Samples were collected from two salmon farm locations in Scotland, namely DUN located close to Oban and LIS located in Loch Linnhe. Sampling was conducted during the mid-production and peak-production period, respectively. Sediment was collected at three sites along a transect from an outer cage edge (CE) to a reference site (REF) in the direction of the prevailing current flow, located ca. 800 m (DUN) and 525 m (LIS) distant from the CE. An intermediate impact zone (Allowable Zone of Effect, AZE) was located at ca. 100 m (DUN) and 109 m (LIS) distance from the cage edge. This sampling design followed a decreasing organic enrichment gradient from the CE towards the REF sites, resulting from the deposits of feed and fish faeces on the sea floor (Keeley et al. 2012; Bannister et al. 2014; Frühe et al. 2021b). For a graphical representation of the sampling design, we refer to Fig. 1. At each site, two biological replicates were taken with a van Veen grab (0.1 m² area, DUN; 0.045 m² area, LIS). From each replicate, we sampled approximately 20–25 g of surface sediment (upper few millimetres) into a sterile 50 ml plastic tube using disposable sterile plastic spatulas. Immediately following collection, we homogenised the samples in the 50-ml collection tube with a spatula. Samples were frozen within a few hours of collection and shipped (< 48 h) from Scotland to our laboratory in Kaiserslautern (Germany) on dry-ice. Samples from farm LIS were taken during routine compliance monitoring and are accompanied by macroinvertebrate inventories which were used to calculate the AMBI Index and ecological quality (Borja et al. 2000; Muxika et al. 2005). The macrofauna was obtained from van Veen grabs after subsampling for DNA analyses. Therefore, the remaining sediment was washed through a 1 mm sieve and the residue was fixed in 4% borax-buffered formaldehyde prior to macrobenthic sorting and counting. The sieve-retained fauna was identified to species level under the National Marine Biological Quality Control Scheme (NMBAQCS) by APEM Ltd., Hertfordshire.

Figure 1.

Schematic representation of the sampling design. Benthic sampling of the two salmon aquaculture installations (DUN and LIS) occurred along a transect from the outermost cage edge of each farm, towards a reference site (REF). Distances of the Allowable Zone of Effect (AZE) and the REF sampling sites are determined according to compliance monitoring regulations separately for each farm prior to stocking. These distances depend on different parameters, such as strength of current and seabed topology. Distances of reference sites are chosen to be outside the influence of seabed depositions originating from aquaculture activities. The inset image (lower left corner) shows the RV Seol Mara of SAMS, Oban, while sampling the CE of DUN farm.

Following our previously described protocol (Frühe et al. 2021a), environmental DNA was obtained from sediment samples using the PowerSoil DNA Kit (Qiagen, Hilden, Germany) according to the manufacturer’s manual. The concentration and quality of extracted DNA were measured with a NanoDrop 2000 spectrophotometer (Peqlab, Erlangen, Germany).

As DNA metabarcodes for Illumina MiSeq sequencing, we obtained the ca. 450 bp long hypervariable V3-V4 region of the bacterial 16S rRNA gene using the Bact_341F (5’-CCTACGGGNGGCWGCAG-3’) and the Bact_805R (5’-GACTACHVGGGTATCTAATCC-3’) primer pair (Herlemann et al. 2011). PCR conditions employed an initial activation step of NEB’s Phusion High-Fidelity DNA polymerase at 98 °C for 30 s, followed by 27 identical three-step cycles consisting of 98 °C for 10 s, 62 °C for 30 s and 72 °C for 30 s; then a final 5-min extension at 72 °C.

For Nanopore sequencing, we obtained the full length 16S rRNA gene using primer pair Bac27f (5’-AGAGTTTGATCMTGGCTCAG-3’) (Frank et al. 2008) and U1492R (5’- ACCTTGTTACGRCTT-3’) (Dawson and Pace 2002), both of which are more universal modifications of the bacterial 27f/1492R primer pair developed by Lane (1991). For amplification, we used the LongAmp Taq polymerase (New England Biolabs, Ipswich, MA, USA). The PCR protocol consisted of an initial activation step at 95 °C for 60 s, followed by 30 identical three-step cycles consisting of 95 °C for 20 s, 55 °C for 30 s and 65 °C for 2 min; then a final 5-min extension at 65 °C.

