We used sequence similarity as inferred from BLAST twice, first to retrieve eukaryotic sequences from the SSD and second to infer the taxonomic affiliation of these sequences. To retrieve the eukaryotic scaffolds from the SSD, we used a reference dataset for each gene chosen as an anchor. We used eight eukaryotic nuclear gene 'anchors', that is, well-conserved genes across the eukaryotic tree of life: 18S rRNA, 28S rRNA, and the genes encoding elongation factor 1a (EF1a), elongation factor 2 (EF2), the large subunit of RNA polymerase II (RPB1), actin, α-tubulin and β-tubulin. Since the genes we selected were well conserved among the eukaryotic lineage, we found little variation in the number of hits between the different species contained in each reference dataset. We even retrieved some prokaryotic scaffolds alongside the eukaryotic ones because of distant conservation with the protein coding genes. We are therefore confident we retrieved all eukaryotic scaffolds containing homologs to these genes using this approach. However, the taxonomic affiliation of these scaffolds as inferred from a local alignment approach has several drawbacks and has been found to be more error prone than phylogenetic based taxonomic affiliation 5. Usually the blast best hit (BBH) against GenBank is the only way to glean information about taxonomic affiliation from most environmental sequences. The reliability of the affiliation depends on the representation of each taxonomic group in GenBank, but there is a high bias towards sequences from Metazoans in this database, with a bias towards larger organisms in general. To exemplify this, we identified no SSD scaffolds found to contain RPB1 matching with a Chlorophyta RPB1, simply because there are no Chlorophyta RPB1 genes in the GenBank protein database yet. Therefore, the taxonomic affiliation is best described for genes sequenced in a large number of species in a broad range of taxa, such as the rRNA sequences. We also checked the taxonomic affiliation by phylogenetic tree reconstruction for the rRNA sequences (see Additional data files 1 and 2 for the 28S rRNA and 18S rRNA supertrees). The taxonomic affiliation of a SSD scaffold as inferred from its BBH was found to be consistent with the tree topology for all rRNA SSD scaffolds for which phylogenetic position could be resolved, that is, for less than half of the scaffolds (Additional data files 1 and 2). However, reducing information to phylogenetic inference is too restrictive for this kind of highly fragmented sequence data. First, because most of the sequences do not contain enough sites for their phylogenetic position to be fully resolved, and second, because highly variable regions have to be discarded from the global alignment, whereas they may contain most of the information (for example, the Internal Transcribed Spacer sequences between ribosomal genes).

Depending on which gene we searched for, we retrieved 4 (EF2) to 41 (28SrRNA) distinct eukaryotic sequences from the SSD (Table 1). This is less than the 69 18S rRNA sequences reported in 12 because we analyzed the assembled sequence data deposited in GenBank, which does not contain the sequences obtained from samples 5 to 8 with larger filter sizes 25 (up to 20 μm; Table S1 in 12). The taxonomic distribution of the sequences, as inferred from BLAST search against GenBank and phylogenetic analysis, is shown in Table 1. Despite the small number of sequences, the species diversity covered is impressive, since the five groups of the tree of eukaryotes 26 are represented for three of the eight nuclear genes (18S rRNA, RPB1, actin). The most abundant high blast score hits were found to sequences from the Dinophyceae (four out of the eight nuclear genes studied). This is consistent with previously reported marine picoeukaryotic diversity studies based on hundreds of 18S rRNA sequences from water filtered through larger pore sizes (5 and 3 μm filter pore size in 1927, respectively). The second most abundant group belongs to the Streptophyta-Chlorophyta (green plants) group, as might be expected for samples collected from surface water.

Since the picoeukaryotic world generally comprises cells smaller than 2 to 3 μm 1927, the available SSD enables a glimpse of the smaller part of the picoeukaryotic fraction (cell size between 0.22 and 0.8 μm). It is not surprising, therefore, that larger Prasinophytes, such as Bathycoccus, with a reported cell diameter around 2 μm, were not found in the data set.

We found two 18S scaffolds and one 28S scaffold matching almost perfectly with an Ostreococcus strain, the smallest photosynthetic picoeukaryotic known so far 2324. The two SSD 18S rRNA sequences do not overlap and these two sequences could thus belong to the same Ostreococcus, closely related to strain RCC143, consistent with previous analysis 28.

