Gilthead seabream larvae were reared at the Aquaculture Research Station (EPPO) of the Portuguese Institute for the Sea and the Atmosphere (IPMA), hereafter termed “EPPO-IPMA,” using methods that were similar to production-scale procedures (Ferreira, 2009). To rear the larvae until their mature stage, an experimentally controlled flow-through system was employed. Water temperature (19.2 ± 1.23°C), salinity (36 ± 1 psu) and dissolved oxygen (7.0 ± 1.05 mg/L) were kept stable throughout the experiment, and light intensity was set at approximately 800 lux within a photoperiod of 14 h light (starting at 9 am) and 10 h dark. Water renewal rate ranged from 20 to 45% per h depending on the type of prey used for feeding (see below), draining through an 80 μm to a 500 μm mesh (Castanho, 2014). The adjacent Ria Formosa lagoon, a highly productive ecosystem well interconnected with coastal seawater, was used as the source of water entering the system.

Gilthead seabream eggs were obtained from broodstock under captivity at EPPO-IPMA, and incubated at 18 ± 0.5°C in 200 L cylindro-conical fiberglass tanks at a density of 0.5 g.L ⁻¹ for 2 days. One day after hatching, fish larvae were distributed across four independent rearing tanks (200 L) at a density of 100 larvae.L ⁻¹. Only live feed was provided to the larvae during the entire rearing period, which comprised 35 days from egg hatching to complete organs' development. Larvae were fed rotifers (Brachionus spp.) and artemia (Artemia sp., nauplii and metanauplii stages) in accordance with larval developmental stage and mouth size (Figure 1, see details below). While rotifers were produced using a batch culture system established at EPPO-IPMA (Ferreira, 2009), artemia nauplii and metanauplii were obtained from Viet Nam Brine Shrimp (VNBS, Golden Lotus Trading LLC, USA) and from Salt Lake Aquafeed (Catvis BV, The Netherlands), respectively. After decapsulation (Ferreira, 2009), artemia cysts were incubated at a density of 4 cysts. mL⁻¹ at 27°C and 27 psu under strong aeration. While artemia nauplii were harvested at hatching to be directly used as food, artemia metanauplii were harvested at hatching to be nutritionally enriched prior to larvae provision. Both rotifers and artemia metanauplii were enriched with the commercial product RedPepper® (Bernaqua NV, Belgium) following the supplier's recommendations for each. Prior to feeding, rotifer and artemia pools were washed with flow-through seawater to minimize the input of allochtonous organic material into the tanks. Rotifers were provided to larvae from 4 DAH (when the latter opened their mouth) until 19 DAH. Artemia nauplii were given from 15 DAH until 19 DAH, and metanauplii from 20 DAH until the end of the rearing trial (34 DAH, Figure 1). Live preys were provided ad libitum, with a minimum concentration of c. five rotifers and one artemia nauplii. mL⁻¹ to compensate for the restricted larval mobility at the earliest developmental stages. The “green water technique” (addition of microalgae to the rearing tanks) was applied using a mixture of Nannochloropsis oculata (1.2 × 10⁵ cells. mL⁻¹) and Isochrysis aff. galbana (3.0 × 10⁴ cells. mL⁻¹) since mouth opening until the end of the trial, totalizing 1.5 × 10⁵ microalgal cells. mL⁻¹ added daily to the rearing tanks. Shortly, this technique helps maintaining live feed nutritional profiles. Castanho (2014) performed assessments of larval wellbeing during the course of the experiment, including larval survival, growth, morphology and development. Fish larvae survival (16.5 ± 2.72%) by the end of the experiment (35 DAH) was considered satisfactory (Castanho, 2014).

The sampling scheme used for bacterial community profiling is depicted in Figure 1. Five host- and particle-associated microhabitats were inspected for bacterial community diversity and composition using 454 pyrosequencing of 16S rRNA gene reads amplified from “total community” DNA (TC-DNA) samples. The microhabitats were (1) rearing-water from 2 and 34 DAH, (2) gilthead seabream larvae from 2 and 34 DAH, (3) rotifers, (4) artemia nauplii and (5) artemia metanauplii, totalizing seven hereafter called “sample categories” to include the temporal analysis (2 DAH vs. 34 DAH) of water and larvae microhabitats. Either three or four independent replicate samples were used in the characterization of each sample category (Figure 1).

