Predicting the Metagenomics Content with Multiple CART Trees

Metagenomics is a technique for the characterization and identification of microbial genomes using direct isolation of genomic DNA from the environment without cultivation. One of the key step in this process is the taxonomic classification and clustering of the DNA fragments, process also known as binning. To date, the most common practice is classifying through alignments to public databases. When a representing specie is present in this database the process is simple and successful, if not, an underestimation of taxonomic abundances is produced. In this work we propose a alignment-free method capable of assign taxa to each read in the sample by analyzing the statistical properties of the reads. Given an environment, we collect genomes from public available databases and generate genomic fragments libraries. Then, statistics of k-mer frequencies, GC ratio and GC skew are computed for each read and stored in an environment-associated dataset used to build a robust machine learning procedure based on multiple CART trees. Finally, for each read the CART trees are asked about their taxa and the most voted ones are selected. The method was tested using simulated and public human gut microbiome data sets. The database was constructed using 98 genera present in Gastrointestinal Tract available at Human Microbiome Project. A multiple CART tree with 558-trees predictor was generated, capable to estimate the genus and abundance in the sample with 47 % of accuracy in read assignments. Performance rates are comparable with those from semi-supervised methods and also the computation times were reduced due to alignment-free methodology. Restricted to 17 early considered genera, our method increases its accuracy to 77 %.