The presence of marine environmental arthropods (BBH is a marine Copepod) and Urochordate sequences (BBH is Ciona) was unexpected, because these organisms are usually much bigger than 0.8 μm. Marine environmental sequences from Copepods (and from Urochordate) have been previously reported in nanoplankton studies (cell size between 2 and 20 μm) but never in picoeukaryotes. Several hypotheses can be proposed to explain the presence of such sequences, one being the presence of gametes or of cell debris from larger organisms. However, even gametes are usually bigger than 0.8 μm and the DNA in cell debris is usually degraded. Another explanation could be the presence of soluble DNA fragments in the Seawater. Finally, a contamination of the filtered batch by non-filtered water cannot be totally ruled out.

Another ecologically relevant issue is the estimation of the relative abundance of phototrophic versus heterotrophic organisms among these picoeukaryotes. Assuming that all Viridiplantae and half of Dinophyceae are phototrophs 29, we nevertheless find 9.5 phototrophs out of 41, that is less than 24%. This is consistent with a higher observed diversity of heterotrophs than autotrophs in picoplankton, suggesting a complex role of heterotrophs in the microbial food web 15. The phylogenetic analysis of the two 18S rRNA Ostreococcus sequences found among the SSD showed that they belong to the deep clade (cladeB in Figure 1 from 30), even though the sea water was collected close to the surface. This observation has also been reported for Prochlorococcus in samples collected from a similar location 31. Since the four Sargasso samples making up the SSD were collected during winter deep-water mixing, this may be a possible explanation for the presence of some deep water features of the SSD, as revealed by a recent study of gene content along the water column 32. Thus, the occurrence of deep microbial strains in surface waters of the Sargasso Sea can probably be explained by frequent upwelling in this ocean area.

The SSD represents an unprecedented and yet unique sequencing effort, since it corresponds to the assembly of a total of 1.7 10⁶ reads from four sea water samples from the Sargasso Sea 12. In this pilot study, three other sea water samples have been sequenced in less depth and left unassembled. One of these additional samples, sample 6, used more conventional filter pore sizes to investigate the picoeukaryotic world, 0.8-3 μm, when compared to the 0.22-0.8 μm range used for three of the four SSD samples. Unfortunately, the sequencing effort of sample 6 was only 5% of the total sequencing effort realized to produce the SSD, or 29% of the smallest SSD sample. As a consequence, this sample contained far less eukaryotic material and enabled us to identify 6 (18S) to 11 (28S) additional eukaryotic paired reads, corresponding to Chromalveolates (Additional data files 1 and 2). We also screened seven additional marine metagenomes from the GOS project, corresponding to samples from seven different open ocean locations, for picoeukaryotic content. These metagenomes are part of the GOS survey 21, and sea water was filtered to collect 0.1-0.8 μm sized organisms. We found out that their picoeukaryotic content was almost negligible (from 0 to 4 reads matching a eukaryotic rRNA sequence). There are at least two reasons for this picoeukaryotic scarcity. First, the sequencing effort was much lower for these locations (4-15% of the SSD sequencing effort), thus reducing the overall diversity of the sample. Second, the collection filters used for these metagenomes were smaller (0.1 μm compared to 0.22 μm for the SSD), which also reduces the eukaryotic versus prokaryotic content. The collection filter size seems to have a major effect on picoeukaryotic sampling, since the one SSD sample collected with a 0.1 μm filter has lower picoeukaryotic content than the three other SSD samples collected with a 0.22 μm filter, despite larger sequencing depth (for example, Table S5 in 12). Therefore, this study focuses on eukaryotic sequence diversity from the largest metagenome from the Sargasso Sea (SSD).

