vClassifier: a toolkit for high-resolution phylogenetic classification of prokaryotic viruses Authors: Kun Zhou, James C Kosmopoulos, Karthik Anantharaman Published: 2026 ABSTRACT AbstractMotivationAs the most abundant and diverse biological entities, prokaryotic viruses play pivotal roles in ecological systems. Their taxonomic classification has been instrumental in elucidating their diversity and ecological functions. However, determination of viral taxonomy remains a considerable challenge. Recently developed approaches succeed in assignment of viral taxonomy at higher ranks, such as at the family level and above, but struggle at the subfamily level and below to the genus and species resolutions.ResultsWe describe a phylogeny-informed methodology to provide species-level taxonomic assignments of viruses. We used single-copy marker genes relevant to specific taxa and reference phylogenetic trees for these groups which facilitates direct comparisons with the taxonomic framework of the International Committee on Taxonomy of Viruses (ICTV). Our method demonstrated significant congruence with the ICTV taxonomy, showing 84%–91% alignment at the subfamily and genus levels. For species-level classification, our strategy was integrated with average nucleotide identity, yielding a high congruence rate of over 92% with the taxonomic data from the NCBI Virus database. This framework is implemented in vClassifier, a high-accuracy toolkit developed for standardized viral taxonomic assignment. Benchmarking comparisons revealed that vClassifier matches or surpasses other available tools regarding assignment rates. By achieving objectivity and high levels of consistency, vClassifier streamlines the taxonomic categorization of prokaryotic viral genomes. Accurate assignments at the subfamily, genus, and species levels will significantly refine the taxonomic resolution of viruses, fostering a deeper understanding of viral diversity in microbiomes and ecosystems.Availability and implementationvClassifier is publicly accessible via https://github.com/AnantharamanLab/vClassifier. 1 Introduction Prokaryotic viruses, which infect bacteria and archaea, constitute the most abundant biological entities on our planet (Koonin et al. 2015). Their influence on ecosystem dynamics is profound, as they significantly shape global biogeochemical cycles and life on Earth (Suttle 2007). Moreover, these viruses play a critical role in regulating microbial populations, promoting genetic diversity through horizontal gene transfer, and driving evolutionary processes (Hendrix and Casjens 2005). By lysing host cells, they release organic matter that fuels nutrient cycles, thereby maintaining ecological balance (Fuhrman 1999). However, despite their ecological importance, the vast diversity of prokaryotic viruses has historically been challenging to categorize (Paez-Espino et al. 2016). Traditional methods for classifying these viruses have been limited by the lack of suitable genetic markers and the immense genetic variability found among viral populations (Hatfull 2008). Recent advancements in computational methodologies for deducing the taxonomy of prokaryotic viruses have revealed bacteriophages and archaeal viruses are vastly diverse. These computational techniques have predominantly utilized gene content, network-based criteria, genome-centric phylogenetic criteria, and pairwise distances and associated thresholds (Gorbalenya and Lauber 2022). Prominent tools for classifying prokaryotic viruses at the family level and above include geNomad and vConTACT (Bolduc et al. 2017, 2025, Bin Jang et al. 2019, Camargo et al. 2024). For classification at the subfamily and genus levels, tools such as vConTACT (Bolduc et al. 2017, Lemoine et al. 2018, Bin Jang et al. 2019), GRAViTy (Aiewsakun et al. 2018, Aiewsakun and Simmonds 2018), VICTOR (Meier-Kolthoff and Goker 2017), ViPTree (Nishimura et al. 2017), and VirClust (Moraru 2023) have been developed. vConTACT generates viral protein clusters and analyzes gene-sharing networks across whole genomes. In contrast, GRAViTy leverages genome features like gene content and orientation for systematic analyses. VICTOR integrates phylogenetic and clustering approaches, utilizing BLAST to compute intergenomic distances with the Genome Blast Distance Phylogeny method (Auch et al. 2010). Similarly, ViPTree also constructs phylogenetic trees based on pairwise genomic distances. Aside from the other tools, VirClust operates without the need for a reference and builds viral genome trees through protein clustering. These tools have proven to be sensitive and accurate in the study of prokaryotic viruses. Nonetheless, challenges remain in taxonomic classification. For instance, VICTOR tends to be slow and lacks scalability, constrained by the number of genomes it can process, and it may not fully account for the monophyly of taxa within trees. Additionally, geNomad, vConTACT, and GRAViTy do not sufficiently consider viral phylogenies, which are pivotal for delineating evolutionary relationships critical to taxonomy (Parks et al. 2018). Moreover, ViPTree and VirClust do not perform bootstrapping for large datasets. Lastly, accurate species-level assignments are crucial for understanding the specific roles and impacts of individual viral species within their ecosystems (Brum and Sullivan 2015). However, the limited availability of software dedicated to the automated species-level assignment of viral metagenomes from human and environmental samples, poses a significant challenge for detailed viral taxonomy and ecological studies. Genome-based phylogenies are ideal for capturing evolutionary relationships, making the criteria of monophyletic