PHIST: fast and accurate prediction of prokaryotic hosts from metagenomic viral sequences.

SUMMARY: PHIST (Phage-Host Interaction Search Tool) predicts prokaryotic hosts of viruses based on exact matches between viral and host genomes. It improves host prediction accuracy at species level over current alignment-based tools (on average by 3 percentage points) as well as alignment-free and CRISPR-based tools (by 14-20 percentage points). PHIST is also two orders of magnitude faster than alignment-based tools making it suitable for metagenomics studies.
AVAILABILITY AND IMPLEMENTATION: GNU-licensed C ++ code wrapped in Python API available at: https://github.com/refresh-bio/phist. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

1 Introduction

Viruses of prokaryotes constitute the vast majority of the global virosphere and play an important role in balancing ecosystems by regulating the composition of bacteria and archaea worldwide. Computational tools allow a routine discovery of previously unknown viruses in metagenomic samples from a wide range of environments (Nayfach ; Paez-Espino ; Ren ). However, little is known about biology of these novel viruses and the hosts with which they interact since isolation and characterization of viruses is remarkably laborious and time-consuming (Edwards ). Computational tools have been developed to predict prokaryotic hosts from metagenome-derived virus sequences based on the molecular signals of virus–host coevolution, including sequence homology (alignment-based tools), sequence composition similarity between viruses and their hosts (alignment-free tools) and matches to host-encoded CRISPR (Clustered Regularly Interspaced Short Palindromic Repeat) spacers (Coclet and Roux, 2021).

Alignment-based tools [e.g. BLASTN (Altschul ), Phirbo (Zielezinski )] produce the highest host prediction accuracy of 25–44% at the species level (i.e. percentage of viruses with a correctly predicted host species; for details, see Supplementary Material). Nevertheless, these tools have considerable time and memory requirements for large datasets, which hinders their use in metagenomic studies. Alignment-free tools [e.g. WIsH (Galiez )] compensate for low speed at the cost of reduced prediction accuracy (18–28% at the species level). Finally, CRISPR-based tools [e.g. SpacePHARER (Zhang )] can provide direct evidence supporting virus–host interactions, but these methods have low prediction accuracy because not many viruses share CRISPR regions with hosts and not many hosts have CRISPR system (Edwards et al., 2016). Here, we introduce Phage–Host Interaction Search Tool (PHIST), a simple tool that allows a fast and accurate host prediction that can be performed on a standard workstation or even a personal laptop computer.

2 Materials and methods

PHIST takes as input two directories containing compressed or uncompressed FASTA files of genomic sequences of viruses and candidate hosts, respectively. The tool links viruses to hosts based on the number of k-mers shared between their sequences. PHIST is built upon the Kmer-db tool (Deorowicz ) using its recently developed mode of determining numbers of shared k-mers (Supplementary Material). PHIST outputs a CSV file with a matrix of common k-mer counts between every virus and prokaryote, and a summary CSV file reporting top scoring hosts for each virus. In addition, PHIST provides P-values for the predicted virus–host pairs (Supplementary Material). Information on reference datasets and benchmarking host prediction tools is provided in Supplementary Material.

3 Results

We evaluated the performance of PHIST on a reference dataset of 2288 viral genomes and 62 493 complete and draft prokaryotic genomes from (Wang ). PHIST correctly identified the hosts species for 29.5% of viruses representing an improvement by 4 percentage points over Phirbo and BLASTN (Fig.  1A). A similar, minor enhancement at the species level (1–2 percentage points) was obtained on two previously published benchmark sets (Edwards ; Galiez ) (Supplementary Material). The tool also had the highest host prediction accuracy at the species level when considering the top five predicted host species (i.e. host species with highest similarity to the query virus) (Fig.  1B). PHIST recovered most of the virus–host interaction pairs that were correctly predicted by other approaches. Specifically, our tool recalled 86% (486 out of 561), 85% (481/563), 93% (353/381), 85% (277/329) and 85% (248/292) virus–host pairs that were correctly predicted by BLASTN, Phirbo, WIsH, SpacePHARER and PILER-CR, respectively (Fig.  1C).

Fig. 1.

