Accurate MAG reconstruction from complex soil microbiome through combined short- and HiFi long-reads metagenomics

Background Advances in high-fidelity long-read (HiFi-LR) sequencing technologies o ffer unprecedented opportunities to uncover the microbial genomic diversity of complex environments, such as soils. While short-read (SR) sequencing has enabled broad insights at gene-level diversity, the inherently limited read length constrains the reconstruction of complete genomes. Conversely, HiFi-LR sequencing enhances the quality and completeness of metagenome-assembled genomes (MAGs), enabling higher-resolution taxonomic and functional annotation. However, the cost and relatively low throughput of HiFi-LR sequencing can limit genome recovery, particularly at the binning stage, where coverage depth is critical.

Here, we present a novel hybrid strategy that differs from classical hybrid assemblies, where SR and LR reads are jointly used at the assembly step. Instead, we use high-depth SR data to improve the binning of HiFi-LR contigs. Using both SR and HiFi-LR metagenomic data generated from a tunnel-cultivated soil sample, we demonstrate that SR-derived coverage information signi ficantly improves the binning of HiFi-LR assemblies. This results in a substantial increase in the number and quality of recovered MAGs compared to using HiFi-LR data alone and an uncomparable improvement compared to SR data alone.

Our findings highlight the power of combining SR and LR in highly diverse environments, such as soil, not for hybrid assembly per se, but to enhance the downstream binning process. The combination of SR and LR data substantially improves the downstream binning process and overall genome recovery. Importantly, this approach underscores the potential of leveraging the vast amount of publicly available Illumina metagenomic datasets. Completing existing SR resources with PacBio HiFi sequencing can maximise assembly contiguity and binning accuracy using massive amounts of SR data already generated. This highlights a practical and forward-looking strategy for microbiome research, where novel LR technologies will bring new value to previous short-read e fforts.

Metagenomic | Soil | HiFi long read sequencing | Short read sequencing | Metagenomic assembled genomes 2 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint

Soil is one of the most diverse microbial ecosystems and remains underexplored (1,2). It hosts an extraordinary diversity of microorganisms, including bacteria, archaea, fungi, protists, and microscopic animals. The work of (3) reports that a single gram of soil may contain up to 10¹⁰ microbial cells covering diverse taxa, and may harbour up to thousands of different species. In comparison to other biomes, such as the human gut or some other host-associated biome, where microbial communities are relatively well-characterised, soil microbiomes show higher alpha-diversity and more unknown taxa. Remarkably, despite this extreme diversity, only ~2% of bacterial phylotypes (approximately 500 out of over 25,000 identi fied) account for nearly 50% of all bacterial sequences in soils worldwide, forming a globally dominant core microbiome shaped primarily by environmental factors such as soil pH, aridity, and plant productivity, rather than by geographic location (4). A comprehensive understanding of soil microbiomes, including rare species, is essential in various applications, ranging from studying biogeochemical cycles (5) to developing sustainable agricultural practices (6). Microbial genomes contribute to the roles and functions of the soil. Indeed, gene contents of the di fferent species and their regulation pathways underlie microbiome activities such as interactions with other microbes or the environment, catalysing essential biochemical reactions or contributing to ecosystem processes (7). Thus, uncovering the diversity of soil microbial genomes is crucial for harnessing the full potential of microbial resources and developing sustainable agricultural practices (6). Due to the complexity of soil microbial ecosystems, bulk culture-independent approaches are the primary strategies employed for their characterisation. Even with the improvement of culturomic methods, many soil microbial species remain di fficult or impossible to grow in laboratory conditions (8). Shotgun metagenomics allows direct DNA sequencing from complex microbial communities without prior cultivation. This approach can provide a genomic snapshot of the microbial community in a given soil sample even if dark matter constituted by rare species remains di fficult to capture. Metagenomics has revolutionised our ability to investigate the diversity and functions of uncultured organisms, providing unprecedented insights into the genetic diversity, functional potential, and ecological roles of microbiota (9). De novo metagenomic assembly allows the reconstruction of genomes for soil species, including uncultured taxa, and the prediction of encoded proteins (10) (11). However, the reconstruction of high-quality genomes from the complex mixture of sequenced reads remains very limited in soil, with usually only a few whole genomes obtained and many unassembled reads (12). Most metagenomic projects from the past decade have relied on short-read (SR) sequencing, generating high-throughput datasets of generally 2x 50–150-base-pair fragments. These 3 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint SR-based metagenomic studies remain largely dominant in terms of publicly available amount of sequence data in public repositories such as the JGI’s IMG/M or the EBI’s MGnify. Such studies face the challenge of reconstructing genomes containing multicopy genes or repetitive sequences, which are inherently di fficult to resolve with SRs. Assemblies from short reads are particularly fragmented because repeats longer than the read length cannot be spanned, creating ambiguous paths in the assembly graph and ultimately impeding genome reconstruction. Highly repetitive elements and the presence of closely related species further complicate genome assembly (13). In the 2022 release of the JGI IMG/M platform, SR-based genome assemblies exhibited high fragmentation, with only 10% of the contigs being at least one kilobase (kb) long. Furthermore, the inconsistency of protein taxonomic assignment within contigs reveals chimeric assemblies (11). Increased read length in metagenomic studies enhances contig lengths by providing long-range information, which allows better resolution of repetitive sequences. In recent years, Oxford Nanopore Technologies (ONT) (14,15) and Pacific Biosciences (PacBio) (16) platforms have significantly improved read accuracy, length, and throughput. The first generations of LR sequencing technologies were characterised by high error rates, which led to the recommendation of combining them with SR for error correction during the assembly step (17). Latest technologies provide high- fidelity (HiFi) sequences, with error rates that are similarly low to those of SR (18). In this study, we assessed the added value of HiFi LR sequencing over SR sequencing in complex environments such as soil, and whether combining both HiFi LR and SR can improve the reconstruction of MAGs (Metagenome-Assembled Genome). We generated and analysed PacBio HiFi long reads (LR), Illumina short reads (SR) shotgun metagenomics and metabarcoding data to explore the microbial community of a cultivated soil sample. We compared the reconstruction of contigs, bins, and MAGs using HiFi long reads alone and short reads alone. To enhance MAG reconstruction, we propose a strategy that leverages the high coverage of SR to complement the relatively low coverage of LR, substantially improving binning and thus genome recovery. Despite many reads remaining unassembled, hybrid binning improved contiguity and yielded hundreds of MAGs. Furthermore, we validated by metabarcoding data that this hybrid approach captured a broader spectrum of soil microbial diversity, including low-abundance taxa. Overall, while HiFi LR sequencing demonstrates strong potential as a standalone technology for metagenomics, the amount of metagenomics data already available is largely dominated by SRs. Our findings highlight the added value of combining LR and SR sequencing for MAG reconstruction in soils. This hybrid strategy unlocks an enormous re-use potential of all the past metagenomics e fforts based on SRs by complementing them with LR data even at relatively low coverage, o ffering 4 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint a practical framework for large-scale exploration of microbial dark matter in complex environments.

