Dimensionality Reduction for Cluster Identification in Metagenomics using Autoencoders

Metagenomics is the study of the genomic content of the microbial organisms extracted from a sample in their natural habitats. These unknown collections of genomic data are analyzed without any prior lab-based cultivation to avoid amplification bias. One of the vital aspects of metagenomics analysis is the visualization of the information that is derived from the genomic sequences of a microbiome sample. In a successful visualization, the congruent reads of the sequences should appear in clusters depending on the diversity and taxonomy of the microorganisms in the sequenced sample. In converting higher dimensional sequence data into lower dimensional data for visualization purposes, preserving the genomic characteristics is given the highest priority. In this process, the demand for precise and efficient methods of dimensionality reduction is crucial. Currently, Principle Component Analysis (PCA) and t-distributed Stochastic Neighbor Embedding (t-SNE) are used for dimensionality reduction purposes in metagenomics, which are linear and non-linear techniques respectively. Although the above-mentioned techniques are widely used, there are shortcomings in accuracy and efficiency in terms of visualizations. In this paper, we explore the possibility of using autoencoders, a deep learning technique, to get a rich dimensionality reduction, overcoming the prevailing impediments of PCA and t-SNE and outperforming them to achieve better metagenomic visualizations.