Profiling of Tumor-Infiltrating Immune Cells and Their Impact on Survival in Glioblastoma Patients Undergoing Immunotherapy with Dendritic Cells

Abstract

Glioblastomas (GBM) are the most common primary malignant brain tumors, comprising 2% of all cancers in adults. Their location and cellular and molecular heterogeneity, along with their highly infiltrative nature, make their treatment challenging. Recently, our research group reported promising results from a prospective phase II clinical trial involving allogeneic vaccination with dendritic cells (DCs). To date, six out of the thirty-seven reported cases remain alive without tumor recurrence. In this study, we focused on the characterization of infiltrating immune cells observed at the time of surgical resection. An analytical model employing a neural network-based predictive algorithm was used to ascertain the potential prognostic implications of immunological variables on patients' overall survival. Counterintuitively, immune phenotyping of tumor-associated macrophages (TAMs) has revealed the extracellular marker PD-L1 to be a positive predictor of overall survival. In contrast, the elevated expression of CD86 within this cellular subset emerged as a negative prognostic indicator. Fundamentally, the neural network algorithm outlined here allows a prediction of the responsiveness of patients undergoing dendritic cell vaccination in terms of overall survival based on clinical parameters and the profile of infiltrated TAMs observed at the time of tumor excision.

Keywords: CD86; PD-L1; dendritic cells; glioblastoma; immunotherapy; patient survival; tumor infiltrate; tumor microenvironment.

Figure 1

Frequency of tumor-associated immune cells and the activation status of TAMs. (A) Correlation between the frequency of CD45+CD3+ and CD45+CD11b+ GBM-infiltrating immune cells (N = 12; * p < 0.05). (B) Frequency of TAM (CD45+HLA-DR+CD11b+CD14+) positive for PD-L1 and CD86 markers. (N = 21; * p < 0.05).

Figure 2

K-means cluster and decision tree analysis. (A) K-means cluster. K was defined as 2 in order to obtain one group (cluster) with long overall survival (cluster 1, in green) and another with short overall survival (cluster 2, in red). The analysis reports the mean and standard deviations of each variable studied in each of the two identified groups (summarized in the table alongside). (B) Decision tree. Partition of overall survival data into 1 level based on the expression of CD45+/PD-L1+ being less than or greater or equal to 40.4%. The analysis refers to the mean of overall survival, standard deviations, and a p-value in terms of LogWorth of the difference between the groups. (C) Similarly, the same type of analysis for survival data since the 1st recurrence coincidentally resulted in 1 level of partition and was based on the expression of CD45+/PD-L1+ but with a slightly different cutoff value (44%), which also reached statistical significance (LogWorth p-value of 1.82).

Figure 3

Survival prediction analysis using neural networks. In Panel (A), the diagram illustrates the analysis with added variables to the model in blue on the left, activation functions in green, and the final variable, overall survival, on the right. The algorithm randomly splits the original data for training and validating the model. Panels (B,C) show graphs comparing the actual overall survival times with the values calculated (predicted) by the training phase (B) and validation phase (C).

Figure 4

Application of the neural network prediction model. In (A), the input of new values for each of the variables of interest, whether they are continuous numerical variables (such as age, percentage of cells positive for HLA-DR, etc.), ordinal variables (KPS, ECOG), or even categorical variables (such as sex and symptomatology). In (B), the reconstructed surface plot with 3 of the variables of interest. In this case, overall survival was kept fixed on the Y-axis, PD-L1+ on the x-axis, and HLADR+ on the z-axis. In (C), the result of the estimated overall survival (“Current Y”, in this case, 60.4 months), with an accuracy of 71%, corresponds to the R-square of the model.