Sampling We collected a soil sample between two rows of salad crops in a tunnel farming setup in Lambesc, southern France. Rapid freezing, followed by subsequent freeze-drying procedures at -40°C, immediately preserved the soil sample. These preservation and preparation steps were meticulously executed at the Conservatoire de Ressources Génétiques (CRG), hosted within the Genosol platform in Dijon, France, to maintain the molecular integrity of the samples in readiness for subsequent analytical procedures. DNA extraction Total genomic DNA was extracted using modi fied IS011063 protocols dor DNA extraction and Nucleospin soil for DNA puri fication (19). From 1g of dry equivalent soil (freeze-dried), DNA was extracted by mechanical lysis using FastPrep 5G (MP Bio) coupled with cell lysis with denaturing detergent (lysis bu ffer : EDTA, Tris ph 8.0, NaCl and SDS 2%). Deproteinisation consisted of precipitating the proteins with a high salt concentration (KAc 5M) followed by high-speed centrifugation to separate separate the DNA (supernatant) and proteins (pellet). DNA is preciptated with isopropanol and washed with ethanol 70%. DNA was quanti fied with NanoDrop technology with a concentration of 114.7 ng/µL (260/280 ratio of 1.84, 260/230 ratio of 1.5). As Nanodrop could overestimate the DNA content, QuBit method, a more reliable method for DNA quanti fication was used and yielded a dosage of 10 ng/µL (total DNA quantity 12.62 µg). Notably, the sample exhibited a straightforward appearance, denoting its suitability for further investigation. Shotgun metagenomic and metabarcoding sequencing and raw data processing Long-read metagenomic sequencing Extracted soil DNA was sequenced on the PacBio platforms using the HiFi sequencing mode. One complete SMRT cell was run on both the Sequel II and Revio platforms, generating two datasets. For the Sequel II sequencing, high-molecular-weight DNA with average fragment size of ~7 Kb was used to prepare an unampli fied, non-multiplexed HiFi library with the Template Prep Kit v2.0, following the PacBio metagenomic shotgun protocol. Sequencing was carried out on a SMRT Cell 8M over 30 h. For the Revio sequencing, the HiFi library was prepared with ~8 Kb not sheared native DNA, using the Template Prep Kit 3.0 . The SMRTBell library was sequenced on a SMRTCell 25M over 24h with the Revio sequencing plate reagents. For both Sequel II and Revio datasets, adapter removal, read quality filtering (minimum Q20), and generation of circular consensus sequencing (CCS) reads were performed using dedicated PacBio pipelines. In total, Sequel II produced 2.9 million high-quality HiFi reads, while Revio sequencing yielded 8.9 million HiFi reads. 5 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint Short-read metagenomic sequencing DNA from the same soil sample underwent a second extraction and was sequenced at ultra-deep coverage on a full NextSeq 2000 Illumina P3 flow cell. This produced 982.5 million paired-end reads of 2x150bp. Adaptor trimming and quality filtering were performed using Fastp v.0.23.4 (20,21), resulting in 903,375,951 high-quality Illumina paired-end short reads. 16S Metabarcoding 16S rRNA gene amplicon sequencing data were generated from the same soil sample through targeted sequencing of the V3-V4 region of the 16S rRNA gene using primers 341F_UDI and 785r_udi on the Illumina NextSeq 2000 platform producing 2x300pb reads. Library preparation and sequencing were conducted with the protocol described by (22). Based on the selected primers, we expected to sequence approximately 444 bp of the 16S gene, with a substantial overlap between R1 and R2. Paired-end reads were merged using FLASH v0.23.4 (23) with flexible overlap settings (--min-overlap 10, --max-overlap 301) to reconstruct full-length amplicons, even with region size variability. Primer sequences were trimmed using Cutadapt v4.9 (24) with parameters optimised for mismatch handling (--error-rate 0.1, --match-read-wildcards, --overlap). Quality filtering was performed using Fastp v0.23.4 (21,25), with the following parameters: a Phred quality score threshold of 30 (--qualified_quality_phred 30), a minimum sequence length of 230 bp (--length_required 230), and an average quality of 30 (--average_qual). Additional trimming at the 5’ and 3’ ends was done using a sliding window approach (--cut_window_size 4, --cut_mean_quality 30) with low-quality bases removed (--cut_front, --cut_right). Adaptor trimming was disabled (--disable_adapter_trimming) to avoid interference with prior primer removal steps. Chimeric sequences were detected and removed using Vsearch v2.9.1 (26) in de novo mode (--uchime3_denovo). Bioinformatic analyses After data acquisition, we performed quality control on each dataset using FastQC software version 0.12.0 (27) with default parameters. The results were merged using MultiQC software version 1.25.1 (Ewels et al., 2016) in default mode. Assembly Paired-end short reads were assembled using MEGAHIT v1.2.9 (29) with the ‘--presets meta-large’ parameter. HiFi long reads were assembled using MetaMDBG v.0.3 with default parameters (30). Binning Sequence features, such as GC content, tetranucleotide frequency, and coverage as a proxy for abundance, were used for clustering contigs that are supposed to belong to the same organism in the bin, toward the reconstruction of MAGs. Both long and short reads were mapped on their respective assemblies using Minimap2 v.2.26 (31) with the parameters -x map-hifi -a --sam-hit-only --secondary=no for long reads, and BWA-MEM2 v2.2 and 6 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint BWA-MEM2 v.2.2 (Vasimuddin et al. 2019 ) with default parameters for short reads, respectively. Resulting alignments were used to compute contig coverage, which served as a proxy for estimating species abundance. Mapping outputs were processed using Samtools v.1.10 (33). Then, contigs were binned using MetaBAT2 v.2.15 (34) in default mode, and SemiBin v.1.5.0 (35) with parameters ‘--environment soil’ and’ --sequencing-type=long_read’ for long-read processing. The obtained results were merged and dereplicated using DAStool v.1.1.6 (36). Subsequently, gene prediction was performed using Prodigal v2.6.3 (37), and the completeness and contamination of bins were assessed using CheckM2 v1.0.1 (38) . Those analyses were performed using a previously published workflow (39), which was tuned to retain as MAGs and taxonomically assign bins with at least 50% completeness and less than 10% contamination, without any filter on the number of contigs (git). Bin and Metagenome-assembled Genome (MAG) evaluation