Hepatitis B viruses (HBVs) are compact viruses with circular genomes of ∼3.2 kb in length. Four genes (HBx, Core, Surface, and Polymerase) generating seven products are encoded on overlapping reading frames. Ten HBV genotypes have been characterised (A-J), which may account for differences in transmission, outcomes of infection, and treatment response. However, HBV genotyping is rarely undertaken, and sequencing remains inaccessible in many settings. We set out to assess which amino acid (aa) sites in the HBV genome are most informative for determining genotype, using a machine learning approach based on random forest algorithms (RFA). We downloaded 5,496 genome-length HBV sequences from a public database, excluding recombinant sequences, regions with conserved indels, and genotypes I and J. Each gene was separately translated into aa, and the proteins concatenated into a single sequence (length 1,614 aa). Using RFA, we searched for aa sites predictive of genotype and assessed covariation among the sites with a mutual information-based method. We were able to discriminate confidently between genotypes A-H using ten aa sites. Half of these sites (5/10) sites were identified in Polymerase (Pol), of which 4/5 were in the spacer domain and one in reverse transcriptase. A further 4/10 sites were located in Surface protein and a single site in HBx. There were no informative sites in Core. Properties of the aa were generally not conserved between genotypes at informative sites. Among the highest co-varying pairs of sites, there were fifty-five pairs that included one of these 'top ten' sites. Overall, we have shown that RFA analysis is a powerful tool for identifying aa sites that predict the HBV lineage, with an unexpectedly high number of such sites in the spacer domain, which has conventionally been viewed as unimportant for structure or function. Our results improve ease of genotype prediction from limited regions of HBV sequences and may have future applications in understanding HBV evolution.
- hepatitis B virus
- machine learning