Article https://doi.org/10.

1038/s41467-024-46906-4

Uveal melanoma immunogenomics predict

immunotherapy resistance and
susceptibility
Received: 20 June 2023 Shravan Leonard-Murali1,2,3,4, Chetana Bhaskarla 1,2,3, Ghanshyam S. Yadav1,2,3,
Sudeep K. Maurya1,2,3, Chenna R. Galiveti1,2,3, Joshua A. Tobin1,2,3, Rachel J. Kann1,
Accepted: 8 March 2024
Eishan Ashwat1, Patrick S. Murphy1,2, Anish B. Chakka 5, Vishal Soman5,
Paul G. Cantalupo5, Xinming Zhuo6, Gopi Vyas6, Dara L. Kozak6, Lindsey M. Kelly6,
Ed Smith 6, Uma R. Chandran1,5, Yen-Michael S. Hsu1,7,8 &
Check for updates Udai S. Kammula 1,2,3
1234567890():,;
1234567890():,;

Immune checkpoint inhibition has shown success in treating metastatic

cutaneous melanoma but has limited efficacy against metastatic uveal mela-
noma, a rare variant arising from the immune privileged eye. To better
understand this resistance, we comprehensively profile 100 human uveal
melanoma metastases using clinicogenomics, transcriptomics, and tumor
infiltrating lymphocyte potency assessment. We find that over half of these
metastases harbor tumor infiltrating lymphocytes with potent autologous
tumor specificity, despite low mutational burden and resistance to prior
immunotherapies. However, we observe strikingly low intratumoral T cell
receptor clonality within the tumor microenvironment even after prior
immunotherapies. To harness these quiescent tumor infiltrating lymphocytes,
we develop a transcriptomic biomarker to enable in vivo identification and
ex vivo liberation to counter their growth suppression. Finally, we demon-
strate that adoptive transfer of these transcriptomically selected tumor infil-
trating lymphocytes can promote tumor immunity in patients with metastatic
uveal melanoma when other immunotherapies are incapable.

Significant advances have been made in the treatment of metastatic improve immunotherapeutic strategies for this large group of unre-
cutaneous melanoma (CM) using immune checkpoint inhibition (ICI) sponsive cancers we focused our studies on uveal melanoma (UM), a
targeting the cytotoxic T-lymphocyte-associated protein 4 (CTLA-4), prototypic ICI resistant cancer with low TMB4,7–10. With an annual
the programmed cell death protein 1 (PD-1), and the lymphocyte- incidence of ~6 per million in Europe and the United States, UM is a
activation gene 3 protein (LAG-3)1–5. Unfortunately, ICI therapy has rare cancer accounting for 3% of all melanomas9,10. Although both CM
not shown comparable activity against most other solid tumors, and UM develop from transformed melanocytes, UM uniquely arises
especially those with low tumor mutational burden (TMB)3,6. To from the pigmented epithelium of the uveal tract, an immune

1
UPMC Hillman Cancer Center, University of Pittsburgh, Pittsburgh, PA, USA. 2Solid Tumor Cellular Immunotherapy Program, UPMC Hillman Cancer Center,
University of Pittsburgh, Pittsburgh, PA, USA. 3Division of Surgical Oncology, Department of Surgery, University of Pittsburgh, Pittsburgh, PA, USA.
4
Department of Epidemiology, University of Pittsburgh, Pittsburgh, PA, USA. 5Department of Biomedical Informatics, University of Pittsburgh, Pittsburgh, PA,
USA. 6UPMC Genome Center, University of Pittsburgh, Pittsburgh, PA, USA. 7UPMC Immunologic Monitoring and Cellular Products Laboratory, University of
Pittsburgh, Pittsburgh, PA, USA. 8Division of Hematology/Oncology, Department of Medicine, University of Pittsburgh, Pittsburgh, PA, USA.
e-mail: [email protected]

Nature Communications | (2024)15:2863 1

Article https://doi.org/10.1038/s41467-024-46906-4

privileged site11, and has an unusual predilection to aggressively Somatic copy number alterations included chromosome 3 loss (46%)
metastasize to the liver which results in a dismal prognosis9. In further and 8q gain (85%) (Fig. 1c). No associations were found between TMB
distinction, immunotherapies demonstrating efficacy against meta- and cohort demographics (Supplementary Fig. 1f and Supplementary
static CM have shown disappointing results against UM7,8, leading to Data 2). In sum, clinicogenomic profiling established this patient
speculation that UM is an immunologically ‘cold’ variant of mela- cohort to be broadly representative of advanced UM and the pro-
noma. However, there has been recent therapeutic progress with the cured metastases to have canonical UM driver alterations and
clinical introduction of tebentafusp, a bispecific glycoprotein 100 low TMB.
peptide-HLA-directed CD3 T cell engager, which has intriguingly
improved overall survival in patients with metastatic UM yet has Unbiased tumor transcriptomics reveals T cell-inflamed uveal
demonstrated only limited ability to mediate tumor regression12,13. To melanoma metastases
reconcile these paradoxical findings and develop more effective Current tumor biomarkers for immunotherapy susceptibility, such as
immunotherapeutics for metastatic UM, we sought to build upon our TMB and PD-L1, are rarely used in metastatic UM due to the uniformly
previous discovery that a subset of UM metastases naturally harbor low expression of these markers in this melanoma variant3,4,17. Thus, we
tumor infiltrating lymphocytes (TIL) with potent autologous anti- first sought to discover alternative immune prognostic metrics by
tumor reactivity14 and that adoptive cell therapy (ACT) administering interrogating the transcriptome of UM metastases using total RNA
such TIL could mediate cancer regression in 35% of patients with sequencing and unbiased computational profiling. Further, to facilitate
metastatic UM, including individuals who were refractory to ICI15. a clinically relevant and minimally invasive biopsy approach for in situ
These observations suggested that occult immune responses exist tumor characterization, we restricted our analysis to a single random
and can be exploited to treat metastatic UM. biopsy from each resected metastasis (~2 mm central core fragment
In this work we perform comprehensive immunogenomic pro- from 93 metastases and ~500,000 cells post tumor dissociation from 7
filing on the largest and most diverse group of human UM metastases metastases). Principal component analysis (PCA) revealed the majority
compiled to date to uncover the tumor microenvironmental prop- of transcriptional variance among the metastases was restricted to PCs
erties that underlie its occult immunogenicity and promote its resis- 1, 2, and 3 (variance contributed: 19%, 13%, 12% respectively) while the
tance and susceptibility to different classes of immunotherapy. We remaining PCs (4–100) each contributed ~5% or less variance (Sup-
find that over half of these metastases harbor TIL with potent auto- plementary Fig. 2a). Metastases were then mapped according to the
logous tumor specificity, despite having low tumor mutational bur- three main PC coordinates (PCs 1, 2, and 3) (Supplementary Fig. 2b, c).
den and resistance to prior immunotherapies, including ICI and the To determine whether specific cellular pathways and processes were
bispecific T cell engager tebentafusp. These T cell infiltrated metas- associated with specific PCs, we correlated PC coordinates (1, 2, and 3)
tases display activated antigen presenting cells, chronic interferon with enrichment scores for each of the canonical hallmark gene sets
signaling, and diverse T cell receptor repertoires. However, we from the Human Molecular Signatures Database (Supplementary
observe strikingly low intratumoral T cell receptor clonality and Data 3)18. Unsupervised clustering (Euclidean distance) of the PC-gene
transcriptionally non-proliferative TIL within the tumor micro- set correlations (Spearman’s rho) identified 4 discrete clusters (A, B, C,
environment even after ICI and tebentafusp therapy, demonstrating and D) with unique biologic motifs (Fig. 2a). Cluster A included cellular
that these immunotherapies were insufficient to induce proliferation metabolism pathways (MYC TARGETS V1, MTORC1 SIGNALING, OXIDA-
of the tumor reactive TIL. To harness the therapeutic potential of TIVE PHOSPHORYLATION), Cluster B included immune and inflamma-
these quiescent TIL, we develop rapid tumor transcriptomic profiling tory signaling pathways (INTERFERON ALPHA RESPONSE, INTERFERON
to enable their selective in vivo identification and ex vivo liberation to GAMMA RESPONSE, ALLOGRAFT REJECTION, IL2 STAT5 SIGNALING),
counter their growth suppression. We demonstrate that adoptive Cluster C included liver dominant physiologic pathways (BILE ACID
transfer of these transcriptomic selected TIL can promote tumor METABOLISM, COAGULATION, CHOLESTEROL HOMEOSTASIS), and
immunity in patients with metastatic UM when other immu- Cluster D included cellular signaling and division (WNT BETA CATENIN
notherapies are incapable. SIGNALING, MYC TARGETS V2, G2M CHECKPOINT). The individual
metastatic samples were further clustered by their relative expression
Results of each of the hallmark gene set clusters to reveal striking variability
Clinicogenomic landscape of metastatic uveal melanoma across the tumor cohort (Fig. 2b). Having identified transcriptomic
One hundred metastases were surgically procured from 84 UM differences among the metastases, we next sought to determine
patients as part of eligibility screening for TIL ACT clinical trials at the whether any of the three PCs independently correlated with the
National Cancer Institute and the University of Pittsburgh Medical expression of the gene set clusters. Average Spearman’s rank correla-
Center between 2013 and 2022 (NCT01814046 and NCT03467516)15,16. tion coefficients (rho) for each of the gene set cluster enrichment
Resected metastases originated from 11 unique anatomic locations scores (A, B, C, and D) were mapped against the individual PCs (1, 2,
(Fig. 1a), with liver as the predominant procurement site (56%) and 3) (Fig. 2c). We observed that cluster A (cellular metabolism) was
(Fig. 1b). Patient demographics revealed a median age of 56 years strongly correlated with PC3 (mean rho = +0.76) but also weakly cor-
(range = 17–78) and an even gender distribution (52% female, 48% related with the negative aspect of PC1 (mean rho = −0.27). Cluster B
male) (Supplementary Data 1). Patients had extensive metastatic dis- (immune and inflammatory signaling) was exclusively correlated with
ease burdens, with 95% having liver involvement, 75% having elevated the negative aspect of PC2 (rho = −0.32). Clusters C and D were not
LDH levels, and 71% with M1B or M1C stage (AJCC 8th edition) (Fig. 1c). found to independently correlate with any of the three PCs. Given the
Metastases were harvested from both treatment naïve patients (24%) independent association of PC2 with Cluster B immune pathways, we
and treatment refractory patients (76%). Notably, 46 patients received postulated that PC2 coordinate position was predominantly driven by
prior ICI therapy (anti-CTLA4 only = 3, anti-PD-1 only = 11, sequential intrinsic immune and inflammatory gene expression in these metas-
therapy = 8, combination therapy = 24) and 9 patients received tases. As support, when the enrichment scores for T cell activation
tebentafusp, of whom none showed objective response (Fig. 1c). gene sets were mapped onto three-dimensional PCA plots of the
Somatic mutational analysis of the metastases confirmed a low TMB metastases, we observed that each gene set had a significant inverse
(median = 0.64 mutations per megabase) with ubiquitous and relationship with the PC2 axis (INTERFERON GAMMA RESPONSE versus
mutually exclusive presence of established UM driver mutations PC2: rho = −0.56, p = 2.52e−9; INTERFERON ALPHA RESPONSE versus
(GNAQ, GNA11, CYSLTR2 or PLCB4) and frequent secondary alterations PC2: rho = −0.56, p = 2.89e−9; ALLOGRAFT REJECTION versus PC2:
of BAP1 (62%) and SF3B1 (42%) (Fig. 1c and Supplementary Fig. 1a–e). rho = −0.49, p = 2.27e−7) (Fig. 2d and Supplementary Fig. 2d).

