MT-MAG: Accurate and interpretable machine learning based taxonomic assignment of metagenome-assembled genomes, with a partial classification option
We propose MT-MAG, a novel machine learning-based taxonomic assignment tool for hierarchically-structured local classification of metagenome-assembled genomes (MAGs). MT-MAG is capable of classifying large and diverse real metagenomic datasets, having analyzed for this study a total of 240 Gbp of data in the training set, and 7 Gbp of data in the test set. MT-MAG is, to the best of our knowledge, the first machine learning method for taxonomic assignment of metagenomic data that offers a "partial classification" option. MT-MAG outputs complete or a partial classification paths, and interpretable numerical classification confidences of its classifications, at all taxonomic ranks. MT-MAG is able to completely classify 48% more sequences than DeepMicrobes to the Species level (the only comparable taxonomic rank for DeepMicrobes), and it outperforms DeepMicrobes by an average of 33% in weighted accuracy, and by 89% in constrained accuracy.