In both cases, V3-V4 region and full-length 16S rRNA gene, we obtained three technical PCR replicates for each of the two grab replicates per sampling site. Resulting PCR products were then purified using Qiagen’s MinElute PCR purification kit. All six PCR replicates of one sampling site (2 sediment grabs × 3 technical PCR replicates per grab) were then pooled in equal amounts of DNA prior to library constructions.

Illumina: From the resulting PCR products, sequencing libraries were constructed using the NEB Next® Ultra™ DNA Library Prep Kit for Illumina (NEB, USA). The quality of the libraries was assessed with an Agilent Bioanalyzer 2100 system. V3-V4 libraries were sequenced on an Illumina MiSeq platform, generating 2 × 300-bp paired-end reads. Illumina short-read amplicons are available in NCBI’s BioProject database under BioProject ID PRJNA768445.

Nanopore: Libraries for Nanopore sequencing were constructed using the Duplex-enabled Native Barcoding Kit 24 V14 (SQK-NBD114.24) (Oxford Nanopore Technologies, ONT), following the manufacturer’s manual. Sequencing of libraries (15 ng DNA total, measured with a Quantus Fluorometer (Promega) was performed on a MinION sequencing device (MIN-101B, ONT) and an R10.4.1 flow cell (FLO-MIN114, ONT). Nanopore sequence reads were deposited in NCBI’s BioProject database under BioProject ID PRJNA1076812.

Sequence data processing of Illumina R1 and R2 output files and of Nanopore pod5 output files were conducted on the high-performance computing cluster (HPC) of the RPTU Kaiserslautern-Landau.

Illumina: Raw sequence reads were quality filtered and trimmed by executing the dada2 (divisive amplicon denoising algorithm) workflow (Callahan et al. 2016) in R Studio 3.5.1 to obtain ASVs (Amplicon Sequence Variants). Truncation length was set to 255 bp (Dully et al. 2021b) and we kept only sequences with a mean Phred quality score ≥ 30 (Q3) corresponding to 99.9% base-call accuracy (Ewing and Green 1998). For maxEE, we chose 1 to maximize downstream sequence quality. The paired-end sequences were merged using a minimum 20 bp overlap and a mismatch of two bases was allowed (Frühe et al. 2021a). Before the construction of the ASV-to-sample matrix, the sequences were checked for chimeras using the uchime_denovo function of vsearch (Rognes et al. 2016). ASVs that were represented by only one single sequence in the complete Illumina dataset (singletons) were removed from the final ASV-to-sample matrix.

Nanopore: Pod5 files were subjected to duplex base calling using ONT’s Dorado duplex base-caller version 0.5.3 with the super-accuracy model (dna_r10.4.1_e8.2_400bps_sup@v4.2.0, https://github.com/nanoporetech/dorado) (Wick 2023). SAMtools (Li et al. 2009) was used to count and extract the duplex reads in the resulting bam files. Barcode demultiplexing from duplex bam files was conducted with Dorado, resulting in one duplex bam file per sample. In this step, the barcodes were trimmed off the sequences. Using SAMtools, bam files were then transformed to fastq files. Using cutadapt v. 1.18 (Martin 2011), sequences < 1400 and > 1600 bp (Girija et al. 2023) were removed from the datasets. Due to noisy signals at the beginning and the end of each sequence, the first and the terminal 50 bases were removed from each sequence with cutadapt. Finally, we extracted all sequences with a mean Phred score ≥ 30 (Q3). As the dada2 pipeline used for ASV calling of Illumina reads is for several reasons not suitable for Nanopore’s long-read 16S rRNA sequences (Zhang et al. 2023), sequences of the high-quality fastq files were binned into operational taxonomic units (OTUs) using NanoCLUST (Rodríguez-Pérez et al. 2021) as implemented in the BugSeq 16S pipeline (Fan et al. 2021; Jung and Chorlton 2021). The sequence similarity threshold for clustering sequences was 98%, corresponding to bacterial species demarcation (Kim et al. 2014).

