a. Proliferative (PRO) and invasive (INV) subpopulations were identified from the ZMEL1 zebrafish melanoma cell line, which was originally isolated from a transgenic zebrafish and can be transplanted into transparent casper zebrafish. b-c. Tracking of individual cells by time-lapse microscopy (both p<0.001 by linear regression, N=4 independent experiments). b. Growth curves (mean ± SE of mean, smoothed with moving window average of 5 time points) and doubling time (mean [95% CI]: 27.0 h [26.9, 27.1] vs. 29.7 h [29.6, 29.9] for ZMEL1-PRO and -INV, respectively). c. (left) Representative displacements of 500 tracks, and (right) model estimates ± 95% CI for alpha, the slope of the log-log plot of mean squared displacement vs. lag time (tau) for each ZMEL1-PRO and -INV. Larger alpha indicates more persistent motion, with α=1 for diffusive and α=2 for projectile motion. d. (left) The INV signature (Hoek et al., 2006) was the top gene set by Gene Set Association Analysis (GSAA) of ZMEL1-INV vs. -PRO RNA-seq. (Right) Dual waterfall plot of GSAA ranked by false discovery rate (FDR). Literature PRO/INV gene sets are indicated with an asterisk and colored according to FDR. e. Heatmap of genes in Hoek INV signature that are differentially expressed between ZMEL1-PRO and -INV (log₂ fold change cutoff ± 1.5, p_adj < 0.05). Human ortholog gene names are used for clarity (see Figure S1e for zebrafish gene names). f. Segmentation of representative images of ZMEL1-PRO and -INV tumors and distant metastases (e.g. to caudal region [box]) at 3 days post-transplant (3dpt). Original images shown in Figure S1e. g. Quantification of caudal metastases seeded by ZMEL1 populations at 3 dpt (OR [95% CI]: 11.62 [1.43, 94.53]; p=0.022 by logistic regression; N=3 independent experiments with PRO/INV 10/10, 31/33, and 13/13 fish per group, respectively; n=110 fish total; plot shows mean ± SD). h. Representative composite images at 1 dpt from time lapse confocal microscopy of ZMEL1 cells transplanted intravenously in larval zebrafish. Arrowhead indicates group of cells invading from the notochord (NC) and caudal hematopoietic tissue (CHT) into the tail fin mesenchyme (TF). Images are representative of n=13 fish per cell type. See Video S1 for full time lapse. i. Quantification of caudal tissue invasion by imaging at 4-6dpt (N=3 independent experiments with PRO/INV 23/23, 21/21, and 19/23 fish per group, respectively; OR [95% CI]: 13.58 [5.56, 33.18]; p<0.001 by logistic regression, plot shows mean ± SD). j. Relative number of ZMEL1-PRO and -INV cells isolated and quantified by flow cytometry from primary tumors and metastases of fish transplanted with a 1:1 mixture of ZMEL1-PRO and -INV (primary tumors from n=6 fish; metastases from n=4 fish; p= 0.51 and p=0.031, respectively, by one-sample two-sided t-test with Bonferroni correction). See also Figure S1, Tables S1, S5, S6, and Video S1.

a. (left) Plot of top Gene Ontology (GO) gene set by GSAA for ZMEL1-INV vs. -PRO RNA-seq. (right) Dual waterfall plot of top/bottom 250 gene sets from GO analysis (for full plot see Figure S2a). Adhesion GO gene sets are indicated with an asterisk and colored according to false discovery rate (FDR) b. Heatmap of genes in adhesion GO gene sets (FDR < 0.05, n=3) that are differentially expressed between ZMEL1-PRO and -INV (log₂ fold change cutoff ± 1.5, p_adj < 0.05). Human ortholog gene names are used for clarity (see Figure S2b-c for absolute expression and zebrafish gene names). c. (top) Schematic of assay and (bottom) quantification of cluster formation in low-bind plates after 2 days (N=6 independent experiments, p<0.001 by two-sided t-test, plot shows mean ± SD). d. Representative images of clusters formed after 3 days. e. Human melanoma cell lines ranked by cluster forming ability (left to right: low to high) with PRO/INV gene expression scores (Hoek et al., 2006) (Spearman correlation and Bonferroni-corrected p-values shown, scale bar 100μm). f. Co-clusters of 1:1 mixture of ZMEL1-PRO and -INV (inset scale 50 μm). g. (left) 3D opacity rendering and (right) central slice (slice 54 of 115) of confocal imaging of co-cluster of ZMEL1-PRO and ZMEL1-INV. h. Co-clusters of 1:1 mixture of ZMEL1-INV and ZMEL1-INV with either control (sg_scr) or cdh1 (sg_cdh1) sgRNA (inset scale 50 μm). f-h are composite images. See also Figure S2, Tables S1, S5,S6, and Videos S2,S3.

