Adaptive sampling with Oxford Nanopore offers a simple way to improve the efficiency of plant metagenomic studies

This is the authors' abstract. We don't add key points for this paper.

Shotgun metagenomics is prohibitively inefficient to implement when host DNA dominates reads, such as sampling for endophytes. Uniquely, sequencers from Oxford Nanopore Technologies can perform adaptive sampling where target sequences are actively enriched during sequencing. In this Letter, we outline how the efficiency of metagenomic studies can be improved by reversing this process, with very large improvements possible when the host DNA is the majority of reads, with microbial reads rising from 1.6% of total reads to 18.7% from maize roots. Importantly, this process is easy and free to implement when the host reference genome is available. High-throughput sequencing has revolutionised microbial ecology. Initially, DNA metabarcoding offered a cost-effective way of profiling plant microbiomes without culturing (Ficetola & Taberlet, 2023). However, the approach suffers from some major limitations, including the introduction of compositional biases by the PCR process, the taxonomic resolution being limited and the approach providing only limited functional insight into plant microbiomes. Alternatively, shotgun metagenomics has also been performed for many years and overcomes these limitations, but it too has its own set of limitations (Huang et al., 2024). For example, to adequately profile the microbiome, much larger quantities of DNA need to be sequenced (millions instead of thousands of reads). This is not a problem when microbial DNA dominates the source material, but often plant samples contain a considerable proportion of host DNA (Khoiri et al., 2021). In these instances, many times more total reads need to be produced, with the majority being discarded in order to generate the millions of microbial reads required to adequately profile the plant microbiome. This makes the approach prohibitively expensive and wasteful for many researchers to perform at scale. To tackle this issue, researchers have developed multiple strategies, including developing specialised protocols, for example extracting DNA from rhizosphere soil instead of roots directly (Quiza et al., 2021). However, these approaches risk missing endophytes completely and biasing the community profile towards loosely adhering microbes. Alternatively, methods have been developed to inhibit the sequencing of highly methylated (i.e. the plant) DNA, but these add cost and complexity to experiments (Heravi et al., 2020). Ultimately, a cheap and simple method to increase the proportion of microbial reads within metagenomic studies is highly desirable. Here, we hypothesise that the feature ‘adaptive sampling’ performed by Oxford Nanopore Technologies (ONT, UK) can fulfil this need. Typically, the feature is used to enrich preloaded target sequences. Here, the sequencer analyses reads in real time by mapping them to the reference genome. The decision to accept or reject a read is made after sequencing the first 400 bases (the ‘adaptive sampling chunk’) when the basecaller determines whether this chunk falls within any target region within the reference file (ONT; ADS_S1016_v1_revM_11Dec2024). If a read is rejected, the polarity of the applied potential on the pore is reversed, and the DNA molecule is ejected from the given pore. The method has been used to enrich both plant genes (Belinchon-Moreno et al., 2025) and specific host-associated microbes (Martin et al., 2022). However, this operation can also be inverted so that sequences matching target regions within the references are rejected (depletion mode), but this is less commonly performed. Despite earlier promising results in cows and animals (Marquet et al., 2022; Ong et al., 2022a,b), this has not been assessed using plants, to the best of our knowledge. We hypothesised that, using the plant genome as a depletion reference, we can reject host-derived reads and increase the proportion of the microbial reads within metagenomic studies. We henceforth refer to this process as adaptive depletion to differentiate it from adaptive sampling (target enrichment). To test this, we utilised maize plants (W22 near isogenic line) (Springer et al., 2018), partitioned into samples of rhizosphere soil that are expected to have a high proportion of microbial reads and the remaining washed root samples that have a very low proportion of microbial reads (i.e. unlikely