Fig. 1: Phenotypic continuum of disease progression in primary UM.

a Distinguishing features of uveal melanoma (UM) with good (blue) and poor (red) prognosis; highlighting two potential models of disease progression. GEP gene expression profile, TCGA the Cancer Genome Atlas. b Patient tissue profiling (metadata summarized in Supplementary Table 1). Immediately following enucleation, six primary tumor specimens were obtained for clinical prognostic and single cell transcriptional profiling. Targeted sequencing using the MSK-IMPACT platform was performed on the formalin-fixed, paraffin embedded enucleation specimens. c Bulk GEP classification assigned to each patient according to the DecisionDx test (Castle Biosciences) (box). Individual tumor cells were likewise assigned to GEP1 (blue) vs. GEP2 (red) clinical prognostic groups according to their average expression of the GEP prognostic gene signatures using a two-component Bayesian Gaussian Mixture Model (BGMM, “Methods”). The fraction of individual tumor cells assigned to GEP1 (blue) and GEP2 (red) per patient is visualized in the bar graphs, where bootstrapping was used to correct for number of cells per patient (bar, mean; whiskers, 95% confidence intervals, 500 tumor cells sampled over n = 20 random subsets of the data). Asterisks, highlight a patient in which there was a discrepancy between DecisionDx bulk classification and the majority GEP classification prescribed by our single cell analysis of resected tumors. Notably, this patient (MSK-UM06) experienced metastatic progression and succumbed to their disease within six months of diagnosis (Supplementary Table 1). Intra-tumoral prognostic heterogeneity was validated in a second, independent cohort recently published by Durante et al.²⁴. d Force-directed layout of all patient tumor cells colored by z-normalized imputed expression of the average GEP2 gene signature. e Individual tumor cells were likewise assigned to one of the four TCGA molecular subtypes of UM according to their average expression of characteristic genes using a four-component BGMM (Methods) in our cohort and a second, independent cohort recently published by Durante et al.²⁴. The fraction of individual tumor cells assigned to TCGA subtype 1 (dark blue), subtype 2 (light blue), subtype 3 (pink) and subtype 4 (red) per patient is visualized in the bar graphs, where bootstrapping was used to correct for number of cells per patient (bar, mean; whiskers, 95% confidence intervals, 500 tumor cells sampled over n = 20 random subsets of the data). f Violin plots showing the distribution of intra-patient phenotypic volume (defined as the pseudo-determinant of the gene expression covariance matrix, detailed in “Methods”), controlled for number of cells and labeled by patient status (alive vs. deceased). Distributions represent 100 random subsamples from the data (n = 150 cells per patient). Overlaid bar and whisker plots reflect the mean and interquartile range.