Taxonomic classification of Illumina-derived ASVs and Nanopore-derived OTUs was conducted using the 16S BugSeq pipeline (Fan et al. 2021; Jung and Chorlton 2021). As the results of taxonomic classifications can be notably influenced by the methodology of assignment and by the reference database (Petrone et al. 2023; Zhang et al. 2023), we chose the same classification pipeline for both the Illumina ASVs and the Nanopore OTUs. BugSeq’s 16S pipeline employs the QIIME2 VSEARCH consensus classifier (Rognes et al. 2016; Hall and Beiko 2018). Sequences were aligned against NCBI’s nucleotide (nr) database with both sequence similarity and alignment length thresholds being set at the default (80%), ensuring that only high-quality alignments are retained. In case of multiple alignments with an equal top score, equally good top hits were collapsed to their lowest common ancestor to ensure an as high as possible accuracy in taxonomic classification (Jung and Chorlton 2021).

For pattern matching between the Illumina-derived ASV-to-sample matrix and the Nanopore-derived OTU-to-sample matrix, we calculated standard measures of diversity. Data analyses were conducted in R v. 4.0.5 using the packages vegan (Oksanen et al. 2022) and ggplot2 (Wickham 2016) for graphical visualisation. Prior to the calculation of alpha- and beta-diversity measures, the number of sequences per sample was rarefied (normalised) to the smallest sample size with the rrarefy function. In case of the Nanopore OTU matrix, this number was 10,986 sequences (LIS_REF, Table 1) and, in case of the Illumina ASV matrix, 60,607 (LIS_REF, Table 2). We calculated the Shannon-Wiener Index H’ for both datasets as a measure of alpha-diversity. The abundance-based Bray-Curtis (BC) index and the incidence-based Jaccard index were used to calculate the similarity between samples (beta-diversity), based on the Hellinger transformed ASV and OTU data. For the calculation of both indices, we considered: (i) the complete ASV and OTU datasets; (ii) only ASVs and OTUs assigned to the taxonomic rank of genus; (iii) only ASVs and OTUs assigned to the taxonomic rank of family. In case of the complete ASV and OTU datasets, four separate distance matrices (ASV Jaccard, ASV BC, OTU Jaccard, OTU BC) had to be calculated, because ASVs and OTUs are taxonomy-independent and, therefore, the two datasets could not be superimposed. In case of beta-diversity indices based on taxonomic ranks, we united the ASV data and the OTU data to a single matrix, one that included all sequences assigned to the rank of families and one to the rank of genera. Thereby, we only considered families and genera in which the sequence read counts accounted for at least 0.5% of all sequence reads within an individual sample. For visualisation of beta-diversity, we then performed hierarchical clustering using the Ward criterion (Murtagh and Legendre 2014).

For a more in-depth taxonomic comparison, we compared the top ten (in terms of sequence reads) bacterial families and genera obtained from both sequence datasets.

The evolution of Nanopore sequence data from raw reads (pod5) per sample to high quality reads which could be taxonomically assigned to the domain Bacteria is shown in Table 1. From an initial nearly 5,200,000 reads after a runtime of 21 hours and 34 minutes (estimated bases: ca. 9 Gbp with an approximated N50 of 1.49 Kbp). The remaining total number of reads after a rigorous quality filtering (see Methods section above) for downstream analyses was 131,494 with the lowest number of reads being in sample LIS_REF (10,986) and the highest in sample DUN_AZE (41,580). The main loss of sequence reads occurred during the extraction of duplex reads and then again after barcode filtering: of the ca. 2 million duplex reads, roughly 500,000 remained that could undoubtedly be assigned to the seven samples (six eDNA samples plus one non-template control sample). None of the samples had a singleton read. In the non-template control, we found eight sequences, none of which returned a hit after aligning to NCBIs nucleotide database. None of these eight sequences occurred in any of the final sample data.