We retrieved 41 distinct scaffolds containing 28S rRNA sequences and 558 distinct scaffolds containing 16S rRNA from the SSD. Assuming an equal distribution of the number of rRNA repeats in the genomes of Eukaryotes and Prokaryotes, that is, assuming that counting the number of rRNA repeats to estimate species richness is biased in the same way in both Eukaryotes and Prokaryotes, we can estimate the eukaryotic/prokaryotic species number ratio, ρ, equal to ρ = 41/558 = 7.3%. The rRNA gene copy number is known to be variable in both prokaryotes 33 and picoeukaryotes 34. Due to the greater occurrence of duplication in eukaryotic genomes, the number of rRNA copies reached in some eukaryotic species is several orders of magnitudes higher than in prokaryotic species. Thus, the above ratio is likely to be an overestimation. The average number of different eukaryotic SSD scaffolds over the 8 nuclear genes is 20, so it seems more realistic to assume ρ = 20/558 = 3.7%. However, this is an underestimate because eukaryotic genomes are, on average, larger than prokaryotic ones. Assuming an equal species abundance, the probability of sequencing orthologous regions of 100 bp in two genomes of size G1 = 10 Mb, that is, of the probability of identifying two distinct species, is one order of magnitude lower than the probability of sequencing two orthologous regions of 100 bp in two genomes of size G2 = 1 Mb (equal to the ratio G1/G2). Thus, this ratio must be corrected by the difference in genome size between prokaryotic and eukaryotic organisms. However, this ratio cannot be estimated precisely, but a minimum of five seems realistic (the Ostreococcus/Synechococcus genome size ratio is 12.6/2.4 = 5.25). Thus, assuming a minimum average difference in picoeukaryotic-prokaryotic genome size of 5, ρ = 3.7 × 5 = 18.5%, which is consistent with recent experimental estimates of relative picoeukaryotic/prokaryotic abundance in surface coastal water 14.

Because some of the anchor genes contained the same SSD scaffolds (for 18S and 28S rRNA, α- and β-tubulin) the total number of distinct eukaryotic scaffolds for all nuclear genes is 128. The nuclear eukaryotic SSD scaffolds have two striking differences to the prokaryotic and organellar scaffolds (Table 2). The first difference is that the nuclear scaffolds are, on average, 25% shorter than the prokaryotic and organellar scaffolds (Student test between SSD scaffolds containing16S rRNA and SSD scaffolds containing 18S rRNA, p value < 10^-7). The shorter length of the eukaryotic nuclear scaffolds can be explained in at least three ways. First it could solely reflect the genome size difference as explained above, since the probability of finding two overlapping sequences and, thus, larger assemblies is smaller for larger genomes. Second, it may also reflect the greater abundance of prokaryotic versus eukaryotic genomes. A greater number of prokaryotic genomes is the direct consequence of a greater number of prokaryotic cells, as estimated experimentally 14, whereas a greater number of organellar genomes could reflect a higher number of genome copies in the organelles compared to the nucleus. Our result suggests that organellar DNA may be present in more copies than nuclear DNA in picoeukaryotes, as observed in the green alga Chlamydomonas 35. Third, the shorter length of eukaryotic scaffolds could also be due to different efficiencies in DNA extraction and sequencing between circular and linear DNA, or between sequences of different base composition.

The second difference is that the AT content of the SSD eukaryotic scaffolds we retrieved is much lower than the average AT content of the SSD (51.4% versus 61.4%; Student test, p value < 10^-15; Figure 2). The few eukaryotic sequences we retrieved from the seven GOS open ocean locations also have a lower AT content (52.2%) than the AT content observed in these metagenomes 36. To test whether this observation is a consequence of a GC biased anchor dataset, we compared the base composition of our anchor dataset to the average GC content in the two complete picoeukaryotic genomes of Ostreococcus tauri and Cyanidioschyzon merolae. The base composition of the eight nuclear anchor genes is actually AT biased in O. tauri (n = 8, f_AT = 45.0% versus n = 7166, f_AT = 39.6, p value = 0.003) and not significantly different from the average AT content of the genes in C.merolae (n = 8, f_AT = 44.4, n = 6699, f_AT = 44.7, p value = 0.79). Foerstner and colleagues 37 argued that the environment shapes the nucleotide composition of genomes because the Sargasso Sea prokaryotic sequences have a higher AT content than sequences from other environments, though the causes responsible for this compositional bias are not clear yet.

We compared the AT composition of the SSD eukaryotic sequences with the AT composition of their GenBank BBH and found no trend in average base composition differences on the alignments (exact test on the difference of AT content between each pair of sequences, p value = 0.93, n = 128); restricting the comparison to non-marine BBH was not significant either (p value = 0.83, n = 30). We also compared the AT composition of 30 of the 128 eukaryotic SSD scaffolds having a blast hit against the soil metagenome (e-value < 10^-6) 3 and found no significant difference in base composition over the alignments (p value = 0.76, n = 30).