clusters and tree topology apt for taxonomic delineation (Simmonds et al. 2023). This approach is well-established in prokaryotes using conserved single-copy genes (SCGs), as demonstrated by tools such as GTDB-Tk (Chaumeil et al. 2019, 2022). For dsDNA prokaryotic viruses within the Caudoviricetes class, SCGs are known to be conserved within certain taxonomic ranks, notably families and subfamilies, such as Ackermannviridae and Ounavirinae (Low et al. 2019). This conservation suggests that members within a family or subfamily possess a similar set of single-copy genes, hinting at the potential universality of SCGs in lower taxonomic ranks of viral taxa. Here, we identified single-copy markers specific to a broad spectrum of taxa, spanning from genera to realms, based on restrictive criteria (see Section 2). We constructed reference genome trees using concatenated sequences of these single-copy proteins. Our findings demonstrated that the phylogenetic topologies of taxa are highly consistent with the ICTV taxonomy at the subfamily and genus levels. We incorporated this phylogeny-based strategy into the taxonomic assignment of prokaryotic viruses and present vClassifier, an automated tool that identifies single-copy markers of viral query genomes and places them into concatenated protein reference trees specific to families or subfamilies for taxonomic assignment. vClassifier assigns taxonomic classifications to genomes of bacteriophages and archaeal viruses based on the monophyly of viral groups within these trees. Moreover, vClassifier integrates average nucleotide identity to determine species-level classifications. When benchmarked against other programs, vClassifier demonstrated comparability or superiority in performance. Overall, we demonstrate that viruses contain single-copy marker genes that can be used to obtain automated accurate species-level assignments using vClassifier. vClassifier is freely available for download at https://github.com/AnantharamanLab/vClassifier. 2 Methods 2.1 Selection of viral taxa and protein markers We accessed the Master Species List (MSL)38 v1 release of the ICTV taxonomy system (25 April 2023). Our selection criteria focused on complete genomes with corresponding GenBank accessions, which were also classified according to the ICTV taxonomy. We compiled a dataset of viral genomes with these selected GenBank accessions from the National Center for Biotechnology Information (NCBI) database comprising 5128 genomes from a diverse range of bacterial and archaeal virus taxa. To ensure robustness in our analysis, we included only those taxa for which at least 10 genomes were available, as detailed in Table 1, available as supplementary data at Bioinformatics Advances online. Retained genomes spanned 212 distinct taxa and were organized into eight taxonomic ranks defined by the ICTV taxonomy. For identifying markers within each taxon, we retrieved hidden Markov models (HMMs) profiles of virus orthologous protein groups from the Virus Orthologous Groups database (VOGdb, release vog216) at http://vogdb.org, accessed on 31 March 2023. We then screened the protein groups from VOGdb against the retained genomes, which spanned 212 taxa, using HMMER version 3.3.2 (Mistry et al. 2013). The screening process adhered to stringent parameters established previously: an E-value threshold of 1 × 10-³ and a minimum alignment coverage of 50% (Low et al. 2019). Ultimately, a range of 1 to 276 VOGdb protein groups that were conserved across 203 taxa were selected as markers, detailed in Tables 1 and 2, available as supplementary data at Bioinformatics Advances online. The selection of these markers was based on specific criteria: (i) occurrence in 50% or more of the genomes of bacterial and archaeal viruses within a taxon; (ii) an average copy number not exceeding 1.2 per genome to ensure the single-copy nature; and (iii) an average protein length greater than 100 amino acids, to facilitate reliable phylogenetic analysis. 2.2 Phylogenetic inference using markers within each taxon For each of the 203 taxa, single-copy marker proteins were aligned using MAFFT version 7.505 (Katoh and Standley 2013). To improve the precision of our data, each Multiple Sequence Alignment (MSA) of markers was individually refined using trimAL version 1.4.rev15 (Capella-Gutiérrez et al. 2009). This refinement involved trimming columns that appeared in less than 50% of the taxa, enhancing the overall quality of the alignments. Following this, we concatenated the refined individual alignments, inserting gaps in the sequences where markers were absent from certain genomes. This concatenated MSA was further processed to exclude genomes that represented less than 5% of the total alignment length in amino acids, ensuring that only highly representative sequences were included in our phylogenetic analysis. The final reference tree was inferred from the refined MSA using Fast-TREE version 2.1.11 (Price et al. 2010) under the LG + GAMMA model with 1000 nonparametric bootstrap replicates. The resultant phylogenetic trees were rooted at their midpoints (Hess and De Moraes Russo 2007), providing a reliable baseline for tree interpretation. In total, 203 phylogenetic trees were generated, each representing the evolutionary relationships within a specified taxon. 