In case of the Illumina dataset (Table 2), the loss of sequence reads after quality filtering was notably lower. Throughout the sequence data processing pipeline, we have lost a per-sample average of 43% of the originally obtained R1/R2 sequence reads. In numbers, we have started with roughly one million reads for the six eDNA samples of which we could use 448,036 for downstream analyses.

The AZTI Marine Biotic Index (AMBI) obtained from the LIS samples in the framework of a compliance monitoring survey resulted in the following finding. Environmental quality was similarly high at LIS_ REF (AMBI: 2.0 for replicate 1 and 1.8 for replicate 2) and LIS_AZE (AMBI: 2.3 for replicate 1 and 2.2 for replicate 2), while LIS-CE was notably more impacted (AMBI: 5.7 for replicate 1 and 5.8 for replicate 2).

The relative trend in Shannon diversity was largely congruent between the Illumina-derived ASV and the Nanopore-derived OTU datasets (Table 3). Both agreed that the lowest Shannon diversity for both the LIS and the DUN farm was at the cage edge sites (CE). Further commonalities between the two datasets were that the Shannon diversity at the CE and the reference (REF) sites were lower at the LIS farm compared to the DUN farm. Noteworthy differences in Shannon diversity between the two datasets were: first, the differences between the two aquaculture installations were notably more pronounced in the Nanopore dataset compared to the Illumina dataset. Second, both datasets disagreed in the obtained Shannon indices for the Allowable Zone of Effect (AZE). For both farms, the Illumina dataset suggested a Shannon diversity at the AZE which was in the same order of magnitude as the one obtained from the REF samples. In contrast, the Nanopore dataset revealed a notably higher Shannon diversity at the AZE compared to the REF for both farms.

The Nanopore and the Illumina datasets showed full agreement when matching their abundance-based Bray-Curtis beta-diversity patterns on an OTU and ASV basis, respectively (Fig. 2). Both datasets showed two distinct clusters. One consisted of the cage edge samples of both aquaculture installation sites (LIS_CE and DUN_CE) plus the AZE sample from the DUN farm. In both cases, the DUN_CE and DUN_AZE were more similar to each other in their bacterial community profiles than any of these two to the bacterial community at the LIS_CE. The second cluster consisted of the reference samples of both farms (DUN_REF and LIS_REF) and the AZE sample of the LIS farm. In both cases, the LIS_REF and LIS_AZE were more similar to each other than any of the two was to the DUN_REF sample. One noteworthy difference was that the Bray-Curtis distances between the bacterial communities were higher in the Nanopore dataset (Fig. 2a) compared to the Illumina dataset (Fig. 2b).

When collating the OTU and ASV matrices on the taxonomic ranks family, genus and species, the pattern described above for ASVs and OTUs remained the same in case of the Nanopore dataset (Figs 2c, e, g). In case of the Illumina dataset, the beta-diversity patterns, based on the taxonomic ranks family (Fig. 2d) and genus (Fig. 2f), changed slightly in one aspect. While the two large clusters remained the same across all beta diversity analyses, the within-group clustering of the “DUN_CE – LIS_CE – LIS_AZE” cluster has changed in the Illumina dataset. Instead of a higher similarity between the bacterial communities of the LIS_REF and LIS_AZE as in all other beta diversity analyses, we found a higher similarity between the LIS_REF and DUN_REF samples in both taxonomy-based Illumina datasets.

While in case of the Nanopore dataset, sufficient bacterial species with high read abundances (> 0.5%) were identified (between 28 in samples LIS_CE and DUN_CE and 39 OTUs in sample DUN_REF, Table 4) to allow for a beta-diversity analysis, this was not the case for the Illumina dataset. Here, only between two (LIS_REF) and nine (LIS_CE) species could be identified. This was an insufficient number to conduct beta-diversity analyses on the species rank.

The beta-diversity results obtained for the incidence-based Jaccard index were identical to the ones described above for the abundance-based Bray-Curtis index (Suppl. material 1: figs S1–S7).

Figure 2.