a. HOMER de-novo motif analysis on genes upregulated in ZMEL1-PRO vs. -INV (log₂ fold change cutoff ± 1.5, p_adj < 0.05, ±500bp of transcription start site [TSS]). b. HOMER de-novo motif analysis of genes differentially expressed between ZMEL1-INV in 3D (clusters) vs. 2D (no clusters) (log₂ fold change cutoff ± 1.5, p_adj < 0.05, ±500bp of TSS). c. Cluster size after 2 days in ZMEL1-PRO with CRISPR-Cas9 inactivation of tfap2a and tfap2e alone or in combination (sg_tfap2a/e) versus control (sg_scr) (p-values by linear regression; N=3 independent experiments). d. Representative images of clusters formed after 2 days from ZMEL1-PRO with sg_scr vs. sg_tfap2a/e. e. Growth of ZMEL1-PRO orthotopic primary tumors with sg_scr vs. sg_tfap2a/e (p=0.011 by linear regression; N=3 independent experiments with sg_scr/sg_tfap2a/e 24/22, 22/22, 24/24 fish per group, respectively; n=138 fish total). f. Representative composite image of extravasated (arrows) and partially extravasated (arrow-head) ZMEL1-PRO cells with sg_tfap2a/e following intravenous transplant in casper fish with kdrl-RFP transgene labeling the vasculature. g. Proportion of larval fish intravenously transplanted with ZMEL1-PRO with sg_scr or sg_tfap2a/e with extravasated cells at 1 dpt, as quantified from confocal time lapse microscopy (OR [95% CI]: 2.20 [1.05, 4.61]; p=0.038 by logistic regression; N=3 independent experiments with sg_scr/sg_tfap2a/e 20/20, 22/23, and 22/22 fish per group, respectively; n=129 fish total). See also Figure S4 and Tables S2,S3.

a. Human melanoma cell lines in the Cancer Cell Line Encyclopedia (CCLE, n=56) plotted as PRO and INV scores (Hoek et al., 2006) and colored according to TFAP2A mRNA expression. Pearson correlation coefficients between TFAP2A and PRO/INV scores are shown on axes. b. TFAP2A mRNA expression in The Cancer Genome Atlas (TCGA) melanoma (SKCM) cohort comparing primary tumors and metastases (p<0.001 by Wilcoxon rank sum test with Bonferroni correction). c. Low-passage human melanoma cell lines ranked by increased cluster forming ability (left to right) with TFAP2A expression quantified by immunofluorescence (plot and top; Spearman correlation shown; scale bar 20 μm) and clustering (bottom, scale bar 500 μm). d. GSAA was run using gene sets and GO gene sets with FDR < 0.05 from INV vs. PRO RNA-seq (n=39 gene sets; cyan points in Figures 1d and and2a).2a). Bars show Normalized Association Score (NAS) for CRISPR (ZMEL1-PRO sg_tfap2a/e vs. sg_scr) and INV vs. PRO for each gene set, with black outline representing FDR<0.05 for CRISPR experiment. e. Plot of INV signature (Hoek et al., 2006) by GSAA for ZMEL1-PRO sg_tfap2a/e vs. sg_scr RNA-seq. f. Heatmap of top genes in Hoek INV and GO Adhesion gene sets that are differentially expressed between ZMEL1-PRO sg_tfap2a/e and sg_scr (log₂ fold change cutoff ± 0.5, p_adj < 0.05). Asterisks (*) indicate genes with associated TFAP2A CUT&RUN peaks. Human ortholog gene names are used for clarity (see Figure S5p for zebrafish gene names). See also Figure S5 and Tables S4,S5,S6.