(A) Host prediction accuracy for 2288 viral genomes against 62 493 candidate hosts, binned by host taxonomic level. (B) Host prediction accuracy at the species level across five highest scoring prokaryotic species for each virus sequence (the prediction was scored as correct if the correct host species was among the first five predicted hosts). (C) Unique and shared correct virus–host predictions among the tools. The bar chart indicates the intersection size of viruses with correct host prediction at the species level. Connected dots on the bottom panel indicate which pair of tools is considered for each intersection. (D) Runtime and memory usage for 2288 viruses and 62 493 prokaryotes. Time and memory measurements of Phirbo also include BLASTN, and SpacePHARER includes PILER-CR analysis. (E) Prediction accuracies for contigs subsampled at various lengths from the 2288 virus genomes. The solid line shows the mean prediction accuracy at host taxonomic level and the grey shade indicates the 95% confidence interval. The horizontal dashed line marks the prediction accuracy using the full-length viral genomes. (F and G) Proportions of congruent predictions for viral contigs between PHIST and those in Paez-Espino et al. and Nayfach et al., respectively. (H) Viral assignments to human gut bacteria and archaea. Distribution of host classes reported by Nayfach et al. (upper bar chart) and a set of PHIST predictions on viruses that were not assigned to hosts by Nayfach et al. (lower bar chart)

PHIST took an hour to process the dataset (approximately 143 million pairwise comparisons between 2288 viruses and 62 492 candidate hosts) on a 16-core 2.1 GHz Intel Xeon, almost 300 times faster than BLASTN and Phirbo, 16 times faster than WIsH and 2–4 times faster than CRISPR-based tools (Fig.  1D).

To evaluate the performance of host prediction for short viral contigs, we randomly subsampled fragments of different lengths (1, 2, 5, 10 and 20 kb) from each of the 2288 viral genomes. The mean host prediction accuracy at the genus and family levels dropped by 9 and 11 percentage points, respectively, for 20–3 kb long contigs (Fig.  1E). At the species level, however, the prediction accuracy fell by only 3 percentage points (from 26% to 23%) for 20 kb to 3 kb long contigs. We also tested PHIST on a set of 125 842 metagenomic viral contigs (MVCs) of 11 kb median length obtained from various environments (Paez-Espino ). The original host prediction used CRISPR-spacer and tRNA sequence matches, and assigned hosts for only 7.9% of the MVCs. PHIST annotated 99% of the MVCs; moreover, the predictions matched the original predictions in 30% at the species level and achieved 80% consistency at the genus level (Fig.  1F).

The resulting accuracies can be considered as lower bounds since metagenomic studies restrict the set of candidate host genomes to those present in the sample (Galiez ). Thus, we used a set of 189 680 metagenomic viral sequences from human gut provided by Nayfach , who predicted the hosts among 286 997 genomes of bacteria and archaea from the Unified Human Gastrointestinal Genome collection (Almeida ). The original host prediction used a combination of CRISPR-spacer matches and whole-genome alignments between viruses and candidate hosts (Nayfach ) and assigned hosts to 90% of the viral contigs (n = 170 072). PHIST annotated 99.95% of the viral contigs (n = 189 586) and provided an overall 81% agreement on host taxonomy with the original predictions (Supplementary Material). Using the Nayfach annotation as reference, PHIST obtained 68 and 81% host prediction accuracy at the species and genus levels, respectively (Fig.  1G). The remaining viral genomes that were not assigned to hosts in Nayfach et al. (n = 19 514) were connected by PHIST mostly to Clostridia and Bacteroidia (Fig.  1H, lower bar chart), which were also two dominant classes of bacteria in Nayfach et al. predictions (Fig.  1H, upper bar chart). Computing all ∼54 billion pairwise scores between 189 680 viruses and 286 997 prokaryotes required just 3.5 h and 25 GB of RAM (Supplementary Material).

4 Conclusion

PHIST predicts hosts species with higher accuracy than alignment-based tools, and allows for rapid analysis of large-scale genomic and metagenomic datasets on non-specialized hardware.

Funding

This work was supported by the National Science Centre, Poland [2018/31/D/NZ2/00108 to A.Z., DEC-2016/21/D/ST6/02952 to A.G., DEC-2019/33/B/ST6/02040 to S.D.].

Conflict of Interest: none declared.

Data availability

All data analysed in this study are available from previously published studies (Supplementary Material).

Click here for additional data file.