Fig. 4: Blocking subsets of PD-L1, CXCR4, PD-1 and CYR61 predicted to drive T-cell infiltration in CRC cohort. | Nature Biomedical Engineering

Fig. 4: Blocking subsets of PD-L1, CXCR4, PD-1 and CYR61 predicted to drive T-cell infiltration in CRC cohort.

From: Identifying perturbations that boost T-cell infiltration into tumours via counterfactual learning of their spatial proteomic profiles

Fig. 4

a, Optimized tumour perturbations aggregated to the patient (row) level (train cohort). The bar graph shows the median relative change in intensity for each molecule across all patients within their cluster. b, Patch-wise correlation between the levels of different molecules and the presence of CD8+ T cells. c, Pie charts show the proportion of patients in each cluster that have FLD; P value from the hypergeometric test. d, Volcano plot comparing molecule levels and cell-type abundance between the two patient cluster using tumour tissues, computed using mean values and Wilcoxon rank-sum test with Šidák correction. Cell types include natural killer (NK) cells, myeloid-derived suppressor cells (MDSCs), dendritic cells (DCs) and others. e, Optimized perturbations aggregated to the level of tissue samples (row). f, UMAP projection of IMC patches. Left: UMAP shows T-cell patches coloured by the tissue samples that they are taken from. Right: UMAP shows counterfactual (perturbed) instances (blue) optimized for tumour patches without T cells (red). g, Line plots show the predicted T-cell infiltration level for each tissue section from the test cohort, before and after perturbation. Bar plots show the predicted mean T-cell infiltration level for each test patient. h, Predicted mean infiltration level across all test patients using perturbation strategies of varying sparsity, obtained by varying β in equation (4). The error bar represents 95% confidence interval.

Back to article page