2.3 Calculation and comparison of the consistency with ICTV and NCBI classification systems We assessed the consistency of our phylogenetic method with the ICTV and NCBI classification systems. To facilitate this assessment, we accessed ICTV taxon information (MSL38 v1 release). Each ICTV/NCBI taxon comprising two or more genome representatives (exemplar sequences designated by the ICTV for each recognized species) was defined as a phyletic group. Moreover, an ICTV/NCBI taxon was considered monophyletic if it included at least two genome representatives and achieved a bootstrap value of 75% or higher, as determined by the topology within the respective tree for each taxon. The consistency of our phylogenetic method was quantified by the ratio of the number of monophyletic groups to the total number of phyletic groups within each taxon (Low et al. 2019). To determine the overall congruence with ICTV/NCBI at a given taxonomic rank, we summed the number of monophyletic groups and divided by the total number of phyletic groups across all analyzed taxa at that rank. A comparative analysis was performed using ViPTree version 1.1.2 (Nishimura et al. 2017), VirClust version 2.0 (Moraru 2023), and vConTACT3 v3.0.0b34 (Bolduc et al. 2025) to evaluate classification performance at the subfamily and genus levels. These tools were applied to the same 36-family (55-subfamily) dataset using default parameters. The resulting outputs were then analyzed to determine their respective levels of congruence with official ICTV taxonomy in comparison to our phylogenetic method. The phylogenetic trees derived from taxon-specific markers for the 36 families and 55 subfamilies were selected to serve as the reference trees in the subsequent vClassifier workflow, as detailed below. 2.4 Workflow and performance of vClassifier 2.4.1 Placement of genomes in reference trees vClassifier processes genomes affiliated with 36 families or 55 subfamilies of viruses, accepting data in FASTA format. The tool uses Prodigal version 2.6.3 with options (-m -p meta) to predict genes (Hyatt et al. 2010). Following gene prediction, single-copy marker proteins from the genome queries are identified and aligned using HMMER version 3.3.2, employing the previously described method. The sequence with the highest bit score is considered the best hit, and in instances of multiple hits, the sequence with the highest score is selected for further analysis. The identified marker genes are then aligned using MAFFT version 7.505 (Katoh and Standley 2013). This alignment process ensures that the sequences are accurately organized according to their evolutionary relationships. The aligned sequences are subsequently concatenated into a comprehensive multiple sequence alignment. This alignment is further refined by trimming columns represented in less than 50% of taxa using trimAL, ensuring that only the most conserved and relevant genetic information is retained. The final step involves placing the genomes into reference phylogenetic trees that were previously generated (see details in Sequence alignment and phylogenetic inference) using pplacer version 1.1.alpha19-0-g807f6f3 (Matsen et al. 2010). Figure 1 describes the overall workflow of vClassifier. Flowchart of phylogenetic reconstruction and vClassifier workflow. The left box represents the initial phase, where phylogenetic trees are constructed based on viral SCGs derived from Viral Orthologous Groups (VOGs). Taxa: biological classification ranks such as species, genera, families, etc. 2.4.2 Subfamily/genus/species assignment The classification of a query genome at the subfamily or genus level depends on its precise positioning within the reference phylogenetic tree. If every reference genome within a monophyletic cluster achieves a bootstrap value of 75% or higher and all belong to the same reference subfamily or genus, the query genome in the cluster is then assigned to that particular subfamily or genus. This method ensures that the classification is both accurate and reflective of established phylogenetic relationships, providing a robust basis for taxonomic assignment. For the classification of a query genome at the species level, alignment coverage and average nucleotide identity (ANI) are critical factors. Species assignments are determined using ANI, calculated with FastANI version 1.33 (Jain et al. 2018), applying customized parameters (—fragLen 500—minFraction 0.8). Specifically, if a query genome is positioned within a genus, it is assigned to the species of the closest reference genome that demonstrates high alignment similarity. This determination requires that the reference genome meet stringent criteria: an alignment fraction of at least 80% and an ANI of 95% or higher. Additionally, the species assignment is based on the highest composite alignment score, calculated as the product of the alignment fraction and the ANI (fraction × ANI), ensuring the most accurate and reliable taxonomic placement. 2.4.3 Requirements vClassifier is engineered to function optimally on a server outfitted with multiple CPUs and 128 GB of RAM. It efficiently handles the classification of thousands of genomes in minutes (∼10 000 genomes in 30–60 mins), utilizing 30 CPUs. For optimal results, it is recommended to use vClassifier with genomes that are of high quality; specifically, those with an estimated completeness of over 90%, as evaluated by CheckV version 1.0.1 (Nayfach et al. 2021). This ensures that the genomes being classified are robust and comprehensive, leading to more accurate and reliable taxonomic assignments. 