Fig. 8: Development and validation of Brain Metastasis-derived Immune Signature (BMIS) associated with immunotherapy outcomes across cancer types.

A Heatmap showing unsupervised clustering of NSCLC patients in the Ravi cohort (n = 122 patients) based on signature scores of six cell populations. Clinical characteristics (gender, age, initial stage) and signature scores are annotated. B, C Kaplan-Meier survival curves comparing overall survival of Cluster 2 (high brain metastasis-enriched populations) versus other clusters in the Ravi cohort (n = 122 NSCLC patients) (B) and mUC cohort (n = 298 mUC patients) (C). P-values from two-sided log-rank tests are shown. D Workflow for BMIS development. Differentially expressed genes (DEGs) from brain metastasis-enriched cell populations were subjected to feature selection using Variable Selection Using Random Forests (VSURF), yielding a 7-gene signature used to construct a logistic regression model. Created in BioRender. Bai, M. (2026) https://BioRender.com/xfyb5zp. E Kaplan-Meier curves for progression-free survival (left) and overall survival (right) stratified by BMIS in the Ravi cohort (n = 122 NSCLC patients). Tick marks indicate censored observations. F Distribution of BMIS scores across immunotherapy response categories (PD, n = 41 patients; SD, n = 32 patients; PR, n = 42 patients; CR, n = 7 patients) in the Ravi cohort. G Proportion of patients achieving objective response (PR/CR) or progressive/stable disease (PD/SD) stratified by BMIS group in the Ravi cohort. Two-sided Fisher’s exact test P-value is shown. H Progression-free survival curves (left) and response rates (right) stratified by BMIS in the Kim cohort (n = 27 NSCLC patients). Two-sided Fisher’s exact test P value is shown. I Overall survival curves (left) and response rates (right) stratified by BMIS in the mUC cohort (n = 298 mUC patients). Tick marks indicate censored observations. Two-sided Fisher’s exact test P-value is shown. J, K Receiver operating characteristic (ROC) curves (left) and forest plots (right) comparing the performance of BMIS with established immunotherapy biomarkers, including tumor mutational burden (TMB), gene expression profile score (GEP), and cytolytic activity score (CYT), in the Ravi (n = 122 NSCLC patients) (J) and mUC (n = 298 mUC patients). Area under the curve (AUC) with 95% confidence intervals are shown. In survival analysis, hazard ratios (HRs) with 95% confidence intervals (CIs) and P-values from two-sided log-rank tests are shown. Survival curves represent Kaplan–Meier estimates; +, censored observations. In box plots (F), center line represents the median, box bounds indicate the 25th and 75th percentiles (IQR), whiskers extend to 1.5× IQR from box bounds, and outliers are shown as individual points.