethal target [23] INPP5A and the UVM-vulnerability gene [24] CDS2 similarly exhibited overall activation by MITF across multiple independent experiments (Figure 3C) and showed direct binding of MITF in their promoters (Figure S3C). Strikingly, among the 29 synthetic-lethal targets, 15 were significantly differentially expressed in 5 or more of the 8 MITF-overexpression experiments; likewise, 15 synthetic-lethal targets showed significant differential expression in 2 or more of the 8 MITF-suppression experiments, with only 7 showing no statistically significant change. Furthermore, MITF CUT&RUN data [32] showed that 16 synthetic-lethal genes were bound by MITF within 1kb of the TSS, with additional four genes bound by MITF in introns (Table S1E). Together, these results demonstrated that MITF directly regulated a large subset of the synthetic-lethal genes in UVM.

Given MITF’s regulation of multiple genes on which UVM critically depended, we next investigated whether MITF also directly regulated GNAQ/11, the two most frequently mutated genes driving UVM progression. MITF CUT&RUN data [32] showed binding in GNAQ introns and GNA11 promoter (Figure 3D,E); the expression levels of MITF and GNAQ, and also GNA11 to a lesser extent, were highly correlated in UVM (Figure 3F, Figure S3D). MITF dosage perturbation data [43] in human primary melanocytes expressing either wild-type BRAF or BRAF^V600E confirmed the transcriptional regulation of many synthetic-lethal genes by MITF, with GNAQ again showing stronger correlation than GNA11 (Figure S3E). In turn, the expression levels of MITF itself and the synthetic-lethal genes regulated by MITF were significantly elevated in choroidal melanocytes engineered [31] to express GNAQ^Q209L (Figure 3C). The increased expression level of MITF may be attributed to the phosphorylation of its regulator CREB by Ca⁺²-dependent protein kinases [67] or Ca⁺²-activated protein kinase A (PKA) [68]. Alternatively, the recently discovered coupling of Ca⁺² sensing and cAMP generation [69] could also contribute to the elevation of MITF expression in GNAQ/11-mutant UVM (Figure 3G, Section 4). These results thus showed that MITF and GNAQ/11 formed a double-positive regulatory feedback loop.

Together, these findings thus explained why MITF was a central hub in the dependency network (Figure 3A), but also raised a paradox that MITF, the master regulator of the synthetic-lethal genes, lost one copy during cancer evolution in ~50% of UVM patients, potentially causing a crisis that cancer cells had to surmount. This paradox also posed an accompanying puzzle: why was MITF not a top hit in Elbatsh et al.’s CRISPR screening, despite being a central hub? We shall now resolve both enigmas by explaining the transcriptional interplay of MITF and MYC in the context of M3 and 8q+.

3.8. MITF and MYC May Form a Transcriptional Regulatory Feedback Loop in Melanocytes

Mean CNA in TCGA-UVM showed frequent global loss of chr3 and whole-arm 8q+ (Figure 4A,B), and these alterations were linked in individual patients (Figure S4A,B), as previously observed [2,3]. Surprisingly, MITF and most of its known transcriptional targets were not differentially expressed between M3 and disomy 3 (D3) genotypes (Figure 4C and Figure S4C), indicating that 8q+ might compensate for MITF hemizygous deletion. One potential mechanism was MYC on 8q24, known to rescue MITF’s general transcriptional activity in melanocytes in the absence of MITF [70]. Consistent with this hypothesis, MYC copy number was significantly increased in patients with MITF hemizygous deletion (Figure 4D); however, MYC expression level again did not correlate with its copy number (Figure 4E). The lack of correlation between expression and copy number for both MITF and MYC indicated the presence of a confounding factor, such as a transcriptional feedback loop for maintaining homeostasis.

Figure 4. Copy number alterations alone cannot explain the expression levels of MITF and MYC in TCGA-UVM. (A) Averaged pattern of chr3 deletions in TCGA-UVM showing that entire chr3 tends to be lost. ABSOLUTE gene-level copy number data are shown. (B) Same as (A), but for chr8 amplifications, showing that only the q arm tends to be amplified. (C) Boxplots of MITF expression levels in M3 vs. D3 patients. Expression level is seen to be sustained when one copy of MITF is lost (Wilcoxon rank-sum test). (D) Boxplots of MYC copy numbers in M3 vs. D3 patients. MYC amplification is significantly associated with MITF deletion (Wilcoxon rank-sum test). (E) Boxplots of MYC expression levels grouped into distinct copy numbers, showing the lack of Pearson correlation.