Hierarchical clustering (based on Bray Curtis index as a measure of beta diversity) as a measure of beta diversity, based on Nanopore-obtained amplicons (left panel) and Illumina-obtained amplicons (right panel) for the two salmon aquaculture installations DUN and LIS. The middle panel shows the (taxonomic) units on which the beta-diversity analyses are based. Colour and shape coding of samples helps visualisation and interpretation of data. AZE = Allowable Zone of Effect; CE = Cage Edge; REF = unimpacted Reference site.

On average, 86% (± 2.8%) of the high-quality Nanopore amplicon reads could be assigned to the taxonomic rank of bacterial family, whereas this was only 52% (± 7.8%) of the high-quality Illumina dataset (Table 4). A maximum of only 13% (LIS_AZE and DUN_REF) of the total bacterial families inferred from Illumina amplicons had a sequence read abundance of > 0.5%. Thus, in the Illumina dataset, the vast majority of families obtained had low read abundances (Table 4). In case of the Nanopore dataset, up to 84% (LIS_CE) of the identified families had a read abundance of > 0.5%. An average of 82% (± 1.8%) of the Nanopore reads could be assigned to the taxonomic rank of bacterial genus, which was only 50% (± 7.1%) in case of the Illumina sequence reads (Table 4). With very few exceptions, all the Illumina-derived genera had low read abundances. In case of the Nanopore dataset, on average, more than half (54% ± 11%) of the detected genera were with high-abundant reads (Table 4). In the Nanopore datasets, a per-sample average of 65% (± 11%) of all sequence reads returned a species hit in the taxonomic analysis, whereas this was only for an average of 15% (± 4%) of the Illumina sequence reads. In the Nanopore dataset, up to 72% of the identified bacterial species had high-abundant reads, whereas this applied to a maximum of only 3% of the Illumina-derived species. Up to > 99% of the ASVs obtained from Illumina sequencing that were assigned a species rank belonged to the rare ASVs.

Seven out of the ten most abundant Nanopore-derived families (combined over all six samples) were also amongst the ten most abundant bacterial families of the Illumina dataset, albeit with distinct relative abundances (Fig. 3). These shared families were Desulfosarcinaceae, Desulfocapsaceae, Desulfobulbaceae, Desulfobacteraceae, Sulfurovaceae, Woeseiaceae and Halieaceae. Thiotrichaceae, Sedimenticolaceae and Sandarinaceae were amongst the top ten Nanopore-derived families, but not amongst the top ten Illumina-derived families.

While Sedimenticolaceae and Sandarinaceae were also present in the Illumina dataset, albeit not amongst the ten most abundant families, Thiotrichaceae entirely escaped Illumina detection (Fig. 4). Chromatiaceae, Thioprofundaceae and Prolixibacteraceae were amongst the top ten Illumina-derived families. While Chromatiaceae was also recorded with the Nanopore protocol, albeit not amongst the ten most abundant, the latter two were not recorded at all with Nanopore sequencing (Fig. 4). In general, only few families which accounted for > 0.5% of the sequence reads in one of the datasets was entirely missing from the other. In addition to Thioprofundaceae and Prolixibacteraceae, two further bacterial families, Desulfosalsimonadaceae and Wenzhouxiangellaceae, were exclusively present in the Illumina dataset, but missing from the Nanopore dataset. In addition to Thiotrichaceae, Thermoanerobaculaceae and Spirochaetaceae escaped detection with the Illumina protocol, but were detected with Nanopore sequencing (Fig. 4).

Comparing the ten genera with the most abundant reads in both sequence datasets (Fig. 5), we found four genera that were shared. These were Desulfobulbus, Sulfurovum, Halioglobus and Woesia. The genera Desulfatiglans, Desulforhopalus, Desulfopila, Thiogranum, Psychromonas and Aminicenantales were exclusively amongst the top ten of the Nanopore dataset (Fig. 5). With the exceptions of Thiogranum and Aminicenantales, these genera were also recorded with Illumina sequencing; however, not amongst the ten most abundant genera. Desulfosarcina, Parahaliea, Thiohalocapsa, Thiolapillus, Thioprofundum and Wenzhouxiangella were amongst the ten most abundant genera in the Illumina dataset, but not amongst the top ten of the Nanopore dataset.