Dereplicated bins were classi fied based on their completeness and contamination levels, as described in the MIMAGs guidelines (40). Bins with 100% completeness and ≤5% contamination were considered complete MAGs. Bins reaching the ≥90% completeness criterion were labelled high-quality (HQ) MAGs. Those comprising between ≤90% and ≥50% completeness were considered medium-quality (MQ) MAGs. Bins below these thresholds were considered low-quality (LQ) MAGs, while those over 10% of contamination were classi fied as ‘contaminated bins’. Metrics regarding assembled and non-assembled reads (e.g., k-mers, GC content, read size) and assembly and binning quality were obtained using Mapler v2.0.0 (41). Taxonomic classification To characterise the taxonomic composition of MAGs, contigs, and reads from each dataset, two complementary taxonomic assignment strategies were applied. For MAGs with more than 50% completeness, taxonomic classi fication was confirmed using GTDB-tk v.2.1.1 (42) with database version 214 providing robust information based on multigene phylogenetic placement (43). Additionally, both Illumina amplicon sequencing and PacBio metagenomic contigs and reads underwent a common taxonomic assignment strategy to facilitate their comparison. Reads (metabarcoding and LR) and contigs (LR and SR metagenomics) were assigned a taxon using MMseqs2 (44) “easy-search” module against the SILVA 138.2 SSURef NR99 database of curated 16S rRNA sequences as the reference. To ensure high-con fidence assignments, homology searches were performed using a stringent threshold of 97% sequence identity and a minimum alignment length of 100 bp (parameters --min-seq-id 0.97 and --min-aln-len 100 respectively). Coverage was set to improve sensitivity in detecting partial sequence matches of fragmented SR and LR, reads and contigs by including overlapping hits (--cov-mode 2). Alignment coverage was analysed using BEDtools (45), 7 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint calculating the proportion of each reference sequence covered by aligned reads. Species with 16S sequences from SILVA covered on more than 97% of their length were retained for further study. Eukaryotic hits against the SILVA database (46) were excluded to ensure consistency in taxonomic inference. The taxonomic lineage was used to increment counts in an abundance matrix for each taxonomic rank from species to domain. These matrices underwent downstream diversity analyses, which were performed using custom Jupyter Notebooks. Phylogenetic visualisations were generated using the Interactive Tree Of Life (iTOL) v7 (47).

HiFi Long-Read Sequencing delivers more comprehensive and continuous metagenome reconstruction. To compare shotgun metagenomic assembly performance of short-read and long-read sequencing technologies, we sequenced a soil sample with Illumina SR and PacBio HiFi LR (Sequel II and Revio) platforms and assembled the resulting datasets. Our data show that LR sequencing, particularly with the Revio platform, yields more comprehensive reconstructions of microbial genomes with enhanced continuity and more complete contigs. Ultra-deep sequencing with paired-end Illumina technology generated 903.3 million reads of 2x150 base pairs in length. In contrast, combining PacBio long-read sequencing using a full SMRTcell on Sequel II and another on Revio platforms yielded a total of 11.2 million reads, with an average length of 7.25 kb, and a maximum length of 40 kb (Table 1). The read length distribution of our metagenome sequencing aligns with DNA fragment size estimates for soil samples, mirroring findings from human gut microbiome studies (16). Long-read assemblies result in more comprehensive genome recovery Reads were assembled using technology-specific pipelines to reconstruct the genomes of soil microorganisms. A first observation is that, before assembly, raw HiFi read lengths substantially surpass the lengths of contigs assembled from short-reads, as previously observed in publicly available soil metagenomes, where the average size of SR contigs was less than 1kb (11). Our dataset con firms these observations, with SR contigs having an N50 of 620 bp and a largest contig size of 192 kb, whereas HiFi unassembled LR have an N50 of 7,618 bp and a longest read length of 40 kb. The assembly of PacBio long reads further increases this length di fference. Sequel II and Revio long reads were both assembled separately, and also pooled (sq+rv) as a third LR dataset, which was assembled to compare the e ffect of technologies and coverage on assembly quality. Revio generated three times more base pairs than Sequel II (Table 1), a trend that persisted at the assembly level, with Revio assemblies containing three times more total assembled base pairs. 8 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint Table 1 - Summary of Illumina Next-seq 2000 short reads and PacBio long reads from Sequel II (sq), Revio (rv) and pooled Sequel and Revio (sq+rv). Category Dataset Short reads Long reads-sq Long reads-rv Long reads-sq+rv Sequencing data information Number of Reads 903,3 M* (x2) 2,9M 8,2 M 11 ,2 M Total bp** 147,371,618,700 20,251,105,762 60,979,354,823 81,230,460,585 Average Read Length (bp) 150 (x2) 6,951 7,358 7,252 Maximum Read Length (bp) 29,991 40,482 40,482 N50 (bp) 7,618 7,969 7,881 Assembly Metrics Number of Contigs 31,071,620 493,072 1,208,875 1,545,396 Total bp 19,141,441,572 3,069,439,982 9,001,313,035 11,591,114,055 Maximum Length (bp) 185,456 1,508,276 5,668,623 6,140,418 N50 Contigs (bp) 620 7,636 9,365 9,553 % Unmapped reads*** 17.5 27.95 16.18 13.95 *’ M’ = Millions, “G”=Giga, **"bp" = base pairs,, ***properly paired Four contigs obtained from Sequel II reads were > 1 Mb, with the longest being 1.5Mb in length. In contrast, assembly of Revio reads yielded 139 contigs of at least 1 Mb, and 87 larger than the longest Sequel II contig. The longest contig obtained with Revio was 5.7 Mb long, demonstrating the capacity of this technology to deliver high-quality, long reads with higher coverage useful for resolving microbial genomes from complex environments. Although Illumina yielded nearly twice the total number of base pairs compared to PacBio at both the read and contig levels, the short-read assembly remained highly fragmented, with an N50 of only 620 bp and a maximum contig length of 185 kb (Table 1, Fig. 1). The Revio platform, by generating more high- fidelity long reads, enables the assembly of contigs longer than those generated not only from Illumina data but also from Sequel II data. The improvement in contigs lengths generated from long