Independent ChIP-seq [35] in melanocyte and CUT&RUN [32] in melanoma revealed that MITF bound the promoter of MYC as well as PVT1, the adjacent long non-coding RNA known to regulate MYC expression [71,72] (Figure 5A); conversely, MYC:MAX bound MITF in melanocyte [39] and 13 different ENCODE [73] cell lines (Figure 5B, Table S1F). Consistent with these findings, MITF and MYC expression levels showed strong correlation in UVM (Figure 5C). Indirect regulation of MYC by MITF was also previously reported, with MITF depletion leading to significant suppression of MYC [70]. Moreover, several independent MITF modulation experiments confirmed its transcriptional activation of MYC (Figure 5D). In particular, MITF and MYC were both significantly elevated in choroidal melanocytes expressing [31] GNAQ^Q209L (Figure 5D).

Figure 5. Evidence for MITF’s coupled double-positive regulatory feedback loops with MYC and GNAQ/11 in UVM. (A) MITF ChIP-seq [35] in melanocyte, showing binding in the promoter of MYC and PVT1. (B) MAX ChIP-seq [39] in melanocyte, as a proxy for MYC, shows binding in the promoters of several MITF isoforms. (C) Scatter plot of MITF and MYC expression levels in TCGA-UVM, showing a strong Pearson correlation. (D) Heatmap of log₂-FC of expression levels of MITF, MYC and MLANA in MITF overexpression and KD/KO/mutation experiments in melanocyte-lineage cells. MITF and MYC are seen to regulate each other, forming a double-positive feedback loop. MLANA serves as a positive control. Black boxes indicate statistical significance (adjusted p-value $<0.05$). (E) Proposed coupled gene regulatory network of GNAQ/GNA11, MITF and MYC in UVM. Depletion of ZBTB17 in UVM switches MYC:MAX into an activator of MITF.

Regulation of MITF by MYC was more complex. To assess whether MITF level changed together with other known targets of MYC, we computed the distribution of correlations between MITF and other coding genes in UVM. Compared to background, MYC targets showed significantly greater correlation with MITF (Figure S5A,B). However, although MYC KD in A375 suppressed MITF in one study [47], treatment of A375 with the MYC-inhibitor Omomyc increased the level of MITF [46] (Figure 5D), informing that MYC could either activate or repress MITF in a context-dependent manner. Supporting this idea, it is known that although the MYC:MAX dimer functions as an activator, ZBTB17 (a.k.a. MIZ1) can interact with MYC and MYC:MAX to turn them into repressors, and this switch depends on the ratio of MYC and ZBTB17 binding [74]. We found evidence that ZBTB17 bound [73] several locations within MITF (Figure S5C), and its expression was anti-correlated with MITF in SKCM (Spearman $\rho =-0.27$, p-value $=$ $2.7\times {10}^{-9}),$ but not in UVM (Spearman $\rho =0.27$, p-value $=0.02$). Importantly, in TCGA-UVM, ZBTB17 located on chr1p was preferentially co-deleted with MITF (binomial test p-value = 0.012), sometimes because of isochromosome (Figure 1B), and deleted preferentially in 8q+ patients (binomial test p-value = $5.6\times {10}^{-5}$). These findings thus explained the previous report that genes typically repressed by the MYC:MAX:ZBTB17 complex were selectively activated in 8q+ patients [3] and supported the transcriptional activation of MITF by MYC. As ubiquitination and subsequent degradation of ZBTB17 via HUWE1 could also help switch MYC activity from repression to activation [75], we examined the effect of HUWE1 on MITF expression. Bivariate regression of MITF expression against MYC and HUWE1 expression showed significant partial correlation with both genes (Figure S5D). These results thus supported that MYC could activate MITF transcription in UVM. In summary, our integrative analysis provided evidence for coupled double-positive feedback loops involving GNAQ/11, MITF and MYC (Figure 5E), which could function together to maintain UVM homeostasis.