to be viable without enrichment) (Fig. 1). We then ran samples on a PromethION P2 Solo (ONT, UK) under standard metagenomic conditions, or with adaptive depletion activated, and compared whether this significantly increases the total proportion of microbial reads in each sample type. Maize (Zea mays subsp. mays L. (W22)) were grown in standard potting soil in a glasshouse in Flakkebjerg (Slagelse, Denmark) and kept at c. 24°C in 16 h : 8 h, day : night cycles. After 5 wk, the root network was extracted and shook to remove loosely adhering soil. It was cut into c. 20 mm lengths and homogenised. Roots under 3 mm in diameter were placed into phosphate buffered saline (PBS) solution (Quiza et al., 2021). These were shaken for 25 min at 250 rpm on a horizontal shaker and subsequently centrifuged for 5 min at 2000 g to pellet the soil. Roots were then transferred into a new tube of PBS, and the process was repeated. Twice-washed roots were then freeze-dried overnight alongside the rhizosphere soil from the first wash. DNA was extracted using a QIAGEN Power Soil Pro Kit (Qiagen), with 250 mg of rhizosphere soil and 100 mg of roots used as starting material. DNA extracts underwent size selection, which was performed using MagBio (MagBio Genomics, Gaithersburg, MD, USA) in a 0.4 bead to sample ratio, removing DNA fragments < 2000 bp. All 20 DNA extracts were then barcoded and built into libraries using the Native Barcoding Kit 24 V14 (SQK-NBD114.24; Oxford Nanopore Technologies, Oxford, UK) following the manufacturer's guidelines. Sequencing was performed using a single flow cell (R10.4.1; FLO-PRO114M) on the PromethION 2 Solo (ONT). Initially, standard sequencing was performed before being stopped after 2 h. The flow cell was then washed and the library reloaded with adaptive sampling enabled (with depletion mode activated), using the W22 genome as the reference (GenBank assembly – GCA_001644905.2). The sequencer was then left to run for c. 72 h. Sequencing data were generated using the MinKnow software (v.24.06.10). The resulting Portable Data 5 (POD5) files were basecalled and demultiplexed using Dorado (0.8.2) with duplex reads enabled. Cutadapt (Martin, 2011) and NanoFilt (v.2.8.0) (De Coster et al., 2018) were used to remove low-quality reads (< 1000 bp and low-quality reads q < 15). Minimap2 (v.2.26) was then used to remove the host DNA using the W22 reference (Li, 2018), with the map-ont flag enabled. Finally, Kraken2 (v.2.1.3) was used to confirm whether reads were microbially derived (Wood et al., 2019). Specifically, the nonhost reads were mapped to the standard Kraken2 microbiome database. At 74.8% (±10.6%), the percentage of nonplant host reads in rhizosphere soil samples, as produced via shotgun metagenomics, was comparable to other studies (Azevedo-Silva et al., 2021; Quiza et al., 2021). Meanwhile, as predicted, washed root samples produced only a very small percentage of nonhost-derived reads, accounting for just 1.8% (±1.5%) of total reads. Interestingly, this rose significantly for both the rhizosphere soil (t = 5.20; P = 0.006) and the roots (t = 4.33; P = 0.012), to 95.8% (±1.6%) and 20.9% (±11.3%) of reads, respectively, with our depletion approach (Fig. 2b). This represented a 1.3-fold increase for the rhizosphere soil and a 13.1-fold increase for the roots (Fig. 2c). Of this nonhost data, the majority was assigned to bacteria (88.2%), with only a fraction unassigned (Supporting Information Table S1). Unfortunately, the shotgun libraries were limited in size, producing a mean of 12 675 (±6688) quality-filtered reads per sample compared with the 837 569 (±1060 793) reads per sample produced in the adaptive sequencing (Fig. 2a). While not sufficient to profile the entire microbiome, this represents many datapoints for the comparisons between methods. Furthermore, these reads had a mean sequence length of 5033 bp, equivalent to 33 times the length of the standard 150 PE Illumina sequencing, which accounted for 62.5 Mbp (±39.3) data per sample. The rhizosphere soils showed no obvious skew in the taxonomic composition between these approaches, suggesting the adaptive depletion approach is not biasing the community profile in a major way (Fig. 2d). We specifically did not use mitochondrial and chloroplast genomes in the adaptive reference sequences due to their high sequence homology to rhizosphere/root bacteria, as this could lead to more closely related microbes to the plasmids being selected against. Further studies are