Resolution of tissue signatures of therapy response in patients with recurrent GBM treated with neoadjuvant anti-PD1

Abstract

The response of patients with recurrent glioblastoma multiforme to neoadjuvant immune checkpoint blockade has been challenging to interpret due to the inter-patient and intra-tumor heterogeneity. We report on a comparative analysis of tumor tissues collected from patients with recurrent glioblastoma and high-risk melanoma, both treated with neoadjuvant checkpoint blockade. We develop a framework that uses multiplex spatial protein profiling, machine learning-based image analysis, and data-driven computational models to investigate the pathophysiological and molecular factors within the tumor microenvironment that influence treatment response. Using melanoma to guide the interpretation of glioblastoma analyses, we interrogate the protein expression in microscopic compartments of tumors, and determine the correlates of cytotoxic CD8+ T cells, tumor growth, treatment response, and immune cell-cell interaction. This work reveals similarities shared between glioblastoma and melanoma, immunosuppressive factors that are unique to the glioblastoma microenvironment, and potential co-targets for enhancing the efficacy of neoadjuvant immune checkpoint blockade.

Fig. 1. Study schema and specimens collected.

a Top panel: Patients with recurrent GBM were treated with off-label neoadjuvant anti-PD1 ~14 days before surgical resection. Tumors were collected for bulk RNA analysis and FFPE sections. Bottom panel: Patients with high-risk resectable melanoma were enrolled in a randomized phase 2 trial (NCT02519322). Tumors were collected at baseline and on-treatment for FFPE sections. Patients were randomized to receive neoadjuvant nivolumab (green) or ipilimumab plus nivolumab (red). b Scheme of micro-tumor analysis. A heterogeneous tumor is analyzed in smaller parts (micro-tumors), where each micro-tumor is deeply characterized and has its own local therapy response (e.g., abundance of T cells). These micro-tumors are used to build a model, which can reveal the important variables that determine the response. c Scheme of immune neighbor analysis. Immune cell–cell interaction is quantified by counting the number of adjacent immune-cell neighbors for each immune cell. The same density of cells can yield different average neighbor numbers. d Molecular response of neoadjuvant anti-PD1 treatment in recurrent GBM determined by bulk gene signatures. The genes and samples were arranged with hierarchical clustering. mNR molecular nonresponder, mR molecular responder, FFPE formalin-fixed paraffin-embedded, RECIST response evaluation criteria in solid tumors, ORR overall response rate, trAEs treatment-related adverse events. Green coloration represents low expression; red coloration represents high expression. Schemes in (a), (b), and (c) were created with BioRender.com. Source data are provided as a Source Data file.

Fig. 2. Regions of interest definition and differential protein expression.

a Fluorescent micrograph of a representative region in a melanoma sample (n = 39) stained for nuclei (blue), CD45 (red), and S100B (green). The immune compartment (contoured with red lines) and the tumor compartment (contoured with green lines) were identified via thresholding the fluorescence intensity of CD45 marker. Scale bar is 100 µm. b–d Volcano plots of two-sided Mann–Whitney U-test comparisons of the protein expressions between RECIST responders and nonresponders in baseline (blue) and on-treatment (red) melanoma samples. The samples were either from the nivolumab (nivo) cohort (b) (n = 9 nonresponders and 3 responders at baseline, n = 7 nonresponders and 2 responders on-treatment), ipilimumab plus nivolumab (ipi-nivo) cohort (c) (n = 2 nonresponders and 8 responders at baseline, n = 3 nonresponders and 5 responders on-treatment), or both arms combined (d) (n = 11 nonresponders and 11 responders at baseline, n = 10 nonresponders and 7 responders on-treatment). Expression of proteins was assessed in the immune compartments of the tissue and quantified as the average count per area of assayed tissue. e Fluorescent micrograph of a representative region in a GBM sample (n = 14) stained for nuclei (blue), CD45 (red), and GFAP (green). Three representative geometric ROIs (in white) were designated as immune-cell rich (immune cell proportion: 37.1% and 50.4%) and immune-cell poor (immune cell proportion: 1.7%). Scale bar is 100 µm. f, g Volcano plots of two-sided Mann–Whitney U-test comparisons of the protein expressions between molecular responders (mR) and molecular nonresponders (mNR) in GBM samples treated with pembrolizumab (pembro). The responses were generated based on a 23-gene molecular signature. Expression of proteins was quantified as average count per area of assayed tissue, which were either in the immune-cell rich regions (f) (n = 8 mNRs and 4 mRs) or immune-cell poor regions (g) (n = 8 mNRs and 6 mRs). For plots in (b–d), (f), and (g), the dashed horizontal lines represent the p value cutoffs (p < 0.10 and p < 0.05) and the dashed vertical lines represent the effect size cutoffs (effect size > |1|). ROIs regions of interest. Source data are provided as a Source Data file.

