Limited consensus of marine viral diversity observed across techniques Authors: Marina Vila-Nistal, Francisco Martinez-Hernandez, Monica Lluesma-Gomez, Oscar Fornas, Simon Roux, Manuel Martinez-Garcia Published: 2025 ABSTRACT AbstractBackgroundViruses are fundamental to many aspects of life influencing ecosystem functions. The `number of lenses´ we use for exploring the viral diversity has expanded, yet each has limitations that constrain our view of the uncultured virosphere. It is fundamental to evaluate the different viromic approaches and sequencing methods on their ability to recover the extant viral diversity and microdiversity present in a sample. The differences in genome recovery between technologies have downstream impacts on subsequent estimates of viral diversity and function within a sample that can limit our comprehension of natural viral assemblages and their interactions with their microbial hosts.ResultsHere, using the same surface seawater sample, we compare short- and long-read viromics (i.e., Illumina, PacBio–HiFi and MinION sequencing) along with high-throughput single-virus genomics (sequencing of 700 uncultured single-viruses) to explore the consensus between approaches to uncover the extant viral diversity (sequencing effort ≈1.6 Tbp). Overall, ≈42,000 viral contigs (> 10 kb) were obtained, resulting in ≈12,500 and ≈23,400 viral OTUs at the genus and species levels, respectively, infecting mostly Flavobacteriaceae and Pelagibacteracea. At the viral family level, single-virus genomics (SVG) recovered viruses with a more distinct taxonomic profile compared to other methods. At lower taxonomic resolution, only < 1% of all species and genera, including some of the most abundant viruses, were captured by all methods; reaching a value of ≈2% when only viromics excluding SVG were considered. The highest pairwise diversity consensus was observed between PacBio-HiFi and Illumina, with approximately ≈11% of PacBio-HiFi species-level vOTUs also detected by Illumina. To understand how different methods resolve the co-occurring genomic microdiversity within species, we used one of the most abundant and microdiverse viruses -the uncultured pelagiphage vSAG 37-F6, proposed to be classified as Pelagimarinivirus ubique- originally discovered by single-virus genomics, as a reference. None of the methods alone were able to assemble the complete genome, which was only achieved by combining all datasets. Similarly, none of the viral clusters at the strain level were recovered by all methods.ConclusionsOur study suggests that the inherent bias of each method still represents a challenge for the recovery of marine viral diversity and potentially for other environmental viral samples. Nevertheless, regarding standard viromic techniques, PacBio HiFi in combination with Illumina seem to perform the best in absolute recovery of viral species and genera.Supplementary InformationThe online version contains supplementary material available at 10.1186/s40793-025-00802-6.Keywords: Viromics, Metagenomics, Single-virus genomics, Pacbio, Illumina, Nanopore, Viruses, Marine, Ocean Introduction Viromic studies have become a staple in environmental microbiology, allowing the study of viral communities directly from environmental samples without the need for culture or isolation [1–6]. Viral metagenomics (hereinafter viromics) also facilitated the study of diversity, abundance, and ecogenomic of relevant viruses in different environments [1], and highlighted the viral mechanisms, such as host metabolic reprogramming, horizontal gene transfer or lysis, that can affect the structure and evolution of microbial populations, such as [2, 5, 7]. Commonly, the process for conducting viromic studies involves the collection of samples, purification of viral fractions, DNA extraction, sequencing using Illumina short-read shotgun technology, and assembly to reconstruct the genomes present in the sample [2, 4, 8]. However, while viromics addresses biases present in culture-dependent techniques, they also have their own limitations. These include for instance, a reliance on databases for identifying newly sequenced viruses, which can be problematic if the viral genomes differ significantly from the sequences present in those databases, and the difficulty to capture rare species or certain viral groups [9, 10]. The genetic viral diversity in co-occurring populations presents a challenge, as viruses with hypervariable regions or populations with high co-occurring microdiversity can hinder the genome assembly of ecologically abundant viral populations, as was the case for the marine, uncultured pelagiphage vSAG 37-F6 discovered through single-virus genomics (SVG) [11–14]. SVG yields sequenced reference genomes from individual sorted virus particles without the need for cultivation or consensus assembly [11, 13]. This method is already capable of targeting both ds- and ssDNA viruses and requires very small sample volumes (≤ 1 ml). Certain steps in the SVG workflow remain challenging, such as the detection of viral particles with very small genomes (e.g., ssDNA and RNA viruses) and the inherent biases introduced by the whole-genome amplification (WGA), albeit a recent WGA method based on primary templated amplification [15] enables nearly complete genome recovery [11,15]. SVG was uniquely able to provide individual genomes for this abundant lytic virus infecting the cosmopolitan SAR11 clade [14], whose capsid protein resulted to be the most abundant viral protein in marine viral proteomes [13, 16]. Further investigations revealed a vast microdiversity of co-existing similar genotypes [17], comprising up to 1422 strains and 10 species in the same location [18], which explain why standard viromics was unable to recover this ubiquitous viral population [13]. The discovery of the uncultured pelagiphage vSAG 37-F6 and other novel viral genera by SVG that have been overlooked by previous viromic studies using short-read sequencing and consensus assembly [13] raises the question of whether different approaches and sequencing technologies, (i.e., long- and short-read sequencing platforms) uncover similar viral diversity, which is critical to build ecologically accurate models. The inherent genetic complexity of virus populations still poses technical difficulties for recovering