reads is expected to positively impact the downstream completeness of gene predictions. De novo gene prediction from long-read assemblies yielded approximately 18 million proteins classi fied as complete, representing 80% of all predictions, compared to fewer than 5 million (12% of all predicted proteins) from short-read contigs. Furthermore, the length distributions of predicted proteins are consistent with those reported by Never et al. (2019) for high-quality predicted proteomes (Supp Fig. 1 B). Indeed in this work, the authors showed that protein length distribution is a reliable indicator of genome assembly quality, with conserved protein 9 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint lengths typically mostly ranging from 236 to 353 amino acids (726 to 1059 base pairs) across different domains of the tree of life (archaea: 242 amino acids, bacteria: 270 amino acids, eukaryotes: 353 amino acids). This increased protein prediction accuracy is observed in both the LR contigs and the predictions from raw LR reads. Figure 1 - Contig size and protein length distribution (log scale) for short reads (blue), SequelII (purple) and Revio (pink) as well as pooled sq+rv (salmon) long reads. Although long reads generated by the Revio platform yield better assembly metrics and gene predictions than short reads, combining the two long-read datasets (Sequel II + Revio) further enhances assembly metrics. While the contig length is only slightly improved (Fig. 1), the number of contigs and total base pairs incorporated in contigs for the pooled sq+rv dataset have signi ficantly increased (Table 1). A comparison of k-mer compositions in the Sequel and Revio datasets further indicated that increasing sequencing depth reduced the number of unique k-mers and improved genome assembly (Fig. 2B). These results con firm the importance of sequencing depth in reconstructing genomes from complex environments. Unmapped reads indicate limitations in coverage and assembly, mainly in low-abundance species. Unmapped reads are those that fail to align with the assembly either because they contain sequencing errors or are not represented in any contig. Characterising unmapped reads is crucial for assessing sequencing completeness and identifying genomic regions (or even species) that are challenging to assemble. In metagenomics, those unmapped reads often originate from low-abundant populations, leading to undetected species or highly fragmented MAGs. 10 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint Here, mapping statistics revealed signi ficant di fferences between sequencing platforms. When Illumina short reads exhibited an unmapped-read rate of 17.5% on their associated contigs, PacBio LR demonstrated unmapped-read rates of 16 to 28% for contigs generated from Revio and Sequel II technologies respectively (Table 1, Fig. 2A). The pooled sq+rv long-read dataset (Sequel II + Revio) resulted in a lower proportion of unmapped reads (14%), suggesting that the higher number of generated contigs and incorporated basepairs increased the diversity of genomic regions captured by the assembly. Further comparisons of reads (Supp Fig. 2) highlighted that unassembled long reads (sq+rv) exhibited slightly shorter lengths than those participating in the assembly, with a median length of 6.3 and 7.0 kb, respectively; the distributions in the two groups di ffered significantly (Mann-Whitney, P < 2.2e -16, two-tailed). Unassembled reads also displayed a lower GC content than the other reads, with a median of 63.6% compared to a median of 65.8%, which is statistically signi ficant (Mann-Whitney, P < 2.2e-16, two-tailed). Additionally, their abundance (as calculated by the median of their k-mer occurrences within the full read dataset) was significantly lower, with a median of 1, compared to 3 for the assembled reads (Mann-Whitney, P < 2.2e-16 two-tailed). It is worth noting that, with an average length of 6.7 kb, the unassembled reads could still contain full-length gene sequences. These sequences could uncover genomic fragments and functions, possibly from low-abundance species still uncaptured in assembly-based analyses. Unmapped reads likely originate from low-abundance or rare microbial taxa not captured by the assembly, although some might also represent sequencing errors. These findings suggest that, due to sequencing depth limitations, long-read assemblies still miss a portion of the metagenomic diversity, resulting in a substantial number of species left unassembled. This underscores the importance of integrating complementary strategies to enhance the recovery of genomes from complex soil communities. Finally, to assess overlap between captured biodiversity across technologies, short reads were mapped on long-read assemblies. We observed 21% of SR still remained unmapped; this proportion increased to 62% under stringent mapping parameters (identity >90% and alignment length ≥130 bp), highlighting partial but incomplete overlap in captured biodiversity. Using short-read coverage improves binning of long-read contigs Assembling reads from a mixture of genomes and species remains a challenging task, particularly in complex environments such as soil, leading to fragmented assemblies (Fig. 1), especially for low-abundance populations. Binning partly addresses these pitfalls by clustering contigs expected to be part of the same genome, based on their similar molecular composition and abundance, as estimated from contig coverage via self-mapping (i.e., 11 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint mapping the reads used for assembly against the resulting contigs). We compared the number and completeness of genomes reconstructed after binning for both SR and LR datasets, considering two di fferent mapping strategies before binning: self SR-SR or LR-LR mapping and combined SR+LR mapping. Since the pooled sq+rv LR from both Revio and Sequel II technologies yielded the best assembly results in terms of contigs length distribution and low % of unmapped reads, we will consider only this dataset as representative of the HiFi LR technology from now on. Comparative analysis of genome bins from Illumina short-read and PacBio long-read assemblies obtained through self-mapping coverage We first considered binning the LR assembly using the mapping of the corresponding long reads to the contigs. Such binning generated more than twice the number of bins (2,587 vs 1,187) and nearly twice as many MAGs (252 vs 145) compared to binning the short-read contigs with short-read mapping (Table 2). This improvement was also re flected by longer contigs (up to 6.7 Mbp vs. 192 Kbp), greater bin lengths after dereplication, and a substantially higher N50 (487 Kbp vs. 9 Kbp), indicating improved contiguity (Table 2). Regarding MAGs, the LR assembly produced substantially more high- and medium-quality bins (75 HQ and 169 MQ vs. 16 HQ and 128 MQ for SR), as well as a greater number of low-quality MAGs (176 vs. 51). Notably, long-read assemblies also generated signi ficantly more complete MAGs (8 vs none), highlighting their potential for near-complete genome reconstruction, even in very complex microbiomes (Fig. 2C). However, these better assembly metrics came with an increase in contamination, as evidenced by a larger number of contaminated bins (>10% contamination estimates) in the long-read dataset (132 vs. 63 bins classi fied as contaminated). This pitfall is likely related to the insu fficient coverage information in LR bins, associated with limited sequencing depth (Table 1) and resulting in relatively homogeneous read abundance when mapped to contigs (Fig. 2A/C). Speci fically, LR self-mapping resulted in low coverage depths on contigs (mean: 3X; range: 0–310X) and on MAGs (mean: 6X; range: 0–63X), limiting the resolution of the binning process. In contrast, the depth of SR sequencing led to broader coverage estimation. Self-mapping short reads on their corresponding contigs yielded substantially higher coverage, with mean depths of 33X on SR-assembled contigs (range: 5–17,156X) and 40X on MAGs (range: 10–254X). Notably, more than half of Illumina reads aligned stringently to LR-assembled contigs (Fig. 2A), with coverage depths reaching up to 27,323X (for bins mean: 9X; for MAGs mean: 16X range: 1–126X). These results suggest that the limited depth of LR self-mapping undermines binning accuracy and results in increased contamination. Consequently, we 12 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint hypothesised that incorporating SR coverage into the LR binning process could improve bin quality by enhancing clustering resolution and reducing contamination. Table 2 - Summary of bins from Illumina and PacBio Sequencing obtained using self-coverage or short-read coverage for clustering of contigs. Binning Strategy Self-coverage SR-coverage Metric Processing Step Illumina Pcb_sequel_revio Pcb_sequel_revio Number of bins Total bins 1 187 2 587 2 518 Dereplicated 258 560 627 MAGs 145 252 313 Binned contigs (all bins) Dereplicated 146,518 83,327 113,232 MAGs 68,669 20,889 31,384 Binned base pairs (Mbp) Dereplicated 987 1 990 2 274 MAGs 548 995 1 196 Max. contig length (Kp) Dereplicated 192 6 697 6 697 MAGs 186 6 140 6 697 N50 (Kb) Dereplicated 9 487 429 MAGs 12 907 750 Comparative analysis of genome bins from PacBio long-read assemblies using self-mapping versus short-read coverage mapping We therefore designed a hybrid binning strategy, using coverage information from Illumina SR mapping on contigs assembled from pooled sq+rv PacBio LR and assessed its bene fit. Although the raw total number of produced bins was slightly lower (2,587 bins for self LR-self coverage vs. 2,518 for SR coverage), SR coverage generated more unique bins. Indeed, dereplication yielded 627 non-redundant bins using SR coverage, representing a more than 10% increase over the 560 bins obtained from self-coverage, which re flects improved resolution of overlapping or ambiguous contigs in the complex samples. 13 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint Clustering improved MAG reconstruction, with SR coverage enabling 313 reconstructed MAGs (+24%) vs. 252 from LR self-coverage (Table 2, Fig. 2C). Figure 2 - A) Average depth of short-reads contigs and long-reads contigs mapped against their respective reads, along with the average depth of LR contigs mapped using short reads (PacBio SRcov). B) Aligned bases percentage (length of reads segments aligned to the assembly divided by total reads length, grouped by bin quality).C) Number of bins and their quality classi fication. Short-read and long-read datasets were binned using self-mapping of their respective reads, while the long-read_SRcov dataset was derived from long-read binning using short-read data. Bin quality categories include Complete, High Quality (HQ), Medium Quality (MQ), Low Quality (LQ), and Contaminated (>10% contamination). Improvement is observed for MAGs, where an additional 30% of contigs participate in high-quality bins (LR: 20,889 vs. SR: 31,384). Therefore, SR coverage allowed for the incorporation of more total base pairs (2,274 Mbp vs. 1,990 Mbp with LR self-coverage) in bins, and within MAG sequence recovery (1,196 Mbp vs. 995 Mbp). However, bins obtained from SR coverage have a slightly reduced median contiguity (MAG N50 decreased from 907.6 Kbp to 750.9 Kbp). The number of complete MAGs remained stable (with eight bins) and the number of high-quality MAGs increased slightly from 75 to 80 bins. In parallel, the number of medium-quality MAGs increased from 169 to 225 bins, and low-quality MAGs increased (176 vs 198 bins). These results suggest that the higher coverage provided by short reads allowed more species to be represented as MAGs, although a substantial proportion remained fragmented. The number of contaminated bins decreased substantially with SR coverage (132 vs 116), suggesting better bin purity with the increase of coverage (Fig. 2A). If direct comparison of genetic diversity between bins is di fficult for assessing genome reconstruction quality in uncultivated species, metrics such as completeness, contamination, and MAG taxonomic assignment can aid in evaluation. The increase in contig recruitment promoted by SR coverage has a primary e ffect on the reconstruction of additional bins, with minimal impact on the completeness of bins already recovered through LR coverage (average completeness estimates of 62.71 and 63.05 for LR and SR coverage, respectively). In 14 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint contrast, this coverage allows for better discrimination by reducing contamination (average contamination estimates of 8.07 and 7.09 for LR and SR coverage, respectively). Taxonomic analysis of MAGs Taxonomic assignment was performed for each dataset on medium- to high-quality MAGs (≥50% completeness) using GTDB-Tk, which leverages multigene phylogenetic placement for robust genome-resolved classi fication. We further compared the MAGs obtained with each of the three protocols by exploring their taxonomic assignments. A total of 34 GTDB lineages were identi fied as being speci fic to the SR-covered LR MAGs dataset. They were not recovered in SR contigs nor LR self-coverage binning (Fig. 3). Recovered LR-SR MAGs spanned over 30 genera or families, of which 20 were also retrieved in Illumina MAGs, supporting the presence of these lineages in the original samples. Although potentially representing binning artefacts, the remaining 10 taxa speci fic to SR coverage may include additional lineages uncovered by improved SR-driven clustering. Only seven LR-self coverage genera were absent in MAGs produced with the SR coverage, including five nonetheless recovered in Illumina-based MAGs (Fig. 3A). Moreover, with SR coverage, some contigs previously clustered in the same bin were redistributed across di fferent bins. Indeed, the low signal from LR mapping originally resulted in the collapse of strain or species-level genomes into a unique bin. Phylogenetic analysis confirmed that LR contigs binned with SR yields the widest taxonomic spectrum (Fig. 3B), recovering previously unbinned clades, such as the bacterial genus SXQI01 from the Verrucomicrobiota phylum. Using SR coverage allowed for the identification of additional MAGs corresponding to 34 taxa, including 17 new phyla, 35 families, and three species. Moreover, the increased number of bins across bacterial phyla, such as Planctomycetota and Acidobacteriota, underscores the e ffectiveness of the hybrid binning strategy in recovering underrepresented lineages. This balance enables broader genome recovery, including low-abundance populations, and yields more taxonomically resolutive results. Despite the substantial reduction of the N50 compared to self-coverage binning, this new binning strategy ensures better grouping of contigs within the same phylum, and improves separation between divergent taxa. SR coverage resolution improvement was observed primarily through an increased number of unique and non-redundant bins, which mainly resulted from the addition of new reconstructed bins and the improvement of some bins that were previously retrieved with LR coverage. In some cases, the number of good-quality bins increased by salvaging contigs and including them in existing bins. SR coverage involved 113,232 contigs, whereas LR coverage allowed 15 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint the recruitment of only 83,327. Clustering capacity improvement resulted from enhanced coverage depth signal with SR, facilitating clustering for more contigs. A detailed analysis of mid-quality bins further illustrates the impact of SR coverage. For example, with LR self-coverage, bin metabat2.163 included 11 contigs, and exhibited a completeness of 93.16% and contamination of 2.98%. When SR coverage was used, 10 of its contigs were regrouped into bin metabat2.1044, with nearly identical metrics (completeness: 93.02%, contamination: 2.70%). Notably, one contig (ctg3929932_19x_l) was excluded from the SR-derived bin. This contig had a signi ficantly lower coverage (17X in LR, 62X in SR) compared to the mean of the SR bin (92X, standard deviation: 12), suggesting it may have been a coverage outlier. The removal of this contig was likely the cause of the reduced contamination e ffect associated with SR-coverage binning. In contrast, the LR bin had a broader coverage dispersion (mean: 27X, standard deviation: 10.1), likely causing erroneous contig inclusion. This example highlights how SR coverage re fines clustering by improving depth resolution and eliminating poorly supported contigs. The overall taxonomic assignment remained consistent between both bins, assigned to the same lineage within the Burkholderiales order (genus JAABQT01), with nearly identical ANI and tree placement results. Thus, a hybrid binning strategy with SR coverage enhances bin completeness and specificity, stabilising bin composition by filtering out ambiguous contigs, which leads to more taxonomically accurate genome reconstructions. 16 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint Figure 3 - A) Taxonomic distribution of metagenome-assembled genomes (MAGs) by phylum across three sequencing and binning strategies, where Illumina data are represented in blue, Self-mapped LR in orange and LR contigs binning with SR coverage in green. B) Phylogenetic tree of taxonomic annotations for MAGs, generated using GTDB-tk [8]. The 17 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint occurrence of each taxon across datasets is visualised as a heatmap. The interactive tree is available at: https://itol.embl.de/export/17616122570468971737508212. Hybrid binning captures a broader soil microbiota diversity To assign taxonomy to shotgun metagenomic data, we applied two complementary strategies: MAG-based classi fication using GTDB phylogenomic approach as previously described and in-silico metabarcoding. In-silico metabarcoding refers to identifying and assigning taxonomy to 16S rRNA gene fragments directly extracted from metagenomic reads or contigs, using the SILVA database as a reference(46). While both methods aim to characterise microbial diversity, MAG-based classi fication provides more accurate resolution at the genome level, whereas in-silico metabarcoding enables direct comparison with taxa obtained by mapping reads (LR and SR) on 16S rRNA reference database. This latter step enables us to evaluate the consistency between taxonomic pro files inferred from metagenomic sequencing and those obtained through standard amplicon-based approaches, which serve as a widely accepted reference in microbial ecology. Additionally, metabarcoding profiles were used as an external benchmark to evaluate the performance of the various binning strategies employed in terms of part of the estimated biodiversity captured in the assemblies. Long-read metagenomics achieves the highest taxonomic overlap with metabarcoding and captures broader microbial diversity While LR sequencing improves assembly contiguity and completeness, SR sequencing provides more informative coverage signals, which are critical for accurate binning. When studying soil samples, a key challenge is to determine how e ffectively sequencing and binning strategies capture the microbial diversity present in the environment. To assess this, we compared in silico metabarcoding obtained from WGS metagenomic data with real metabarcoding data obtained from the same uncharacterised soil sample. The metabarcoding approach, targeting the V3-V4 region of the 16S rRNA gene, served as a

