a) Schematic showing the study workflow. Tumors extracted from patient liver metastases were subjected to scRNA-seq, formalin fixation for immunohistochemical (IHC) analysis and PDX generation. b) UMAP dimensionality-reduced expression profiles of single cells in biopsies from 14 patients with UM. scRNA-seq data from all samples were integrated using FastMNN, Louvain clustering performed and clusters annotated using cell type marker genes compiled from the literature. Some cell types were annotated after additional subclustering (Supplementary Fig. 1a-f). c) Cluster proportions within each sample, for all cell types. Subcutaneous biopsies are indicated with asterisk, remaining samples being liver biopsies.

a) UMAP dimensionality reduction and more detailed re-annotation of the CD8+ T cell subsets from Fig. 1b, using the marker genes shown in (c). b) CD8+ T cell cluster proportions within each sample. c) Expression levels of markers used for CD8+ T cell subset annotation. d) Expression of marker genes signifying early (NR4A1) activation, naïve/memory-like phenotypes (TCF7, IL7R), late activation (HLA-DRA, PDCD1, ICOS) and progression towards exhaustion / dysfunction (LAG3, TIGIT, HAVCR2). Additional statistically identified genes differing between clusters can be found in Supplementary Table 1. e) Multiplex IHC staining for T cell (CD3, ICOS, PD-1) and cancer marker genes (β-catenin) in a patient biopsy (UM24).

a) IHC staining of HMB45, SOX10, S100, MELAN-A (MART1) and PRAME expression in one PDX (See also Supplementary Fig. 2 for additional PDX models), as well as Sanger sequencing to verify mutation of GNA11 at Q209L. Color reactions were with magenta. b) Tumor growth curves of established PDX models. c-d) TIL therapy experiments in PDX models. UM1 and UM22 were transplanted subcutaneously into NOG or hIL2-NOG mice. When tumors were palpable, 20 million autologous TILs were injected in the tail vein. Tumor growth was monitored by a caliper. Tumor sizes are average size ± SEM. e) Images of tumor spheroids created (scale bar represents 100uM). UM23 was unable to form spheres. f-g) Activation (41BB+) and degranulation (CD107a+) status for spheres and TIL co-cultures, (f) and (g) respectively, using n = 2 or 3 replicates.

a) Proportions of cells from each sample that match clonotypes found in experimentally identified 4- 1BB+ and MART1-reactive T cells and which are present in a given biopsy CD8+ T cell cluster. The difference between matching and non-matching cells in each cluster is shown on the right, highlighting subsets that are enriched among the reactive T cells. b) Biopsy CD8+ T cells with clonotypes matching identified reactive T cells, highlighted in the UMAP representation. c) As in (b), but highlighting cells with clonotypes shared between biopsies and TILs. d) Proportions of biopsy CD8+ T cells in each cluster matching either clonotypes from TIL cultures or experimentally identified reactive T cells. Statistical differences in frequency were determined using binomial tests between the frequency of the latter in each cluster relative to the background frequency of all TIL clonotypes present in the same cluster. Frequencies were calculated as number of cells from a given category in a given cluster divided by the total number of cells from that category, where category refers to either TILs or 4- 1BB+/MART1-reactive cells. p-values were adjusted for multiple testing using Bonferroni correction. e) Distributions of cells matching the two different categories of experimentally identified reactive cells among biopsy CD8+ T cells clusters. f) Cells from scRNA-seq of TIL cultures that have clonotypes matching experimentally identified reactive cells. g) Shared clonotypes among all biopsy and TIL samples. h) GLIPH2 clusters of significantly similar and HLA-restricted clonotypes43, mapped to known antigens in public databases using TCRMatch42, based on CDR3β sequences. i) Biopsy CD8+ T cells with clonotypes in either high or low confidence GLIPH2 clusters that match MART1 motifs in public databases.

a) Bulk unsorted TILs or MART1 selected TILs were injected into hIL2-NOG mice carrying liver tumors from patient UM22. b) Flow Cytometry analysis of single cell suspension liver metastasis, comparing treatment of UM22 TILs and UM22 MART1-specific TILs for CD3+, CD3+CD8+CD69+ or CD3+CD8+CD137+. c) IHC with diaminobenzidine (DAB) showing tumor (SOX10) and TILs (CD3) within a liver metastasis, d-e) corresponding analysis of 4-1BB+ TILs in a section (d) and in image analysis comparing both treatments (e). Statistical tests in b and e were unpaired two-tailed t-tests, assuming equal variance. *: p < 0.05; **: p < 0.01. f) Samples of tumor and TILs from the liver and spleen, respectively, were sequenced with scRNA-seq. n = 3 biological replicates were performed for each group of liver samples, and n = 2 for spleen samples (out of which one spleen sample for MART1-selected TILs was the pooled material of two independent mice). Sequencing reads mapping to human and mouse were separated with XenoCell66, after which cells were clustered and annotated as described in Methods. g) Subclustering of CD8+ T cells identified three overall clusters, one of represented a mixed profile of the other two, but with marked cell cycle activity. h) Contributions of the different experimental conditions to each CD8+ T cell cluster. i) Markers distinguishing CD8+ T cell clusters, identified using the FindAllMarkers function of Seurat (the union of the top 25 genes per condition, ranked by log2 fold change). Expression per experimental condition is shown below. j) All TCRý chains identified in each experimental condition. Subsets found in TIL culture scRNA-seq data are highlighted, as are any matches to antigens in public databases. k) Differentially expressed genes between bulk TIL mixtures or MART1-selected TILs present in the livers of mice. A pseudo-bulk approach was used, summing read counts across all cells within a given replicate, and statistical testing performed with DESeq281. Genes with q < 0.05 after Benjamini-Hochberg correction were considered significant. l-m) Expression of KRT86, DUSP4 and LAYN in biopsy CD8+ T cells (l) and the phenotypic clusters they are members of (m).