2.4.4 Benchmarking and comparison We evaluated vClassifier’s proficiency in classifying viruses from diverse environments, including aquatic, engineered, host-associated, sediment, and soil contexts. To assess taxonomic accuracy, we benchmarked the tool against four widespread genera: Chivirus, Choctavirus, Hinehbovirus, and Tequatrovirus. Viral genomes were sourced from the IMG/VR v4 database (Camargo et al. 2023) to establish genus-specific benchmarking datasets. Each dataset comprised a set of positive controls (the target genus) and negative controls (non-target genomes) derived from relevant environments:Chivirus: 13 positive and 2754 negative controls (aquatic, engineered, and host-associated).Choctavirus: 17 positive and 1799 negative controls (aquatic and engineered).Hinehbovirus: 16 positive and 1896 negative controls (aquatic, engineered, and soil).Tequatrovirus: 56 positive and 2711 negative controls (aquatic, engineered, and soil). Chivirus: 13 positive and 2754 negative controls (aquatic, engineered, and host-associated). Choctavirus: 17 positive and 1799 negative controls (aquatic and engineered). Hinehbovirus: 16 positive and 1896 negative controls (aquatic, engineered, and soil). Tequatrovirus: 56 positive and 2711 negative controls (aquatic, engineered, and soil). We further assessed the alignment between vClassifier and the IMG/VR classification system. Overall consistency at the genus and species ranks was calculated as the ratio of sequences with taxonomic assignments matching the IMG/VR framework to the total number of assigned sequences. To statistically validate these results at the genus rank, we compared vClassifier’s performance against vConTACT3 (run with default parameters), allowing for a direct assessment of our method’s effectiveness relative to established bioinformatic tools. Additionally, utilizing a dataset of 66 231 viral scaffolds sourced from the IMG/VR database (Camargo et al. 2023), we determined the taxonomic assignment rate for prokaryotic viruses across 36 families and 55 subfamilies based on the outcomes generated by vClassifier. The final output file generated by vClassifier included taxonomic information for each query viral scaffold at the species, genus, subfamily, family, order, class, phylum, kingdom, and realm levels. To statistically compare the classifications at the subfamily and genus levels, we selected vConTACT3 and PhaGCN2 version 2.3.0 (Jiang et al. 2023) and employed them with their default parameters, enabling a comprehensive analysis of vClassifier’s effectiveness relative to established tools. Furthermore, we presented a use case to evaluate the performance of vClassifier in categorizing viruses extracted from metagenomes, highlighting its broad applicability. We sourced raw Illumina reads from various metagenomes (Table 3, available as supplementary data at Bioinformatics Advances online), including three associated with biofilms (Zhang et al. 2019, Zhou et al. 2023), three linked to sponges (Zhou et al. 2019, 2021), 95 human gut metagenomes (Ning et al. 2023), two chimney metagenomes from deep-sea hydrothermal vents, 3 freshwater metagenomes (Okazaki et al. 2019), and 13 human skin metagenomes (Nguyen et al. 2024). These reads were obtained from NCBI, with the following BioProject accessions: PRJNA983852 for biofilms, PRJNA552185 for sponges, and PRJEB15371 for humans, PRJNA445001 (JGI Genome Database 2018b) and PRJNA445000 (JGI Genome Database 2018a) for vent chimneys, PRJDB7309 for freshwater, and PRJNA763232 for human skins. The retrieved reads underwent trimming using Trimmomatic version 0.36 (Bolger et al. 2014) with customized parameters (ILLUMINACLIP: TruSeq3-PE.fa: 2:30:10, LEADING: 3, TRAILING: 3, SLIDINGWINDOW: 4:15, MINLEN: 40). The trimmed reads from the sponge and human microbiomes were assembled employing MEGAHIT version 1.2.9 (Li et al. 2015) with specific modifications (with k-mer sizes ranging from 51 to 91 and a 10-mer step size). To identify viral sequences in the metagenomes and generate taxonomic classifications at the family level, we utilized geNomad version 1.7.411 (Camargo et al. 2024). The viral genomes identified across 36 reference families were subsequently input into vClassifier, which operated on 30 CPUs during the analysis, for classification at lower taxonomic ranks. Simultaneously, the same viral genomes were examined using vConTACT3, also utilizing 30 CPUs, and PhaGCN2, which was run with default parameters, for comparative analysis purposes. 3 Results 3.1 Numerous single-copy marker genes are present in viral taxa at low ranks Based on our search against VOGdb, we identified a total of 3009 protein families serving as single-copy markers prevalent across the selected 203 taxa spanning five viral realms (as detailed in Table 1, available as supplementary data at Bioinformatics Advances online and illustrated in Fig. 2A). Interestingly, aggregating the single-copy genes from each taxon across the five realms revealed that no markers were universally shared among all realms. However, one marker was shared between the realms Varidnaviria and Duplodnaviria, and four markers were shared between Adnaviria and Duplodnaviria (shown in Fig. 2B). This finding underscores the absence of universal single-copy markers, akin to the prokaryotic 16S ribosomal RNA genes, across viral realms (Roux et al. 2021). Nonetheless, identifying these single-copy marker genes across a broad spectrum of viral taxa provides a valuable resource for taxonomic and evolutionary studies (Khot et al. 2020). This dataset highlights the diversity in marker gene distribution, reflecting the varied evolutionary histories of these viruses. Such heterogeneity in marker gene distribution has significant implications for the depth and reliability of taxonomic classifications, enhancing our understanding of viral phylogeny and taxonomy. Single-copy genes (SCGs) of 203 taxa across prokaryotic viruses. (A) Number of single-copy marker genes of taxa. (B) Distribution of single-copy marker genes in viruses at realm level. The numbers in brackets indicate the total genomes analyzed for SCG identification within each realm. (C) Number of identified SCGs of taxa at each rank. This graph represents the number of SCGs identified in each taxon, depicted by individual circles. The median number of SCGs per taxon is shown by the horizontal line within each box. The upper and lower boundaries of the box indicate the upper and lower quartiles, respectively. The whiskers extend to 1.5× the interquartile range, and data points beyond the whiskers are considered potential outliers. There is a discernible pattern of fewer identified marker genes in selected taxa at higher taxonomic ranks (Fig. 2C), as detailed in Table 1, available as supplementary data at Bioinformatics Advances online. All taxa possess no more than five genes at the order level or above, with only two at this level exhibiting more than five markers. In contrast, a significantly higher count of markers is observed in taxa at the family, subfamily, and genus ranks (Fig. 2C). Among taxa at these lower ranks, 87.6% contain at least 10 markers, and 81.4% contain at least 30 markers. This inverse relationship between taxonomic rank and the number of identified single-copy marker genes can likely be attributed to the broader genetic variation encompassed within higher ranks. These ranks aggregate a wide array of genetically diverse groups, which dilutes the presence of common markers. Conversely, lower ranks, which represent more closely related entities, demonstrate greater genetic specificity, thus allowing for the identification of a higher number of unique markers. The scarcity of markers at higher taxonomic ranks challenges the resolution of deeper phylogenetic relationships, potentially necessitating supplementary methods for more accurate classification. The abundance of markers at lower ranks not only facilitates detailed taxonomic resolution within these groups but also enhances our understanding of their evolutionary relationships. For example, in the case of the family Herelleviridae, phylogenetic topologies based on 14 single marker genes effectively reveal evolutionary relationships at the subfamily and genus levels (Barylski et al. 2020). This underscores the utility of these markers in enriching our understanding of viral taxonomy and phylogeny. 3.2 Our phylogenetic method demonstrated high congruence with ICTV taxonomy at subfamily and genus ranks and with NCBI taxonomy at species rank Based on the identified markers across the selected taxa, we constructed 203 phylogenetic trees, one for each of the 203 taxa analyzed. Despite this extensive effort, only the tree topologies for taxa at the family and subfamily levels, encompassing 36 families and 55 subfamilies, demonstrated high consistency with the ICTV classification system. This notable alignment underscores the robustness of the marker selection and tree construction methodologies at these taxonomic ranks. For a visual representation of this consistency, refer to the examples of phylogenetic tree topologies presented in Fig. 3, and for further details, refer to Table 4, available as supplementary data at Bioinformatics Advances online. This specificity highlights the effectiveness of our approach in reflecting recognized taxonomic relationships at higher hierarchical levels, providing a strong foundation for reliable taxonomic classification within the framework established by the ICTV. Representatives of phylogenetic tree topologies. (A) ssDNA representative: Inoviridae (100% consistent with ICTV taxonomy). (B) dsDNA representative: Herelleviridae (100% consistent with ICTV taxonomy). Solid blue circles indicate bootstrap values of 0.5 or higher. The tree is further detailed with color-coded rings; the outer ring represents viral subfamilies, and the inner ring denotes genera. Both subfamilies and genera within the Inoviridae and Herelleviridae families are identified as well-supported monophyletic groups, with bootstrap values of 0.75 or greater. This indicates a high level of congruence between the phylogenetic tree topologies and the ICTV classification, affirming the robustness of the phylogenetic structure in reflecting accurate taxonomic relationships. For the 36 families analyzed, we calculated the overall congruence with ICTV at the subfamily and genus levels by summing the number of monophyletic groups and dividing by the total number of phyletic groups across all taxa. Our results reveal substantial support for monophyly, with 85.9% of the taxa at the subfamily level (Fig. 4A) and 84.3% at the genus level (Fig. 4B) confirmed as monophyletic. This high level of monophyly is particularly pronounced among dsDNA viral families. For instance, families such as Ackermannviridae, Herelleviridae, and Mesyanzhinovviridae demonstrated exceptional congruence with ICTV taxonomy, with several achieving a perfect 100% alignment. Conversely, some families showed lower congruence values, such as Aliceevansviridae with only 33.3% congruence and Peduoviridae with 81.3%. The mixed-category Pleolipoviridae (dsDNA + ssDNA) exhibited a moderate congruence of 66.7%. In the realm of RNA viruses, congruence was notably variable, with Fiersviridae and Steitzviridae recording lower congruence values of 58.3% and 56.4%, respectively. These figures indicate classification challenges and discrepancies that could stem from the underlying phylogenetic methodologies or the existing ICTV taxonomic framework. The generally high congruence observed in many dsDNA viral families suggests that the employed phylogenetic method is reliable for this type of virus. However, the noted variability, especially among RNA viruses and specific dsDNA families, points to areas that could benefit from further refinement of the phylogenetic approach or necessitate updates to the ICTV taxonomic criteria to enhance accuracy and resolve existing discrepancies. Congruence of phylogenetic tree topologies with ICTV subfamily and genus classification. Percentage bar charts in green color showing the consistency of tree topologies with ICTV subfamily (A) and genus (B) classification based on identified SCGs at family rank. Bars represent subfamily/genera with two or more genome representatives within a family (A and B). Percentage bar charts in blue show the consistency of tree topologies with ICTV genus (C) classification based on identified SCGs at subfamily rank. Bars represent genera with two or more genome representatives within a subfamily (C). Bar sizes represent the percentage of monophyletic subfamilies within a family or monophyletic genera within a family/subfamily. For the 55 subfamilies analyzed, we determined the overall congruence with ICTV at the genus levels by summing the counts of all monophyletic groups and dividing by the sum of all total phyletic groups. Our results indicate strong monophyletic support at the genus level for 91.9% of the taxa (as shown in Fig. 4C). Notably, a perfect congruence rate of 100% was observed among the majority of dsDNA virus subfamilies, such as Asinivirinae, Azedovirinae, and Bastillvirinae. This high level of congruence underscores the consistency of the employed phylogenetic method with the ICTV taxonomy for these subfamilies, indicating a robust alignment. However, some subfamilies showed significantly lower congruence values, suggesting discrepancies in classification alignment with ICTV standards. Specifically, Gorcegavirinae exhibited a congruence of 66.7%, Hendrixvirinae had only 20%, Pclasvirinae had 50%, and Sepvirinae had 33.3%. These figures reflect a lesser degree of taxonomic alignment, which may prompt further investigations or necessitate revisions in the phylogenetic classification approach or the current ICTV taxonomy for these groups. The overall findings demonstrate a strong congruence for most dsDNA virus subfamilies, affirming the reliability and accuracy of the phylogenetic classification method utilized. However, the subfamilies with congruence values below 80% highlight areas of potential taxonomic uncertainty or ambiguity. These may stem from factors such as evolutionary divergence, lateral gene transfer, or other genetic complexities that are common in viral phylogenies, suggesting the need for continual refinement and adaptation of classification methodologies to accommodate evolving taxonomic insights. Overall, the phylogenetic analysis of SCGs shows outstanding performance in classifying taxa at the subfamily and genus levels. However, when extending this approach to species-level classification, the congruence with ICTV taxonomy ranges between 67% and 75%. This indicates that relying solely on the phylogeny of SCGs may not be sufficiently effective for viral clustering at the species level. To enhance species classification accuracy, we adopted a combined approach that integrates the phylogeny of SCGs with ANI. For benchmarking, we selected GenBank genomes that contain species information and are affiliated with the specified 36 families or 55 subfamilies. Our method involved first placing these query genomes into reference trees to generate genera based on monophyly. Subsequently, we assigned species to these query genomes based on their ANI compared to reference genomes within the same genus. Since no other programs currently assign taxonomy at the species level, we exclusively utilized this combination approach to benchmark against NCBI taxonomy. Our results for the taxonomic assignment of 3279 viral genomes (encompassing 1920 species) from the 36 families showed that 1772 out of 1920 species (92.3%) aligned consistently with NCBI taxonomy (Fig. 5A). Similarly, the assignment of 2703 viral genomes (1450 species) from the 55 subfamilies demonstrated that 1404 out of 1450 species (96.8%) were consistent with NCBI taxonomy (Fig. 5B). These findings underscore the effectiveness of the combined approach of SCG phylogeny and ANI in achieving high congruence at the species level, thereby significantly advancing the precision and reliability of taxonomic classifications for viral genomes. Comparative analysis of taxonomic consistency across different programs. Consistency of phylogenetic tree topologies with ICTV and NCBI classifications for 36 viral families (A) and 55 viral subfamilies (B) that have ICTV classifications at the subfamily (A) and genus (A, B) levels as well as NCBI classifications at the species level (B). 3.3 Our phylogenetic method was comparable to, or outperformed other tools For a comparative analysis of taxonomic assignment, 2564 genomes (representing 64 subfamilies within 36 families) were processed using vConTACT3, VipTree, VirClust, and our phylogenetic method (Fig. 5A). This allowed for a direct comparison of classification consistency across multiple taxonomic ranks using a consistent reference dataset. For subfamily classifications, our phylogenetic method showed the highest congruence (85.9%), followed by ViPTree (82.8%), VirClust (81.3%), and vConTACT3 (53.1%). For genus classification, ViPTree led with the highest congruence (88.2%), followed by our approach (84.3%), VirClust (68.5%), and vConTACT3 (48.9%). For species classification, our phylogenetic method was the only tool that can assign species and showed significant congruence (92.3%). Our phylogenetic method consistently exhibited the high congruence with the ICTV taxonomy at the subfamily and genus levels, and it was the only approach to demonstrate significant congruence at the species level. This suggests that our phylogenetic method may be more