Rearing-water samples consisted of 2 L volumes taken separately from each of the four experimental tanks at 2 and 34 DAH using disinfected (70% ethanol) plastic beakers, thereby comprising four independent replicates from each sampling point. Rearing-water samples were first passed through a disinfected 150 μM nylon mesh, and then filtered through 0.22 μM pore-sized nitrocellulose filters (Millipore, Billerica, MA, USA) with the aid of a vacuum pump. Filters were stored at −80°C until TC-DNA extraction. Active gilthead seabream larvae were also harvested at 2 and 34 DAH. Each sample consisted of about 0.2 g (wet weight) larval pools, corresponding to approximately 50 larvae from 2 DAH and 6 larvae from 34 DAH, respectively, taken from each of the four experimental rearing tanks using a disinfected 150 μM nylon mesh. Larval pools were gently rinsed three times with sterile artificial seawater to remove microbial cells not firmly attached to the larvae. Larval samples (independent replicates from 2 and 34 DAH, Figure 1) were then transferred to sterile, 2 mL polypropylene tubes and stored at −80°C until TC-DNA extraction. To determine the structure of bacterial communities added each day to the rearing tanks through the provision of feed organisms, three independent replicates of the live feed were sampled (prior to their introduction to the tanks), at random days during the rearing trial (depiction of live feed collection dates was simplified in Figure 1 for the sake of clarity). Live feed replicate samples consisted of 0.2 g pools of each rotifers, artemia nauplii and metanauplii placed into sterile, 2 mL polypropylene tubes after harvesting with disinfected beakers and gentle rinsing (3x) with sterile artificial seawater. Samples were kept at −80°C until TC-DNA extraction.

TC-DNA extraction from all sample categories was carried out using the Ultra Clean® Soil DNA isolation kit (MO BIO Laboratories Inc., Carlsbad, CA, USA). For rearing-water samples, filters obtained as above were first cut into smaller pieces with sterile scissors prior to TC-DNA extraction following the manufacturer's instructions. Larvae, rotifer and artemia (nauplii and metanauplii) samples were first thoroughly homogenized in 500 μL artificial sterile seawater using a 10 cm³ Potter-Elvehjem PTFE pestle and glass tube (Scherf-Präzision Europa GmbH, Meiningen, Germany) before being subjected to TC-DNA extraction. Here, an enzymatic lysis step was introduced after mechanical shearing—via bead beating—of the sample material to enable higher DNA yields. This consisted of two successive, 1 h incubation periods with 10 mg/mL lysozyme (Merck-Millipore, Billerica, MA, USA) at 37°C (Pangastuti et al., 2010) and 2 mg/mL proteinase K (Merck-Millipore) at 55°C (Sáenz De Rodrigáñez et al., 2009; Bakke et al., 2013).

For bar-coded 454 pyrosequencing, a nested PCR approach was employed to enable standardized 16S rRNA gene amplification from all TC-DNA samples, including 2 DAH fish larvae which presented the lowest DNA yields. In the first PCR, c. 10 ng template DNA were used for the amplification of near full-length bacterial 16S rRNA genes using 30 thermal cycles and the universal primer pair F27 (AGAGTTTGATCMTGGCTCAG)—R1492 (TACGGYTACCTTGTTACACTT) (Weisburg et al., 1991), as described elsewhere (Hardoim et al., 2012). The resulting amplicons (2 μL) were used as template in a second PCR with the Ribosomal Database Project (RDP) primer set (V4_titF-AYTGGGYDTAAAGNG and V4_titR-TACNVRRGTHTCTAATYC), which targets the V4 hypervariable region of bacterial 16S rRNA genes, generating amplicons of around 248 bp in length. PCR amplification took place using the PCR Master Mix Kit (QIAGEN GmbH, Hilden, Germany), containing 2.5 units Taq DNA polymerase, 1.5 mM MgCl₂, 1X QIAGEN PCR Buffer and 0.2 mM dNTPs (final concentrations), to which 0.2 μM of each primer were added. Each sample was tagged by different 8-mer barcodes attached to the reverse primer (Appendix S1, Supplementary Material). Thermal cycling involved a touchdown procedure to improve the retrieval of amplicons of the correct size (especially needed for 16S rRNA gene amplification from TC-DNA of 2 DAH larvae), with initial denaturation at 94°C for 4 min, followed by 10 cycles of 30 s at 94°C, 45 s at 65–55°C and 1 min at 68°C. Further 20 cycles followed as described above, except for the use of a constant annealing temperature of 55°C and a final elongation step at 68°C for 10 min. Two 25 μL amplifications were carried out per sample. The final 50 μL amplicon mixtures of each sample were delivered for pyrosequencing on a 454 Genome Sequencer GS FLX Titanium platform (Roche Diagnostics Ltd, West Sussex, UK) at Genoinseq (Biotechnology Innovation Center, Cantanhede, Portugal). For more details on sequencing procedures, see Appendix S1.