complete virus genomes from natural assemblages. Long-read sequencing technologies, such Oxford Nanopore Technologies (ONT) [19, 20] and Pacific Biosciences high-accuracy long-read (PacBio-HiFi) sequencing [21], have improved on short-read consensus assembly by generating individual longer reads ranging from several kb to 100 kb or even longer [22–31]. In marine viral communities, different sequencing and metagenomic approaches have been used [22–31], albeit a comprehensive comparison of short- and long-read viromics (i.e., Illumina, PacBio–HiFi and nanopore sequencing) along with high-throughput single-virus genomics have not been performed. Although long reads reduce computational requirements for assembly and are remarkable in closing gaps and assembly of complex genomic regions [23, 27, 28, 32–36], they also have drawbacks compared to short-read Illumina sequencing, such as lower sequencing depth, and sometimes higher error and indel rates compared to Illumina [37–39]. This can be partially solved by combining reads from both long and short sequencing platforms to generate hybrid assemblies [24–26, 31, 40–43]; albeit this strategy increases computational requirements, making them less practical in certain scenarios [43]. Therefore, it is fundamental to evaluate the existing approaches and sequencing technologies on their ability to recover the extant viral diversity and microdiversity present in a sample. The differences in genome recovery between technologies have downstream impacts on subsequent estimates of viral diversity and function within a sample that can limit our comprehension of natural viral assemblages and their interactions with their microbial hosts. In this study, using the same seawater sample, we compare high throughput single-virus genomics and viromics using long- and short-read sequencing platforms (Illumina, PacBio HiFi and ONT MinION) to evaluate their efficacy in capturing the viral population diversity and microdiversity within co-occurring populations. This holistic and comprehensive approach assesses the level of consensus between different strategies for exploring marine viruses, and explores the severity of our myopia in marine viromics when looking at uncultured viral assemblages. Materials and methods Marine sample collection processing for viromics A 300 L surface seawater sample was collected from the North-Western Mediterranean Sea at the Blanes Bay Microbial Observatory (BBMO) (41º40’13.5’’N, 2º48’00.6’’ E, 2.7 miles offshore) on 10th of March, 2020. The seawater sample was filtered through a 0.45 μm and a 0.2 μm polyethersulfone (PES) membrane filter, and the viruses present in the samples were concentrated using tangential flow filtration with six 30 kDa PES Vivaflow 200 membranes (ref. VF20P0, Sartorius, Germany) connected in two parallel groups of three Vivaflow units in series. Next, the samples were filtered again through a 0.22 μm PES syringe membrane and stored at 4 °C until the next day, when the purified viral fractions were further concentrated to 250 µL using Amicon Ultra-15 mL centrifugal filter units (ref. UFC901024D, Merck Millipore, Massachusetts, USA). The concentrated samples were then treated with 0.45% volume of Turbo DNase I (ref. AM2238, Thermo Fisher Scientific, Massachusetts, USA) and a 9% volume of DNase Buffer at 37 °C for 1 h with agitation at 900 rpm to ensure the elimination of external DNA, followed by inactivation with 9% volume of inactivation buffer at room temperature for 5 min. To destabilize the viral capsids, the samples were incubated with a 1% volume of Proteinase K (ref. EO0491, Thermo Fisher Scientific) and 10% volume of TE 10X by incubating 1 h at 65 °C agitation 1200 g, followed by an inactivation through incubation of the samples on ice for 5 min. DNA extraction was performed using the QIAamp MinElute Virus Spin Kit (ref. 57704, Qiagen, Germany) according to the instructions of the manufacturer, using 50 µL of AVE buffer for the elution of the viral DNA. The quantity of extracted DNA was measured using a Qubit fluorometer (Thermo Fisher Scientific). The samples were frozen at -80 °C until use. Sample processing for single virus genomics From the above mentioned 300 L of seawater collected, approximately 20 ml was filtered through 0.2 μm PES filter, and then 1 mL of sample was stained with 5 µL of SYBR-Gold 1000X, incubated at room temperature in the dark for 30 min, and washed three times with 500 µL of 0.02 μm filtered 1X sterile TE buffer per wash in Amicon Ultra-0.5 mL centrifugal filter units (ref. UFC5030, Merck Millipore, Massachusetts, USA) as previously described [13]. The sample was then resuspended in 500 µL of 0.02 μm filtered 1X TE buffer and processed for fluorescence activated virus sorting (FAVS) as described by [13], with the only modification being the option of index sorting activated in the Influx sorter (BD, California, USA). This enabled the association of each virus with their position in the cytometers plot. The sample was sorted to obtain one viral particle in each well, with the exception of four wells per plate in which 10 and 20 viruses were sorted to be used later on as a control. The whole-genome amplification of single viruses was performed using multiple-displacement amplification (MDA) with EquiPhi29 DNA polymerase (ref. A39390; Thermo Fisher Scientific) as described in [13, 44, 45]. The wells containing 10 (2 wells) and 20 viruses (2 wells) were used as positive controls to confirm that the wells with higher DNA concentrations showed a faster fluorescence increase and amplification than the wells containing a single virus. Sequencing and read treatment The resulting viral DNA extracted from the viral fraction virome was sequenced using three different sequencing platforms: Illumina NovaSeq S4, PacBio Sequel IIe and MinION. For single virus-genomics (SVG), 700 positive single amplified viral genomes (vSAGs) from the MDA were selected at random for Illumina NovaSeq S4 short read sequencing. The library preparation method and sequencing was performed as per manufacturer´s recommendation at the Joint Genome Institute (California, USA), and read treatment protocols and metagenome assembly for each sample are provided in Supplementary Table 1. The virome was sequenced with Pacbio HiFi and Illumina method at the Joint Genome