Nature Communications | (2024)15:2863 2

Article https://doi.org/10.1038/s41467-024-46906-4

Fig. 1 | Clinicogenomic landscape of metastatic uveal melanoma. a Diversity of metastases. Each column represents a single metastasis. BAP1 mRNA z-scores were
source tissues of resected metastases. Created with BioRender.com. b Distribution calculated using log2(normalized counts).
of source tissues of resected metastases. c Clinicogenomic annotation of individual

Nature Communications | (2024)15:2863 3

Article https://doi.org/10.1038/s41467-024-46906-4

Fig. 2 | Unbiased tumor transcriptomics reveals T cell-inflamed uveal mela- metastases. Each heatmap was clustered by metastases separately to display tumor
noma metastases. a Unsupervised clustering of Spearman’s rank correlation heterogeneity within each hallmark cluster. Z-scores were calculated per row.
coefficients derived from correlating PC coordinates (columns) with enrichment of c Matrix of mean Spearman’s rank correlation coefficients for each cluster-PC
hallmark signatures (row) for each individual metastatic sample (n = 100). Natural combination. d Three-dimensional PCA plots displaying enrichment scores for
clusters identified by the row dendrogram are split, labeled (A, B, C, D), annotated, selective hallmark immune related pathways identified in cluster B. Euclidean dis-
and color coded for visualization. b Heatmaps illustrating heterogeneity of hall- tance was used for hierarchical clustering (a, b). Statistical comparisons were
mark signature enrichment across UM metastases (n = 100). Rows correspond to performed using Spearman’s rank correlation (a–d).
hallmark signatures listed in (a). Columns within each heatmap represent individual

Collectively, unbiased computational profiling revealed PC2 coordi- First, we defined the 2394 genes that positively correlated with
nate mapping as an effective initial approach to segregate UM immune and inflammatory hallmark gene set enrichment (those with
metastases with T cell-inflamed transcriptomic attributes. negative PC2 loadings). Rather than biasing this gene list with
supervised filtering, we utilized the entire list of 2394 genes to
Development of Uveal Melanoma Immunogenomic facilitate discovery of novel biologic processes. Further, to enable
Score (UMIS) single-sample prospective analysis, we employed a cohort-
We next sought to refine the rudimentary PC2 variable into a more independent rank-based gene set scoring method (singscore18) to
specific and clinically applicable immune metric for UM metastases. calculate enrichment scores for individual biopsies based upon

Nature Communications | (2024)15:2863 4

Article https://doi.org/10.1038/s41467-024-46906-4

Fig. 3 | Development of Uveal Melanoma Immunogenomic Score (UMIS). UMIS scores across the cohort of 100 metastases. f Gene set enrichment analysis of
a Workflow for the development of UMIS. Created with BioRender.com. differentially expressed genes between high UMIS and low UMIS UM metastases.
b Correlation of UMIS scores calculated by the cohort-independent method, The ten pathways with the lowest FDR are displayed. g Comparison of UMIS by
singscore, with UMIS scores calculated by the cohort-dependent method, gene set source tissue of resected metastases (n = 100 biologically independent samples;
variation analysis (GSVA). c Annotation of UMIS genes using Human Genome liver = 56, subcutaneous = 20, lung = 6, other = 18). h Correlation of UMIS with TMB.
Organization (HUGO) Gene Nomenclature Committee (HGNC). d Functional Statistical comparisons were performed using Spearman’s rank correlation with
annotation of protein-coding genes within UMIS using Database for Annotation, overlaid simple linear regression to illustrate linearity (b, h), DAVID modified
Visualization and Integrated Discovery (DAVID) and Human Molecular Signatures Fisher’s exact test (d), fast preranked gene set enrichment analysis (f), or
Database Gene Ontology Biological Process gene set collection. e Distribution of Kruskal–Wallis test by ranks (g).

transcript abundance (transcripts per million; TPM). Using this Fig. 3c). Thus, UMIS represented a unique single-sample gene
approach, we generated a single continuous variable for each expression score derived from an unbiased mixture of coding, non-
metastasis called Uveal Melanoma Immunogenomic Score (UMIS) coding, and unannotated transcripts that could rank UM metastases
which reflected the concordance and mean percentile rank of our list based upon the expression level of immune and inflammatory genes.
of 2394 genes within the sample transcriptome (Fig. 3a). Cohort-
independent UMIS (singscore) correlated strongly with the corre- UMIS uncovers in vivo drivers of T cell recruitment and
sponding cohort-dependent score calculated using established exclusion
pipelines (GSVA), supporting that our single-sample rank-based tool To characterize the tumor microenvironmental cellular attributes
could be used for prospective evaluation of tumor biopsies without contributing to UMIS, we performed whole-tumor single cell tran-
batch artifact (Fig. 3b)18. Of the 2394 genes that constitute UMIS, 1527 scriptomics of six UM metastases with disparate UMIS values; 3 high
were protein-coding and the remaining 867 were a mixture of non- UMIS (0.300, 0.268, 0.264) versus 3 low UMIS (0.199, 0.178, 0.162)
coding, unclassified, and pseudo genes (Fig. 3c and Supplementary (Supplementary Fig. 4a, b). We cataloged the 93,670 analyzed cells by
Data 4). Functional annotation of UMIS coding genes confirmed building a unique metastatic UM single cell atlas using a two-step
pathways related to immune and inflammatory response (Fig. 3d). process that first categorized cells into large buckets (tumor, immune,
UMIS values ranged from 0.114 to 0.347 across the 100 metastases and stroma) then assigned specific cellular and lineage labels (myeloid
with a median score of 0.237, which was used as a cutoff to define versus lymphoid) to the immune cellular fraction (Fig. 4a, b; Supple-
high and low UMIS groups for categorical comparisons (Fig. 3e). mentary Fig. 4c–e and Supplementary Data 8 and 9)19–21. Our single cell
Gene set enrichment analysis of high versus low UMIS metastases analysis of low UMIS tumors had expectedly low numbers of immune
demonstrated that the most significantly enriched pathways were in cells. However, to maintain the true proportional landscape of specific
the high UMIS group and involved T cell activation (Fig. 3f and cell types and avoid manipulation induced transcriptomic changes, we
Supplementary Fig. 3a). UMIS level was observed to be independent profiled the tumor digests without an additional enrichment step. We
of metastatic site (Fig. 3g and Supplementary Data 5), TMB (Fig. 3h observed more lymphoid cells (proportion ratio = 10.50, p = 0.047)
and Supplementary Data 4), somatic mutations and copy number and fewer tumor cells (proportion ratio = 0.88, p = 0.047) in high UMIS
alterations (Supplementary Fig. 3b and Supplementary Data 6 and 7), versus low UMIS metastases (Fig. 4b and Supplementary Fig. 4f). Fur-
and class I human leukocyte antigen (HLA) alleles (Supplementary ther, the composition of these lymphoid fractions differed, with the

Nature Communications | (2024)15:2863 5

Article https://doi.org/10.1038/s41467-024-46906-4

high UMIS metastases being enriched with CD8+ T cells (proportion p = 1.13e−5) (Fig. 4d). Intriguingly, we observed that 9% of the CD8+
ratio = 6.31, p = 6.15e−7) and the low UMIS metastases being enriched exhausted and 20% of the CD8+ cytotoxic TIL retained transcriptomic
with CD4+ T cells (proportion ratio = 0.63, p = 3.32e−4) and T helper/ expression of TCF7, suggesting possible progenitor capability (Sup-
Th17 T cells (proportion ratio = 0.25, p = 0.024) (Fig. 4c). A granular plementary Fig. 5a). Differential gene expression of the lymphoid cells
analysis of the lymphoid cells revealed that the high UMIS metastases revealed the high UMIS metastases had upregulation of genes invol-
were enriched for CD8+ exhausted T cells (proportion ratio = 40.73, ving T cell activation (TNFRSF9, TNFRSF4), T cell exhaustion (PDCD1,
p = 4.04e−7) and CD8+ cytotoxic T cells (proportion ratio = 4.70, CTLA4, LAG3, HAVCR2, VSIR), lymphocyte activation (STAT1),

Nature Communications | (2024)15:2863 6

Article https://doi.org/10.1038/s41467-024-46906-4

Fig. 4 | UMIS uncovers in vivo drivers of T cell recruitment and exclusion. grouped by the UMIS level of their metastasis. The genes included had log2(fold
a Uniform manifold approximation and projection (UMAP) plot of all cells analyzed change) ≥ |0.5| and FDR < 0.05. Z-scores were calculated per row. i Selected genes
from 6 UM metastases. Magnified panel shows immune subset of cells after from differential gene expression analysis of high UMIS versus low UMIS tumor
reclustering. Cell labeling is with a broad classification. b Proportion of overall cell cells. Bars indicate medians and log2fc refers to log2(fold change). UMAP plots
types within UMIS groups. Fold enrichment refers to proportion ratio (high UMIS/ display all cells within each UMIS subset. j Correlation of UMIS with immune
low UMIS). c Proportion of lymphoid broad cell types within UMIS groups. Fold resistance program scores33 in UM metastases (n = 100). k Comparison of immune
enrichment refers to proportion ratio (high UMIS/low UMIS). d Volcano plot of resistance program scores33 by UMIS level in UM metastases (high UMIS n = 50, low
lymphoid granular cell types within UMIS groups. Fold enrichment refers to pro- UMIS n = 50; total n = 100 biologically independent samples). l Correlation of
portion ratio (high UMIS/low UMIS). e Selected genes from differential gene SNHG7 with CTNNB1 transcript expression in UM metastases (n = 100). Units are
expression analysis of high UMIS versus low UMIS lymphoid cells. Bars indicate log2(normalized counts) from bulk RNAseq. m Correlation of SNHG7 with cano-
medians and log2fc refers to log2(fold change). f Proportion of myeloid broad cell nical melanoma marker transcripts (S100A1, SOX10, MITF) in UM metastases
types within UMIS groups. Fold enrichment refers to proportion ratio (high UMIS/ (n = 100). Units are log2(normalized counts) from bulk RNAseq. Statistical com-
low UMIS). g Selected genes from differential gene expression analysis of high parisons were performed using propeller (arcsin square root transformation of
UMIS versus low UMIS myeloid cells. Bars indicate medians and log2fc refers to proportions) (b–d, f), Wilcoxon rank-sum test (two-tailed) (e, g, i, k) and Spear-
log2(fold change). h Heatmap of differentially expressed genes between high UMIS man’s rank correlation with overlaid simple linear regression to illustrate linear-
and low UMIS tumor cells. Columns are individual cells, rows are genes. Cells are ity (j, l, m).

