ISPRF: a machine learning model to predict the immune subtype of kidney cancer samples by four genes

Abstract

Background: Clear cell renal cell carcinoma (ccRCC) is the most common type of renal cell carcinoma (RCC). Immunotherapy, especially anti-PD-1, is becoming a pillar of ccRCC treatment. However, precise biomarkers and robust models are needed to select the proper patients for immunotherapy.

Methods: A total of 831 ccRCC transcriptomic profiles were obtained from 6 datasets. Unsupervised clustering was performed to identify the immune subtypes among ccRCC samples based on immune cell enrichment scores. Weighted correlation network analysis (WGCNA) was used to identify hub genes distinguishing subtypes and related to prognosis. A machine learning model was established by a random forest (RF) algorithm and used on an open and free online website to predict the immune subtype.

Results: In the identified immune subtypes, subtype2 was enriched in immune cell enrichment scores and immunotherapy biomarkers. WGCNA analysis identified four hub genes related to immune subtypes, CTLA4, FOXP3, IFNG, and CD19. The RF model was constructed by mRNA expression of these four hub genes, and the value of area under the receiver operating characteristic curve (AUC) was 0.78. Subtype2 patients in the independent validation cohort had a better drug response and prognosis for immunotherapy treatment. Moreover, an open and free website was developed by the RF model (https://immunotype.shinyapps.io/ISPRF/).

Conclusions: The current study constructs a model and provides a free online website that could identify suitable ccRCC patients for immunotherapy, and it is an important step forward to personalized treatment.

Keywords: Renal cell carcinoma (RCC); immune subtypes; machine learning; online website.

Figure 1

CC for the ccRCC by combining six datasets GSE15641, GSE36895, GSE40435, GSE46699, GSE53757, and TCGA). (A) Consensus matrix heatmap plots when k=2. (B) Tracking plot for k=2 to 6. In the tracking plot, the colors in each row represented the samples in different subtypes (C) Five-year Kaplan-Meier curves for OS of ccRCC patients stratified by the immune subtypes. (D) Five-year Kaplan-Meier curves for PFS of ccRCC patients stratified by the immune subtypes. The log-rank test calculated the P value among subtypes. CC, consensus clustering; ccRCC, clear cell renal cell carcinoma; TCGA, The Cancer Genome Atlas; OS, overall survival; PFS, progression-free survival.

Figure 2

The gene expression scores of 28 immune signatures in two subtypes are displayed by heatmap.

Figure 3

Identification of key modules connected with clinical features and immune subtypes through WGCNA. (A,B) The scale-free fit index and the mean connectivity for various soft-thresholding powers, respectively. When the soft-thresholding powers (β) equaled three, the average degree of connectivity was close to zero. (C) The cluster dendrogram of 5,000 module eigengenes from the TCGA dataset. Each branch in the figure represented one gene, and every color below represented one co-expression module. (D) Heatmap of the correlation between module eigengenes and clinical traits, including molecular subtypes. The color of cells in the heatmap represented the correlation coefficients of different sizes. Specifically, red colors represented the positive correlations, and green colors stood for the negative correlations. The figure without brackets in each cell indicated the clinical feature correlation coefficients. The corresponding P value was shown below in parentheses. WGCNA, weighted correlation network analysis; TCGA, The Cancer Genome Atlas.

Figure 4

PPI network of genes in selected modules. The color intensity and the size of nodes were positively correlated with the degree score. (A) Turquoise module. (B) Blue module. PPI, protein-protein interaction.

Figure 5

The pipeline of machine learning (RF) workflow. RF, random forest; TCGA, The Cancer Genome Atlas; CV, cross-validation; AUC, area under the receiver operating characteristic curve.

Figure 6

Parameter tuning and model validation. (A) The ‘mtry’ with the highest AUC was selected as the optimal value of the RF algorithm. (B) The ‘ntree’ with the highest AUC was selected as the optimal value of the RF algorithm. (C) Validation of model in the testing dataset. (D) The correlation of predicted immune subtype with the response rate to immunotherapy in the IMvigor210 dataset. (E) The correlation of predicted immune subtype with the survival analysis in the IMvigor210 dataset. AUC, area under the receiver operating characteristic curve; RF, random forest; CR, complete response; PR, partial response; SD, stable disease; PD, progressive disease.

Figure 7

The workflow and homepage of Shiny APP. (A) The workflow of RF model in Shiny APP (https://immunotype.shinyapps.io/ISPRF/). (B) The interface shows an example of predicting the immune subtype by four genes expression. RF, random forest.