Institute (California, USA) following manufacture´s protocols and JGI standards (see Supplementary Table 1 for methods details on library sequencing, QC treatment, and assembly). Hybrid assemblies were performed using both short reads from the Illumina sequencing and long reads from the MinION and the PacBio HiFi, using SPAdes (v.3.15.5) [46] with the options: --meta -k 21,33,55,77,99,127 --only-assembler. Hybrid assemblies were performed using both short reads from the Illumina sequencing and long reads from the MinION and the PacBio HiFi, using SPAdes (v.3.15.5) [46] with the options: --meta -k 21,33,55,77,99,127 --only-assembler. All assembled contigs shorter than 500 bp were excluded from the analyses. The prediction of ORFs was done in Prodigal (v.2.6.3) [47] using the option -p meta. Bioinformatic analysis To compare the different sequencing methods, we took into consideration several factors, such as the number and size of contigs, the species and genera recovered, the number of viral contigs and taxonomic groups identified and their ability to capture the overall viral existing diversity in the sample and, in particular, the microdiversity of the uncultured virus vSAG 37-F6 as a reference model, since it is likely one of the most microdiverse populations in seawater [17, 18]. The identification of viral contigs, quality, completeness and taxonomic classification was performed using a combination of Virsorter2 (v.2.2.3; default parameters targeting ssDNAphage, NCLDV, RNA, ssDNA virus, andl avidaviridae) [48], CheckV (default parameters) [49], and geNomad (v.1.3.3; default parameters) [50]. Taxonomic classification obtained by geNomad was used to calculate a PCA plot based on the relative percentage of each taxonomy recovered by each sequencing method. The PCA analysis was conducted in R (v.4.0.2). To evaluate the range of species and genera recovered by each sequencing method, we clustered all contigs over 10 kb based on identity thresholds to group them in genus (70% identity and 50% coverage) and species (95% identity and 85% coverage), as previously recommended [51], using the scripts anicalc and uniclust, which are part of the CheckV package [49]. Next, we explored the overlapping diversity resulting from the different methodologies in order to explore which taxonomic groups were recovered by each sequencing technology. Finally, we predicted the host taxonomy of more than 90% of the viral clusters at the level of genus and species using the software iPHoP (v.1.3.3) [52], considering the longest contig present in each cluster and the host database of iPHoP program. Some viral clusters were tentatively assigned to some taxa, such as Prevotella, which are typical contaminants in extraction and sequencing reagents that were removed from the analysis [53]. In addition to assessing the relatedness of all obtained viral contigs over 10 kb from all methods, we used vConTACT2 (v.0.11.3) [54] to build a viral protein-sharing network as described in [13]. The network was visualized using Cytoscape [55] Later, we also evaluated the microdiversity recovered by each sequencing approach using the viral cluster containing the uncultured pelagiphage vSAG 37-F6. We used this virus as a reference model, since it is one of the most microdiverse marine viruses [18]. Then, all viral contigs [55] pertaining to the vSAG 37-F6 network (see Fig. 4B) were further clustered by genus and species as described before. In addition, they were also clustered at the strain level (>99% nucleotide identity). Proportion of species (blue) and genera (orange) recovered by each method, normalized by the number of base pairs (Mbp) assembled in contigs over 10 kb. (A) All species and genera detected by each method. (B) Species and genera with only one recovered member (singletons) detected by each method Finally, we aimed to recover the complete or nearly complete consensus genome of vSAG 37-F6 using a combination of long and short sequencing reads, and data from single-virus genomics. Illumina reads were subsampled to discard those with less than 1 kmer match to the reference vSAG 37-F6 genome using the software bbudk (v.39.03) [56] with the options - kfilter maskmiddle = f. Next, we performed multiple iterations of hybrid assembly with SPAdes, first combining Illumina short reads with the PacBio long reads and using the original vSAG 37-F6 genome recovered in [13] as the trusted contig following SPAdes manual using the command “--trusted-contigs”. Next, to achieve the completeness of vSAG 37-F6 genome, we perform a subsequent assembly iteration with the Minion nanopore data and SVG reads using as trusted contig the longest contig obtained by the previous iteration repeating the process in SPAdes using the option “--trusted-contigs”. Once the longest possible contig was recovered, we imported the resulting contig into Geneious program [57] and performed a mapping by extension focusing on both genome ends using the Geneious mapper option [57] with a custom sensitivity of 200 bp min length overlapping, and a minimum of 99% of nucleotide identity. This resulted in a 59 kb genome. A total of less than 10 nucleotide positions that remained undetermined in the complete genome were solved by mapping the original Illumina reads to the resulting genome using Bowtie2 [58] implemented in Geneious program to determine the most common nucleotide in those positions. To ensure that no artifacts were introduced during assembly, we verified the absence of autohits and regions with internal repetitions generated by wrongful assembly. For that, we split the genome into pieces of 1 kb in Geneious program, and then a database was generated to perform a BLASTn analysis of all vs. all in order to check for autohits and repetitions. Finally, ORF prediction and gene annotation was performed using GeneMark-hmm (v 3.25) and manual annotation using BLASTp (query coverage >70% and identity >40%), BLAST of conserved domains (e value < 1.0E-3) against the nr NCBI database. In addition, gene detection and annotation implemented in Virsorter2 DRAM-v2 program was used to complement gene annotation. Results and discussion Method performance: assembly results This study addresses a fundamental question in virology of whether different viral recovery approaches might bias our understanding of the