Fig. 3. Micro-tumor analyses reveal proteins that correlate with CD8 presence and treatment response.

a–c PLS regression analysis of immune-rich regions of interest with CD8A as output in melanoma samples treated with nivolumab (nivo) (a), melanoma samples treated with ipilimumab plus nivolumab (ipi-nivo) (b), and GBM samples treated with pembrolizumab (pembro) (c). Top panels: plots of actual observations versus model predictions of CD8A expression. Bottom panels: plots of variable importance in the projection (VIP) versus model coefficients. The dashed horizontal line represents the VIP > 1.0 cutoff and the dashed vertical line is where the model coefficient is 0. d Heat map summarizing the important (VIP > 1.0) CD8A predictors that were measured in both melanoma and GBM. Red coloration represents positive coefficients, and blue coloration represents negative coefficients. e–g PLS discriminant analysis of immune-rich regions of interest with treatment response as output in melanoma samples treated with nivolumab (e), melanoma samples treated with ipilimumab plus nivolumab (f), and GBM samples treated with pembrolizumab (g). The predictors of lower importance (VIP < 0.8, coefficient < |0.1|) were iteratively removed to determine the minimum set of predictors that enables accurate (AUC > 0.90) classification of RECIST or molecular responder and nonresponder ROIs. Shown are the model prediction equations and box plots of the final prediction model versus treatment response for individual ROIs. The horizontal line in each box represents the median sample value, the ends of the box represent the 25th and 75th percentiles, and the whiskers extend from the ends of the box to the outer most data points. Inset, the area under the receiver operating characteristics curve (AUC). ROIs regions of interest. Response for melanoma and GBM are based on RECIST criteria and a 23-gene molecular signature, respectively. Source data are provided as a Source Data file.

Fig. 4. Micro-tumor analysis reveals correlates of tumor proliferation.

a Box plots of Ki67 expression in the immune-poor regions of interest in GBM samples treated with pembrolizumab (Pembro) (n = 8 mNR and 6 mR). The horizontal line in each box represents the median sample value, the ends of the box represent the 25th and 75th percentiles, and the whiskers extend from the ends of the box to the outer most data points. The comparison was made using a two-sided Mann–Whitney U-test (U = 7). b, c PLS regression analysis of immune-poor regions of interest with Ki67 as output in GBM samples treated with pembrolizumab. Plot of actual observations versus model predictions of Ki67 expression (b). Plot of variable importance in the projection (VIP) versus model coefficients (c). The dashed horizontal line represents the VIP > 1.0 cutoff and the dashed vertical line is where the model coefficient is 0. d Plot of B7-H3 versus IDO-1, color-coded with Ki67 level (red: high, blue: low). mNR molecular nonresponder, mR molecular responder. Molecular responses are based on a 23-gene molecular signature. Source data are provided as a Source Data file.

Fig. 5. Immune neighbors are associated with treatment response in melanoma.

Individual immune cells were identified and mapped out in the tumor, and the number of immune cells adjacent to each immune cell was enumerated and averaged across the tumor. a Immune cell density maps of representative melanoma samples from responder and nonresponder at baseline and on-treatment. The average number of immune neighbors for each tissue is indicated at the bottom left. b Box plots of the average number of neighbors in melanoma samples at baseline (left) and on-treatment (right) (n = 10 nonresponders 11 responders at baseline, n = 10 nonresponders and 7 responders on-treatment). Comparisons were made using two-sided Mann–Whitney U-tests (U = 28 for baseline, U = 4 for on-treatment). c Immune cell density maps of representative GBM samples from molecular responder and nonresponder. The average number of immune neighbors for each tissue is indicated at the bottom left. d Box plots of the average number of neighbors in GBM samples (n = 8 mNR and 6 mR). Comparisons were made using two-sided Mann–Whitney U-tests (U = 21) e, Heat map summarizing PLS regression analyses of immune-rich regions of interest with the average neighbor number of those regions as output. Shown are important (VIP > 1.0) neighbor number predictors that were measured in both melanoma and GBM. Red coloration represents positive coefficients, and blue coloration represents negative coefficients. For box plots in (b) and (d), the horizontal line in each box represents the median sample value, the ends of the box represent the 25th and 75th percentiles, and the whiskers extend from the ends of the box to the outer most data points. Response for melanoma and GBM are based on response evaluation criteria in solid tumors (RECIST) criteria and a 23-gene molecular signature, respectively. VIP variable importance in the projection. Source data are provided as a Source Data file.