Abstract Bulk RNA-seq is now among the standard of care for patients with newly diagnosed glioblastoma (GBM), but whether this data can predict therapeutic response is unclear. In silico deconvolution of bulk RNA-seq data allows for quantification of cell types comprising the tumor and its microenvironment. Our study tests whether cell type proportions predict therapeutic response in patients with GBM and high-grade glioma (HGG) enrolled in trials of investigational agents. 127 patients with GBM/HGG underwent post-resection/biopsy whole exome RNA-seq and reads were post-processed using standard computational pipelines. In silico deconvolution was performed using GBMDeconvoluteR, which generates cellular proportions comprising the RNA-seq reads. Multivariate Cox regression models were used to examine each cell proportion as a predictor of progression-free (PFS) and overall (OS) survival. Statistical significance was ascertained using Bonferroni correction for multiple comparisons. Among 54 patients treated with a PD-L1 inhibitor (nivolumab) and VEGF inhibitor (bevacizumab), the proportions of NK cells (p=0.0024), radial glial cells (p=0.0010), and mural cells (p=0.0025), was associated with increased PFS, while proportion of endothelial cells (p=0.0008) was associated with decreased PFS. Astrocyte (p=0.0022), macrophage (p=0.0015), NK cell (p<0.0001), and mural cell (p=0.0004) proportions were associated with longer OS, while endothelial cell proportion (p=0.0001) was associated with shorter OS. Among 21 patients treated with a tyrosine kinase inhibitor (ibrutinib) combined with radiation and temozolomide (RT/TMZ), cell proportions assessed at the time of diagnosis did not predict PFS or OS. In 52 patients with HGG treated with a JAK inhibitor (ruxolitinib) with RT/TMZ, only mural cell proportion was associated with a longer OS (p=0.0017). Our results highlight how bulk RNA-seq data can be used to predict therapeutic response. In GBM patients treated with immunotherapies, the proportion of NK cells may predict response. Further work will examine which gene programs are predictive of therapeutic response.
Read full abstract