With the exception of Desulfosarcina, none of the other five Illumina-derived top ten genera was recorded with the Nanopore dataset. Only one genus (Thiogranum) which had a sequence read abundance of > 0.5% in the Nanopore dataset escaped detection with the Illumina dataset (Fig. 6). In contrast, five genera (Wenzhouxiangella, Thioprofundum, Thiolapillus, Thiohalocapsa, Parahaliea and Marimicrobium) with a sequence read abundance of > 0.5% in the Illumina dataset could not be detected with the Nanopore protocol (Fig. 6)

Figure 3.

The ten most abundant (in terms of sequence reads) bacterial families and their relative abundances obtained from the Nanopore and Illumina datasets.

Figure 4.

Bacterial families which were detected exclusively with Nanopore or with Illumina sequencing. The violin-and-box-plots show the relative abundances. Violins (coloured areas) show the relative abundance distribution across the individual samples. Boxes show median, 25%- and 75%-quartiles and min-max values.

Figure 5.

The ten most abundant (in terms of sequence reads) bacterial genera and their relative abundances obtained from the Nanopore and Illumina datasets.

Figure 6.

Bacterial genera which were detected exclusively with Nanopore or with Illumina sequencing. The violin-and-box-plots show the relative abundances. Violins (coloured areas) show the relative abundance distribution across the individual samples. Boxes show median, 25%- and 75%-quartiles and min-max values.

Our results witness a notably improved data quality of Nanopore sequences compared to previous studies (e.g. Older et al. (2023); Petrone et al. (2023); Zhang et al. (2023)). This improved sequence quality is a result of the recently introduced sequencing chemistry, which, in combination with the new R10.4.1 flow cell, enables duplex reads, whereas the previous chemistry with the same flow cell allowed only for single-molecule reads (simplex reads) only. For example, in our LIS_CE dataset, 63% of the duplex reads with the correct length have passed the Phred30 quality filter. Previous Nanopore sequencing studies analysing bacterial mock or natural communities did not even attempt to use such a high Phred score quality filter as is standard for Illumina MiSeq data. Older et al. (2023) found an average of 59% of 16S rRNA gene sequence reads that have passed a Phred20 filter. Petrone et al. (2023) reported an average Phred score of 18.1 with an average of 79% of reads passing a Phred15 filter. Similarly, Zhang et al. (2023) obtained an average Phred score of 18.8 for their bacterial 16S rRNA gene dataset, but did not apply any Phred quality filter for downstream analyses.

The proportion of duplex reads obtained in our study was 36% (Table 1) and, thus, notably higher compared to other studies. Genome-sequencing different bacterial species, Lerminiaux et al. (2024), as well as Sanderson et al. (2023), reported a duplex rate of only ca. 6% and 7%, respectively, using the same chemistry and flow cell type we have used in our study. It is not unlikely that the success rate of duplex sequencing correlates with the length of the target gene. Nanopore genome sequencing produces up to > 2 Mbp, with 10–30 Kbp being common (Amarasinghe et al. 2020; Lerminiaux et al. 2024). There may be a higher likelihood for shorter DNA fragments binding at the pore while waiting to follow the first strand through the pore. This assumption, however, needs to be verified. Thus far, no further study is published (to the best of our knowledge) that has analysed bacterial community structures using 16S rRNA gene Nanopore duplex sequencing (or any other targeted gene with a size < 2Kbp), which would allow a confirmation of this assumption.

