A machine learning approach for viral genome classification

Abstract

Background: Advances in cloning and sequencing technology are yielding a massive number of viral genomes. The classification and annotation of these genomes constitute important assets in the discovery of genomic variability, taxonomic characteristics and disease mechanisms. Existing classification methods are often designed for specific well-studied family of viruses. Thus, the viral comparative genomic studies could benefit from more generic, fast and accurate tools for classifying and typing newly sequenced strains of diverse virus families.

Results: Here, we introduce a virus classification platform, CASTOR, based on machine learning methods. CASTOR is inspired by a well-known technique in molecular biology: restriction fragment length polymorphism (RFLP). It simulates, in silico, the restriction digestion of genomic material by different enzymes into fragments. It uses two metrics to construct feature vectors for machine learning algorithms in the classification step. We benchmark CASTOR for the classification of distinct datasets of human papillomaviruses (HPV), hepatitis B viruses (HBV) and human immunodeficiency viruses type 1 (HIV-1). Results reveal true positive rates of 99%, 99% and 98% for HPV Alpha species, HBV genotyping and HIV-1 M subtyping, respectively. Furthermore, CASTOR shows a competitive performance compared to well-known HIV-1 specific classifiers (REGA and COMET) on whole genomes and pol fragments.

Conclusion: The performance of CASTOR, its genericity and robustness could permit to perform novel and accurate large scale virus studies. The CASTOR web platform provides an open access, collaborative and reproducible machine learning classifiers. CASTOR can be accessed at http://castor.bioinfo.uqam.ca .

Keywords: Prediction; Sequence classification; Virus classification.

Fig. 1

Overview of CASTOR kernel architecture. The kernel is composed of two main units (classifier construction and prediction). White rectangles represent input and output data; grey and curved rectangles represent processes. TS and VS are training set and validation set, respectively

Fig. 2

Class cohesion of three virus datasets. The four columns illustrate the separability and compactness of three virus complete genomes datasets based on 172 restriction enzyme cuts. The first column shows heatmaps of CUT clustered by x-axis. The samples in the y-axis are grouped by studied classes followed by intra-class clusterings. The second column shows MDS of the CUT distances between samples. The third and fourth column represent, respectively, the Cohesion and Silhouette indexes of the classes. a Classes in HPV are Alpha species, Beta and Gamma genera. b Classes in HBV are A-H genotypes c Classes in HIV-1 are M pure subtypes and CRFs

Fig. 3

Learning algorithm evaluation on five datasets. This figure illustrates the F-measure distribution (boxplot) of seven learning algorithms on the prediction of a HPV genera, b HPV Alpha species, c HBV genotypes, d HIV-1 M subtypes with complete genomes e HIV-1 M subtypes with pol fragments. HPV and HBV datasets are complete genomes. The number below each boxplot corresponds to the statistically discriminative rank of the algorithms. The ranking is performed with paired Student’s t test. μ, σ are the mean and the standard deviation of the overall F-measures, respectively. p is the p-value of the statistically significance of the weighted F-measure mean differences among the algorithms computed with the Wilcoxon/Kruskal-Wallis test

Fig. 4

Performance of CASTOR with COMET and REGA predictors on HIV-1 datasets. The panels a and b show the percentage of correct classifications for HIV-1 complete genomes and HIV-1 pol fragments, respectively. The number of instances and the associated classes for each sampling is presented above the panels. Complete sampling corresponds to 10% of Los Alamos HIV data selected randomly. In specific subtypes sampling, the predictors are assessed against their trained classes. In common subtypes sampling, the predictors are assessed against the intersection of the classes of the three trained predictors