Processing and analysis of 454 pyrosequencing data followed the approach and scripts of Hardoim and Costa (2014) and Hardoim et al. (2014), with a few modifications. In summary, raw data were handled with AmpliconNoise (Quince et al., 2011) for the stringent retrieval of high-quality sequences and removal of homopolymers and chimeras. Sequences were subsequently trimmed using Galaxy (http://usegalaxy.org/) to obtain reads between 150 and 260 bp in length. Processing of quality-filtered sequences was performed with the Quantitative Insights Into Microbial Ecology (QIIME) software package (Caporaso et al., 2010). Operational taxonomic units (OTUs) were defined at ≥97% 16S rRNA gene sequence similarity using the UCLUST method (Edgar, 2010). Representative sequences of each OTU were picked using QIIME default parameters, and aligned employing Infernal (Nawrocki et al., 2009) using a STOCKHOLM file of pre-aligned sequences and secondary structures. Taxonomic assignment of representative sequences was carried out with the BLAST taxonomy assigner method using the latest Greengenes database (release 13_05) within the QIIME environment. After OTUs unclassifiable at the domain level or identified as mitochondria and chloroplasts were removed, a final OTU vs. samples table was generated and used for downstream analyses. These comprised (i) estimates of bacterial richness (Chao1) and diversity (Shannon's index) across microhabitats, (ii) phylum- and genus-level bacterial composition in individual and pooled samples per microhabitat, (iii) determination of OTUs specific to and shared by microhabitats using Venn diagrams, and (iv) multivariate analysis of OTU data. The latter was performed via UPGMA clustering and Principal Coordinate Analysis (PCoA) of OTU profiles using both the weighted and unweighted Unifrac metric. Analyses (i) to (iv) were undertaken using two datasets, with and without singleton OTUs, and results from the first dataset are shown unless otherwise stated. We used size-normalized sample libraries to perform analyses (i), (ii), and (iv), whereas the exploration of the full (non-normalized) quality-filtered dataset was employed in Venn Diagram constructions, and to create absolute abundance ranks of OTUs per sample categories and across the whole dataset. The taxonomy of the most differentiating OTUs (see below) was verified and refined, if needed, using the latest SILVA database (version 123.1 of March 29 2016, http://www.arb-silva.de/download/archive/release_123_1) and custom phylogenetic assessments within the software package ARB (Ludwig et al., 2004) as reported elsewhere (Costa et al., 2013; Keller-Costa et al., 2014).

Sequencing data were deposited in the European Nucleotide Archive (ENA) under the study accession number PRJEB9367, with sample accession numbers ERS726185-ERS726201 (host-associated samples) and ERS726303–ERS726310 (rearing-water samples). A sample vs. quality-filtered OTUs table with the corresponding taxonomic assignment of each OTU, including singleton OTUs with verified phylogenetic validity, is provided as Supplementary Material (Table S1, Supplementary Material).