required to investigate whether specific microbial taxa or regions of microbial genomes are suppressed in the technique. Meanwhile, as only a fraction of reads were nonhost derived from the root samples, this led to very few microbial reads from the shotgun dataset being assigned to bacteria, explaining the variability in the community compositions produced between the shotgun and adaptive depletion approaches (Fig. 2d). Therefore, we performed an additional check to determine whether differences in read depths were biasing our results. This was performed by subsetting the adaptive depletion data to 10 000 reads per sample to be comparable to our shotgun dataset, and then the subset data were reprocessed. The proportion of nonhost reads was significantly higher in the subset data than in the complete dataset (t = −8.55; P = 0.001) (Table S2), rising from 95.8% to 97.7% being nonhost. However, both of these values deviate much more from the percentage of nonhost reads produced with the shotgun method (74.8%), suggesting the lower read depths in the shotgun metagenomic data was very unlikely to be the main cause of the differences in nonhost reads observed between methods. While our results suggest the adaptive depletion is powerful, it is not without limitations. First, a reference genome is required, and this is not available for many species (Marks et al., 2021). While the approach is likely to be effective across plant species and genotypes, this needs confirmation. More likely, lower genome quality and contiguity will reduce the effectiveness of the host depletion. Related genomes could potentially be substituted, but this too will likely impact the efficiency by which host reads are rejected. Furthermore, adaptive depletion cannot be performed on DNA contaminated by a mixture of hosts, such as bulk soil samples. Second, ONT sequencers continue to have higher error rates than other sequencing platforms (such as PacBio and Illumina) (Sevim et al., 2019; Cook et al., 2024). Error rates have reduced in recent years to the extent that it was possible to assemble microbial and plant genomes without supplementing with more accurate short-read data (Sereika et al., 2022). For example, reads here had a mean q-score of 19.7, or c. 1% error rate. However, it remains much less accurate and thus requires more sequencing depth than other methods. While further testing is needed, these results suggest adaptive depletion is highly beneficial in scenarios in which host genomic data need to be filtered and a reference host genome is available. Importantly, adaptive depletion does not require additional laboratory work, is free and easy to implement and does not seem to bias the community composition. Furthermore, the approach is beneficial at both high and low microbial DNA proportions. Given the ease of adaptive depletion to implement and the lack of adverse effects on the composition, we highly recommend researchers deploy the approach where possible. The authors would like to thank the funders for making this work possible. Specifically, the Villum Fonden as part of their Villum Young Investigator programme (60770) and by the Aarhus Universitets Forskningsfond (AUFF-E-2024-9-14). None declared. JTPV, ASM and CJB performed bioinformatic analyses. NR, IBN and SM prepared samples and processed them in the laboratory. TKN, IBN and SSTM helped to calibrate equipment and generate data from the sequencer. CJB conceived of the concept. All authors contributed to the initial manuscript draft and to its revisions. JTPV and ASM contributed equally to this work. All data are freely available on the NCBI under the BioProject accession no. PRJNA1257062. Table S1 Metagenomics data were produced from maize roots/rhizosphere soil via standard shotgun metagenomics or via adaptive sampling, and the ratio of host and nonhost reads were determined. Table S2 Adaptive data from maize roots/rhizosphere soil contained many more reads than the shotgun data; therefore, it was subsampled to a comparable size to the shotgun data and reanalysed for potential biases. Please note: Wiley is not responsible for the content or functionality of any Supporting Information supplied by the authors. Any queries (other than missing material) should be directed to the New Phytologist Central Office. Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article. The New Phytologist Foundation remains neutral with regard to jurisdictional claims in maps and in any institutional affiliations.