Fig. 3: Combinatorial chemokine therapy predicted to drive T-cell infiltration in patients with metastatic melanoma.

a, Whole-tumour perturbations optimized across IMC images of patients (row) from the training cohort, with bar graph showing the median relative change in intensity for each molecule. b, Distribution of cancer stages among patients within two clusters; grey indicates unknown stage. c, Volcano plot comparing chemokine level and cell-type abundance from patient clusters 1 and 2, computed using mean values and Wilcoxon rank-sum test with Šidák correction. Grey indicates non-statistical significance. Non-significant chemokines not shown: CXCL12 (fold change = 0.96, P = 1) and CCL8 (fold change = 0.93, P = 0.91). d, Patch-wise chemokine profile (left); one-dimensional heatmap (right): infiltration status (light/dark = from infiltrated/deserted tumour), tumour cell (light/dark = present/absent), CD8⁺ T cells (light/dark = present/absent). e, Patch-wise correlation between chemokine signals and the presence of CD8⁺ T cells. f, Top: UMAP projection of tumour patches (chemokine channels) shows a clear separation of masked patches with and without T cells. Bottom: coloured arrows connect UMAP projection of patches without T cells and their corresponding counterfactual (perturbed) patch, where the colours correspond to k-nearest neighbour clusters (i–iv) of the counterfactual patches. Pie charts (i–iv) show the distribution of patients whose original tumour patches are found in the corresponding cluster regions in the UMAP. g, Cell maps computed from a patient’s IMC image, showing the distribution of T cells before and after perturbation. h, Original versus perturbed (predicted) mean infiltration level across all patients (test cohort) with 95% confidence interval (only shown for patients with more than two samples). i, Mean infiltration level across all patients (test cohort) for optimized perturbation strategies of varying sparsity. The error bar represents 95% confidence interval.