Normality (Shapiro-Wilk) and equal variance tests were performed to inspect the distribution of the OTU richness and diversity measures, as well as of relative abundance values of the most dominant bacterial phyla and genera found across the seven sample categories, all estimated from 454-pyrosequencing data. One-Way Analysis of Variance (ANOVA) was performed on log-transformed alpha diversity measures (OTU richness, Chao1 and Shannon indices), all of which showing normal data distributions, to test whether mean values obtained for all sample groups were equal, followed by all pair-wise multiple comparison procedures using the Holm-Sidak method to determine significance between groups, in our case the seven sample categories. The Kruskal-Wallis test (One-Way ANOVA on Ranks) was employed to test whether the relative abundances of the most dominant bacterial phyla and genera changed significantly across the seven sample categories, given the absence of normal data distributions in most cases. A post-hoc Dunn's test was used to verify differences among sample categories in a pair-wise manner. Analyses were conducted using SigmaPlot 11 (Systat Software Inc., London, UK). Jackknifed beta-diversity procedures were run within the QIIME environment (jackknifed_beta_diversity.py) to test the statistical validity of sample groups generated by cluster and ordination (PCoA) analyses of OTU data, and thus whether bacterial community profiles generated by 454 pyrosequencing could discriminate between the seven sample categories defined in this study. The Similarity Percentage (SIMPER) test (Clarke, 1993) was run on PAST software (Hammer et al., 2001) version 3.10 to identify which bacterial OTUs contributed the most to the (Bray-Curtis) dissimilarities observed among microhabitats.

In total, 113,260 raw 16S rRNA gene V4 sequence reads amplified from 24 TC-DNA samples were obtained. Of these, 82,400 passed quality filtering with AmpliconNoise (Quince et al., 2011). Further removal of post-filtering reads representing OTUs non-classifiable at the Domain level, or identified as chloroplasts or mitochondria, resulted in a total of 80,353 reads that constituted the analytical dataset. Altogether, these high-quality sequences were assigned to 1,953 OTUs at a 97% gene similarity cut-off, including 1,068 singleton OTUs classifiable at least at the Domain level (Table S1). These comprised 1.33% of the total number of analyzed reads.

Differences in bacterial richness were statistically significant among the seven inspected sample categories (One Way ANOVA, P < 0.001, DF = 23, Figure 2A), with rearing-water samples displaying significantly higher values than host-associated samples (Figure 2A). Of note was a striking decrease in rearing-water bacterial richness during the trial, with averages (± standard deviation) of 286.5 ± 55.45 vs. 134.75 ± 34.05 OTUs detected per sample at 2 DAH and 34 DAH, respectively (Holm-Sidak t = 5.982, P < 0.001). In contrast, a subtle, non-significant (Holm-Sidak t = 2.371, P = 0.06) increase in richness was observed for fish larvae from 2 DAH (40 ± 10.14 OTUs) to 34 DAH (53.75 ± 7 OTUs) (Figure 2A). The live feed of fish larvae—rotifers, artemia nauplii and metanauplii—were similar in terms of bacterial richness (pairwise Holm-Sidak t < 1.830, P > 0.08, Figure 2A). Chao1 richness estimates retrieved for both 2 DAH and 34 DAH rearing-water were about 3-fold higher than the corresponding, observed bacterial richness values (Figure 2A). The difference between observed and estimated richness was not as pronounced, and often negligible, for the host-associated bacterial communities (Figure 2A). As expected due to their high richness values, bacterial communities from 2 and 34 DAH tank water were the most susceptible to the removal of singleton OTUs (Table 1, Figure 2B), averaging 193.5 ± 41.16 and 97.25 ± 25.90 OTUs per sample, respectively (Figure 2B). Reduction in richness values were not as pronounced in host-associated samples (Figure 2B) and, in comparative terms, both datasets with and without singletons revealed the same trends concerning shifts in bacterial richness across sample categories (Figures 2A,B). Shannon diversity indices—which consider not only the number of bacterial phylotypes (i.e., OTUs) but also their relative abundances in each sample—obtained for fish larvae (2 and 34 DAH) were, usually, significantly higher than those registered for the live feed (Figure 2C), suggesting greater equitability among community members in fish larvae than in the live feed. Further, at 34 DAH bacterial community diversity in the rearing-water was even lower than fish larvae diversity (Holm-Sidak t = 3.247, P = 0.005; Figure 2C) in spite of the much higher richness values registered for 34 DAH rearing-water in comparison with larvae (Figure 2A). Diversity estimates did not change significantly after removal of singleton OTUs (Figure 2C).