for estimation of the microbial diversity. Taxonomic comparisons were performed based on matches of 16S sequences between metabarcoding data and contigs or reads from the di fferent metagenomic datasets (short reads, long reads, and their corresponding contigs). Given the limited reliability of abundance estimation using metagenomics, these analyses focused solely on the presence/absence of taxa (Fig. 4A). Across all methods, both bacterial and archaeal lineages were consistently identi fied. However, signi ficant di fferences emerged when comparing richness at various taxonomic ranks: 18 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint (i) Short-read metagenomics showed the lowest diversity across most taxonomic levels, except for species (Fig. 4F), where an unusually large number of taxa (7,897 vs. 3,593 in metabarcoding) were detected. This in flation is likely due to low speci ficity of SR mapping to 16S libraries due to their short length, resulting in overestimation. (ii) Long-read data consistently showed the highest richness among metagenomic datasets and captured taxonomic pro files closer to metabarcoding. At the phylum level, LR assemblies and raw long reads identi fied 24 to 29 phyla compared to 33 identi fied in metabarcoding. (iii) At higher resolution (e.g. genus level), long-read assemblies recovered ~33% of genera detected by metabarcoding. In contrast, raw long reads recovered nearly two-thirds of the genera, confirming that the reads forming contigs represent a subset of the total diversity (Table 2, Fig. 4A). To better visualise taxonomic overlaps, an Upset plot at the genus level (Fig. 4B) illustrates shared and unique genera across the five datasets (short reads, SR assembly, long reads, LR assembly, and metabarcoding). Metabarcoding recovered the highest number of genera (784), followed by the raw LR dataset. However, only a small subset of genera was shared across all approaches, re flecting detection thresholds and method-speci fic biases due to the use of databases (i.e., GTDB/SILVA) constructed with different lineage structures. Notably, over 200 genera were uniquely detected by metabarcoding, indicating that a fraction of the microbial community, likely corresponding to rare taxa, remains elusive to metagenomics, even with long reads. The taxonomic pro file provided by metabarcoding sequencing was used to assess the taxonomic reliability of reconstructed genomes. We compared GTDB annotations of MAGs with 16S-based annotations of their contigs using the SILVA database (Supp Fig. 1). Our

was to assess whether LR-based and SR-based coverage at the binning step resulted in di fferences in MAG quality. Across all taxonomic ranks, annotations from bins generated using SR coverage were consistently more concordant with SILVA annotations than those from LR coverage (Fig. 3 supplementary data). At the domain level, annotations were consistently assessed across all methods. As the level of resolution increased (from phylum to species), discrepancies became more frequent, due in part to di fferences in database content and lineage structure. Despite this, SR-informed binning always yielded a higher number of MAGs with consistent taxonomic placement. Out of 313 bins, 200 MAGs could be compared (due to the absence of 16S in ~1/3 of high-quality bins), con firming that SR coverage significantly improves both detection and phylogenetic resolution. 19 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint Figure 4 - A) Comparison of taxonomic richness across sequencing strategies and taxonomic ranks. Bar plots show the number of distinct taxa, richness, identi fied at six taxonomic levels (Domain, Phylum, Order, Family, Genus, and Species) using five sequencing approaches: MG short-reads assembly, MG short-reads, MG long-reads assembly, MG long-reads, and MB amplicons, where MG refers to data produced by shotgun metagenomic sequencing and MB: refers to metabarcoding data of 16S V3-V3 region. B) Taxonomic overlap at the genus level across five datasets: MG short-reads assembly, MG short-reads, MG long-reads assembly, MG long-reads, and MB amplicons, where MG refers to data produced by whole genome metagenomic sequencing and MB refers to metabarcoding data of the 16S V3-V3 region. Horizontal bars on the left of the upset plot represent the total number of genera detected by each method. Vertical bars above indicate the size of each intersection set, 20 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint corresponding to speci fic combinations of techniques shown by filled circles below. Overall, the plot illustrates both shared and method-speci fic taxonomic recovery at the genus level (N = 914).ii. Taxonomic reference con firmed the annotation accuracy improvement through short-read coverage-informed binning LR metagenomic microbial detection correlates strongly with taxon abundance To assess the relationship between microbial abundance and capture e fficiency in shotgun metagenomics, we used metabarcoding relative abundance information to compare genera detected and undetected in the LR assembly. Figure 5A illustrates the distributions of relative abundances. Genera recovered through metagenomics exhibit signi ficantly higher abundance in the metabarcoding dataset (Wilcoxon test, p = 5.5 × 10⁻⁸⁰), suggesting a strong positive correlation between microbial abundance and detection likelihood. Consequently, metagenomic capture and reconstruction is biased towards dominant taxa. Consistently, the comparison of community pro files across amplicons, bins, and MAGs shows that while dominant phyla such as Actinomycetota, Gemmatimonadota, and Planctomycetota are robustly recovered across approaches, low-abundance phyla (e.g., Chloro flexota, Bacillota) are better represented in amplicon data but tend to be underrepresented in MAGs (Fig. 5B). However, integrating short-read SR binning enhances sensitivity, enabling the recovery of several lower-abundance taxa. This approach e ffectively reduces the detection gap between metagenomics MAGs-identified taxa and metabarcoding profiles. Figure 5 - A) Distribution of relative abundances for genera assessed by metabarcoding analysis that were either detected or not detected by PacBio combined assembly (Sequel + Revio). Density plots show the relative abundance of genera, classi fied as either detected (green) or not detected (blue), on a log10 scale using long-reads. Genera detected by metagenomics (MG) exhibit signi ficantly higher abundances compared to those not detected, with a highly signi ficant di fference (Wilcoxon test, p = 6 × 10⁻⁸⁰). B) Relative 21 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint abundance of phyla among reconstructed bins and MAGs and in the 16S rRNA gene amplicon sequencing dataset (“amplicons”). Phyla accounting for less than 5% of relative abundance were merged into “Others”.