3.9. MYC Amplification Rescues UVM from Accidental MITF Hemizygous Deletion

Given the aforementioned regulatory feedback, we next examined the effect of MYC amplification on MITF expression. Whereas MITF expression diminished upon MITF deletion with MYC copy = 2, MYC amplification restored the expression to D3 levels, with no significant difference between MITF copy = 1 with MYC amplification and MITF copy = 2 (Figure 6A). Conditioned on MITF deletion, MITF, RASGRP3 and INPP5A all showed statistically significant elevation with MYC amplification (Figure 6B–D), restoring or even slightly increasing the levels of RASGPR3 and INPP5A relative to the D3 condition (Figure 6E,F). MYC expression could also be predicted by MYC amplification, conditioned on MITF copy = 1 (Figure S5E), supporting that the negligible marginal correlation was confounded by the regulatory feedback between MITF and MYC (Figure 4E). These results strongly supported that MYC amplification effectively restored the expression levels of MITF and its key target genes critical for UVM oncogenesis. Our findings also informed that the accidental deletion of MITF in M3 UVM was initially sensed by MYC through its diminished activation by MITF [70], leading to 8q+ and 1p- as compensatory mechanisms to maintain the feedback-mediated homeostasis of UVM (Figure 5E).

Figure 6. The regulatory feedback loop between MITF and MYC restores the expression of MITF and synthetic-lethal genes in M3 patients. (A) Distributions of MITF expression levels in patients grouped by MITF and MYC copy numbers. Single-copy MITF with WT copy number MYC results in significantly lower expression relative to either single-copy MITF with amplified MYC or WT copy number MITF. However, MYC amplification effectively compensates for MITF deletion. The t-test p-values are shown. (B) Conditioned on MITF deletion, MITF expression level can be explained by MYC copy number, supporting that MYC amplification rescues MITF expression in M3 patients (Pearson $r$ = 0.42). Dotted red line indicates median MITF expression in the D3 group. (C,D) Same as (B), but for MITF-regulated RASGRP3 and INPP5A, respectively. (E) Boxplots of RASGRP3 expression levels in M3 vs. D3. The difference between the two groups is not significant (Wilcoxon rank-sum test). (F) Same as (E), but for INPP5A. The M3 group had a slightly higher level than D3 (Wilcoxon rank-sum test).

In terms of modeling, synergistic dual positive feedback loops with nonlinear interactions are commonly found in genetic networks to create bistability [76]. Moreover, coupling multiple double-positive feedback loops, similar to the MITF/MYC/GNAQ interactions, was shown to be the ideal network architecture for creating bistable states robust against stochastic fluctuations [77]. The existence of bistable states was supported by the bimodal distributions [49] of GNAQ, MITF and MYC expression in M3 patients harboring GNAQ/11 mutations, jointly exhibiting a high(GNAQ, MITF, MYC) or a low(GNAQ, MITF, MYC) state (Figure S5F), akin to the MITF^low and MITF^high states in SKCM [78,79]. By contrast, GNA11 did not fit a simple bimodal distribution and contained an extra low(GNAQ, MITF, MYC)/high(GNA11) state that could not be explained by MITF regulation alone (Figure S5F).

3.10. MYC Amplification Is a Significant Predictor of Survival Within M3 UVM Patients

UVM patients with co-occurring M3 and 8q+ are known to have poor prognosis [2,3]. Even within the M3 context, we found that TCGA-UVM patients harboring MYC amplification exhibited significantly reduced overall survival compared to those with MYC copy = 2 (log-rank test p-value $=$ $0.018$; Figure 7A). In contrast, Johansson et al. found no statistically significant relationship between 8q+ and relapse-free survival in UVM patients with M3 (log-rank test p-value $=$ $0.11$) [2]. However, aggregating their data with M3 patients in TCGA-UVM showed a significant reduction in relapse-free survival for M3 patients with 8q+, regardless of whether patients with chr3 copy-number-neutral or partial LOH were removed (log-rank test p-value $=$ $0.0095$; Figure 7B) or retained (log-rank test p-value $=$ $0.011$; Figure S6A). These results thus demonstrated that 8q+ synergized with M3 to significantly reduce the survival of M3 patients.