interferon response (HLA-A, HLA-B, HLA-C, B2M, IFNGR1, IRF1, IFI27, IFI6, several cancers (Fig. 4h, i and Supplementary Data 12)34–38. Our findings
IFITM1, IFITM2, IFITM3), T cell memory (IL7R), lymphocyte trafficking confirmed a strong association between SNHG7 and CTNNB1 expres-
(CXCL13, CCR7, SELL, CXCR3), and T cell progenitor capability (TCF7) sion level in UM metastases (n = 100) (Fig. 4l) that was independent of
(Fig. 4e; Supplementary Fig. 5b and Supplementary Data 10)22–28. Taken tumor cell abundance as measured by melanoma-specific gene
together, these data demonstrate that the TIL found in high UMIS expression (S100A1, SOX10, MITF) and total RNA quantity (ACTB,
metastases had undergone activation and effector differentiation GAPDH) (Fig. 4m and Supplementary Fig. 6c).
consistent with an in vivo adaptive anti-tumor response and indicative In sum, single cell transcriptomics demonstrated that UMIS was a
of a T cell-inflamed microenvironment. holistic metric that reflected the gene expression of the lymphoid,
We next investigated the myeloid cells found in high UMIS versus myeloid, and tumor compartments within the tumor microenviron-
low UMIS metastases (Fig. 4f and Supplementary Fig. 5c). Although ment. Further, UMIS classification of metastases revealed increased
there was no enrichment of specific myeloid cell types (macrophages, CTNNB1 expression by low UMIS tumors cells as a putative driver of
dendritic cells, mast cells) in either group, differential gene expression immune exclusion. In contrast, high UMIS metastases displayed lower
revealed the high UMIS myeloid cells had upregulated genes involving tumor cell CTNNB1 expression, more activated APCs, greater CD8+ T
CD8+ T cell recruitment (CXCL10, CXCL9), tumor phagocytosis cell recruitment, and robust interferon signaling.
(SLAMF7), antigen processing (TAP1, TAP2), antigen presentation (HLA-
A, HLA-B, HLA-C, B2M, HLA-DPB1, HLA-DQB1, HLA-DRB1) and interferon UMIS predicts anti-tumor potency of ex vivo expanded TIL
response (IRF1, IRF8, IFI27, IFI6) (Fig. 4g; Supplementary Fig. 5d and To validate the transcriptomics demonstrating T-cell inflamed gene
Supplementary Data 11)24,25,28–31. These findings support that the T cell- expression, we next interrogated the specific anti-tumor potency of
inflamed microenvironment found in high UMIS metastases also the endogenous TIL from each of the UM metastases. Currently, the
included more active myeloid lineage antigen presenting cells (APCs) assessment of TIL tumor reactivity requires patients to undergo sur-
capable of recruiting CD8+ T cells. gical resection of metastases followed by several weeks of ex vivo TIL
Finally, since UMIS was derived using unbiased whole-tumor expansion and finally resource intensive coculture with autologous
transcriptomics we postulated that this score may also reflect intrinsic tumor cells. Thus, we also investigated whether UMIS could serve as a
differences among the tumor cells within high versus low UMIS rapid and minimally invasive clinical tool to predict the tumor specific
metastases. Upon reclustering of tumor cells, we confirmed distinct potency of endogenous TIL. We compared UMIS values, derived from
separation of cells derived from high versus low UMIS metastases a single random biopsy from each source metastasis (n = 100), with the
(Supplementary Fig. 6a). Differential gene expression revealed high level of TIL anti-tumor reactivity found after conventional ex vivo
UMIS tumor cells had significantly increased expression of several expansion (Fig. 5a). TIL cultures (n ~ 24) were initiated from each
interferon-inducible transcription factors and elements (IRF1, IFI27, freshly resected metastasis using a standardized ex vivo tumor frag-
IFI6, IFITM1, IFITM2, IFITM3) and each of the major histocompatibility mentation approach to address tumor heterogeneity, as previously
complex (MHC) class I molecule heterodimer components (HLA-A, described15. The individual TIL fragment cultures were tested for
HLA-B, HLA-C, B2M) (Fig. 4h, i; Supplementary Fig. 6b and Supple- tumor specificity by coculture with autologous tumor digest (versus
mentary Data 12)23–25. These findings suggested that high UMIS normal tissue controls) followed by measurement of 4-1BB upregula-
metastases were composed of IFN-γ primed tumor cells that had tion on CD3+ cells (flow cytometry) and IFN-γ release (ELISA), which
upregulated MHC class I expression in response to chronic IFN-γ were found to be strongly correlated (Fig. 5b, c). The percentage of TIL
secretion from tumor specific CD8+ T cells23,24. In contrast, low UMIS cultures having tumor-specific reactivity from each metastasis was
tumor cells had 1.44-fold higher expression of CTNNB1 (log2(fold used as a standardized reactivity metric for comparing the level of anti-
change) = −0.53, FDR ~ 0) which encodes the beta-catenin protein tumor TIL responses across tumors (Fig. 5d). We found the frequency
(Fig. 4h, l and Supplementary Data 12). Activation of the Wnt/beta- of tumor reactive TIL cultures varied significantly among the total
catenin pathway has been implicated in T cell exclusion and may cohort (median = 6%; range = 0–100%) with 55 metastases having
explain the paucity of CD8+ T cell infiltrate in low UMIS metastases30,32. measurable anti-tumor reactivity and the remaining 45 metastases
In support, we found a significant inverse relationship across the total with no detected reactivity (Fig. 5e). Further, the metastases that had
metastatic cohort (n = 100) between UMIS and the expression of a undergone prior ICI (n = 53) and tebentafusp therapy (n = 12) showed
previously reported immune resistance program (Fig. 4j, k)33. Inter- no difference in the mean percentage of tumor reactive TIL cultures
estingly, the most upregulated gene in low UMIS tumor cells was the when compared to samples that had not undergone these treatments
long non-coding RNA, SNHG7, which was 3.48-fold upregulated in low (ICI 22% vs. no ICI 23%, p = ns; tebentafusp 32% vs. no tebentafusp 21%,
UMIS tumor cells (log2(fold change) = −1.80, FDR ~ 0) and has been p = ns) (Fig. 5e and Supplementary Data 13). The percentage of tumor
previously reported as a positive regulator of CTNNB1 expression in reactive TIL cultures was also independent of metastatic site, TMB,

Nature Communications | (2024)15:2863 7

Article https://doi.org/10.1038/s41467-024-46906-4

Fig. 5 | UMIS predicts anti-tumor potency of ex vivo expanded TIL. a Workflow f Correlation of UMIS with percent tumor reactive TIL cultures. Color of each point
for parallel analysis of source tumor transcriptomics and expanded TIL anti-tumor denotes tumor digest viability percentage. g Correlative benchmarking of UMIS
reactivity. Created with BioRender.com. b Example of TIL culture anti-tumor against published gene expression profiles and tumor biomarkers (n = 100 metas-
reactivity screening from source tumor UM #100. From left, tumor fragments tases). All correlations are with percent tumor reactive TIL cultures. h Predictive
(n = 24) are cultured individually for ~2 weeks before overnight coculture with benchmarking of UMIS against published gene expression profiles and tumor
autologous tumor cells and measurement of 4-1BB (CD137) expression by flow biomarkers (n = 100 metastases). Receiver operating characteristic (ROC) curves
cytometry and IFN-γ release by ELISA. Final reactivity measurement subtracts and accompanying statistics are for variables’ prediction of ≥33% tumor reactive TIL
background reactivity of TIL (TIL alone) and non-specific reactivity (TIL + cultures. i Disparate UMIS and TIL culture reactivity from synchronous hepatic
autologous APCs). c Correlation between %4-1BB + CD3+ cells and IFN-γ release metastases in UM patient #1. Created with BioRender.com. j Validation of UMIS’
among the 24 fragment cultures from UM #100 after overnight tumor coculture. ability to predict ex vivo TIL reactivity in an independent metastatic biopsy cohort
d Individual TIL fragment culture anti-tumor reactivity as assessed by 4-1BB upre- (n = 20 metastases). Receiver operating characteristic (ROC) curve and area under
gulation and IFN-γ release from source tumor UM #100. TIL cultures were classified curve (AUC) value is for UMIS’ prediction of ≥33% tumor reactive TIL cultures.
as tumor reactive if their 4-1BB expression was >1% (dotted line) and twice back- Statistical comparisons were performed using Spearman’s rank correlation with
ground or IFN-γ release was >100 pg/ml (dotted line) and twice background. Per- overlaid simple linear regression to illustrate linearity (c, f, g) or univariate logistic
centage tumor reactive TIL cultures was defined as 100*(tumor reactive TIL regression (h, j). Gene expression profiles were calculated using singscore to best
cultures)/(total TIL cultures). e Distribution of percent tumor reactive TIL cultures assess their cohort-independent predictive ability (Supplementary Data 16)30,39–41.
among the cohort of 100 metastases color-coded by pre-harvest treatments.

specific mutation expression, copy number alterations, and class I HLA as a preoperative threshold to avoid futile surgical resection of non-
alleles (Supplementary Fig. 7a–d and Supplementary Data 13–15). inflamed UM metastases. Interestingly, we did observe a small subset
When UMIS of each source metastasis was compared to the percen- of discordant metastases (n = 16) with high UMIS values that yielded
tage of tumor reactive TIL cultures that were generated several weeks TIL with no detectable anti-tumor reactivity based upon coculture with
later, we found a strong positive correlation (rho = +0.47, p = 7.06e−7) autologous tumor digest (Fig. 5f). However, upon assessing the quality
(Fig. 5f). Notably, reactive TIL cultures were rarely expanded from of these specific tumor digest samples, we found they had significantly
metastases with a UMIS less than 0.2, suggesting the use of this cutoff lower viability when compared to digests (n = 34) that yielded