structure and biodiversity of uncultured viral assemblages in nature. For that, we performed a comprehensive comparison of existing technologies to characterize the genomic diversity in the marine virosphere using single-virus genomics and three of the most common sequencing strategies in viromics using short- and long-read sequencing platforms. To compare the performance of the different approaches, we Illumina sequenced a total of 700 randomly selected uncultured viral single-amplified genomes (vSAGs) obtained by SVG (1–5 Gpb of sequencing effort per vSAG; 1.5 Tpb in total) from the viral fraction in a Mediterranean surface seawater sample. In addition, the same seawater sample was also studied through viromics using Illumina (64,3 Gpb), PacBio HiFi (12.11 Gpb), and MinION (2.3 Gpb sequencing platforms (Supplementary Table 2) along with a combination of hybrid assemblies with long- and short read datasets. Critically, the same viral DNA extract was used for all viromic strategies. In total, we obtained 903,553 assembled contigs (≥500 bp) from all datasets identified as viruses with geNomad program (see methods; Supplementary Table 2). Current methods for virus sequence identification struggle to reliably identify short viral sequences (< 10 kb; [59, 60] and therefore they have been discarded for downstream analysis. The method that showed the largest absolute number of pair bases assembled into viral contigs over 10 kb was PacBio HiFi (4.02E8 bp), followed by the hybrid assembly Illumina – PacBio HiFi (2.69E8 bp), and Illumina (1.47E8 bp) (Fig. 1A). The three remaining methods recovered an order of magnitude less. When normalizing this value by the total assembled data obtained from each method (Fig. 1B), the two methods with the highest recovery proportion were SVG and PacBio HiFi followed by Oxford MinION sequencing, which means that some methods are more prone to deliver very fragmented viral genomes (< 10 kb) in the assembled dataset. An analysis of the length distribution of the contigs recovered by each technology (Fig. 1C) showed that the hybrid assemblies and Illumina had significantly lower average contig length than the third-generation long-read technologies and SVG (p-value < 0.05; ANOVA and Levene´s test). Furthermore, all methods in general yielded fragmented and incomplete genomes since < 10% of all assembled viral contigs obtained from each method were classified as high-quality (>90% completeness) genomes by using CheckV program [61] (Supplementary Table 3). Most sequences over 10 kb were categorized as genome fragments (< 90% complete or no estimated genome size) with up to a maximum of ≈8% being classified as high quality (Pacbio HiFi and the hybrid assembly Illumina – Pacbio), depending on the sequencing technology (Supplementary Table 3). In other environments, relative proportion of high quality viral sequences varied from 0.2 to 54%, with the average of 3.4% in oral samples and 7.4% in stool samples [1]. Significant differences in GC content in the assembled viral contig datasets (≥10 kb) were not observed between methods (37.6–40.4%) within the range of marine viruses. Viral genome recovery by method. (A) Comparison of contigs recovered by each sequencing approach: Illumina, PacBio Hifi, MinION, Hybrid assembly Illumina – PacBio Hifi, Hybrid assembly Illumina – MinION and SVG. Number of base pairs assembled in contigs, depicting all the contigs > 500 pb (grey) and the subset of contigs over 10 kb (stripped pattern). (B) Comparison of contigs recovered by using normalized number of base pairs assembled in contigs over 10 kb by number of base pairs assembled in all contigs (> 500 bp) for each methodology. (C) Length comparison of the contigs recovered by each sequencing approach, considering contigs > 500 pb (top) and contigs under 10 kb (bottom). (D) PCA plot of the taxonomy of the recovered contigs by each method, considering only those within the Caudoviricetes clade (viral family taxa) We found that over 83% of all assembled contigs over 10 kbs from all methods (n = 42,197) were identified as bona fide viral genomes in all methods with PacBio HiFi showing the largest proportion (97%) and SVG the lowest one (83%) (Supplementary Fig. 1). These differences could be attributed to the potential sorting and recovery of free cellular DNA, gene transfer agents or DNA present in marine vesicles [11, 62–64]. Some vesicles might contain sufficient DNA to be visible following staining with the SYBR fluorescent DNA dyes, such as the one used in our study for detecting and sorting virus-like particles [65]. There is also the possibility that these non bona fide contigs could actually have a viral origin, but not be reliably predicted as viral, as it has been reported for viruses missing some viral hallmark genes, viral contigs similar to prophages or viruses with auxiliary metabolic genes [62, 66, 67]. Differences in the recovery of higher viral taxonomic ranks by method To explore the possible bias on the viral diversity recovery at higher viral taxonomic ranks (i.e., family, order, and class) present in each method, we classified the viral contigs from each method and studied the range of taxonomies. As the vast majority of viruses identified (>80%) pertained to the class Caudoviricetes, we focused our attention on lower taxonomic ranks within this clade mainly at the family level (Supplementary Fig. 2). A principal component analysis revealed that PacBio and MinION methods recovered viruses with a more similar taxonomic composition than SVG, which appeared more dispersed in the clustering. Similar clustering was observed for Illumina and hybrid strategies, while SVG recovered a more distinct profile compared to the rest of methods (Fig. 1D). For instance, SVG recovered a much larger relative proportion of viral families belonging to Nucleocitoviricota and a lower proportion of Caudoviricetes compared to other methods (Supplementary Table 4; Fig. 1D). Furthermore, SVG and PacBio HiFi were able to resolve viruses in different viral families (Supplementary Fig. S2), while Illumina and MinION are less effective at capturing viruses from different families likely due to assembly limitations and higher error rates, respectively [25, 26]. Detailed differences in the recovery of some of the most abundant families obtained in