Analysis of patch-based GCN risk score distribution corresponding to gene mutations in colorectal cancer using whole slide images.

Abstract

e15650 Background: Colorectal cancer (CRC) is one of the most common malignancies, with the third highest incidence and the second highest mortality in the world. Risk scores are typically employed to predict outcomes such as disease risk, prognosis, or treatment response. Risk scores are usually calculated by combining biomarkers or genetic variants with statistical models, such as logistic regression or Cox proportional hazards models, and then observe the relationship between genetic variations and survival time through survival analysis. We have constructed a deep learning model based on Graph Convolutional Networks (GCNs), which is capable of predicting risk scores using Whole Slide Images (WSIs). Statistical analyses were applied to explore the distribution of risk scores associated with different variants in colorectal cancer. Methods: We performed risk scores prediction on a total of 384 samples from TCGA-COAD and TCGA-READ using our patch-based GCN model. We analyzed the distribution of risk scores in colorectal cancer samples and investigated the effects of gene mutations in key pathways on these risk scores. Results: Risk score predictions by our patch-based GCN model were conducted on 384 samples, revealing that the maximum, minimum, median, mean, and standard deviation of risk scores were -0.78, -1.89, -1.34, -1.34, and 0.20, respectively. The data overall conformed to a normal distribution ( p < 0.05 ). We selected core genes in key pathways combined with top ten genes with the highest mutation frequency in colorectal cancer for statistical analysis to explore the correlation with the predicted risk scores. These gene mutations were found in a total of 346 patients. The results indicated that mutations in the TP53 and OBSCN genes led to significant differences in risk scores as predicted by our model ( p < 0.05 ). The risk scores based on gene mutations also have significant differences and consistent trend in these genes. Conclusions: Our results indicate the risk scores generated by the patch-based GCN model have the consistent trend with the risk scores generated by mutation information of core genes in colorectal cancer. Whether risk scores based on digital pathology prediction can replace mutation detection in predicting patient survival status requires further exploration.