Three bacterial phyla, Proteobacteria, Bacteroidetes and Firmicutes (Table 1), accounted for more than 98.5% of all the retrieved quality-filtered sequence reads, with their relative abundances varying at a larger extent than would be expected by chance across the seven sample categories (Kruskal-Wallis, overall P < 0.03, DF = 6, Figure 3A. See Table S2, Supplementary Material, for details). Proteobacteria was clearly the dominant phylum in larvae- (2 and 34 DAH), rotifers-, and artemia metanauplii-associated bacterial communities, with average relative abundances of 83.81 ± 17.8%, 76.44 ± 2.25%, 95.46 ± 0.80%, and 91.07 ± 3.68%, respectively. Conversely, shared dominance between Bacteroidetes and Proteobacteria was observed in 2 DAH rearing-water (50.49 ± 8.25% and 43.38 ± 7.31%, respectively) and artemia nauplii (61.33 ± 1.65% and 34.57 ± 0.96%, respectively) samples (Figure 3A). A marked shift in community composition at the phylum level was registered for rearing-water during the trial as high Bacteroidetes dominance (87.39 ± 5.02%) was observed at 34 DAH. In fish larvae we detected a pronounced increase in Firmicutes abundance from 2 (3.82 ± 5.27%) to 34 DAH (19.55 ± 3.51%, Figure 3A). Finally, the until-recently candidate phylum GN02, now formally recognized as Gracilibacteria (Rinke et al., 2013), was apparently (see below) a characteristic phylum of early-stage fish larvae (2 DAH), occurring also–at moderate levels (2.70 ± 1.40%)—in artemia nauplii samples. In 2 DAH larvae, GN02 was represented mainly by one single OTU (OTU 2192, Table S1) and displayed high variability in relative abundance among samples (11.89 ± 18.83%, Figure 3A), showing therefore an inconsistent pattern of occurrence in this sample category. The proportions mentioned above remained largely unchanged when we explored relative abundances of phyla using the non-normalized dataset (Table 1).

The composition of the most dominant bacterial genera and unclassified families was markedly different among sample categories (Figure 3B), with relative abundances showing greater variations than could be predicted by chance (Kruskal-Wallis, overall P < 0.007, DF = 6. See Table S2 for details). A reduction in the number of genera was observed in rearing-water samples during the trial, coinciding with observations made at the OTU level (Figure 2). Indeed, most of the Bacteroidetes abundance in 34 DAH rearing-water could be attributed to a single OTU (OTU 268, Table S1) of the family Cryomorphaceae, which could not be classified at the genus level (Figure 3B). In contrast, 2 DAH rearing-water samples displayed a more balanced share between five genera in the Bacteroidetes and Alphaproteobacteria clades (Figure 3B), besides harboring several other, low abundance genera (Figure 3B, Table S1). Further, specific proteobacterial assemblages were identified in different larval developmental stages. Larvae sampled at 2 DAH were characterized by the genera Marinomonas, Acinetobacter and Pseudoalteromonas in the Gammaproteobacteria class, along with the genus Acidocella and an unclassified lineage (OTU 166, Table S1) in the Alphaproteobacteria class. Conversely, the genera Pseudomonas, Actinobacillus (Gammaproteobacteria), Paracoccus (Alphaproteobacteria) and Streptococcus (Firmicutes), in addition to a taxon (OTU 928, Table S1) of the Oxalobacteraceae family (Betaproteobacteria) tentatively affiliated with the genus Massilia (Table S3, Supplementary Material), prevailed in 34 DAH larvae (Figure 3B). Among live feed organisms, rotifers were dominated by a single OTU of the genus Loktanella (OTU 1801, Table S1), whereas artemia nauplii showed high abundances of unclassified Flavobacteriaceae (OTU 1902, Table S1), unclassified Rhodobacteraceae (OTUs 85 and 708, Table S1) and Paracoccus (Alphaproteobacteria, OTU 2374). The latter was the dominant genus in artemia metanauplii samples (Figure 3B). Usually, the most dominant genera / OTUs in the entire dataset (Table 2) displayed sharply variable abundance patterns, and therefore significantly contributed to differences in taxonomic composition among microhabitats (Table S3). Bacterial communities from early-stage larvae (2 DAH) displayed the highest level of variability in genus-level composition, especially regarding the relative abundances of Pseudoalteromonas, Marinomonas, unclassified Alphaproteobacteria and GN02 (Figure 3B).