High-fidelity long-read sequencing sets a new standard for genome-resolved soil microbiome analysis Soil represents one of the most microbially diverse ecosystems on Earth, yet remains vastly under-characterised at the genomic level. Amplicon sequencing and short-read metagenomics have provided substantial insight into microbial composition and functional potential (1,2); however, a genome-resolved understanding has lagged behind methodological limitations. Short-read assemblies typically su ffer from high fragmentation, with contigs rarely exceeding 1 kb (11), and struggle to resolve closely related taxa or repetitive genomic regions (13). These limitations hinder the recovery of complete MAGs, obscuring functional inference and ecological interpretation. Recent advances in long-read technologies, particularly PacBio HiFi sequencing, are transforming this landscape. HiFi reads o ffer both high accuracy and substantial length (6–20 kb), enabling more contiguous assemblies and more accurate gene recovery. Studies on gut microbiomes have demonstrated their ability to yield longer contigs, higher N50 values, and more complete metagenome-assembled genomes (MAGs) at lower sequencing depths than conventional approaches (16,48). However, the extent to which these bene fits apply to highly complex environments, such as soil, remained unclear. Here, we demonstrate that PacBio HiFi sequencing signi ficantly enhances the recovery of MAGs from soil, even at moderate depth. While only 313 bins met the criteria for high-quality metagenome-assembled genomes (MAGs), a total of 2,518 bins were recovered, likely representing distinct species or strains. This highlights the power of long-read technologies in retrieving genomic content that would otherwise be inaccessible through short-read sequencing alone. Assemblies constructed from HiFi data achieved substantially higher contiguity and completeness compared to those built from Illumina sequencing alone, even with very high sequencing depth. HiFi-based MAGs also contained more complete protein-coding sequences, consistent with recent observations that HiFi reads can recover full-length genes and operons even from complex samples (39). This enables more precise taxonomic and functional annotations, providing a robust foundation for studying microbe–microbe and microbe–environment interactions in soil. 22 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint Hybrid binning with short-read coverage further improves MAG resolution and specificity Despite these improvements, long-read-based binning remains limited by low coverage depth and limited abundance signal, particularly for rare taxa. In our study, the average coverage of long-read contigs was insufficient to support accurate abundance-based binning, resulting in some contamination and collapsed bins. To overcome this limitation, we incorporated high-coverage Illumina data to compute coverage pro files for HiFi-assembled contigs. This hybrid binning strategy signi ficantly improved bin quality, as evidenced by reduced contamination rates, increased dereplication, and a 24% increase in MAG yield. Additionally, it enabled the resolution of additional MAGs that included low-abundance and taxonomically underrepresented lineages, which were previously inaccessible through long-read self-coverage or SR-only binning. These results are consistent with previous hybrid studies in host-associated microbiomes (14), and con firm that short-read-supported binning of long-read contigs is an e ffective and scalable strategy in more complex and less-characterised environments. Significantly, this approach increased taxonomic resolution across multiple clades, including underrepresented phyla such as Acidobacteriota and Planctomycetota. We recovered 34 genera not observed in any other binning strategy, many of which had no close matches in the current reference databases. This highlights the utility of short-read-enhanced binning not only for improving MAG quality but also for broadening the accessible phylogenetic space in environmental metagenomics. With this work, we provide a rich metagenomic dataset from a soil sample that can serve as a benchmark and calibration for future sequencing e fforts on such complex microbiomes. A total of 81 Gbp of PacBio HiFi long-read sequences (61 Gbp from Revio and 20 Gbp from Sequel II) is made available, along with 147 Gbp of Illumina short reads from the same soil sample. We also produced 16S rRNA gene amplicon sequencing data, which can be used as a measure of the microbial ecosystem’s diversity. Datasets that integrate multiple sequencing platforms, including HiFi long-read data, remain scarce (49–51). More generally, complex ecosystems (Richy et al., 2024) are still undercharacterized compared to the extensively studied gut microbiome (39).

and priorities for future soil genome reconstruction Although we demonstrate that HiFi sequencing and hybrid binning improve genome reconstruction, the number of recovered MAGs remains signi ficantly lower than the taxonomic richness estimated by amplicon-based pro filing. While this discrepancy has been reported in other environments (12,52), it is likely to be more pronounced in soil probably 23 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint because of their extreme heterogeneity and the prevalence of low-abundance and refractory taxa. One major limitation is the small length of extracted DNA from soil, which rarely exceeds 6 kb, far below the 30–40 kb routinely recovered from human gut or aquatic samples (53). This restricts the full potential of HiFi or other longer-read technologies. Enhancing DNA extraction protocols for soil, by minimising shearing and improving inhibitor removal, should be a top priority. Improved fragment length distributions would increase assembly contiguity and support full-length repeat resolution, enabling higher-quality MAGs from more taxa. A second constraint is the incomplete nature of reference genome databases for soil microbes. Tools such as CheckM2, which are e ffective in human and model systems, become less reliable when reference phylogenies are sparse or uneven. This hampers con fidence in MAG quality assessments and limits downstream functional and evolutionary analyses. Systematic e fforts to catalogue and curate soil microbial genomes, including poorly characterised lineages, are essential for supporting the next generation of genome-resolved environmental studies (10). Our results demonstrate that HiFi LR sequencing, particularly when combined with short-read coverage information, is a powerful approach for enhancing MAG reconstruction in complex and under-characterised environments, such as soil. This hybrid strategy balances contiguity and coverage, enabling better binning performance and deeper ecological insight than either approach alone. It reduced the resolution gap between short-read metagenomics and amplicon pro filing, allowing the recovery of genomes from previously inaccessible taxa through assembly-based approaches. Looking forward, scaling HiFi sequencing through multiple Revio runs, optimising long-fragment extraction protocols, and expanding environmental genome catalogues will be essential to harness the full power of long-read metagenomics. These improvements will provide the foundation for more comprehensive and functionally resolved characterisations of microbial communities across terrestrial ecosystems, informing models of nutrient cycling, soil health, and climate resilience.

Collectively, our results demonstrate that integrating HiFi long-read sequencing with short-read coverage for binning markedly enhances the recovery, resolution, and diversity of reconstructed microbial genomes from a complex soil environment. This hybrid strategy bridges the gap between the microbial diversity estimated using amplicon-based community 24 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint profiling and that captured by shotgun metagenomics assembly. HiFi LR sequencing benefits from SR coverage with hybrid LR-SR binning. Improvement in MAG reconstruction facilitates access to previously undetected microbial lineages, enabling more comprehensive taxonomic and functional characterisation of soil microbiomes. