Deep Learning Approach to Taxonomic Classification with CNN-ELM

Abstract

In the fight against viruses, DNA sequencing plays a critical role. When a patient gets infected, sequencing a sample of their viral genome allows scientists to: identify the virus, trace its origin, and develop vaccines. This paper presents a novel approach for classifying bacterial DNA sequences. It utilizes modified convolutional neural networks (CNNs) that extract features from DNA sequences represented as images using Frequency Chaos Game Representation (FCGR). These features are then classified by a combination of two extreme learning machines (ELMs) following the final fully connected layer of the CNN model. ELMs offer advantages such as rapid learning, straightforward convergence, and reduced randomness. The proposed CNN-ELM system achieved an impressive 98.5% accuracy for full-length sequences and 81.2% accuracy for 500bp-length sequences using the Ribosomal Database Proj ect, Release 11 database.