accurate in capturing taxonomic nuances. The lack of congruence at the species level for vConTACT3, VirClust, and ViPTree could indicate limitations in their algorithms or databases, which might not be as comprehensive or recently updated as those used by our phylogenetic method. In our comparative study of genus and species classification for the 55 subfamilies, we analyzed the standardized dataset of 1438 genomes across 172 genera using four different approaches: vConTACT3, VipTree, VirClust, and our in-house developed phylogenetic method (illustrated in Fig. 5B). At the genus level, VipTree exhibited the highest congruence at 97.1%. Our approach also performed impressively, achieving a congruence of 92.4%. VirClust followed with a commendable congruence of 86.7%, while vConTACT3 demonstrated the lowest congruence among the tools at 79.1%. For species-level classification, our phylogenetic method stood out as the sole approach offering this level of detailed classification, achieving a high congruence rate of 96.8%. The results at the genus level show a generally high level of congruence across all tools. While VipTree shows a slightly higher congruence than our phylogenetic method at the genus level, it is important to note that VipTree does not incorporate bootstrap support in its analyses. The absence of this statistical support could lead to incorrect assignments, giving a false impression of higher congruence (Lemoine et al. 2018). In contrast, the robust performance of our phylogenetic method is achieved at both the genus and species levels and highlights its standardized analytical capabilities. The notable 96.8% congruence at the species level underscores our method’s advanced utility for fine-scale taxonomic resolutions, which is essential for thorough viral characterization and research. The absence of species-level assignment capabilities in the other tools could point to limitations in their current versions or a deliberate focus on broader taxonomic levels. This distinction emphasizes the unique position of our phylogenetic method in providing detailed taxonomic insights, which can greatly enhance the precision of viral identification and classification studies. Accordingly, we integrated this phylogenetic framework into vClassifier, our in-house tool for viral classification and taxonomic assignment. 3.4 Taxonomic assignment of viruses in diverse environments with vClassifier When evaluated against the IMG/VR framework, vClassifier demonstrated high taxonomic accuracy and consistency across diverse environments. Across all tested genera (Chivirus, Choctavirus, Hinehbovirus, and Tequatrovirus), the tool achieved 100% accuracy (Table 5, available as supplementary data at Bioinformatics Advances online). Furthermore, taxonomic assignments showed high alignment with the IMG/VR framework, with overall consistency with IMG/VR taxonomy ranging from 95% to 100% at the genus rank and 87.8% to 100% at the species rank (Fig. 6A and Table 6, available as supplementary data at Bioinformatics Advances online). We further compared vClassifier against vConTACT3, a widely recognized virus taxonomic assignment tool. vClassifier demonstrated superior overall consistency with IMG/VR taxonomy at the genus level across all analyzed environments (Fig. 6A). vClassifier allows taxonomic assignment of viral genomes from various environments. (A) Taxonomic consistency of vClassifier, vConTACT3, and PhaGCN2 with IMG/VR classifications across diverse environmental habitats. (B) Taxonomic assignment performance of vClassifier, vConTACT3 and PhaGCN2 across various environmental samples. (C) Taxonomic assignment rates of metagenome-derived viruses sourced from human gut, marine biofilms, freshwater, vent chimneys, human skins, and marine sponges. Additionally, we analyzed the assignment rates across various environmental samples (illustrated in Fig. 6B; see details in Table 7, available as supplementary data at Bioinformatics Advances online). vClassifier demonstrated superior taxonomic assignment rates at the genus level for most environments. This was particularly notable in engineered and soil samples, where it achieved assignment rates of 68.5% and 63.1%, respectively. vClassifier excelled at the subfamily level in sediment samples with an impressive 83.1% assignment rate, significantly outperforming other methods in this environment. The lower assignment rates observed for subfamily classifications in other environments, such as soil, can be attributed to the limited availability of subfamilies in the ICTV database. For example, the subfamilies of more than 50% of the soil viruses included in the study have not been published by the ICTV. Consequently, this restriction prevents numerous sequences from being classified at the subfamily level. As a result, it is generally challenging to attain high assignment rates for subfamilies across diverse environments. Although vClassifier’s performance at the species level is not as robust as its performance at the genus and subfamily levels, it represents a significant advancement in the classification of species among environmental and metagenome-derived viruses. The observed limitations in species-level classifications are likely a consequence of our limited knowledge regarding viral diversity (Carroll et al. 2018, Hatano et al. 2021), as evidenced by the fact that public databases contain species information for only a small fraction of viruses. For instance, it is estimated that only 1% of zoonotic virus species have been identified so far (Carroll et al. 2018). The scarcity of available sequences with detailed species information significantly hampers efforts to classify viruses at