Regional reproducibility of immune expression programs in glioblastoma and implications for single-biopsy assessment

Tissue immune measurements can guide immunotherapy only if one biopsy represents the rest of the tumor. We quantified this in a secondary analysis of public data from 75 magnetic resonance imaging (MRI)-localized glioblastoma (GBM) biopsies from 19 patients. Fourteen transcriptomic modules scored supportive, inhibitory, and checkpoint programs. The intraclass correlation coefficient (ICC) gave the share of variance between patients rather than regions, before and after removing the component explained by a 12-gene leukocyte composite. An ICC of 1 indicates no variation in module scores among biopsies within each patient, whereas an ICC of 0 indicates that biopsies from the same patient are no more similar, on average, than biopsies from different patients. Two hepatocellular carcinoma (HCC) cohorts and a microdissected GBM atlas served as comparisons. CD8 had the lowest unadjusted ICC among abundance scores (0.42; 95% patient-bootstrap interval, 0.10 to 0.68; leukocytes, 0.65). Within patients, CD8 correlated with supportive and inhibitory modules alike (unadjusted mean r, 0.37 and 0.36) but not with major histocompatibility complex class I (MHC-I; r = 0.06). Leukocyte adjustment lowered the median module ICC from 0.56 to 0.34 but interferon (IFN) only from 0.74 to 0.70. IFN had the highest adjusted ICC and ranked first in 86% of patient-bootstrap resamples (post hoc). MHC-I was second (0.47) and not separable from IFN (difference 0.23; interval, -0.04 to 0.43); after purity residualization in 67 biopsies, IFN and MHC-I were similar (0.66 and 0.62). Median adjusted ICCs were 0.34 in bulk GBM, 0.62 and 0.74 in HCC, and 0.08 in microdissected GBM. In bulk GBM, only IFN, and MHC-I after purity residualization, had adjusted ICC point estimates above 0.5. The data support multiregion sampling with CD8, IFN, MHC-I, and checkpoint readouts from the same regions.