our study are illustrated in Supplementary Fig. 2. Most of the abundant viral families in the sample were represented and recovered by all methods (e.g., Autographiviridae), albeit differences in their relative representation were observed. For instance, the family Kyanoviridae is equally recovered (i.e. similar relative frequency) by all methods, while other families, such as Autographiviridae and Demerecviridae were more preferentially recovered by some of the methods (Supplementary Fig. 2). The former family includes common marine viruses of Vibrio, Pelagibacter, and Prochlorococcus. These differences in composition can be explained by the technical biases of each technique. To determine which viruses were ‘missed’ due to either lack of sequencing (or low coverage) or poor assembly performance, we conducted a read mapping analysis obtained from each sequencing dataset against all dereplicated viral contigs assembled by the different methods. Nanopore read mapping showed that most of the viral contigs present in the sample were either not captured or underrepresented while other viruses were clearly over sequenced, with a mean read identity value unexpectedly low at 80%. However, PacBio and Illumina read datasets mapped to most of the assembled viral contigs with high accuracy (>97.5% mean read nucleotide identity) and exhibited greater evenness than nanopore data (Fig. 2). This suggest that the assembled viral contigs from Nanopore contained much higher number of misassembles likely due to low read quality values. In this study, we have not focused on the comparative analysis of potential errors and indels introduced by the long-read sequencing platforms since it has been previously demonstrated using a mock bacteriophage community and natural viral fractions from fecal samples [25, 26]. These studies showed that previous Nanopore and PacBio sequencing performed poorly in comparison to Illumina in both genome recovery and error rates manifested as spurious SNPs and indels at frequencies from 12 to 157% higher than those found in Illumina assemblies [26]. Similar results have been found in mock bacterial communities [24]. Normalized mapping reads obtained from each one of the datasets against all concatenated dereplicated assembled viral contigs obtained from all sequencing datasets. Mapping 2 Gbp reads from each one of the sequencing methods (A) Nanopore data (MinION platform), (B) Pacbio data, and (C) Illumina data using dereplicated viral contigs (≥10 kb) obtained from all methods (X axis in all panels). Mapping of long read sequencing data was performed with Geneious bioinformatic program using the Geneious mapper with custom sensitivity employing a minimum read overlap of 500 bp and 80% of nucleotide identity for long reads sequencing datasets. Mean identity of mapped reads was estimated in Geneious program. Mapping of Illumina reads was performed with Bowttie2 using default parameters. Normalized mapping data obtained from each one of the methods is shown in panel D (Gbp of mapped data against all dereplicated viral contigs per Gbp of sequencing data) Tackling the species and genus diversity recovery and viral-host assignments To further study the recovered viral diversity, we grouped all obtained contigs over 10 kb (n = 42,197) from all approaches in viral OTUs (vOTUs) corresponding to genera (70% identity and 50% coverage) and species levels (95% identity and 85% coverage) following a previous reported consensus [61]. A total of 12,502 and 23,412 different vOTUs at the genera and species respectively, were recovered in our study from a single sample (Fig. 3). Long-read viromics using Nanopore sequencing combined with Illumina identified over 2,000 populations from 186 genera in the oligotrophic Sargasso Sea [31]. At the Station ALOHA (Pacific Ocean), using a very similar approach, a total of 1864 viral draft genomes were obtained [29]. Recovery and consensus agreement of viral species and genera by all methods. Overlap of clusters over 10 kb recovered by each sequencing method at the genus (A) and species (B) levels. The horizontal bar plot shows the number of clusters detected by each method, where at least one contig was recovered, and the vertical bar plot indicates number of clusters with contigs recovered by each method or combination of methods colored in the matrix diagram. Clusters recovered by all methods are highlighted in yellow, while those recovered by only one method are highlighted in red. Clusters recovered by some of the methods are indicated in blue. An example of how to read and interpret the image. The first bar in graphic of panel A (left side; red dot “hybrid Illumina + Pacbio”) shows that a total of 3233 unique genera were only obtained from the hybrid assembly dataset using Illumina + Pacbio HiFi and not by the other strategies. The second bar shows that a total of 1952 unique genera were oonly btained from the hybrid assembly of Illumina + Nanopore and not by the other strategies. The third bar shows that the same 1267 genera were found by both approaches, hybrid Illumina + Nanopore dataset and hybrid llumina + PacBio dataset. The fourth bar shows that the same 1034 genera hwere found in the Pacbio dataset, Illumina dataset, and also in the dataset obtained by the corresponding hybrid datasets. Table depicts total number of genera or species found by each strategy Our study revealed that only 0.48% and 0.081% of the total obtained viral species and genera respectively, were detected by all methods suggesting a very low consensus agreement (i.e., overlapping diversity region) between the different approaches (Fig. 3). When only viromic approaches were considered (excluding SVG from the analysis), the consensus agreement of genera and species recovered by all three methods—PacBio HiFi, Illumina, and MinION—regardless of whether their corresponding hybrid strategies were included or not, was < 2.19% (Fig. 3). PacBio HiFi recovered the highest number of total species (37.80%) (Fig. 3) obtained in our study followed by the hybrid assembly Illumina – PacBio HiFi (20.69%) in agreement with previous studies comparing PacBio HiFi with Illumina sequencing [36]. For genera, the hybrid assembly Illumina – PacBio HiFi detected 79.63% of total genera while the hybrid Illumina – MinION detected 11.80% of total