In addition, while Sanderson et al. (2023), as well as Lerminiaux et al. (2024), have used ONTs Guppy tool for duplex base-calling, we in this study have used ONTs new Dorado base-caller, which relies on a bi-directional Recurrent Neural Network (RNN) algorithm that was optimised for duplex reads in contrast to the Guppy base-caller (https://github.com/nanoporetech/dorado, Wick (2023)). Despite a high loss of sequence reads during duplex base-calling and another noteworthy loss due to erroneous barcode reads, the data per sample that we have retained for analyses exceeded by far the number of reads required for a sample when subjected to DNA-based marine biomonitoring. Dully et al. (2021b) tested several different marine DNA-based biomonitoring scenarios using microbial communities and identified a per-sample sequence depth of 3,000–5,000 sequences as sufficient. Any further increase in sequence numbers did not affect the monitoring results. Wilding et al. (2023) analysed the signal-to-noise ratio in 16S rRNA gene amplicon dataset obtained from DNA-based monitoring of coastal salmon aquaculture installations. The authors found that between 10 and 100 top abundant bacterial ASVs are optimal for biomonitoring purposes. Both recommendations are met in our Nanopore sequence datasets, thus, allowing for a comparison of the results obtained by Nanopore sequencing with the ones obtained from the sequencing of the same samples with Illumina MiSeq.

Both sequencing protocols produced largely consistent alpha- and beta-diversity results. This agrees with other studies which also observed consistent ecological trends in 16S rRNA gene data obtained from the two sequencing platforms (e.g. Nygaard et al. (2020); Lemoinne et al. (2023); Older et al. (2023); Stevens et al. (2023); Zorz et al. (2023)). On the level of OTUs (Nanopore) and ASVs (Illumina), both datasets showed the same clustering of the six different samples collected from the two aquaculture installations. The two main clusters grouped the samples according to environmental impact into a “high impact cluster” (samples DUN_CE, DUN_AZE, LIS_CE) and a “low impact cluster” (samples DUN_REF, LIS_REF, LIS_AZE). This corroborates well with previous findings which suggested that the structures of benthic bacterial communities are a robust indicator of environmental impact arising from fish farming (Dowle et al. 2015; Keeley et al. 2018; Stoeck et al. 2018; Dully et al. 2021a; Frühe et al. 2021b; Leontidou et al. 2023; Wilding et al. 2023).

Both LIS_REF and LIS_AZE had AMBI values, obtained from macroinvertebrate-based compliance monitoring of the LIS farm, which indicated a good ecological status (Muxika et al. 2005). Thus, we can conclude that the LIS_AZE was hardly impacted by organic pollution through the salmon farm. This explains the clustering of LIS_AZE together with the two reference samples of both farms (LIS_REF and DUN_REF). These reference samples were chosen at a distance far enough from the fish pens to remain unimpacted by any deposits resulting from the aquaculture activities (Wilding et al. 2023). Furthermore, we can conclude that the DUN_AZE was notably more impacted than the LIS_AZE, because DUN_AZE together with DUN_CE clustered together with LIS_CE, which had an averaged AMBI of 5.75, indicating a bad ecological status and a heavily disturbed site (Muxika et al. 2005). A further interesting finding from the OTU/ASV-based beta-diversity analyses is that the Bray-Curtis distances between the bacterial communities were higher in the Nanopore datasets compared to the Illumina dataset. This suggests a higher sensitivity of the Nanopore dataset in distinguishing bacterial communities of different samples. This finds support in the results obtained for the Shannon index. Even though the general trend is highly similar, the differences between individual samples are notably more pronounced in the Nanopore data compared to the Illumina data. A higher sensitivity in bacterial community diagnostics is of advantage in biomonitoring to indicate more subtle environmental changes.

When collating sequence data across the taxonomic levels family and genus, the branching of samples within the Illumina-derived high-impact clusters changed compared to all other beta-diversity analyses. Short hypervariable 16S rRNA gene regions often lack a differentiation between genera of the same family and, likewise, between species of the same genus (Callahan et al. 2019; Klair et al. 2023; Older et al. 2023; Zhang et al. 2023). Therefore, it is not surprising that several studies reported differences in the presence/absence of individual taxa in bacterial communities that were sequenced with both Illumina and Nanopore (Nygaard et al. 2020; Klair et al. 2023; Lemoinne et al. 2023; Stevens et al. 2023; Tandon et al. 2023). In the Nanopore dataset, we only considered sequences with ≥ 98% similarity to a deposited sequence in the NCBI database, which allows for more accurate taxonomic classifications compared to short reads (Benitez-Paez et al. 2016; Klair et al. 2023; Zhang et al. 2023). More than one third (36%) of the here-obtained high-quality Nanopore sequences used in downstream analyses had a sequence similarity of > 99% to NCBI reference sequences which corresponds to the species- and strain-discrimination boundaries in bacterial 16S rRNA gene sequences (Benitez-Paez et al. 2016). This makes taxonomic classifications, based on long Nanopore reads, more robust and reliable compared to the short Illumina reads, particularly on lower taxonomic levels.