Nature Communications | (2024)15:2863 8

Article https://doi.org/10.1038/s41467-024-46906-4

concordance between high UMIS and co-culture reactivity (median ability of this bispecific T cell engager to recruit T cells to these
digest viability: 78% vs. 93%, p = 0.046) (Fig. 5f and Supplementary metastases (Fig. 6a, c and Supplementary Fig. 10a, c, d). However,
Fig. 7e). Since UMIS quantitation was neither associated with nor neither prior ICI nor tebentafusp therapy were associated with an
dependent upon tumor digest viability (Supplementary Fig. 7f), the increase in TCR clonality (Fig. 6a, c and Supplementary Fig. 10a, c, d),
discordance with anti-tumor reactivity observed with this outlier indicating that these immunotherapies were incapable of inducing
subset likely stemmed from insufficient stimulatory capacity of these in vivo proliferation of the endogenous TIL. When specific TRB and
low viability tumor digests. To further characterize the performance of TRA sequences were compared across the metastases (n = 100), we
UMIS in our discovery cohort of 100 UM metastases, we benchmarked found that most of the sequences were private, with rare and limited
its ability to predict co-culture anti-tumor reactivity against several public expression suggesting unique, rather than shared, antigen tar-
other tumor biopsy metrics including TMB, percentage of infiltrating geting (Supplementary Fig. 10e). Cumulatively, these TCR repertoire
CD8+ T cells, and several published gene expression profiles for T cell studies demonstrated that although high UMIS metastases were infil-
inflammation (Fig. 5g, h and Supplementary Data 16)30,39–41. We found trated with a unique polyclonal population of TIL, these T cells
that UMIS was the strongest performer as both a correlative metric remained quiescent, even after receiving ICI and tebentafusp therapy.
(rho = +0.47, p = 7.06e−7) and classification metric (AUC = 0.85) for To determine whether the deficient proliferation of the intratu-
predicting ex vivo TIL reactivity (Fig. 5g, h). Not surprisingly, TMB had moral T cells was due to T cell exhaustion or other intrinsic pro-
no predictive value (rho = +0.01, p = ns; AUC = 0.51) (Fig. 5g, h). In liferative defects, we performed clinical scale ex vivo rapid expansion
further support of UMIS as a preoperative biomarker, we found that (REP) of TIL from UM metastases that were either naïve to ICI and
UMIS level could identify metastases with the greatest yield of tumor tebentafusp (n = 3), or refractory to ICI (n = 10), tebentafusp (n = 4), or
reactive TIL among synchronous metastases in individual UM patients both therapies (n = 2) (Fig. 6d, e and Supplementary Fig. 11a–c). We
(Fig. 5i). In a prospective and independent validation cohort of meta- observed that TIL from each of the metastases demonstrated
static UM biopsies, we corroborated the predictive ability of UMIS for approximately 5-log expansion, reaching massive cell counts (med-
ex vivo TIL reactivity (n = 20, AUC = 0.76) (Fig. 5j). Additionally, we ian = 7.31e10, range = 1.00e10–1.12e11) (Fig. 6d, e and Supplementary
validated that UMIS remained consistent across spatially distinct areas Fig. 11a). Further, these expanded TIL demonstrated a significant
of individual tumors (Supplementary Fig. 8a) and could also be decrease in TCR diversity (p = 4e−6) and a significant increase in TCR
obtained from minimally invasive core biopsies (Supplementary clonality (p = 4e−6) as compared to their source metastases using
Fig. 8b). Taken together, these findings establish that UMIS, obtained highly specific targeted TCR sequencing (Fig. 6d, e and Supplementary
from a metastatic biopsy, could serve as a minimally invasive pre- Fig. 11a–c). These findings suggested that the endogenous TIL were not
operative biomarker to both identify UM metastases harboring tumor limited by intrinsic proliferative deficiencies, but instead their growth
reactive TIL and predict the percentage of tumor reactive TIL cultures was likely suppressed by the tumor microenvironment. Taken toge-
that could be expanded without the limitations associated with con- ther, we observed the quiescence of endogenous TIL in UM metastases
ventional coculture assays. was not reversed with ICI or tebentafusp but could be revived with
ex vivo liberation and expansion.
UMIS identifies quiescent TIL resistant to ICI and tebentafusp Based upon the observation that UMIS from a metastatic biopsy
but sensitive to ex vivo expansion and adoptive transfer could predict the ex vivo potency of quiescent endogenous TIL, we
Having found that UMIS strongly correlated with the level of TIL anti- postulated that UMIS might also predict the clinical efficacy of adop-
tumor reactivity in metastases, we were surprised to find that high tive transfer of these TIL after ex vivo liberation and expansion (Fig. 6f).
UMIS status was not significantly associated with improved survival in Of the 100 UM metastases profiled, 19 had been used to manufacture
our UM cohort (Supplementary Fig. 9a). Furthermore, despite dis- TIL for a previously reported ACT trial in patients with metastatic UM
covering tumor reactive TIL within the metastases of 23 UM patients (NCT01814046)15. Among this treatment cohort, which included 6
(50%) who received prior ICI and 7 patients (78%) who received responders and 13 nonresponders, we observed a strong correlation
tebentafusp, we noted that none of these patients showed objective between source tumor UMIS and the ex vivo anti-tumor reactivity of
tumor regression with these therapies. To investigate these para- the post-REP TIL infusion product (n = 17, rho = +0.61, p = 0.011; 2
doxical findings, we analyzed the intratumoral TCR repertoire within infusion products were not tested due to insufficient tumor) (Fig. 6g).
the source metastases42. We found that the in situ diversity of the TCR Additionally, we found that UMIS as a continuous variable strongly
beta (TRB; n = 88) and TCR alpha (TRA; n = 82) chains varied sig- correlated with magnitude of clinical tumor regression after adoptive
nificantly across the total cohort of metastases (Shannon index ranges; transfer in patients with metastatic UM, including ICI refractory indi-
TRB = 0.08–4.84, TRA = 0.69–4.72) (Fig. 6a and Supplementary viduals (n = 19, rho = −0.68, p = 0.001) (Fig. 6h). To help define a UMIS
Fig. 10a). UMIS was found to strongly correlate with both TRB diversity threshold value that might have future clinical utility in predicting
(rho = +0.54, p = 5.40e−8) (Fig. 6a) and TRA diversity (rho = +0.45, RECIST objective responses (≥30% reduction), we utilized the median
p = 2.14e−5) (Supplementary Fig. 10a) suggesting that high UMIS UMIS value of the non-responder group (UMIS = 0.246) as a response
metastases had more polyclonal T cell infiltrates. In contrast, TRB and threshold (Fig. 6i). We observed that patients having source metas-
TRA clonality, an in vivo surrogate for relative TIL clonal expansion, tases above this threshold had significantly improved progression-free
was low and minimally variant across the samples (1 − Pielou’s index and overall survival after TIL ACT versus those below the threshold
ranges; TRB = 0–0.89, TRA = 0–0.23) (Fig. 6a and Supplementary (Fig. 6j). In sum, these findings demonstrate that UMIS, performed on a
Fig. 10a). We expectedly found no correlations between UMIS and the pre-treatment metastatic biopsy, correlated with the clinical outcome
clonality of the TRB (rho = +0.02, p = ns) (Fig. 6a) and TRA (rho = −0.07, after adoptive transfer of TIL and may serve as a future predictive
p = ns) chains (Supplementary Fig. 10a). The in vivo quiescence of biomarker for the treatment of metastatic UM with ACT.
these TIL was further corroborated by single cell TCR repertoire ana-
lysis demonstrating low clonality (Supplementary Fig. 10b) and single Discussion
cell transcriptomics which found that the percentage of proliferative Here we profiled the immunogenomic landscape of metastatic UM
T cells was equivalently low in high and low UMIS metastases (Figs. 4c using bulk and single cell transcriptomics, TCR repertoire analysis, TIL
and 6b). Interestingly, prior ICI therapies (n = 53) had no influence on reactivity assessment, and clinical adoptive cell therapy. Our findings
TCR diversity compared with untreated samples (Fig. 6a, c and Sup- establish that metastatic UM is not an immunologically ‘cold’ cancer,
plementary Fig. 10a, d). In contrast, prior tebentafusp treatment but instead over half of the analyzed UM metastases harbored
(n = 12) was associated with greater TCR diversity, consistent with the tumor reactive TIL, despite having one of the lowest mutational

Nature Communications | (2024)15:2863 9

Article https://doi.org/10.1038/s41467-024-46906-4

burdens of any solid cancer4,9,43,44 and an equally limited responsive- landscape of this rare cancer and uniquely suited to answer two critical
ness to approved immunotherapies, including ICI therapy and questions that have significant therapeutic implications for UM: what
tebentafusp4,7,8,12,13,45,46. To avoid sampling bias in our study, we ana- factors drive T cell inflammation in metastatic UM, and why does
lyzed a large clinically representative group of UM patients (n = 84) and metastatic UM respond so poorly to currently approved
their metastases (n = 100) which were procured from a diverse array of immunotherapies?
organ sites (n = 11). Further, the metastases in the current study were To define drivers of immune response against metastatic UM, we
genomically validated to be of uveal origin by expression of canonical utilized bulk total RNA sequencing of metastatic biopsies and found
UM somatic alterations and low TMB4,9,43,47,48. Thus, we believe our that T cell-inflamed metastases naturally segregated from T cell exclu-
study cohort to accurately represent the metastatic immunogenomic ded metastases based upon an unsupervised transcriptomic signature

Nature Communications | (2024)15:2863 10

Article https://doi.org/10.1038/s41467-024-46906-4

Fig. 6 | UMIS identifies quiescent TIL resistant to ICI and tebentafusp but tebentafusp = UM #59, both = UM #49). Schematic created with BioRender.com.
sensitive to ex vivo expansion and adoptive transfer. a T cell receptor beta (TRB) f Schematic for evaluation of UMIS in the context of NCT01814046 (ACT of TIL for
repertoire analysis of bulk RNAseq of UM metastases (n = 88). Immune checkpoint metastatic UM)15. Created with BioRender.com. g Correlation of source metastasis
inhibition (ICI) refers to treatment history prior to metastatic biopsy. b Proportion UMIS with TIL infusion product reactivity (n = 17). h Correlation of source metas-
of proliferative T cells in UMIS groups by single cell atlas. c Comparisons of TRB tasis UMIS with maximum percent change in tumor size from baseline (RECIST v1.1)
diversity and clonality in ICI or tebentafusp untreated versus treated metastases after TIL ACT. RECIST response line is drawn at −30% (n = 19). i Comparison of UMIS
(ICI: 48 treated, 40 untreated; tebentafusp: 12 treated, 76 untreated). d Ex vivo TIL between responders (R; n = 6) and nonresponders (NR; n = 13) to TIL ACT. The
expansion from treatment naïve and refractory patients (n = 19). Listed therapies median UMIS of the NR group (0.246) was used as a clinical response threshold for
were received prior to metastatic biopsy. Changes in TIL cell counts (left), T cell outcome analyses. j Time-to-event curves of post-ACT survivals by UMIS response
receptor beta (TRB) diversity (middle), and TRB clonality (right) are shown for thresholds (n = 19). Progression-free survival used progressive disease as the event
source metastases and corresponding TIL cultures post rapid expansion protocol (median follow-up (months): high = 5.09, low = 2.07). Overall survival used death as
(post-REP TIL). Metastases’ TIL counts were conservatively estimated to be ≤106. the event (median follow-up (months): high = 20.97, low = 4.90). Hazard ratios (HR)
TRB repertoires were characterized with targeted TCR repertoire analysis. Sche- are for above versus below threshold groups. Statistical comparisons were per-
matic created with BioRender.com. e Examples of TRB dynamics with ex vivo TIL formed using Spearman’s rank correlation with overlaid simple linear regression to
expansion. Bubble plots represent unique TRB clonotypes (color coded) with illustrate linearity (a, g, h), Fishers exact test (b), Wilcoxon rank-sum test (two-
bubble size indicating percentage of total clonotypes. Shown are representative tailed) (c, i), Kruskal–Wallis test by ranks (c), Wilcoxon signed-rank test (two-tailed)
examples for each pre-harvest treatment group (neither = UM #73, ICI = UM #50, (d) or logrank test (j).