Figure 7. Chr3 deletion worsens prognosis and enhances selective dependence of UVM on MITF and MYC. (A) Kaplan–Meier overall survival curve for TCGA-UVM patients with M3, split into groups with WT or gain in MYC copy number (log-rank test p-value = 0.018). The shaded region indicates a 95% confidence interval; crosshairs mark right-censored data. (B) Same as (A), but for relapse-free survival of combined patients from TCGA-UVM and Johansson et al. [2], only including patients with M3 marked by having hemizygous chr3p ploidy (log-rank test p-value = 0.0095). (C) DepMap UVM cell lines with available CRISPR gene effect scores for MITF and MYC. Chr3 log₂ mean expression is a proxy for chr3 deletion status. UPMM3 and WM3772F (red box) have the lowest chr3 expression and show the highest vulnerability to both MITF and MYC KO. (D–F) DepMap CRISPR gene effect scores in UVM cell lines for MYC, MITF, GNAQ/11, Elbatsh et al.’s [23] synthetic-lethal genes, and the UVM-vulnerability gene CDS2 [24]; genes with missing data are excluded. UPMM3 has M3, while UPMD1 has D3. The dashed line at −1 is a recommended threshold for assessing high dependency (lower values).

3.11. M3 Enhances the Dependency of UVM on MITF, MYC, and Synthetic-Lethal Genes

We reasoned that even though UVM patients with co-occurring M3 and 8q+ had worse prognoses than those with either D3 or D8, they might benefit from a therapy targeting the specific regulatory network on which this UVM subtype critically depended. To identify potential synergistic targets, we examined DepMap’s CRISPR-Cas9 KO gene effect scores [26] for one UVM cell line with M3/8q+ (UPMM3 [80,81,82,83]), one with unverified chr3 status but with MYC amplification (WM3772F [84]), two with D3/D8 (MEL270 [83,85,86] and MEL290 [80,85,87]), and four with D3/8q+ (UPMD1, MEL202, MEL285, and OMM25) [80,81,82,85,87,88] (Figure 7C–F and Figure S6B–F).

While all eight UVM cell lines showed strong dependence on MYC, a common essential gene, the three with the highest MITF dependency were also either GNAQ-mutant (WM3772F, Figure 7D; UPMM3, Figure 7E) or GNA11-mutant (UPMD1, Figure 7F) [80,81,89] and depended on either RASGRP3 or INPP5A as well as CDS2, consistent with their direct regulation by MITF (Figure 3C and Figure S3C). Two (UPMM3, WM3772F) of these cell lines vulnerable to MITF KO had MYC amplification and the lowest levels of overall chr3 expression (Figure 7C), with UPMM3 previously verified as being M3 [80,81,82,83], and they were strongly dependent on both MYC and MITF (Figure 7D,E). However, UPMD1 possessing D3/8q+, although dependent on MITF, showed the least dependency on MYC among the 8 UVM cell lines (Figure 7C,F). Moreover, MITF KO did not have a pronounced effect on other D3 UVM cell lines (Figure S6B–F); likewise, RASGRP3, INPP5A and GNAQ KO also did not have a major effect on two D3 GNAQ-mutant cell lines (MEL202, MEL270) (Figure S6B,C). These results thus supported that chr3 inactivation in the context of GNAQ/11 mutations sensitized UVM cells to targeting MITF and MITF-regulated synthetic-lethal genes and elevated their dependency on MYC, needed to compensate for MITF suppression (Figure 5E). These enhanced dependencies could be explained by our prediction that an acute therapeutic reduction in MITF or MYC in M3 cells would significantly decrease the levels of GNAQ/11 and synthetic-lethal genes regulated by MITF, recreating the crisis encountered by M3 UVM during early oncogenesis. These findings also shed light on why a CRISPR screening using D3 cell lines might not detect MITF as a top candidate [23].

4. Discussion

Chr3p deletion occurs in multiple cancers [6,12,51,90] and diseases [91,92], but understanding the function of this chromosome arm in complex diseases has been challenging. By comparing the pattern of 3p deletion across cancers, we have identified SETD2 deletion to be shared (Figure 1C) and strongly implicated in forming isochromosomes in human cancers (Figure 2A–C and Figure S2A), as recently proposed [14]. In addition to SETD2, which establishes H3K36me3, the reader PSIP1 of this epigenetic modification is also frequently deleted in patients with isochromosomes (Figure 2C and Figure S2B). Although PSIP1 is the only H3K36me3-binding protein known to be involved in homologous recombination and DSB repair [63], our analysis suggests that chr9q may also harbor a gene involved in isochromosome formation (Figure S2B). Incidentally, PHF19 residing on 9q encodes another reader of H3K36me3 which functions as a component of the Polycomb repressive complex 2 (PRC2) [93]. Further investigation is needed to assess whether PHF19 aberration promotes isochromosomes.

This work has analyzed only the copy number LOH information about SETD2 and PSIP1 and did not consider their expression levels. It is, however, plausible that suppression of SETD2 and PSIP1 via aberrant transcriptional regulation might also affect their proper function and cause isochromosomes in human cancers [94] and other diseases [95,96]. In cancer, isochromosomes can have a profound functional impact on genome evolution by creating a transient state of genomic instability responsible for rapid, large-scale genomic rearrangements, likely as an early step during oncogenesis [9,14,59,60,64,97]. By comprehensively analyzing isochromosomes in 33 TCGA cancers, we have shown that dicentric isochromosomes may shatter during cell division, giving rise to massive DNA fragmentation. Despite the complex process of fracture mechanics upon being pulled, the average number of reassembled fragments increases linearly as a function of the distance between fusion point and centromere (Figure 2F), and the size distribution follows the prediction from dynamic fragmentation theory of solids under impulse (Figure 2G and Figure S2D).

This study has also uncovered several hitherto-undocumented critical roles of MITF in both UVM and melanocyte biology. The first role resides in its function as a multifarious regulator of Ca⁺² homeostasis, required for both GNAQ/11-driven UVM survival and normal melanogenesis. The GNAQ/11-mediated production of IP3 in UVM parallels the pigmentation pathway in normal melanocytes. A well-established function of αMSH that initiates melanogenesis is to bind and activate the melanocortin 1 receptor (MC1R), which predominantly interacts with the G-protein GNAS to generate cAMP, which in turn activates the transcription of MITF and other pigmentation-related genes [25]. A less well-understood function of αMSH is to also activate the phospholipase Cβ (PLCβ) [98] which hydrolyzes phosphatidylinositol 4,5-bisphosphate (PIP₂) into DAG and IP3 (Figure 3G)—the same two messengers that are constitutively produced in GNAQ/11-driven UVM; it is currently unknown which melanocortin receptor directly interacts with GNAQ and/or GNA11 to initiate this signaling cascade. In normal melanocytes, the elevated level of IP3 then induces Ca⁺² release from the endoplasmic reticulum (ER) and oligomerization of STIM1 at the ER-plasma membrane junction, which in turn leads to Ca⁺² influx into melanocytes from the environment and activates the adenylyl cyclase ADCY6 to maintain the generation of cAMP [69] (Figure 3G). This positive feedback loop between cAMP and Ca⁺² sensing is thought to be essential for sustaining the high level of cAMP and transport of L-phenylalanine into the cell for robust melanogenesis [98,99]. These phenomena highlight the importance of maintaining the balance of Ca⁺² concentration for normal melanocyte function.

A recent study has shown that MITF transcriptionally activates STIM1 [99], the recipient of the message from IP3. We here have shown that MITF also regulates the annihilator of the message, INPP5A. This MITF-mediated Ca⁺² homeostasis mechanism built into melanocytes for melanogenesis may create a strong dependency of melanocyte-derived melanomas, such as UVM, on the clearing potential of INPP5A when evolutionary forces select for potent oncogenic mutations, such as the GNAQ/11 mutations. Melanocytes therefore may have been ready to resolve the constitutive IP3 production and Ca⁺² flux changes arising as an inadvertent byproduct of GNAQ/11 mutations, but this acute dependency also creates vulnerability and a valuable opportunity for therapy. In the absence of INPP5A, constantly accumulating IP3 triggers p53-dependent apoptosis and thus selective synthetic lethality in GNAQ/11-driven UVM [23]. Supporting this vulnerability, calcium channel blockers have shown efficacy in treating UVM [100].

The second role of MITF our study has discovered is its transcriptional regulation of GNAQ/11, as well as RASGRP3 encoding the guanine nucleotide exchange factor activated by DAG, one of the two messengers constitutively produced by the GNAQ/11 mutations. In GNAQ/11-mutant UVM, RASGRP3 is a critical mediator of MAPK pathway activation [22] and a synthetic-lethal target [23]. The aforementioned feedback loop of sustaining the cAMP level initiated by constitutive IP3 production implies a positive feedback loop between MITF and mutant GNAQ/11 gene regulation, explaining the dominant dark skin phenotype of GNAQ/11 mutations observed in mice [101]. Consistent with our model, choroidal melanocytes engineered to express GNAQ^Q209L express a significantly higher level of MITF [31]. In turn, our finding that MITF directly regulates RASGRP3 further explains why GNAQ/11 mutations increase the RASGRP3 mRNA level, even in SKCM [22]. The positive feedback loop between MITF and GNAQ/11 thus not only helps sustain the level of GNAQ/11 but also elevates the level of the key effector protein RASGRP3.

The third critical role of MITF lies in its central position at the hub of the co-dependency network of previously reported synthetic-lethal genes (Figure 3A). We have shown that the centrality of MITF in the network can be partly attributed to the direct transcriptional regulation of many synthetic-lethal genes by MITF itself (Figure 3C–F, Figure S3C,D). MITF is thus a comprehensive modulator of the synthetic-lethal gene network. These observations predict that targeting MITF will perturb several synthetic-lethal genes simultaneously in UVM, suggesting an efficacious combinatorial therapy of targeting MITF together with other co-dependent genes using small molecules [23,102,103]. For some UVM subtypes, targeting MITF alone may even suffice to profoundly inhibit the proliferation of cancer cells (Figure 7C–F).

Namely, we have found a source of evolutionary selection force that can explain the preferential co-occurrence of M3 and 8q+ in UVM, and this subtype is particularly vulnerable to targeting MITF. Being at the hub of the synthetic-lethal gene network and directly regulating many of the essential genes, the inadvertent hemizygous deletion of MITF in the M3 background is disadvantageous for UVM, requiring a dosage compensation by amplifying its transcriptional regulator MYC on 8q and deleting the repressor ZBTB17 on 1p. MYC amplification is sufficient to bring MITF expression back to the D3 level and thereby rescue UVM cells from the inadvertent reduction in synthetic-lethal genes caused by M3 (Figure 6). This recurrent pattern of genomic evolution found in ~50% of UVM, representing the most aggressive subtype (Figure 7A,B), thus imposes acute co-dependency between the two oncogenes and suggests that reduced-dose combinatorial targeting of MYC and MITF together with RASGRP3, INPP5A and CDS2 should create potent synthetic lethality in patients harboring both M3 and 8q+ (Figure 7C–E), while minimizing toxicity. Although traditionally considered undruggable, new inhibitors targeting MITF, MYC, or MYC:MAX heterodimerization are under active development [102,103,104] and will facilitate the exploration of this therapeutic strategy.

Further study is needed to investigate the role of the regulatory feedback loop between MITF and MYC in SKCM. Unlike UVM, SKCM cancers tend to activate the MAPK pathway by mutating BRAF, and MYC activation has been shown to be sufficient and necessary to develop resistance to BRAF inhibition [105]. As BRAF inhibition leads to MITF accumulation and enhances its transcriptional activity (Figure 3C), our finding that MITF binds MYC/PVT1 and activates MYC expression helps explain the transcriptional therapeutic response leading to a MYC-mediated resistance mechanism. Consistent with our hypothesis, MITF suppression increases the sensitivity of SKCM to BRAF inhibition [102].

5. Conclusions

Our study has identified the evolutionary force driving the genomic alteration landscape of ~50% of UVM. We have shown that the nested double positive feedback loops involving MITF, MYC and GNAQ/11 are paramount to melanocyte function and that this co-dependency provides an integrated conceptual framework for understanding the previously reported synthetic-lethal targets and the key effectors of GNAQ/11 mutations.