Shorter genome sizes and the higher cost of synthesis of G and C compared to A or T nucleotides have been invoked as possible explanations for base composition differences between genomes, because of their indirect influence on growth rate 38. Global environmental features (nutrient availability, organism density, ecosystem complexity) may induce different pressures on growth rates and, thus, on genomic base composition 37. This analysis suggests that base composition in picoeukaryotes is not subjected to the same selective or neutral forces as prokaryotic sequences in the Sargasso Sea.

The SSD sequence data was retrieved from GenBank (accession number AACY01000000, Locus CH004737 to CH236877). These sequence data are the database of scaffolds not associated with any particular organism. It was obtained from samples 1-4, prefiltered through 0.8 μm and collected on one 0.1 and three 0.22 μm filters (Table S1 in 12). The reads corresponding to this assembly, the reads obtained from sample 6, prefiltered through 3 μm and collected on 0.8 μm filters, and the reads corresponding to the seven other open ocean locations were downloaded from the CAMERA database 4546. The O. tauri gene content was retrieved from GenBank (accession numbers CR954201-CR954220).

To assess picoeukaryotic diversity, we used eight eukaryotic nuclear gene 'anchors', that is, well-conserved genes across the eukaryotic tree of life: 18S rRNA, 28S rRNA, and genes encoding EF1a, EF2, RPB1, actin, α-tubulin and β-tubulin. For each of the six nuclear protein coding genes, we retrieved the seven corresponding genes from the KOG database 47, corresponding to the genes of Arabidopsis thaliana, Caenorhabditis elegans, Drosophila melanogaster, Homo sapiens, Saccharomyces cerevisiae, Schizosaccharomyces pombe, Encephalitozoon cuniculi and the corresponding O. tauri gene. We then extended each reference dataset by searching GenBank for representatives of these genes in each of the supergroups of the eukaryotic tree of life 26. The total number of genes in each dataset was 17 (EF1a), 15 (EF2), 21 (RPB1), 22 (actin), 20 (α-tubulin) and 20 (β-tubulin).

We also used one chloroplast gene, that encoding the large subunit of ribulose carboxylase (rbcL), and one mitochondrial gene, that encoding the first subunit of cytochrome oxydase (cox1). For each gene, we retrieved 21 and 31 genes from GenBank, respectively, randomly sampling representatives in each of the five supergroups of the eukaryotic tree of life.

To assess prokaryotic diversity on the same dataset, we used 16S rRNA. The reference dataset for 16S rRNA was retrieved from the RDPII database 48 and contained 4,409 sequences. We randomly chose one sequence for each sequence sharing the same taxonomic affiliation (given by the first name of the organism, for example, Persephonella), which reduced the number of sequences to 906.

The reference datasets for 18S and 28S RNA were retrieved from GenBank using the ACNUC retrieval system 49 excluding sequences from metazoans. As for the 16S rRNA dataset we randomly chose one sequence when several organisms shared the same taxonomic affiliation. We thus obtained a reference dataset of 252 18S and 246 28S sequences.

To assess the diversity and abundance of picoeukaryotes in this dataset, we performed a BLAST search 50 of the ten eukaryotic 'anchor' genes against the SSD, blastn for RNA and tblastn for proteins. We retrieved all Sargasso Sea scaffolds matching these genes with E-values smaller than 10^-14 for blastn and 10^-7 for tblastn. We then retrieved these SSD scaffolds and performed a BLAST search against GenBank for taxonomic affiliation. We used blastn against GenBank for scaffolds containing one of the two rRNA genes, and blastx against GenBank's protein database for the scaffolds containing one of the eight 'anchor' protein genes. We deduced the taxonomic affiliation of the environmental sequence from the taxonomic affiliation of the BBH when the E-value of the BBH was smaller than 10^-18 (blastn) and 10^-10 (blastx). Otherwise, we considered it as unknown.

The SSD scaffolds matching a gene of the anchor 18S and 28S datasets, the corresponding anchor gene and the GenBank BBH, were aligned by MAFFT version 5 5152 and the alignment was checked by eye with Se-Al v2.0a11 53. Ambiguous regions were deleted from the alignment, for a final length of 3,045 bp for the 28S rRNA dataset (90 sequences in total) and 1,374 bp for the 18S rRNA dataset (61 sequences in total).