The low taxonomic resolution of short Illumina reads is furthermore of disadvantage regarding the effect of sequencing errors on taxonomic assignment accuracy. A MiSeq-platform inherent weakness is the accumulation of substitution errors (Schirmer et al. 2015; Schirmer et al. 2016). This may lead to sequences that do not exist in nature and, eventually to spurious taxa or ASVs (false positives) which logically cannot be detected with Nanopore sequencing (Xue et al. 2018; Stevens et al. 2023). Additionally, Nanopore sequencing is generating errors. However, if the same Phred sequence quality filter is applied, the effects of these errors on taxonomic assignment accuracy are much lower for the (near) full-length 16S rRNA gene compared to a short gene fragment.

Different protocols for sequence generation may distort relative abundance comparisons, in particular due to PCR primer bias and stochasticity (Balint et al. 2016; Lemoinne et al. 2023). The primer pair we used in this study to amplify the V3-V4 16S rRNA gene region is very popular in bacterial diversity research in combination with Illumina MiSeq sequencing. Amongst others, this primer pair is frequently used as standard in biomonitoring of aquaculture installations (Stoeck et al. 2018; Dully et al. 2021a; Dully et al. 2021b; Frühe et al. 2021a; Frühe et al. 2021b; Wilding et al. 2023). A recent study demonstrated that numerous bacterial taxon groups, amongst others on family and genus level, but even on phylum-level, are missed by this primer pair (Leontidou et al. 2023). For the amplification of the full-length 16S rRNA gene for Nanopore sequencing, we also used a primer pair (27F/1492R) which is very popular in bacterial diversity research (e.g. Galkiewicz and Kellogg (2008); Klindworth et al. (2013); Johnson et al. (2019); Fujiyoshi et al. (2020); Older et al. (2023); Tandon et al. (2023)). Likewise, this primer pair 27F/1492R, or variants thereof, discriminates against several bacterial taxon groups as demonstrated in previous studies (see, for example, Older et al. (2023); Tandon et al. (2023)).

In addition, differences in relative read abundances typically affect ecological trends inferred from Nanopore and Illumina sequencing of the same bacterial (mock and natural) communities (e.g. Klair et al. (2023); Older et al. (2023); Stevens et al. (2023); Tandon et al. (2023)). Many diversity measures, used with ASVs or OTUs, such as the Bray Curtis Index, are abundance-weighted and, therefore, largely influenced by the most abundant ASVs or OTUs rather than by the less abundant ones (Chiarello et al. 2022). Differences in relative abundances of individual bacterial taxon groups reflect an inherent technical bias of all PCR-based high-throughput sequencing methods resulting from the stochastic process of target-gene amplification (Balint et al. 2016; Lemoinne et al. 2023). One possibility to exclude the two above-mentioned sources of error that apply equally to Nanopore and Illumina amplicon protocols is PCR-independent metagenome sequencing (Leontidou et al. 2023). This is, however, too expensive to apply in routine DNA-based biomonitoring practice.

The authors have declared that no competing interests exist.

No ethical statement was reported.

The research leading to these results received funding from the Deutsche Forschungsgemeinschaft (DFG grant STO414/15-2).

Conceptualization: TS. Data curation: TS. Formal analysis: TS, SNK, HWB, VR. Funding acquisition: TS. Investigation: TS. Methodology: TS. Project administration: TS. Validation: VR, TS, SNK. Visualization: SNK. Writing - original draft: TS. Writing - review and editing: HWB, VR, SNK.

Thorsten Stoeck https://orcid.org/0000-0001-5180-5659

Verena Rubel https://orcid.org/0000-0001-9630-9050

All of the data that support the findings of this study are available in the main text or Supplementary Information.