Identifies Immune Feature Genes for Prediction of Chemotherapy Benefit in Cancer

Abstract

Chemotherapy is still the most fundamental treatment for advanced cancers so far. Previous studies have indicated that immune cell infiltration (ICI) index could serve as a biomarker to predict chemotherapy benefit in breast cancer and colorectal cancer. However, due to different responses of tumor infiltrating immune cells (TIICs) to chemotherapy, the prediction efficiency of ICI index is not fully confirmed by now. In our study, we first extended this conclusion in 7 cancers that high ICI index could certainly indicate chemotherapy benefit (P<0.05). But we also found the fraction of different TIICs and the interaction of TIICs were varies greatly from cancer to cancer. Therefore, we executed correlation and causal network analysis to identify chemotherapy associated immune feature genes, and fortunately identified six co-owned immune feature genes (CD48, GPR65, C3AR1, CD2, CD3E and ARHGAP9) in 10 cancers (BLCA, BRCA, COAD, LUAD, LUSC, OV, PAAD, SKCM, STAD and UCEC). Base on this, we developed a chemotherapy benefit prediction model within six co-owned immune feature genes through random forest classifying (AUC =0.83) in cancers mentioned above, and validated its efficiency in external datasets. In short, our work offers a novel model with a shrinking panel which has the potential to guide optimal chemotherapy in cancer.

Keywords: Chemotherapy benefit; Immune cell infiltration; Network analysis; prediction model.

Figure 1

Data screening and analysis process.

Figure 2

Chemotherapy information of cancer patients in TCGA. (A) Number of chemotherapy and non-chemotherapy patients in different cancers. (B) Percentage of chemotherapy and non-chemotherapy patients in the total patients in different cancers. (C) We screened cancers with a proportion of chemotherapy patients >30% and number of chemotherapy patients >100 for subsequent analysis.

Figure 3

ICI index indicates chemotherapy benefit in multiple cancers. Each row represents a type of cancer. Each column represents a different grouping. A Kaplan-Meier Plotter was used to show the survival prediction ability of ICI index in multiple cancers (BLCA, BRCA, COAD, SKCM, LUSC, UCEC, OV, LUAD, STAD and PAAD). chemo+: chemotherapy patients; chemo-: non-chemotherapy patients.

Figure 4

The composition and interaction of TIICs are diversity among different cancers. (A) Heat map showed the composition of TIICs in chemotherapy patients among cancer. (B) The expressions of Macrophage M0 (marker gene: CCR2) and CD8+ T cells (marker gene: UCHL1) in the preoperative puncture samples of NSCLC chemotherapy and the corresponding histogram of immunostaining score. CR: complete response; PR: partial response; PD: progressive disease. (C, D) The interaction between TIICs in chemotherapy and non-chemotherapy patients of COAD and SKCM.

Figure 5

Identifies six co-owned chemotherapy associated immune feature genes through correlation and causal network analysis. (A) The flow chart showed how we determined chemotherapy associated immune feature genes by correlation and causal network analysis. (B-K) The volcano showed differential immune feature genes of 10 cancers. Blue indicates down-regulated genes, and red indicates up-regulated genes. (L) The histogram showed the number of genes screened by the combined score >0.9 after correlation network analysis of differential immune feature genes. (M) Upset displayed top 20 out-degree and top 20 in-degree genes obtained after causal network analysis. (N) The network relationship of chemotherapy associated immune feature genes in all chemotherapy patients.

Figure 6

Construction of a chemotherapy benefit prediction model and potential function analysis of immune feature genes. (A) Based on six immune feature genes to construct a multivariate cox model, chemotherapy patients were divided into three groups according to the risk score. (B) The learning curve of n_estimators from 0 to 200. (C) ROC curve of this model in the test set. (D) ROC curve of this model in validation sets of GSE25055. (E) Boxplots show the functional similarity of six immune feature genes. The line in the box represents the average value of functional similarity. Proteins with high average functional similarity (cut-off >0.6) are considered interacting proteins. The dashed line indicates the cut-off value. (F, G) Display of biological functions and signal pathways enriched by immune feature genes.