Virus genome sequence classification using features based on nucleotides, words and compression

TL;DR

This study evaluates various genome sequence features and classifiers for virus taxonomy, achieving high accuracy in classifying viruses into ICTV Orders using sequence-derived features and machine learning.

Contribution

It introduces a comprehensive comparison of nucleotide, word, and compression-based features with classifiers for virus genome classification, highlighting the effectiveness of 4-mer counts and SVM.

Findings

Best classifier: 4-mer counts with SVM (error rate 0.006)

Genome length and k-NN perform worst but still reasonably

Predicted orders for unclassified viruses with high accuracy

Abstract

The ICTV develops, refines and maintains a universal virus taxonomy; Order is the highest taxon in the branching hierarchy of recognised viral taxa. Historically, ICTV (sub)committees have classified viruses on the basis of morphological characteristics and various other attributes. Today, virtually all new viral genomes are assembled from metagenomic datasets and are not linked directly to biological agents. Thus, placing a virus into a taxonomic scheme solely from primary genome structure is an increasingly important problem. Various simple descriptive statistics of a viral genome sequence have been used successfully for virus classification. Here, we use the NCBI's viral and viroid reference sequence collection (RefSeq) and a common experimental framework to compare the performance of different genome sequence-derived features and classifiers in the task of assigning a virus to one of seven ICTV Orders. The nucleotide-, word-, and compression-based features we consider include genome length, the k-mer Natural Vector (k = 1, ..., 6) and its derivatives, return time distribution, and general-purpose and DNA-specific compression ratios; the classifiers used are the k-NN and SVM. The combination of genome length and k-NN has the worst, yet still respectable, performance (mean error rate of 0.137); the best performance is achieved using 4-mer counts and SVM (mean error rate of 0.006). We investigate the main causes of misclassification, explore which viruses are more difficult to classify, and use the best performing combination to predict the Orders of 1,834 unclassified viruses. A subsequent version of RefSeq assigned Orders to 17 of these previously unlabelled viruses. Since 16 of our predictions match these assignments, our approach could aid virologists dealing with viruses that are known only from sequence data.