An efficient comparative machine learning-based metagenomics binning technique via using Random forest

Metagenomics is the study of microorganisms collected directly from natural environments. Metagenomics studies use DNA fragments obtained directly from a natural environment using whole genome shotgun (WGS) sequencing. Sequencing random fragments obtained from whole genome shotgun into taxa-based groups is known as binning. Currently, there are two different methods of binning: sequence similarity methods and sequence composition methods. Sequence similarity methods are usually based on sequence alignment to known genome like BLAST, or MEGAN. As only a very small fraction of species is available in the current databases, similarity methods do not yield good results. As a given database of organisms grows, the complexity of the search will also grow. Sequence composition methods are based on compositional features of a given DNA sequence like K-mers, or other genomic signature(s). Most of these current methods for binning have two major issues: they do not work well with short sequences and closely related genomes. In this paper we propose new machine learning related predictive DNA sequence feature selection algorithms to solve binning problems in more accurate and efficient ways. In this work we use Oligonucleotide frequencies from 2-mers to 4-mers as features to differentiate between sequences. 2-mers produces 16 features, 3-mers produces 64 features and 4-mers produces 256 features. We did not use feature higher than 4-mers as the number of feature increases exponentially and for 5-mers the number of feature would be 1024 features. We found out that the 4-mers produces better results than 2-mers and 3-mers. The data used in this work has an average length of 250, 500, 1000, and 2000 base pairs. Experimental results of the proposed algorithms are presented to show the potential value of the proposed methods. The proposed algorithm accuracy is tested on a variety of data sets and the classification/prediction accuracy achieved is between 78% - 99% for various simulated data sets using Random forest classifier and 37% - 95% using Naïve Bayes classifier. Random forest Classifier did better in classification in all the dataset compared to Naïve Bayes.