Combining autoencoder and artificial neural network for classifying colorectal cancer stages

The genomics era produced a large amount of molecular data. Many efforts have been made in the last decade to sequence as many types of tumours as possible. The Genome Data Commons (GDC) is the largest repository of cancer molecular and clinical data. To date, the challenge is to use them to improve tumour classification and therapeutic approaches. Bioinformatics and data science became even more important to the aim to develop algorithms for translating genomic data into clinical practice. Colorectal cancer (CRC) is one of the deadliest malignancies in the world and, despite the therapeutic advances, much more is far from being known to better address patients. In the present study, we aimed to classify CRC tumour stages through gene expression data. Autoencoder and ANN are combined in a CRC grades classification framework based on gene expression. After performing differential expression analysis, we evaluated different strategies for features reduction. Since the autoencoder allowed to transform the feature space from 3213 genes to 64 features, it was used as input to an ANN. The robustness of the designed classifier was evaluated training and testing the ANN 250 times, randomly splitting data into training (80 %) and test (20 %) sets. Results are reported as mean accuracy, sensitivity and specificity, showing about 84 % for accuracy, 89 % for sensitivity and 78 % of specificity. In conclusion, the proposed approach could be useful in the molecular classification based on transcriptomic data of the pathological stages of CRC.