Most SSD scaffolds are of different sizes, together covering almost all 18S and/or 28S rRNA. These sequence length differences made it difficult to recontruct a phylogenetic tree directly from the whole matrix of aligned sequences. Thus, overlapping subsets of sequences were defined for the maximum possible number of species, given that the aligned sequences were long enough to reconstruct well-supported phylogenetic trees. The trees issued from these datasets will hereafter be named 'subtrees'. They were reconstructed by Bayesian analysis with MrBayes 3.1.2 54. The reconstruction used four chains of 10⁶ generations with the best evolutionary models chosen via hierarchical likelihood ratio test by MrModelTest 2.2 5556 (the MrModeltest 2.2 program is distributed by the author, Evolutionary Biology Centre, Uppsala University). The Burnin value was set to 20% of the sampled trees (1% of the number of generations) and only clades with at least 90% posterior probability support were kept as conservative estimates in the final consensus tree. Thirty-one subtrees (28S rRNA) and 23 subtrees (18S rRNA) were constructed.

All subtrees were combined in a supertree with the use of RadCon 57, using matrix representation with parsimony with the Baum 58 and Ragan 59 coding scheme 6061. The combined matrix was subjected to a parsimony analysis with the heuristic algorithm implemented in PAUP* 62, using 500 random addition replicates and the tree bisection-reconnection branch-swapping algorithm, holding a maximum of 1,000 trees for each replicate. The 498,000 (28S) and 423,000 (18S) most parsimonious trees obtained were combined in a majority-rule consensus. Supertrees computed from subtrees obtained via Bayesian analysis and maximum likelihood were not significantly different (p < 0.01, symmetric-difference test 63, computed with PAUP* 4.0), and only supertrees computed from Bayesian inferred subtrees are presented. To assign a SSD scaffold to a taxonomic group, the branch support of this sequence within a taxonomic group had to be over 80%; otherwise, we assumed that the taxonomic affiliation of the SSD scaffold was unresolved by the supertree topology.

The 18S rRNA sequences from several Ostreococcus strains 30 and the corresponding first blast hit of the O.tauri 18S on the SSD were aligned manually. This alignment was used to build a phylogenetic tree by Bayesian analysis with MrBayes 3.1.1 54. The reconstruction used four chains of 10⁶ generations with the best evolutionary models chosen via hierarchical likelihood ratio test by MrModelTest 2.2 56. The best model was Hasegawa-Kishino-Yano (HKY+Γ) for 18S rRNA. Several analyses were independently run from random trees and to assess convergence. The tree was rooted using related prasinophyte taxa: Bathycoccus.

For each SSD scaffold, we computed the length; the number of gaps, the distance between gaps and the base composition using home made computer programs (C language). Statistical analysis was performed with R software 64.

To compare the AT frequency between the SSD scaffolds and the AT frequency of the corresponding BBH, we derived the variance, V, of the average of the difference in AT frequency between the two sequences, M. Under the null hypothesis of no difference in AT composition, M follows a normal distribution of mean 0 and variance V:

V = 1 n 2 ∑ i = 1 n f i ( 1 − f i ) + f ' i ( 1 − f ' i ) k i MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2Caerbhv2BYDwAHbqedmvETj2BSbqee0evGueE0jxyaibaiKI8=vI8GiVeY=Pipec8Eeeu0xXdbba9frFj0xb9Lqpepeea0xd9q8qiYRWxGi6xij=hbbc9s8aq0=yqpe0xbbG8A8frFve9Fve9Fj0dmeaabaqaciaacaGaaeqabaqabeGadaaakeaacaWGwbGaeyypa0tcfa4aaSaaaeaacaaIXaaabaGaamOBamaaCaaabeqaaiaaikdaaaaaaOWaaabCaKqbagaadaWcaaqaaiaadAgadaWgaaqaaiaadMgaaeqaaiaacIcacaaIXaGaeyOeI0IaamOzamaaBaaabaGaamyAaaqabaGaaiykaiabgUcaRiaadAgacaGGNaWaaSbaaeaacaWGPbaabeaacaGGOaGaaGymaiabgkHiTiaadAgacaGGNaWaaSbaaeaacaWGPbaabeaacaGGPaaabaGaam4AamaaBaaabaGaamyAaaqabaaaaaWcbaGaamyAaiabg2da9iaaigdaaeaacaWGUbaaniabggHiLdaaaa@4DFB@

with n the number of SSD scaffolds used, k_i the length of the alignment over which the AT frequencies of the SSD scaffold, f_i, and the corresponding BBH, f'_i, was computed.