Using machine learning to characterize lung cancer microenvironment and the development of a model to predict the presence of similar microenvironment in other cancers.

Abstract

2634 Background: Poor response to immune checkpoint inhibitors (ICI) in colorectal cancer (CRC) is believed to be due to lack of immune suppressive tumor microenvironment (TME). In contrast, lung cancer TME is believed to be significantly more immunologically active and responsible for the relative success of ICI in lung cancer. We evaluated the TME in lung cancer and CRC using 43 immune biomarkers quantified using RNA sequencing and developed a model to classify TME into immunologically active (similar to lung cancer) vs inactive (similar to colorectal). These 43 immune biomarkers included B- and T-cell markers, cytokines and chemokines. Methods: RNA was extracted from FFPE samples from 707 patients with lung cancer, 227 patients with CRC, 131 patients with breast cancer, 111 patients with ovarian cancer, and 72 patients with pancreatic cancer. The expression levels of the 42 immunological markers were quantified using next generation sequencing (NGS) as a part of larger targeted RNA sequencing panel of 1408 genes. Using a machine learning algorithm, we first selected the relative genes that distinguish between two classes using two criteria: performance of each gene with K-fold cross-validation (K=12) and second based on stability measure using statistical significance tests. The selected genes were then used to predict one class from the other using random forest classifier. Samples were divided into a training set (67%) and testing set (33%). Results: A Bayesian-based algorithm selected the expression of 20 genes that were significantly relevant in differentiating between immunologically active and inactive TME. Using these 20 genes in Random Forest model, we can distinguish between lung and CRC with AUC of 0.997 (95% CI: 0.992-1.00) in the training set and AUC of 0.923 (95% CI: 0.880-0.966) in the testing set. Testing 131 breast cancer samples showed 23 (18%) with TME that can be classified as immunologically active. Of 111 ovarian samples 13 (12%) showed immunologically active TME and of 72 pancreatic samples, 17 (24%) showed microenvironment classified as active. The 20 genes that were adequate to distinguish between TME active vs inactive are: CD74, FCGBP, IL1R1, CD44, CD274, FCGR2B, IL21R, IL1RAP, IL7R, CD79A, CCL2, CYFIP2, CD19, IL2RA, CD8A, CD79B, ID1, CD22, FZD10, and IL1B, listed in order of their importance. Conclusions: This data shows that lung cancer TME is significantly different from that of CRC. Only 20 immune biomarkers adequate to distinguish between the two TME. The relevant biomarkers included CD274 (PD-L1) as well as one marker for T-cells (CD8A) but three markers for B-cells (CD19, CD22 and CD79B). This suggests that B-cells play a significant role in immunologically active TME and should be explored further as biomarkers for predicting response to ICI therapy.