Construction of artificial neural network model for predicting the efficacy of first-line FOLFOX chemotherapy for metastatic colorectal cancer

Abstract

Objective: To explore and establish an artificial neural network (ANN) model for predicting the efficacy of first-line FOLFOX chemotherapy for metastatic colorectal cancer. Methods: A set of FOLFOX chemotherapy data from a group of patients with metastatic colorectal cancer (mCRC) (GSE104645) was downloaded from the GEO database as a training set. According to the FOLFOX protocol, the efficacy was divided into two groups: the chemo-sensitive group (including complete response and partial response) and the chemo-resistant group (including stable disease and progressive disease), including 31 cases in the sensitive group and 23 in the resistant group. Then, chip data (accessible number: GSE69657) from Fujian Medical University Union Hospital were chosen as a test set. A total of 30 patients were enrolled in the study, including 13 in the sensitive group and 17 in the resistant group. The batch effect correction was performed on the expression values of the two sets of matrices using the R 3.5.1 software Combat package. The gene expression difference of sensitive and resistant group in GSE104645 was analyzed by the GEO2R platform. P<0.05 and the absolute value of log(2)FC>0.33 (FC abbreviation of fold change) were used as the threshold value to screen the drug resistance and sensitive genes of the FOLFOX regimen. An ANN was constructed using the multi-layer perceptron (MLP) to perform the FOLFOX regimen on the GSE104645 dataset. The GSE69657 expression matrix and clinical efficacy parameters were then used for retrospective verification. Receiver operating characteristic(ROC) curves were used to evaluate the test results and predictive power. Results: A total of 2, 076 differentially expressed genes in GSE104645 were selected, of which 822 genes were up-regulated and 1, 254 genes were down-regulated in the chemo-resistance group. The down-regulated genes were sensitive genes. GO analysis of the biological processes in which the differentially expressed genes were involved, revealed that they were mainly involved in the regulation of substance metabolism. A total of 39 genes were included in the final model construction. This was a neural network model with two hidden layers. The accuracy of predicting training samples and test samples was 75.7% and 76.5%, respectively, and the area under the ROC curve was 0.875. The chip data set of our department (GSE69657) was set as the test set, and the area under the ROC curve was 0.778. Conclusions: In this study, an artificial neural network model is successfully constructed to predict the efficacy of first-line FOLFOX regimen for metastatic colorectal cancer based on the microarray, and an independent external verification is also conducted. The model has good stability and well prediction efficiency. Besides, the results of this study suggest that the gene functions related to oxaliplatin resistance are mainly enriched in the regulation process of substance metabolism.