Comparative Study Using Neural Networks for 16S Ribosomal Gene Classification

Abstract

Bacterial 16S ribosomal gene was used to classify bacteria because it consists of both highly conservative region, as well as a hypervariable region, in its sequence. This hypervariable region serves as a discriminative factor to differentiate bacteria at taxonomic levels. In the past, many efforts have been made to correctly identify a bacterial species from environmental samples or human gut microbiome samples, yet this identification and subsequent classification task is challenging. For such bacterial taxonomic classification, several studies in the past have been performed based on k-mer frequency matching, assembly-based clustering, supervised/unsupervised machine learning models, and a very few studies with deep learning architectures. In this article, we study and propose six different deep learning architectures involving recurrent neural networks (RNNs) and convolutional neural networks to classify bacteria at a family, genus, and species taxonomic level using ∼12,900 16S ribosomal DNA sequences. The best classification accuracies achieved are 92%, 86%, and 70% at family, genus, and species taxonomic level, respectively, by variants of RNN.