Development of Machine Learning based Colorectal Cancer Subtype Classificator*

Abstract

The 4 Consensus Molecular Subtypes (CMS1-4) determined by the Colorectal Cancer subtyping Consortium (CRCSC) could have been identified by high-priced methods so far. This study aimed at building a model which can reliably classify patients into the same subtypes with high accuracy using data from publicly available datasets and less expensive clinical procedures. The gene expression data from The Cancer Genome Atlas (TCGA) database was used as a basis for classifying the patients. Our objective was to decrease the number of considered genes from 20000 to around 100 without significant deterioration of the predictive ability of the model. In order to perform the classification, Artificial Neural Networks were trained for the labeled data of the total number of dimensions checking the goodness of the patient classification. Then dimensionality reduction was used, paying attention not to decrease the integrity of the classification significantly. We managed to reduce the number of genes to 100, while we did not deteriorate the accuracy of the classification drastically. The final model on the reduced geneset produced a result of 82% accuracy. The developed software can be used for classifying patients with colorectal cancer. The 100 genes have to be provided for each patient, and the software returns 4 probabilities as a result: the probabilities of belonging to either of the 4 subtypes. The subtype with the highest probability is the final result of the classification.