Extended Data Figure 5: Peripheral T-cell exhaustion, reinvigoration, CD8/T_reg ratio, and tumour PD-L1 predict response to radiation + immune checkpoint blockade.

a, Heat map showing the relative proportions of PD-1⁺ CD8 T cells that are Ki67⁺GzmB⁺ or Eomes⁺ and the CD8/T_reg ratio for each sample (columns) subtracted from the average values of untreated controls. Black hatches indicated complete response (CR) and radiation treatment (RT) + anti-CTLA4 (C4) ± anti-PD-L1 (P1). From these data, a multivariable random forest predictor for complete response was developed. Boxplot shows bootstrap distributions of variable importance scores (more predictive variables have higher values), and of b, minimal depth (MD), a statistic to measure predictiveness. Bar plot shows % bootstrap models for which the MD for the indicated variable was significant. Bootstrap mean ± s.d. for the out-of-bag prediction error rate is listed on top. c, Probability of complete response vs change (treated vs untreated control) in CD8/T_reg ratio for mice with a high (blue dots) or low (red dots) change in % PD-1⁺ splenic CD8 T cells that are Eomes⁺. d, Heat map similar to a except using T cells from peripheral blood. e, Percent peripheral blood PD-1⁺ CD8 T cells that are Eomes⁺ vs Ki67⁺GzmB⁺ after RT + checkpoint blockade. Values are subtracted from average of untreated controls. Each circle represents a mouse. Probability of complete response (proportional to circle size), prediction error rate, and quadrant boundaries are estimated from the random forest model. f, Representative contour plots examining splenic CD8 T cells from B16-F10 or Res 499 tumours for PD-1 and Eomes (top), followed by examination of the PD-1⁺Eomes⁺ subset for Ki67 and GzmB (bottom). g, Ratios of PD-1⁺Eomes⁺ splenic CD8 T cells that are Ki67⁺GzmB⁺ (reinvigorated) compared to Ki67⁻GzmB⁻ (exhausted) from mice with Res 499 tumours. P value by two-tailed t-test.