species. The rest of methods recovered each one less than ≈20% of the total genera or species. The highest pairwise diversity consensus between Illumina, Pacbio-HiFi, and Nanopore -without considering hybrid strategies- was observed between PacBio-HiFi and Illumina, with approximately ≈11% of PacBio-HiFi species-level vOTUs also detected by Illumina. The rest of possible pairwise comparisons showed lower consensus values. Only < 6% of total vOTUs recovered by SVG were found by the other methods. To compare and assess the overall efficiency of each method to recover the viral diversity, we normalized the number of species and genera obtained from each method by the megabase pairs sequenced that assembled in contigs longer than 10 kb (Fig. 4). This allows us to assess how much of the viral diversity is assembled into interpretable long viral contigs, in contrast to datasets that deliver short viral DNA fragments and very fragmented genomic information. This is particularly important for ecosystem studies to resolve the characterization of existing viral biodiversity into meaningful viral genomes. The hybrid assembly Illumina – MinION method recovered the largest normalized values with ≈49 and ≈41 species and genera per each megabase pair (Mbp) assembled in contigs larger than 10 kb respectively, while the PacBio HiFi assembly recovered the fewest in both taxonomic categories (32.64 and 15.38 species and genera, respectively). The differences between methods were more pronounced at the genus level, where the gap between the highest and lowest-performing methods was 25.57, while at the species level, the difference was 16.74. There was a significant number of viral clusters at the level of species (45.04% of total clusters) and genus (24.09% of total clusters) comprised of a single viral contig (hereinafter named as singletons), suggesting that overall this study was able to sequence a promising wide spectrum of taxonomically different viruses including rare viruses (Fig. 3). When comparing the number of recovered singleton species and genera normalized per Mbp sequenced from each method that assembled in contigs longer than 10 kb, the SVG method had the highest rate of recovery for both species (28.54) and genera (21.61) followed by long-read sequencing technologies. We would like to emphasize that, although SVG alone captured less than 4% of the total species and genera present in the sample (see above) without normalizing the data by base pairs assembled, taking into consideration that we sequenced a total of 700 vSAGs, the efficiency in the viral recovery in large viral contigs representing different genera (n = 637) was significant. Of the 23,412 viral species obtained in this study, a total of 3305 (14%) viruses were confidently assigned to a host using in silico approaches, and most of them (92.6%) were classified as Bacteria, with only 7.4% as Archaea (Fig. 5), which is in agreement with the relative distribution of those microbes in surface seawater [68]. Of these 3,305 viral species assigned to hosts, a total of 358 viruses were not represented in databases, as they showed a genome-wide average amino acid identity lower than 95% to the closest viral reference [52]. This agrees with the microbial composition and abundance in surface seawater, which is dominated by bacteria [68]. In total, there were 281 different predicted host families divided into 494 different host genera. The two most abundant families were Flavobacteriaceae and Pelagibacteraceae, followed by Cyanobiaceae, similar to findings from previous studies using a Mediterranean seawater sample [22, 36]. At the genus level, the most predominant was Pelagibacter followed by Prochlorococcus and the marine flavobacterium Joostella [69]. Taxonomy of the most abundant viral hosts assigned to the recovered viral contigs over 10 kb. Relative abundance was determined based on the number of clusters assigned to each host and their size or number of contigs that contained. In the left, the viral host assignment at the family level is shown, while in the right the assignment at the genus level is depicted. All families belong to the Bacteria domain, except for the archaeal family ARS21. The most abundant 25 genera assigned to viruses within each one of the detected families are shown with connectors: Flavobacteriaceae (orange), Pelagibacteracea (red) and Cyanobiaceae (blue), while families represented by only one genus are shown in grey Resolving co-occurring genetic microdiversity: the power of all? As previously demonstrated with virome simulation data using different levels of co-occurring diversity and microdiversity of vSAG 37-F6 viral populations [13], genetic microdiversity hinders metagenomic assembly, making it difficult, in particular for viromics relying on short reads, to assemble complete genomes from highly microdiverse viral populations, such as vSAG 37-F6 [13]. Similar results have been found for microbial communities [70]. In order to explore how the different methods capture microdiverse co-occurring populations in the assembled datasets, we built a viral network with all the contigs over 10 kb (n = 42,197; Fig. 6) and analyzed in detail those clusters containing high microdiversity, in particular viral clusters related to the uncultured pelagiphages vSAG 37-F6 [13]. We used vSAG 37-F6 as a reference model since it is known to harbor highly microdiverse population, and it is ecologically relevant as one of the most abundant and globally ubiquitous marine viruses, making it an ideal representative for studying viral diversity and population structure across different environments [13, 17, 18, 71]. Recovery of viral microdiversity by different methods, considering only contigs over 10 kb. (A) Viral network with each contig coloured depending on the sequencing method that recovered it (Illumina – yellow, MinION – dark green, PacBio HiFi – dark blue, SVG – purple, hybrid Illumina – MinION – light green, hybrid Illumina – Pacbio Hifi – light blue, control genomes introduced by the vConTACT2 software – grey). (B) vSAG 37-F6 subnetwork, with the position of the 37-F6 genome in the network signaled by a red cross. Contigs depicted in panel B are contained within the apparent “dot” highlighted in part A. Nodes represent viruses and edges between nodes indicate a statistically