Cancer Classification with Deep Learning using Genomics Data

Abstract

Deep learning has been monumental in Computer Vision, Natural Language Processing, Machine Translation task and so on. In bioinformatics, Deep learning is playing an important role in drug discover and protein structure prediction. In cancer diagnosis, thanks to advances in Computer Vision Deep Learning models are able to accurately classify cancer. However, not much work has been done in the field of Cancer diagnosis with genomic data. Several authors attempted to use genomic data using machine learning, however it was restricted to single cancer subtypes. In this thesis, we explored classification of all types of cancer using miRNA genome data by creating new model architectures. We are proposing two new architectures a basic ANN and a novel architecture based on ResNet called CResNet. We have trained 4 different kinds of model. LSTM, Artificial Neural Network, CResNet (Variant of ResNet Architecture) and Ensemble models using model averaging. Our models have achieved MCC (Mathew’s correlation coefficient) value of 0.8596,0.9625,0.9745 and 0.9439 which is greater than the SOTA model’s MCC model demonstrating that our architecture performed better than the current architecture.