At the approximate “species” level of taxonomic resolution (OTUs), Principal Coordinates Analysis (PCoA) and Jackknifed UPGMA clustering performed on weighted (Figures 4A,B) and unweighted (Figures 4C,D) Unifrac measures were used to inspect the continuous (PCoA) vs. discrete (UPGMA) grouping of samples according to their degrees of OTU-community (dis)similarity. Altogether, these analyses revealed discrete grouping of replicates from each sample category into separate clusters with statistical support (Figures 4B,D), corroborating trends revealed by genus-level inspection of taxonomic composition (Figure 3B). The only exception to discrete clustering per microhabitat was one replicate from 2 DAH larvae (2L4), which first grouped with one cluster encompassing all rotifer replicates in the weighted cluster analysis (Figure 4B). Although grouping patterns were the same in both weighted and unweighted analysis, the extent of dissimilarity between the sample categories was consistently larger in the latter comparison. Trends revealed by community ordination remained unchanged when analyses were undertaken in the absence of singleton OTUs (Figure S1, Supplementary Material).

To determine how many and which OTUs were common or specific to each sample category, we explored the full (non-normalized) quality-filtered dataset using Venn diagrams where replicate samples per category were pooled (Figure 5). Quite surprisingly, only five bacterial OTUs were common to all 2 and 34 DAH rearing-water and fish larvae sample categories (Figure 5A). Larvae at 2 DAH hosted 29 specific OTUs within these four sample categories and shared 39 bacterial OTUs with its primary surrounding environment, that is, 2 DAH rearing-water (Figure 5A, Table S4, Supplementary Material). Only 17 OTUs were common to 34 DAH larvae and rearing-water from 696 OTUs detected in both sample types (Figure 5A, Table S5, Supplementary Material). Although several bacterial phylotypes shared by rearing-water and fish larvae corresponded to low or only moderately abundant OTUs across the data (Tables S4, S5), four of the 10 most abundant OTUs in the dataset (Table 2) were common to fish larvae and rearing-water samples. While OTUs 1296 (Pseudoalteromonas) and 2374 (Paracoccus) appeared to be enriched in 2 and 34 DAH fish larvae, respectively, presenting only very low numbers in the corresponding rearing-water samples, OTUs 268 (Cryomorphaceae) and 708 (Rhodobacteraceae) occurred in higher abundance in rearing-water, and thus appeared to be de-selected in the fish host (Table 2, Tables S4, S5). Further, the divergence between bacterial community structures in rearing-water from 2 and 34 DAH (see e.g., Figure 4) was also well illustrated, as a much larger pool of bacterial OTUs specific to, than shared by, both sample categories could be depicted (Figure 5A). The dichotomy between 2 and 34 DAH larvae-associated bacterial communities was also evident, as only 24 OTUs were common to both sample categories, while 58 and 110 OTUs from 2 and 34 DAH larvae, respectively, remained exclusive to each (Figure 5A).

To more specifically address the relative contribution of live feed and rearing-water as bacterial vectors for mature fish larvae, a Venn diagram was constructed using OTUs detected in the live feed (pooling of rotifers and artemia nauplii and metanauplii samples), 34 DAH rearing-water, and 34 DAH fish larvae as discrete OTU pools (Figure 5B). We further built one diagram in which the live feed was divided into rotifers, artemia nauplii and artemia metanauplii as single categories to determine their own unique contribution to shaping bacterial communities in 34 DAH larvae (Figure 5C). We found that a minority portion (36) of the OTUs documented in the live feed was also present in fish larvae approaching the juvenile stage (Figure 5B, Table S6, Supplementary Material), with only 10 such OTUs being simultaneously present in fish larvae and each of the live feed used in larval rearing (Figure 5C, Table S6). Nevertheless, we were able to single out several examples of strong selection of bacterial phylotypes in fish larvae from within this fraction of shared OTUs. This was certainly the case for OTUs 928 (unclassified Oxalobacteraceae), 705 and 1953 (Pseudomonas), 84 (unclassified Enterobacteriaceae), 807 (Bradyrhizobium), 1387 (Sphingomonas), and 1970 (Acidocella). All these OTUs displayed enhanced numbers in 34 DAH fish larvae but occurred at very low abundances in the live feed (Table S6).

Because OTUs represented by one read are by definition sample-specific, removal of singletons from the data led to a substantial reduction in the number of OTUs exclusive to each sample category, whereas the number of OTUs common to all sample categories remained unchanged (Figures S2A–C). However, the effect caused by singleton exclusion did not erode the picture of a larger pool of OTUs specific to than shared by sample categories (Figures S2A–C).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.