significant weighted pairwise similarity between the protein profiles of each node with similarity scores ≥ 1. Proteins from all assembled viral contigs from all datasets were predicted and were compared through all-verses-all BLASTP with an E-value threshold of 10− 5 and 50 for bit score. Protein clusters (PCs) were then defined using Markov Clustering Algorithm (MCL), using default parameters and 2 for an inflation value. vContact 2.0 was then used to calculate a similar score between every pair of genomes based on the number of PCs shared between two sequences and all pairs using the hypergeometric similarity. MCL was applied to the similarity scores using a threshold of 1 and MCL inflation of 2 to generate viral clusters. See more details in reference [13] In our study, when examining the resulting viral protein-sharing network considering all viral contigs, we observed a large subnetwork composed of 38,007 contigs (91.5% of all contigs; Fig. 6) that included the uncultured pelagiphage vSAG 37-F6, with a total of 107 viruses having a direct connection with vSAG 37-F6 genome (70–95% nucleotide identity). This subnetwork depicted in Fig. 6B contained the largest microdiversity with more than 20,000 estimated different viral species, some of them without connections with other viruses and located in the periphery of the subnetwork, and others were densely connected with nearby viruses, such as the case of vSAG 37-F6 (Fig. 6). Indeed, this large subnetwork containing thousands of contigs is represented as a “dot” in Fig. 6A. To further explore this subnetwork, we clustered them by nucleotide identity at different taxonomic ranks up to the strain level using two sets of thresholds (99% identity and 85% query coverage and 99% identity and 95% query coverage) (Fig. 7). The majority of resultant clusters (≥79%) at the strain level were singletons containing only 1 viral contig obtained from only one of the methods. At the species level, 67% of viral contigs in this microdiverse subnetwork were singletons. All clusters at the strain and species level had less than 50 contigs and none of the clusters contained at least one sequence recovered by all methods. In other words, none of the viral clusters at the strain and species levels from this large subnetwork has been recovered by all used methods. Our results show that the methods used present certain biases as to what they recovered, and as discussed above, the different indel error rates of long-read sequencing platforms likely introduce a large spurious microdiversity variation precluding a minimum consensus in the viral strain clustering. It is important to remark that the different sequencing methods contributed a similar number of contigs to this viral subnetwork. Contribution of each methodology to the microdiverse vSAG 37-F6 subnetwork with only contigs over 10 kb, indicating the proportion of contigs recovered by each method that form the vSAG 37-F6 clusters at four different taxonomic levels: strain (99% identity and 85% query coverage or 95% identity and 85% query coverage), specie (95% identity and 85% query coverage) and genus (70% identity and 50% query coverage) To study how each sequencing method captured such microdiversity, we examined the degree of connection of the contigs recovered by each method. We found that PacBio HiFi was the method that recovered contigs with the highest number of direct neighbors (p value < 2E-16; ANOVA), while SVG recovered the greatest number of singletons or contigs with only 1 direct neighbor (p value < 1E-5; ANOVA). Furthermore, to explore the ecological relevance of the viruses forming this subnetwork, we analyzed the relationship between the degree of connection and the relative abundance of the contigs in the sample (measured using kilobase of reads per kilobase of the genome per Gigabase of reads from Illumina dataset), and found a lack of statistical connection, as some of these contigs with no or very little connections showed a high abundance in the sample. These contigs likely belong to distinct, unrelated viral lineages that bloom in the collected marine sample. Remarkably, in our study, only a combination of all datasets from all approaches resulted in the recovery, for the first time, of the complete or nearly complete consensus genome of the uncultured microdiverse pelagiphage vSAG 37-F6 (59,799 bp and GC content of 37%; 109 ORF; Supplementary Fig. 3). Recently, the International Committee on Taxonomy of Viruses has proposed to be classified as a new species and a viral family (Pelagimarinivirus ubique; family Mariniviridae). Mapping the Illumina reads against this genome resulted in a complete coverage, indicating that the genetic information from this virus was already recovered in the short-read dataset but the metagenomic consensus assembly approach was later unable to assemble the genome, delivering only partial genome fragments. Limitations of different methods and suggestions SVG clearly provides a unique window into viral diversity compared to viromics, yet its lower throughput remains a limiting factor. Some of the most promising methods to alleviate this limitation is the use of microfluidics in droplets [72] or semi-permeable capsules and gel capsules [73, 74]; which has already proven efficient for single viral particles [72]. Some of these technologies are already commercially available and offer simultaneous genomic libraries and sequencing for ≈10,000 single amplified genomes in a single microfluidic run at a very effective cost [74] using a minute amount of sample (a few milliliters), which is a clear advantage over long-read sequencing platforms. Meanwhile, a significant sampling and processing effort (240 L of seawater) was needed to have enough viral DNA for both Pacbio and Nanopore sequencing, which is not practical in many surveys like in an oceanic campaign if several samples have to be processed. During SVG, as with single-cell genomics [75–78], viral stained particles are sorted at random and therefore, the uncultured viruses sorted here, represent a random subset of, likely, abundant DNA viral species and genera [13, 45, 79]. Indeed, SVG workflow is already able to target both ds- and ssDNA viruses, while RNA viruses clearly remain a technical challenge [11, 13, 80]. The virome protocols used in this