A novel semi-supervised algorithm for the taxonomic assignment of metagenomic reads

Taxonomic assignment is a crucial step in a metagenomic project which aims to identify the origin of sequences in an environmental sample. Among the existing methods, since composition-based algorithms are not sufficient for classifying short reads, recent algorithms use only the feature of similarity, or similarity-based combined features. However, those algorithms suffer from the computational expense because the task of similarity search is very time-consuming. Besides, the lack of similarity information between reads and reference sequences due to the length of short reads reduces significantly the classification quality.

This paper presents a novel taxonomic assignment algorithm, called SeMeta, which is based on semi-supervised learning to produce a fast and highly accurate classification of short-length reads with sufficient mutual overlap. The proposed algorithm firstly separates reads into clusters using their composition feature. It then labels the clusters with the support of an efficient filtering technique on results of the similarity search between their reads and reference databases. Furthermore, instead of performing the similarity search for all reads in the clusters, SeMeta only does for reads in their subgroups by utilizing the information of sequence overlapping. The experimental results demonstrate that SeMeta outperforms two other similarity-based algorithms on different aspects.

By using a semi-supervised method as well as taking the advantages of various features, the proposed algorithm is able not only to achieve high classification quality, but also to reduce much computational cost. The source codes of the algorithm can be downloaded at http://it.hcmute.edu.vn/bioinfo/metapro/SeMeta.html.