a) Establishment of another uveal melanoma liver metastasis model. UM1 PDX cells were injected in the tail vein after having been serially transplanted in spleen followed by harvesting from liver. Ultrasound confirmed growth in liver before injection of autologous UM1 TILs or HER2 CAR-T cells as controls. b) Response to TILs as assessed by ultrasound monitoring. c) UMAP of T cells in the UM1 liver metastatic model, showing 15 different cell populations. A list of marker genes was used to annotate clusters. The marker genes list was compiled from differential expression analysis and literature. d) Dot plot showing an average expression of marker genes and detection rate of cells in which the marker gene is detected across 15 cell populations.

a) Bar plot showing top 10 the most abundant clonotypes for liver and spleen samples detected by the TCR-seq b) UMAP showing where clonotypes 302, 11155 and 21797 are residing. c) Volcano plot highlighting differentially expressed genes between liver and spleen detected for the clonotype 302 using pseudo bulk approach. log2FC positive means liver is upregulated relative to the control (spleen) d) UMAP showing where TRLs (from Fig.3 f,g, Fig.4) are residing.

a-b) Subclustering of cells that initially grouped together with endothelial cells (a). Shown per tissue of origin (b). c-d) Subclustering of monocyte-like cells. Shown per tissue of origin (d). e-f) Marker genes used to label clusters in (a) and (c), respectively. g) Sample contribution to each cluster from Fig. 1b. h) Copy number profiles for UM cells identified in each sample, as determined by inferCNV80 analysis. Blue represents copy number loss and red copy number gain. i) Sample contribution to each cluster from Fig. 2a. j-k) Cell contribution to each CD8+ T cell cluster from biopsies of liver and subcutaneous tumors, respectively.69

IHC showing expression of HMB45, SOX10, S100, MELAN-A, PRAME expression in PDX models. Staining color used was either DAB (top four samples) or magenta.

a) Gating strategy for Far Red Dye tagged tumor cells (APC) and Jade dye tagged TILs (FITC). b) Far Red fluorescent dye incorporated spheres co-cultured with Jade dye labeled TILs (green) co-culture with a co-localized population (yellow) at 24-hour time-point.

a) Biopsy CD8+ T cells with clonotypes that match scRNA-sequenced TILs, as in Fig. 4c but shown separately for each sample. b) TIL cells with clonotypes that match any corresponding biopsy. c) Proportions within each CD8+ T cell cluster that match either 4-1BB+ or MART1-selected TILs from UM1, UM9 and UM46, respectively. d) Overlap among TCRβ chains from the reactive cells sorted out from these three samples. e) Shared TCRβ chains among the subset of cells from biopsies and TIL cultures that also have a match to any TCRβ chain among experimentally identified reactive TILs. f-g) Matches to any CDR3β-antigen pair in public databases, as inferred by TRCMatch, for CD8+ T cells from biopsies and TILs. Samples from liver and subcutaneous metastases shown in (f) and (g), respectively. h) As in (f-g), for experimentally identified reactive TILs. i) Cells from biopsy CD8+ T cells that were included in any high-confidence GLIPH2 cluster.

a) All cells from scRNA-seq analysis of PDX samples, shown separately per experimental condition. b) Copy number profiles of UM cells in PDX samples, as determined by inferCNV80 analysis. Blue represents copy number loss and red copy number gain. c) Markers discriminating CD8+ T cell clusters in biopsies, contrasted with CD8+ T cell clusters identified in PDX samples. d) TCRβ chains identified in the PDX samples that match clonotypes in a separate scRNA-seq experiment of UM22 TILs. e) Shared unique TCRβ chains between PDX samples from each experimental condition. f) As in (e), shown relative to PDX CD8+ T cell clusters. g) As in (f), but only including cells that also match any clonotype found in scRNA-seq from UM22 TIL cultures. h) Arithmetic difference between (g) and (f), showing cells with TCRs uniquely detected PDX samples compared to the sequenced UM22 TIL culture. i) All TCRβ chains identified in each PDX CD8+ T cell cluster. Clonotypes found in TIL cultures are highlighted, as well as any matches to public antigen databases. j) Differentially expressed genes between bulk TIL mixtures residing in liver and spleen of PDX models. A pseudo-bulk approach86 was used, summing read counts across all cells within a given replicate, and statistical testing performed with DESeq281. Genes with q < 0.05 were considered significant.

a) Expression of marker genes related to a previously described tumor-reactive set of tissue-resident CD8+ T cells45, among CD8+ T cells found in PDX models. b) Expression of KRT86, DUSP4, LAYN, HAVCR2 and IL7R in TIL culture from UM22. Whether a given cell has a clonotype identified in PDX samples is indicated.

a) UMAP split on spleen and liver. Liver has 20,418 cells and spleen has 18,208 cells. b) Bar plot showing number of cells across each cell type comparing between spleen and liver. c) Violin plot showing expression distribution for CD4 and CD8A genes across clusters between liver and spleen. In liver, CD8A gene is mainly expressed in CD8+ cytotoxic KLRG1+ and CD8+ cytotoxic clusters, and partially in proliferative and cell cycle clusters. CD4 gene is expressed in all cell types except for cluster CD8 cytotoxic KLRG1+ and CD8 cytotoxic clusters. Two UMAPs showing expression level of CD4 and CD8A genes. d) Heatmap showing gene expression (z-score) values of pseudo bulk (“summed counts”) for each cell population. e) Feature plot showing gene expression of two common marker genes for each cell population.