study exclude large viruses, RNA and likely ssDNA viruses as well [60]. Furthermore, chemical modifications of viral DNA [81, 82] can contribute to skewing the recovery of certain viral groups in viromic datasets. Another factor is the extraction of viral DNA from the capsids, which is critical to guarantee the overall success in downstream analyses. During SVG, a combination of KOH buffer and liquid nitrogen shocks proved to be efficient for lysing the viral capsids from marine samples [13, 18], though sub-optimal lysis of some virus groups might have led to them to be underrepresented. Similar sub-optimal lysis effects can apply to viral DNA extraction protocols in viromics that commonly rely on commercial kits [13] or a combination of enzymatic and chemical extraction [83]. Virome replicates were not performed in our study, such as in Tara Ocean expedition, due to the large volume and enormous effort required, making replication prohibitively expensive. It is important to remark that in seawater, replicate viromes using different methodologies were precise as evidenced by minimal variation between replicates which allows for quantitative cross-comparison [84, 85]. Our study suggests that the inherent bias of each method still represents a challenge for the recovery of marine viral diversity and although does not draw universal conclusions for other viral samples with potentially varying biases, we believe our findings from a representative, common marine sample are significant enough to raise awareness in the broader scientific community. The choice of method and sequencing platform largely will depend on the research questions and technical and funding constrains. Commonly, in viral ecology, researchers use the Illumina platform as the standard or combine it with one of the long read-sequencing platforms. Using an ultra-deep sequencing coverage for all techniques and the combination of all possible sequencing platforms will result in a better dataset and consensus agreement. However, since funds and manpower are veery limited, this is unfeasible and impractical. Not to mention the amount of sample and DNA required if all method are used. It is critical to remark that data suggest that all methods under-sample natural viral communities at least in the ocean surface and the sequencing coverage is likely an important factor contributing to this disagreement between approaches. According to our viromic data shown in Fig. 3, PacBio HiFi alone or in combination with Illumina seem to perform the best in absolute recovery of viral species and genera, albeit the consensus level between both techniques was 11%. Results also might vary depending on the type of sample chosen. While methods are rapidly advancing, in the meantime, the fact that natural viral communities harbor an extensive diversity unlikely to be fully captured by any existing approach needs to be considered when interpreting results of viromics analyses. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 2 Acknowledgements We thank the technicians at the FACS unit of PRBB (Barcelona, Spain) for supporting this research, Dr. Ben Temperton for his support and advise on sequencing viromes using nanopores, and Dra. Beatriz Camara for technical support. Author contributions M.V.N performed experiments, analysed data, prepared figures and wrote the main manuscript textF. M-H. performed experiments and reviewed the manuscriptM.L.G. performed experiments and reviewed the manuscriptO.F. performed experiments and reviewed the manuscriptS.R. analysed data and wrote the main manuscript textM.M.G. analysed data, prepared figures and wrote the main manuscript text. Funding This research has been funded by Generalitat Valenciana (ref. “Virhost” CIPROM/ 2021/006PROMETEO2022). This work was supported by Joint Genome Institute-Department of Energy of EEUU (CSP award no. 506015), the Generalitat Valenciana ACIF2020 grant and by the research grants funded by the Spanish Ministry of Science and Innovation and Agencia Estatal de Investigación (PID2021-125175OB-I00 and RTI2018-094248-B-I00). Data availability Viromic data (Illumina, Pacbio, and MinION metagenomes) are publicly available at the SRA database (Genbank) under the Bioproject number PRJNA1248896. Genbank accession number for the complete or near complete consensus genome of vSAG 37-F6 is PQ625738. All assembled viral contigs from SVGs are available as Supplementary dataset 1. In addition, data (raw reads, assembled data and annotation of viral single amplified genomes, Pacbio and Illumina metagenomes) are also available at the JGI GOLD (https://gold.jgi.doe.gov/studies) under the following Gold Study ID Gs0154162: Pacbio virome: GOLD Analysis Project ID Ga0496899; IMG Submission ID 283325. Illumina virome: GOLD Analysis Project ID Ga0497141; IMG Submission ID 283142. viral single amplified genomes: GOLD Analysis Project IDs from Ga0505711 to Ga0506408. Furthermore, another manner to access the data is throughout Zenodo repository under the following link: https://zenodo.org/uploads/14055532 (MinION data and assembled metagenomes (contigs ≥500 bp) obtained from all viromic datasets). Declarations Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Supplementary Material 2 Data Availability Statement Viromic data (Illumina, Pacbio, and MinION metagenomes) are publicly available at the SRA database (Genbank) under the Bioproject number PRJNA1248896. Genbank accession number for the complete or near complete consensus genome of vSAG 37-F6 is PQ625738. All assembled viral contigs from SVGs are available as Supplementary dataset 1. In addition, data (raw reads, assembled data and annotation of viral single amplified genomes, Pacbio and Illumina metagenomes) are also available at the JGI GOLD (https://gold.jgi.doe.gov/studies) under the following Gold Study ID Gs0154162: Pacbio virome: GOLD Analysis Project ID Ga0496899; IMG Submission ID 283325. Illumina virome: GOLD Analysis Project ID Ga0497141; IMG Submission ID 283142. viral single amplified genomes: GOLD Analysis Project IDs from Ga0505711 to Ga0506408. Furthermore, another manner to access the data is throughout Zenodo repository under the following link: https://zenodo.org/uploads/14055532 (MinION data and assembled metagenomes (contigs ≥500 bp) obtained from all viromic datasets).