composed of coding, non-coding, and unannotated transcripts. Rather variance in TCR clonality, with higher pre-treatment clonality being
than biasing this gene list with supervised filtering, we integrated the associated with improved response to PD-1 blockage55–58. Interestingly,
entire 2394 gene set into a gene expression score called UMIS. Based we found the quiescent TIL from ICI and tebentafusp treated UM
upon a unique single cell transcriptomic atlas that we developed spe- metastases could demonstrate significant ex vivo expansion, indicat-
cifically for metastatic UM, we found that UMIS could holistically reflect ing that these T cells were not limited by intrinsic factors such as
the multiple cellular components of the tumor microenvironment49. exhaustion, but rather by extrinsic constraints within the UM tumor
Metastases with low UMIS (versus high UMIS) had a paucity of TIL and microenvironment. In support, we previously observed that adoptive
were composed of tumor cells with higher beta-catenin transcript transfer of tumor reactive TIL could mediate objective regression in ICI
expression (CTNNB1), which has been described as a transcriptional refractory UM patients15. In sum, these findings reveal that occult T cell
repressor of BATF3-lineage dendritic cell recruitment of CD8+ responses do exist against metastatic UM but require therapeutic
T cells30,50,51. In contrast, high UMIS metastases had lower tumor cell strategies such as ACT to overcome their growth-suppressed state
expression of CTNNB1, increased APC expression of T cell chemoat- within the tumor microenvironment.
tractant ligands (CXCL10 and CXCL9), greater tumor reactive TIL Finally, our study revealed the importance of UMIS as a tumor
recruitment, and markedly elevated MHC expression on multiple cell intrinsic biomarker to predict TIL potency and clinical response after
populations within the tumor microenvironment, suggesting prominent adoptive transfer in UM patients. Whereas recent reports have pro-
interferon signaling. Thus, we believe Wnt/beta-catenin signaling to play posed phenotypic and transcriptomic markers for the purpose of
an important role in promoting immune exclusion in metastatic UM, defining neo-antigen specific TCR sequences from TIL27,59,60, we believe
similar to prior reports in metastatic CM30,32,50–52. Surprisingly, we iden- UMIS represents a unique tumor biomarker for the identification of
tified only a single metastasis with a possible activating somatic muta- tumor reactive TIL capable of ex vivo expansion for clinical adoptive
tion of the Wnt/beta-catenin pathway, suggesting hotspot mutations are transfer. Importantly, a UMIS level of less than 0.2 identified metas-
not a common driver of beta-catenin overexpression in UM tases that were unlikely to yield potent TIL, suggesting that pre-
metastases53. However, we did observe a strong correlation between the operative UMIS measurement could prevent futile invasive surgical
expression of the long non-coding RNA, SNHG7, and CTNNB1. Based harvests. We found UMIS performed significantly better as a tumor
upon several reports that SNHG7 is a positive regulator of CTNNB1 and intrinsic biomarker of TIL potency when compared to several focused
compelling evidence that in vitro knockdown of SNHG7 leads to gene expression signatures of T cell inflammation. We postulate that
downregulation of the Wnt/beta-catenin pathway in various other the superior performance of UMIS was a result of its unique derivation
cancers34–38 we are investigating the mechanistic role of this non-coding from a large unbiased mixture of coding and non-coding transcripts.
RNA in driving T cell exclusion in UM metastases and therapeutic stra- Further, rather than narrowly reflecting the gene expression of only
tegies to potentially abrogate its effect in low UMIS metastases. immune cells, UMIS was developed as a whole-tumor metric that
To better understand why UM metastases rarely regress with reflected the gene expression of the lymphoid, myeloid, and tumor
approved immunotherapies, we evaluated TIL from ICI and tebenta- compartments within the tumor microenvironment.
fusp resistant patients. From our total cohort of TIL samples (n = 100), Potential limitations of our study include selection bias of the
we observed that 55% of UM metastases harbored tumor reactive TIL patients and metastases analyzed. The rare nature of UM limited our
and there was no difference in the percentage of tumor reactive TIL sample size to 100 metastases. A subset of patients contributed mul-
cultures expanded from ICI and tebentafusp treated metastases versus tiple metastases to the study (14 patients with multiple metastases
untreated. Yet, despite the presence of potent TIL in these metastases, included). We chose to include these metastases to maximize sample
we found they were strikingly quiescent with an absence of in vivo TIL size in this rare cancer but recognize that this may be a source of
expansion using TCR clonality analysis and single cell transcriptomics. selection bias. Given that UM patients presented with the intent of ACT
Interestingly, we observed prior tebentafusp therapy was associated screening, which has strict eligibility requirements, the analyzed cohort
with increased in vivo TCR diversity in our samples, demonstrating its may not represent elderly individuals or those demonstrating rapidly
ability as a T cell recruiter12,13. However, neither tebentafusp nor ICI progressive metastatic disease and declining performance. Further,
therapy were associated with an increase in TCR clonality. The quies- although UMIS was found to correlate with TIL reactivity across 24
cence of these T cells within the tumor microenvironment may explain geographically unique tumor fragments, the predictive ability of UMIS
the low rates of objective tumor regression and the intriguing decou- in larger and more heterogenous lesions needs further study. Finally,
pling of overall response rate as a surrogate for overall survival in UM while our ongoing independent validation cohort (n = 20) has corro-
patients treated with tebentafusp12,13,54. In contrast, TCR repertoire borated the ability of UMIS to predict TIL reactivity, longer clinical
studies of cutaneous melanoma metastases have reported significant follow up is required to evaluate survival in this group of patients.

Nature Communications | (2024)15:2863 11

Article https://doi.org/10.1038/s41467-024-46906-4

Methods determined by percent propidium iodide negative cells by flow cyto-

Patient samples and clinical annotation metry or percent trypan blue negative cells by manual cell counting
Patients were screened and tumor samples were obtained after (UM #47). After ~2 weeks of growth, individual TIL fragment cultures
informed consent in conjunction with tumor procurement banking were tested for tumor specificity by coculture with autologous tumor
protocols associated with two adoptive TIL transfer clinical trials: cells (versus normal tissue controls) followed by measurement of
NCT03467516 (Hillman Cancer Center, UPMC, Pittsburgh, PA, USA) 4-1BB upregulation on CD3+ cells by flow cytometry using anti-human
and NCT01814046 (Surgery Branch, NCI, Bethesda, MD, USA). Patients CD137 (4-1BB)-APC (BD Biosciences) and IFN-γ release by ELISA14,15.
gave informed consent in accordance with the Declaration of Helsinki, Tumor single cell suspensions (digests) and peripheral blood mono-
and the trials were reviewed and approved by the NCI and UPMC nuclear cells were cryopreserved in freezing media and stored long
Institutional Review Boards. There was no requirement for previous term in liquid nitrogen. Monocytes were isolated from peripheral
systemic therapy, given the lack of known effective systemic treat- blood mononuclear cells using the CD14+ MicroBead isolation kit
ments for metastatic UM at the time of study. If patients did receive (Miltenyi Biotec). All flow cytometry data was analyzed with FlowJo
previous systemic treatment, more than 4 weeks must have elapsed v10.8 Software (BD Life Sciences).
before initiation of the current trial therapy, and patients’ toxicities Clinical scale ex vivo rapid expansion protocol (REP) involved
must have recovered to a grade 1 or less (except for toxicities such as selection of individual fragment T cell cultures for further expansion
alopecia or vitiligo). All patients were required to have progressive and based on proliferative capacity and evidence of autologous tumor
measurable metastatic disease with an Eastern Cooperative Oncology reactivity. Final large-scale expansion of selected TIL cultures was done
Group performance status of 0 or 1 and life expectancy greater than with anti-CD3 antibody (30 ng/ml, Ortho Biotech or 50 ng/ml, Miltenyi
3 months at the time of enrollment. Patients were required to have Biotec) and recombinant interleukin-2 (3000 IU/ml; Clinigen) in the
adequate hematological, renal, and hepatic function. Patients were presence of irradiated peripheral blood mononuclear feeder cells15.
excluded if they had active systemic infections, coagulation disorders, The specific anti-tumor reactivity of the infused TIL from
or other active major medical illnesses of the immune system15. NCT01814046 (Surgery Branch, NCI, Bethesda, MD, USA) was assessed
Clinical information, including demographics and treatments, by ELISA-based assays. Following overnight co-culture of the TIL with
were collected relative to the date of metastatic tumor harvest. Sex was their autologous source tumor, the supernatant from these respective
self-reported by patients and consistent with biological sex deter- co-cultures was assessed by ELISA to determine the tumor-induced
mined by genomic analysis. Time-to-event data was collected relative IFN-γ production as described previously15.
to multiple dates: date of primary diagnosis, date of metastatic diag-
nosis, and date of ACT (if applicable). In cases of patients with multiple DNA extraction, library preparation, sequencing, and somatic
metastatic biopsies an algorithm was adopted for selection of a analysis
representative biopsy for the purposes of patient-centered time-to- Whole genome sequencing of the majority of samples was performed
event analysis (in descending order of priority: metastasis harvested at the UPMC Genome Center (n = 93). Genomic DNA was isolated from
prior to any ACT, metastasis whose TIL was utilized for subsequent tumor samples or peripheral blood mononuclear cells on the auto-
ACT, more recently harvested metastasis. Response to ACT was eval- mated Chemagic 360 (PerkinElmer) instrument according to the
uated using RECIST v1.1 criteria61. manufacturer’s instructions. Extracted DNA was quantitated using
Qubit dsDNA BR Assay Kit (Thermo Fisher Scientific). DNA libraries
Tumor procurement, ex vivo TIL culture and tumor reactivity were prepared using the KAPA Hyper Plus Kit (KAPA Biosystems).
testing Genomic DNA was processed through fragmentation, enzymatic end-
All patients had surgical metastatectomies as screening for clinical repair and A-tailing, ligation, and quality check a Standard Sensitivity
trials NCT03467516 (Hillman Cancer Center, UPMC, Pittsburgh, PA, NGS Fragment Analyzer Kit (Agilent). Libraries with an average size of
USA) and NCT01814046 (Surgery Branch, NCI, Bethesda, MD, USA) to 450 base pairs (range = 300–600 base pairs) were quantified by qPCR
procure tumor tissue to generate autologous TIL for therapy16. After on the LightCycler 480 (Roche) using the KAPA qPCR quantification kit
surgical procurement of a metastatic lesion, the fresh tumor under- (KAPA Biosystems). The libraries were normalized and pooled as per
went sterile dissection in the UPMC Immunological and Cell Products manufacturer protocol (Illumina). Sequencing was performed using
Laboratory (Pittsburgh, PA, USA) or the Surgery Branch Cell Produc- the NovaSeq 6000 platform (Illumina) with 151 base pair paired end
tion Facility (Bethesda, MD, USA). Representative samples of tumor reads to an average target depth of 70X coverage. The sequencing data
were sent for formal pathological confirmation of UM. Tumor samples was demultiplexed with bcl2fastq2 v2.20 (Illumina) to produce the
procured at the Surgery Branch, NCI, were transferred via a Material fastq files.
Transfer Agreement to the University of Pittsburgh where the analyses The samples were mapped with Sentieon v1.3.462. Somatic var-
were conducted. To develop a clinically relevant core biopsy approach iants were called by TNhaplotyper2 on tumor-normal mode with the
for in situ tumor characterization, a single random biopsy was best-practice recommended whole genome sequencing setting. Var-
obtained from each resected metastasis (~2 mm central core fragment iants were annotated with Funcotator from GATK v4.0.563. Copy
from 93 metastases and ~500,000 cells post tumor dissociation from 7 number alterations were called with an in-house developed ensemble
metastases) and were snap-frozen in liquid nitrogen and stored long method (CNVsenate) with mapped BAM and somatic SNV VCF files.
term at −80 °C for future DNA and RNA extraction. CNVsenate gathers calling results from GATK v4.0.563, CNVkit v0.9.564,
TIL cultures were initiated from geographically discrete 1–2 mm3 CNVnator v0.2.765, Manta v1.3.266, Sentieon CNV (201911)62 and com-
tumor fragments (n ~ 24) that were placed individually in wells of a 24- bines calling with SURVIVOR2 v1.0.367, then uses machine learning and
well culture plate containing complete media with human AB serum event size for filtering. The filtered results were annotated with
and recombinant interleukin-2 (6000 IU/ml; Clinigen). Remaining AnnotSV v1.1.168 for affected genes. The denoised copy ratio for
fresh tumor was processed by mechanical and enzymatic digestion chromosomal segments from GATK was primarily used. A customized
with the human Tumor Dissociation Kit (Miltenyi Biotec) and gentle- script was used to calculate the denoised copy ratio for chromosomal
MACS Dissociator (Miltenyi Biotec) to provide a single cell suspension arms. Copy number gain was defined as chromosomal arm denoised
of autologous tumor targets for TIL reactivity testing. Tumor digests copy ratio ≥1.25, while copy number loss was defined as chromosomal
underwent flow cytometric phenotyping with propidium iodide fol- arm denoised copy ratio ≤0.80.
lowed by the following anti-human monoclonal antibodies: CD3-APC- Somatic SNV VCF files were converted to MAF format with
Cy7, CD8-PE-Cy7, CD4-PE (BD Biosciences). Tumor digest viability was vcf2maf v1.6.1969 and annotated with VEP v10270. The MAF cohort was

Nature Communications | (2024)15:2863 12

Article https://doi.org/10.1038/s41467-024-46906-4

filtered with a genomic data commons-like strategy, including for genes that positively correlated with immune and inflammatory hall-
population allele frequency <2%, coding regions, and presence in mark gene set enrichment (negative PC2 gene loading) (Supplemen-
dbSNP71 and COSMIC72. Further filtering was done for only somatic tary Data 4). A cohort-dependent version of UMIS was also calculated
mutations with variant allele frequency ≥5%. Mutations were manually using gene set variation analysis (GSVA) with GSVA v1.42.087 using
tabulated for one sample that was unable to be processed into the MAF default settings and the same list of genes (Supplementary Data 4).
format (UM #20). All computational processes above were performed This was only done for the purposes of comparison to the cohort-
on a linux-based amazon web services ec2 instance on the DNAnexus independent implementation with singscore and was not used else-
platform (DNAnexus). The MAF file was then processed and summar- where. Functional annotation of genes within UMIS (n = 2394) was
ized using maftools v2.10.0573. performed with the Database for Annotation, Visualization and Inte-
For six samples (UM #4, #13, #22, #23, #26 and #30) without grated Discovery (DAVID) online tool (accessed March 20th, 2022)88
sufficient tumor tissue for whole genome sequencing we utilized after filtering for protein coding genes using Human Genome Orga-
whole exome sequencing performed as described previously14 to nization (HUGO) Gene Nomenclature Committee (HGNC) complete
assess for canonical UM somatic mutations. For one sample (UM #53) set annotation (accessed October 6th, 2021). Similarly, functional
without sufficient tumor tissue for whole genome sequencing DNA and annotation of differentially expressed genes between UMIS levels was
RNA were extracted from paraffin embedded tumor tissue and pro- performed with clusterProfiler v4.2.289 using the fgsea v3.16 method
cessed with the Oncomine Comprehensive Assay v3 DNA and RNA on only protein-coding genes using HGNC complete set annotation
primer sets (Thermo Fisher Scientific) according to the manufacturer’s (accessed October 6th, 2021). Correlation and clustering analysis of
protocol. Alterations assessed were per the UPMC Oncomine panel PCs used the Human Molecular Signatures Database Hallmark gene set
which has been described previously74. collection90 while functional annotation used the Human Molecular
Signatures Database Gene Ontology Biological Process gene set
RNA extraction, library preparation, sequencing, and read collection91.
alignment Human leukocyte antigen (HLA) typing of patients was performed
Total RNA was isolated from tumor samples on the automated Che- using tumor bulk total RNAseq data. The arcasHLA v0.5.092 package
magic 360 (PerkinElmer) instrument according to the manufacturer’s was run with default settings to produce an output of genotypes for
instructions. Extracted RNA was quantitated with the Qubit RNA BR samples. Representative data for patients with multiple tumor samples
Assay Kit (Thermo Fisher Scientific) followed by an RNA quality check was selected using the same algorithm as described previously in the
using Fragment Analyzer (Agilent). For each sample, RNA libraries survival analysis. TCR repertoires of tumors were analyzed from bulk
were prepared from 100 ng of RNA using the KAPA RNA HyperPrep Kit total RNAseq data using MiXCR v3.0.1293 with allowPartialAlignment-
with RiboErase (Kapa Biosystems) according to the manufacturer’s s=true as recommended for bulk RNAseq data. Counts were tabulated
protocol, followed by a quality check using Fragment Analyzer (Agi- per amino acid CDR3 clonotype and used to calculate diversity
lent) and quantification by qPCR with the Kapa qPCR quantification kit (Shannon index) and clonality (1–Pielou’s index) for TRB and TRA
(Kapa Biosystems). The libraries were normalized, pooled, and chains42. Samples with one or zero detected unique clonotypes were
sequenced using the NovaSeq 6000 platform (Illumina) to an average excluded from diversity and clonality analysis due to mathematically
of ~50 million 101 base pair paired end reads. The sequencing data was undefinable clonality; this resulted in exclusion of 12 metastases from
demultiplexed with bcl2fastq2 v2.20 (Illumina) to produce the TRB analysis and 18 metastases from TRA analysis. Public versus pri-
fastq files. vate repertoire analysis was performed using immunarch v0.6.994.

Bulk transcriptomic computational analyses Targeted TCR repertoire library preparation, sequencing, and
Sequencing data was quality controlled with FastQC v0.11.775 before analysis
and after adapter trimming with cutadapt v1.1876 along with assess- Targeted TCR repertoires of paired tumors and post-REP TIL were
ment of estimated ribosomal content with sortmerna v4.3.477. Trim- derived from respective total RNA. Libraries were prepared using the
med reads were then aligned with STAR v2.7.5a78 using the Gencode QIAseq Immune Repertoire RNA Library Kit (Qiagen) per manu-
v38 GTF and GRCh38 fasta references79. Uniquely mapped percentage facturer’s instructions. Libraries underwent quality check using a
of reads and total uniquely mapped reads metrics after STAR mapping Standard Sensitivity NGS Fragment Analyzer Kit (Agilent) and quanti-
were used as further quality control metrics. The BAM file was indexed fication by qPCR with the Kapa qPCR quantification kit (Kapa Biosys-
with samtools v1.1080. Gene counts from the STAR BAM files were tems). The libraries were normalized, pooled, and sequenced using the
calculated with htseq-count v0.13.581. MiSeq platform (Illumina) to an average of ~2.5 million 251 base pair
Gene names were converted from Ensembl v10382 to HUGO gene paired end reads. The sequencing data was demultiplexed with
symbols with biomaRt v2.50.383. Redundant gene counts after name bcl2fastq2 v2.20 (Illumina) to produce the fastq files. Sequencing data
conversion were summed. Transcripts per million (TPM) were calcu- was processed using the Qiagen Biomedical Genomics Analysis 23.0
lated in standard fashion using gene lengths calculated with Feature- (Qiagen) per default recommended settings. Counts were tabulated
Counts v1.6.284. Raw counts were normalized with DESeq2 v1.34.085 from output files per amino acid CDR3 clonotype and used to calculate
using default and recommended parameters. Variance stabilizing diversity (Shannon index) and clonality (1–Pielou’s index) for TRB and
transformation was performed on the normalized counts and used for TRA chains.
principal component analysis (PCA) with PCAtools v2.10.086. PCA was
performed using the 10% most variant genes (n = 5942) in the dataset. Single cell RNA sequencing library preparation and sequencing
Differential gene expression by UMIS level was performed without any Selected tumors and previous treatments were UM #72 (tebentafusp),
adjustment parameters with default and recommended settings. 83 (ICI and tebentafusp), 100 (ICI), 46 (liver directed therapy), 79
Enrichment scores of gene sets were calculated with singscore (cytotoxic chemotherapy and ICI), and 80 (liver directed therapy,
v1.14.018 using TPM input. Calculations utilized the unidirectional kinase inhibition, antiangiogenic therapy and ICI). Cryopreserved sin-
expected-upregulated mode, with the exception of the immune gle cell suspensions of selected tumors were prepared for input into
resistance program score33 which was calculated using the bidirec- the Chromium Next GEM Single Cell 5’ Reagent Kit v2 (10X Genomics)
tional mode using separate expected-upregulated and expected- by thawing in complete media with human AB serum, sequential fil-
downregulated gene sets. UMIS was calculated with singscore using tration through 70 mm and 30 mm MACS SmartStrainers (Miltenyi
the unidirectional expected-upregulated mode using with the 2394 Biotec) and removal of dead cells using the Dead Cell Removal Kit

Nature Communications | (2024)15:2863 13

Article https://doi.org/10.1038/s41467-024-46906-4

(Miltenyi Biotec) per manufacturer’s protocol. Cell suspensions were univariate logistic regression and mapping of true positive 1 − spe-
inspected to confirm adequate viability (≥70%). Each tumor sample cificity versus sensitivity. Areas under the ROC curves were calcu-
was processed in a separate lane of the Chip K, with ~35,000 cells lated using the trapezoid rule. Time-to-event curves using the
loaded per sample. The 5’ gene expression and TCR V(D)J libraries were Kaplan–Meier method were generated with survminer v0.4.9105 and
then prepared per manufacturer’s instructions. Prior to sequencing, comparisons between categorical groups were assessed with the
libraries underwent quality check using a Standard Sensitivity NGS logrank test. Clustering analysis was performed with Complex-
Fragment Analyzer Kit (Agilent) and quantification by qPCR with the Heatmap v2.10.0106 and used the default method of Euclidean dis-
Kapa qPCR quantification kit (Kapa Biosystems). The libraries were tance. Where appropriate, multiple comparison adjustment was
normalized, pooled, and sequenced using the NovaSeq6000 platform performed with the false discovery rate (FDR) method using the
(Illumina) to an average of ~30,000 paired-end reads per 5’ gene p.adjust function with method = “fdr” in R. In R the lowest possible
expression library per cell and ~5000 paired-end reads per TCR V(D)J numeric value is roughly 1e−324. Thus, we presented values less than
library per cell with parameters per manufacturer’s protocol. The 1e−324 as ~0 rather than listing arbitrary lower limit numbers.
sequencing data was demultiplexed with bcl2fastq2 v2.20 (Illumina) to
produce the fastq files. Utility visualization software
Aside from software previously mentioned, the following were used
Single cell RNA sequencing computational processing for various visualizations throughout the manuscript: tidyverse
Sequencing data was processed using 10X Genomics Cell Ranger multi v1.3.2107, ggplot2 v3.4108, RColorBrewer v1.1-3109, ggprism v1.0.4110,
v6.1.295 using 10X Genomics Cloud Analysis with introns excluded and patchwork v1.1.2111, packcircles v0.3.4112, plotly v4.10.0.9001. Illustra-
an estimated expected cell count of 20,000. Bioinformatic processing tions were created with BioRender.com.
of each sample involved adjustment for ambient RNA contamination
with SoupX v1.5.296 (default settings), normalization with the sctrans- Reporting summary
form v2 method within Seurat 4.1.197,98 (default settings), and estima- Further information on research design is available in the Nature
tion and removal of doublets with DoubletFinder v2.0.399 (default Portfolio Reporting Summary linked to this article.
settings). Cells remaining after quality control and removal of doublets
were then input into our cataloging algorithm. This involved first Data availability
assigning cells to large buckets using UCell v2.1.020 in the following Bulk total RNAseq raw sequencing data generated in this study have
order: immune (UCell score >0 for gene set of PTPRC), tumor (UCell been deposited in the database of Genotypes and Phenotypes
score >0 for gene set of SOX10, S100A1, MITF, MLANA, PMEL, TYR), (dbGaP) under accession number phs003330.v1.p1 [https://www.ncbi.
stroma (all remaining cells). A TIL atlas was created using a published nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs003330.v1.
dataset19 with harmony v0.1.0100 and symphony v0.1.021 using settings p1]. These data are available under restricted access for patient con-
appropriate for the normalization method of the published dataset. fidentiality reasons and access can be obtained by request via the
Our samples’ immune fractions were then mapped onto the atlas using dbGaP system by following the instructions provided by the website.
symphony settings recommended for data normalized with sctrans- Approval is determined by the National Cancer Institute Data Access
form v2. The higher resolution “level 2” annotation, which included 31 Committee, which can be emailed at [email protected]. Access to
phenotypes, was utilized. The addition of tumor and stroma cells to data is generally granted within a month of successful application and
these immune cells completed our cellular cataloging and various available indefinitely thereafter. Selected raw data are protected and
levels (overall, broad, granular) were also assigned to the cells (Sup- are not publicly available due to data privacy laws but may be shared
plementary Data 9). For pooled analysis, samples were integrated with upon request. Source data are provided with this paper.
Seurat 4.1.1 using settings appropriate for sctransform v2-normalized
data. Dimensionality reduction and differential gene expression were Code availability
performed on the integrated Seurat object. Counts of specific cell Software packages were implemented as described in the “Methods”
types were derived from this integrated Seurat object. Comparison of and no custom packages were created.
cell type proportions by UMIS level was performed with the propeller
function within the speckle v0.0.3101 package using the arcsin square References
root transformation of proportions method. 1. Hodi, F. S. et al. Improved survival with ipilimumab in patients with
TCR V(D)J repertoires were filtered for most frequent TRA and metastatic melanoma. N. Engl. J. Med. 363, 711–723 (2010).
TRB chains using scRepertoire v1.4.0102. Counts were tabulated for 2. Larkin, J. et al. Five-year survival with combined nivolumab and
cells with paired TRA and TRB chains per unique TRA-TRB amino acid ipilimumab in advanced melanoma. N. Engl. J. Med. 381,
CDR3 clonotype and used to calculate diversity (Shannon index) and 1535–1546 (2019).
clonality (1–Pielou’s index). 3. Samstein, R. M. et al. Tumor mutational load predicts survival after
immunotherapy across multiple cancer types. Nat. Genet. 51,
Statistical analysis 202–206 (2019).
Statistics were calculated using R v4.1.2103 (R Core Team) with RStu- 4. Yarchoan, M., Hopkins, A. & Jaffee, E. M. Tumor mutational burden
dio v.2022.12.0 + 353104 (Rstudio Team) or GraphPad Prism v9.5.0 and response rate to PD-1 inhibition. N. Engl. J. Med. 377,
(GraphPad Software), and specific statistical analyses used are high- 2500–2501 (2017).
lighted in the respective figure legends. In general, continuous- 5. Tawbi, H. A. et al. Relatlimab and nivolumab versus nivolumab in
continuous associations were assessed with the Spearman’s rank untreated advanced melanoma. N. Engl. J. Med. 386,
correlation with simple linear regression with 95% confidence inter- 24–34 (2022).
vals only to illustrate linearity. Unpaired categorical-continuous 6. Kalbasi, A. & Ribas, A. Tumour-intrinsic resistance to immune
associations were assessed with the Wilcoxon rank-sum test (two- checkpoint blockade. Nat. Rev. Immunol. 20, 25–39 (2020).
tailed) or Kruskal–Wallis one-way analysis of variance test as appro- 7. Algazi, A. P. et al. Clinical outcomes in metastatic uveal melanoma
priate. Paired categorical-continuous associations were assessed treated with PD-1 and PD-L1 antibodies. Cancer 122,
with the Wilcoxon signed-rank test (two-tailed). Categorical- 3344–3353 (2016).
categorical associations were assessed with the Fisher’s exact test. 8. Piulats, J. M. et al. Nivolumab plus ipilimumab for treatment-naïve
Receiver operating characteristic curves were generated using metastatic uveal melanoma: an open-label, multicenter, phase II