the species level. To overcome this challenge, future research should focus on expanding the catalog of viral sequences with comprehensive species annotations. This would not only improve species-level identification but also enhance our overall understanding of viral diversity and its implications for public health, ecology, and evolutionary biology. In comparison to vConTACT3 and PhaGCN2, vClassifier consistently demonstrates higher assignment rates and shorter processing times for job execution (Table 7, available as supplementary data at Bioinformatics Advances online). For instance, when analyzing 21 252 genomes of soil viruses, vClassifier completed the task in 1 hour and 9 minutes using 30 CPUs, successfully classifying 63.12% of the viruses at the genus level. In contrast, vConTACT3 took 8 hours and 40 minutes with the same CPU allocation and managed to classify only 0.43% of the viruses at the genus level. Moreover, vClassifier demonstrates the ability to classify metagenome-derived viruses with relatively high assignment rates ranging from 6% to 45% (Fig. 6B), highlighting its broad applicability. In contrast, vConTACT3 and PhaGCN2 face challenges in classifying viral metagenomes at both the subfamily and genus levels, reporting assignment rates of less than 2.5%. This indicates that viral classification at lower taxonomic ranks is particularly difficult; even at the family level, assignment rates remain very low, as evidenced by the widely used tool geNomad, which has assignment rates of only 2% to 9% (Fig. 6B). This comparative analysis emphasizes the overall effectiveness of vClassifier in providing efficient taxonomic classifications across various environmental contexts. 4 Discussion Prokaryotic viruses display diverse evolutionary modes and rates, leading to notable differences in genome composition (Mavrich and Hatfull 2017). This variability impacts the conservation of markers across these viruses; notably, most single-gene markers are prevalent in fewer than 20% of viruses, which limits the phylogenetic signals they can provide (Low et al. 2019). This aligns with our observations that fewer single-copy marker genes are shared among taxa at higher ranks. While there are no universal single-copy genes across all viruses, specific sets of single-copy markers are frequently found within lower taxonomic ranks, such as families and subfamilies. Phylogenetic analyses based on concatenated markers at the family or subfamily level have consistently demonstrated strong monophyly of subfamilies and genera within our reference trees. These classifications correlate well with the standards set by ICTV, exhibiting over 80% consistency. However, the congruence between tree topologies and ICTV taxonomy drops below 80% when it comes to species classification. This discrepancy is likely due to the limited number of viral genomes that have been classified at the species level within the ICTV framework (Pollock et al. 2002, Zwickl and Hillis 2002), which may introduce biases in species classification assessments. To address this issue, we utilized an ANI approach for species classification, which has shown high consistency with the taxonomy provided by NCBI. The successful application of ANI, alongside phylogenetic analyses, suggests that combining these approaches offers a promising way to normalize rank assignments based solely on genomic data (Parks et al. 2018, Rinke et al. 2021). This combination leverages robust phylogenetic frameworks and precise genomic comparisons, providing a comprehensive method for accurate and consistent viral taxonomy. In this study, we developed vClassifier which leverages reference phylogenetic trees and ANI comparisons for determination of species-level taxonomy of viruses. vClassifier showed a high degree of consistency, exceeding 80%, with both ICTV and NCBI classifications. Given vClassifier’s exceptional performance in analyzing host-associated samples, it is particularly recommended for classifying viruses that inhabit host-associated environments, such as those found in humans. However, it is crucial to acknowledge certain limitations of vClassifier regarding its classification capabilities. The tool is not designed for direct taxonomic classification of viral genomes at the family level or higher. Nevertheless, this information can be indirectly inferred for a query genome annotated by vClassifier, based on the taxonomy of closely related reference genomes. Another limitation of vClassifier is that its capability for taxonomic classification only extends to the specific 36 families or 55 subfamilies recognized by the tool. While these constitute a large fraction of the families and sub-families on ICTV, vClassifier may not be able to generate useful outputs on novel viruses or viruses from other families outside of those represented in vClassifier’s database. To minimize these limitations, vClassifier will be updated regularly to ensure that its databases and taxonomy reflect the state of the art and to ensure users can fully exploit vClassifier’s capabilities to classify viruses at the subfamily, genus, and species ranks, thereby enhancing the overall scope and accuracy of viral taxonomy assignments. Regarding the input selection, we recommend the following pre-processing step to identify query sequences belonging to the 36 families and 55 subfamilies for input into vClassifier: Utilize external tools (e.g. geNomad) to obtain preliminary family and subfamily taxonomic classifications. To streamline and automate this procedure, custom scripts and troubleshooting guidance are available on our GitHub repository (https://github.com/AnantharamanLab/vClassifier). Supplementary Material