Nature Communications | (2024)15:2863 14

Article https://doi.org/10.1038/s41467-024-46906-4

trial by the Spanish Multidisciplinary Melanoma Group (GEM- 31. Duraiswamy, J. et al. Myeloid antigen-presenting cell niches sus-
1402). J. Clin. Oncol. 39, 586–598 (2021). tain antitumor T cells and license PD-1 blockade via CD28 costi-
9. Jager, M. J. et al. Uveal melanoma. Nat. Rev. Dis. Primers 6, mulation. Cancer Cell 39, 1623–1642.e20 (2021).
24 (2020). 32. Luke, J. J., Bao, R., Sweis, R. F., Spranger, S. & Gajewski, T. F. WNT/
10. Singh, A. D., Turell, M. E. & Topham, A. K. Uveal melanoma: trends β-catenin pathway activation correlates with immune exclusion
in incidence, treatment, and survival. Ophthalmology 118, across human cancers. Clin Cancer Res. 25, 3074–3083 (2019).
1881–1885 (2011). 33. Jerby-Arnon, L. et al. A cancer cell program promotes T cell
11. Niederkorn, J. Y. Ocular immune privilege and ocular melanoma: exclusion and resistance to checkpoint blockade. Cell 175,
parallel universes or immunological plagiarism? Front. Immunol. 984–997.e24 (2018).
3, 148 (2012). 34. Yu, F. et al. Loss of lncRNA-SNHG7 promotes the suppression of
12. Carvajal, R. D. et al. Clinical and molecular response to tebenta- hepatic stellate cell activation via miR-378a-3p and DVL2. Mol.
fusp in previously treated patients with metastatic uveal mela- Ther. Nucleic Acids 17, 235–244 (2019).
noma: a phase 2 trial. Nat. Med. 28, 2364–2373 (2022). 35. Chen, Y. et al. Knockdown of lncRNA SNHG7 inhibited cell pro-
13. Nathan, P. et al. Overall survival benefit with tebentafusp in liferation and migration in bladder cancer through activating Wnt/
metastatic uveal melanoma. N. Engl. J. Med. 385, β-catenin pathway. Pathol. Res. Pract. 215, 302–307 (2019).
1196–1206 (2021). 36. Bian, Z. et al. The role of long noncoding RNA SNHG7 in human
14. Rothermel, L. D. et al. Identification of an immunogenic subset of cancers (Review). Mol. Clin. Oncol. 13, 45 (2020).
metastatic uveal melanoma. Clin. Cancer Res. 22, 37. Najafi, S. et al. Oncogenic roles of small nucleolar RNA host gene 7
2237–2249 (2016). (SNHG7) long noncoding RNA in human cancers and potentials.
15. Chandran, S. S. et al. Treatment of metastatic uveal melanoma Front. Cell Dev. Biol. 9, 809345 (2021).
with adoptive transfer of tumour-infiltrating lymphocytes: a sin- 38. Ren, J. et al. Long noncoding RNA SNHG7 promotes the pro-
gle-centre, two-stage, single-arm, phase 2 study. Lancet Oncol. gression and growth of glioblastoma via inhibition of miR-5095.
18, 792–802 (2017). Biochem. Biophys. Res. Commun. 496, 712–718 (2018).
16. Crompton, J. G., Klemen, N. & Kammula, U. S. Metastasectomy for 39. Rooney, M. S., Shukla, S. A., Wu, C. J., Getz, G. & Hacohen, N.
tumor-infiltrating lymphocytes: an emerging operative indication Molecular and genetic properties of tumors associated with local
in surgical oncology. Ann. Surg. Oncol. 25, 565–572 (2018). immune cytolytic activity. Cell 160, 48–61 (2015).
17. Javed, A. et al. PD-L1 expression in tumor metastasis is different 40. Fehrenbacher, L. et al. Atezolizumab versus docetaxel for patients
between uveal melanoma and cutaneous melanoma. Immu- with previously treated non-small-cell lung cancer (POPLAR): a
notherapy 9, 1323–1330 (2017). multicentre, open-label, phase 2 randomised controlled trial.
18. Foroutan, M. et al. Single sample scoring of molecular pheno- Lancet 387, 1837–1846 (2016).
types. BMC Bioinform. 19, 404 (2018). 41. Ayers, M. et al. IFN-γ-related mRNA profile predicts clinical
19. Nieto, P. et al. A single-cell tumor immune atlas for precision response to PD-1 blockade. J. Clin. Invest. 127, 2930–2940 (2017).
oncology. Genome Res. 31, 1913–1926 (2021). 42. Chiffelle, J. et al. T-cell repertoire analysis and metrics of diversity
20. Andreatta, M. & Carmona, S. J. UCell: robust and scalable single- and clonality. Curr. Opin. Biotechnol. 65, 284–295 (2020).
cell gene signature scoring. Comput. Struct. Biotechnol. J. 19, 43. Robertson, A. G. et al. Integrative analysis identifies four molecular
3796–3798 (2021). and clinical subsets in uveal melanoma. Cancer Cell 32,
21. Kang, J. B. et al. Efficient and precise single-cell reference atlas 204–220.e15 (2017).
mapping with Symphony. Nat. Commun. 12, 5890 (2021). 44. Field, M. G. et al. Punctuated evolution of canonical genomic
22. Smith-Garvin, J. E., Koretzky, G. A. & Jordan, M. S. T cell activation. aberrations in uveal melanoma. Nat. Commun. 9, 116 (2018).
Annu. Rev. Immunol. 27, 591–619 (2009). 45. Carvajal, R. D. et al. Advances in the clinical management of uveal
23. Ivashkiv, L. B. IFNγ: signalling, epigenetics and roles in immunity, melanoma. Nat. Rev. Clin. Oncol. 20, 99–115 (2023).
metabolism, disease and cancer immunotherapy. Nat. Rev. 46. Pelster, M. S. et al. Nivolumab and ipilimumab in metastatic uveal
Immunol. 18, 545–558 (2018). melanoma: results from a single-arm phase II study. J. Clin. Oncol.
24. Dersh, D., Hollý, J. & Yewdell, J. W. A few good peptides: MHC 39, 599–607 (2021).
class I-based cancer immunosurveillance and immunoevasion. 47. Karlsson, J. et al. Molecular profiling of driver events in metastatic
Nat. Rev. Immunol. 21, 116–128 (2021). uveal melanoma. Nat. Commun. 11, 1894 (2020).
25. Diamond, M. S. & Farzan, M. The broad-spectrum antiviral func- 48. Shain, A. H. et al. The genetic evolution of metastatic uveal mel-
tions of IFIT and IFITM proteins. Nat. Rev. Immunol. 13, anoma. Nat. Genet. 51, 1123–1130 (2019).
46–57 (2013). 49. Durante, M. A. et al. Single-cell analysis reveals new evolutionary
26. Chandran, S. S. et al. Tumor-specific effector CD8+ T cells that can complexity in uveal melanoma. Nat. Commun. 11, 496 (2020).
establish immunological memory in humans after adoptive 50. Spranger, S., Dai, D., Horton, B. & Gajewski, T. F. Tumor-residing
transfer are marked by expression of IL7 receptor and c-myc. Batf3 dendritic cells are required for effector T cell trafficking and
Cancer Res. 75, 3216–3226 (2015). adoptive T cell therapy. Cancer Cell 31, 711–723.e14 (2017).
27. Krishna, S. et al. Stem-like CD8 T cells mediate response of 51. Sharma, P., Hu-Lieskovan, S., Wargo, J. A. & Ribas, A. Primary,
adoptive cell immunotherapy against human cancer. Science adaptive, and acquired resistance to cancer immunotherapy. Cell
370, 1328–1334 (2020). 168, 707–723 (2017).
28. Chow, M. T. et al. Intratumoral activity of the CXCR3 chemokine 52. Wang, Y. et al. Multimodal single-cell and whole-genome
system is required for the efficacy of anti-PD-1 therapy. Immunity sequencing of small, frozen clinical specimens. Nat. Genet. 55,
50, 1498–1512.e5 (2019). 19–25 (2023).
29. Chen, J. et al. SLAMF7 is critical for phagocytosis of haemato- 53. Bugter, J. M., Fenderico, N. & Maurice, M. M. Mutations and
poietic tumour cells via Mac-1 integrin. Nature 544, mechanisms of WNT pathway tumour suppressors in cancer. Nat.
493–497 (2017). Rev. Cancer 21, 5–21 (2021).
30. Spranger, S., Bao, R. & Gajewski, T. F. Melanoma-intrinsic β-cate- 54. Mariani, P. et al. Immunohistochemical characterisation of the
nin signalling prevents anti-tumour immunity. Nature 523, immune landscape in primary uveal melanoma and liver metas-
231–235 (2015). tases. Br. J. Cancer 129, 772–781 (2023).

Nature Communications | (2024)15:2863 15

Article https://doi.org/10.1038/s41467-024-46906-4

55. Valpione, S. et al. The T cell receptor repertoire of tumor infil- 79. Frankish, A. et al. GENCODE 2021. Nucleic Acids Res. 49,
trating T cells is predictive and prognostic for cancer survival. Nat. D916–D923 (2021).
Commun. 12, 4098 (2021). 80. Li, H. et al. The sequence alignment/map format and SAMtools.
56. Yusko, E. et al. Association of tumor microenvironment T-cell Bioinformatics 25, 2078–2079 (2009).
repertoire and mutational load with clinical outcome after 81. Anders, S., Pyl, P. T. & Huber, W. HTSeq—a Python framework to
sequential checkpoint blockade in melanoma. Cancer Immunol. work with high-throughput sequencing data. Bioinformatics 31,
Res. 7, 458–465 (2019). 166–169 (2015).
57. Riaz, N. et al. Tumor and microenvironment evolution during 82. Cunningham, F. et al. Ensembl 2022. Nucleic Acids Res. 50,
immunotherapy with nivolumab. Cell 171, 934–949.e16 (2017). D988–D995 (2022).
58. Tumeh, P. C. et al. PD-1 blockade induces responses by inhibiting 83. Smedley, D. et al. The BioMart community portal: an innovative
adaptive immune resistance. Nature 515, 568–571 (2014). alternative to large, centralized data repositories. Nucleic Acids
59. Hanada, K. I. et al. A phenotypic signature that identifies Res. 43, W589–598 (2015).
neoantigen-reactive T cells in fresh human lung cancers. Cancer 84. Liao, Y., Smyth, G. K. & Shi, W. featureCounts: an efficient general
Cell 40, 479–493.e6 (2022). purpose program for assigning sequence reads to genomic fea-
60. Lowery, F. J. et al. Molecular signatures of antitumor neoantigen- tures. Bioinformatics 30, 923–930 (2014).
reactive T cells from metastatic human cancers. Science 375, 85. Love, M. I., Huber, W. & Anders, S. Moderated estimation of fold
877–884 (2022). change and dispersion for RNA-seq data with DESeq2. Genome
61. Eisenhauer, E. A. et al. New response evaluation criteria in solid Biol. 15, 550 (2014).
tumours: revised RECIST guideline (version 1.1). Eur. J. Cancer 45, 86. Blighe, K. PCAtools: PCAtools: Everything Principal Components
228–247 (2009). Analysis. R package version 2.10.0 (2022).
62. Kendig, K. I. et al. Sentieon DNASeq variant calling workflow 87. Hänzelmann, S., Castelo, R. & Guinney, J. GSVA: gene set variation
demonstrates strong computational performance and accuracy. analysis for microarray and RNA-seq data. BMC Bioinform. 14,
Front. Genet. 10, 736 (2019). 7 (2013).
63. Van der Auwera, G. A. et al. From FastQ data to high confidence 88. Sherman, B. T. et al. DAVID: a web server for functional enrichment
variant calls: the Genome Analysis Toolkit best practices pipeline. analysis and functional annotation of gene lists (2021 update).
Curr. Protoc. Bioinform. 43, 11.10.11–11.10.33 (2013). Nucleic Acids Res. 50, W216–221 (2022).
64. Talevich, E., Shain, A. H., Botton, T. & Bastian, B. C. CNVkit: 89. Wu, T. et al. clusterProfiler 4.0: a universal enrichment tool for
genome-wide copy number detection and visualization from tar- interpreting omics data. Innovation 2, 100141 (2021).
geted DNA sequencing. PLoS Comput. Biol. 12, e1004873 (2016). 90. Liberzon, A. et al. The Molecular Signatures Database (MSigDB)
65. Abyzov, A., Urban, A. E., Snyder, M. & Gerstein, M. CNVnator: an hallmark gene set collection. Cell Syst. 1, 417–425 (2015).
approach to discover, genotype, and characterize typical and 91. The Gene Ontology Consortium. The Gene Ontology Resource: 20
atypical CNVs from family and population genome sequencing. years and still GOing strong. Nucleic Acids Res. 47,
Genome Res. 21, 974–984 (2011). D330–D338 (2019).
66. Chen, X. et al. Manta: rapid detection of structural variants and 92. Orenbuch, R. et al. arcasHLA: high-resolution HLA typing from
indels for germline and cancer sequencing applications. Bioin- RNAseq. Bioinformatics 36, 33–40 (2020).
formatics 32, 1220–1222 (2016). 93. Bolotin, D. A. et al. MiXCR: software for comprehensive adaptive
67. Jeffares, D. C. et al. Transient structural variations have strong immunity profiling. Nat. Methods 12, 380–381 (2015).
effects on quantitative traits and reproductive isolation in fission 94. Nazarov, V. et al. immunarch: Bioinformatics Analysis of T-Cell and
yeast. Nat. Commun. 8, 14061 (2017). B-Cell Immune Repertoires R package version 1.0.0 (2022).
68. Geoffroy, V. et al. AnnotSV: an integrated tool for structural var- 95. Zheng, G. X. et al. Massively parallel digital transcriptional profil-
iations annotation. Bioinformatics 34, 3572–3574 (2018). ing of single cells. Nat. Commun. 8, 14049 (2017).
69. Kandoth, C. mskcc/vcf2maf: vcf2maf v1.6.19 (2020). 96. Young, M. D. & Behjati, S. SoupX removes ambient RNA con-
70. McLaren, W. et al. The Ensembl variant effect predictor. Genome tamination from droplet-based single-cell RNA sequencing data.
Biol. 17, 122 (2016). Gigascience 9, 12 (2020)
71. Smigielski, E. M., Sirotkin, K., Ward, M. & Sherry, S. T. dbSNP: a 97. Choudhary, S. & Satija, R. Comparison and evaluation of statistical
database of single nucleotide polymorphisms. Nucleic Acids Res. error models for scRNA-seq. Genome Biol. 23, 27 (2022).
28, 352–355 (2000). 98. Hao, Y. et al. Integrated analysis of multimodal single-cell data.
72. Tate, J. G. et al. COSMIC: the Catalogue Of Somatic Mutations In Cell 184, 3573–3587.e29 (2021).
Cancer. Nucleic Acids Res. 47, D941–D947 (2019). 99. McGinnis, C. S., Murrow, L. M. & Gartner, Z. J. DoubletFinder:
73. Mayakonda, A., Lin, D. C., Assenov, Y., Plass, C. & Koeffler, H. P. doublet detection in single-cell RNA sequencing data using arti-
Maftools: efficient and comprehensive analysis of somatic variants ficial nearest neighbors. Cell Syst. 8, 329–337.e4 (2019).
in cancer. Genome Res. 28, 1747–1756 (2018). 100. Korsunsky, I. et al. Fast, sensitive and accurate integration of
74. Paniccia, A. et al. Prospective, multi-institutional, real-time next- single-cell data with harmony. Nat. Methods 16, 1289–1296
generation sequencing of pancreatic cyst fluid reveals diverse (2019).
genomic alterations that improve the clinical management of 101. Phipson, B. et al. propeller: testing for differences in cell type
pancreatic cysts. Gastroenterology 164, 117–133.e17 (2023). proportions in single cell data. Bioinformatics 38,
75. Andrews, S. FastQC: a quality control tool for high throughput 4720–4726 (2022).
sequence data (2010). 102. Borcherding, N., Bormann, N. L. & Kraus, G. scRepertoire: an
76. Martin, M. Cutadapt removes adapter sequences from high- R-based toolkit for single-cell immune receptor analysis.
throughput sequencing reads. EMBnet J. 17, 10–12 (2011). F1000Res 9, 47 (2020).
77. Kopylova, E., Noé, L. & Touzet, H. SortMeRNA: fast and accurate 103. The R Development Core Team. R: A Language and Environment
filtering of ribosomal RNAs in metatranscriptomic data. Bioinfor- for Statistical Computing (2021).
matics 28, 3211–3217 (2012). 104. RStudio Team. RStudio: Integrated Development for R (2020).
78. Dobin, A. et al. STAR: ultrafast universal RNA-seq aligner. Bioin- 105. Kassambara, A. et al. survminer: Drawing Survival Curves using
formatics 29, 15–21 (2013). ‘ggplot2’. R package version 0.4.9 (2021).

Nature Communications | (2024)15:2863 16

Article https://doi.org/10.1038/s41467-024-46906-4

106. Gu, Z., Eils, R. & Schlesner, M. Complex heatmaps reveal patterns Competing interests
and correlations in multidimensional genomic data. Bioinfor- The authors declare no competing interests.
matics 32, 2847–2849 (2016).
107. Wickham, H. et al. Welcome to the tidyverse. J. Open Source Ethical approval
Softw. 4, 1686 (2019). We have complied with all relevant ethical regulations in the conduct
108. Wickham, H. ggplot2: Elegant Graphics for Data Analysis and reporting of this study.
(Springer-Verlag, 2016).
109. Brewer, C. RColorBrewer: ColorBrewer Palettes. R package ver- Additional information
sion 1.1-3 (2022). Supplementary information The online version contains
110. Dawson, C. ggprism: A. ‘ggplot2’ Extension Inspired by ‘GraphPad supplementary material available at
Prism’. R package version 1.0.4 (2022). https://doi.org/10.1038/s41467-024-46906-4.
111. Pedersen, T. patchwork: The Composer of Plots. R package version
1.1.2 (2022). Correspondence and requests for materials should be addressed to
112. Bedward, M., Eppstein, D. & Menzel, P. packcircles: Circle Packing. Udai S. Kammula.
R package version 0.3.5 (2022).
Peer review information Nature Communications thanks Duc-Hau Le,
Acknowledgements Yong-Chen Lu, Lars Ny and the other, anonymous, reviewer(s) for their
We thank the members of the UPMC clinical trial and research teams for contribution to the peer review of this work. A peer review file is
their efforts in this study and the Surgery Branch, NCI, for providing available.
selective tumor samples. We thank Clinigen for providing interleukin-2
for clinical and research studies and thank all the patients who partici- Reprints and permissions information is available at
pated in this study. This research was supported by the UPMC Immune http://www.nature.com/reprints
Transplant and Therapy Center and in part by the University of Pittsburgh
Center for Research Computing, RRID:SCR_022735, through the Publisher’s note Springer Nature remains neutral with regard to
resources provided. Specifically, this work used the HTC cluster, which jurisdictional claims in published maps and institutional affiliations.
is supported by NIH award number S10OD028483. This work utilized the
UPMC Hillman Cancer Center Immunologic Monitoring and Cellular Open Access This article is licensed under a Creative Commons
Products Laboratory, a shared resource at the University of Pittsburgh Attribution 4.0 International License, which permits use, sharing,
supported by the CCSG P30 CA047904. S.L.-M. was supported by a adaptation, distribution and reproduction in any medium or format, as
National Institutes of Health National Cancer Institute training grant long as you give appropriate credit to the original author(s) and the
(T32CA113263). source, provide a link to the Creative Commons licence, and indicate if
changes were made. The images or other third party material in this
Author contributions article are included in the article’s Creative Commons licence, unless
S.L.-M., C.B., G.Y., S.M., C.G., J.T., E.A. and P.M. carried out the experi- indicated otherwise in a credit line to the material. If material is not
ments. S.L.-M., R.K. and U.K. analyzed data and prepared visualizations. included in the article’s Creative Commons licence and your intended
S.L.-M., P.M., A.C., V.S., P.C. and U.C. designed the RNAseq bioinfor- use is not permitted by statutory regulation or exceeds the permitted
matic pipeline. S.L.-M., P.M., E.S., G.V., L.K., D.K. and X.Z. performed use, you will need to obtain permission directly from the copyright
next-generation sequencing. U.K. conducted associated TIL ACT clinical holder. To view a copy of this licence, visit http://creativecommons.org/
trials. Y.-M.H. and U.K. supervised TIL manufacturing associated with TIL licenses/by/4.0/.
ACT clinical trials. S.L.-M. collected clinicogenomic patient and metas-
tasis data. S.L.-M. and U.K. conceived the project, designed the experi- © The Author(s) 2024
ments, and wrote the manuscript.

Nature Communications | (2024)15:2863 17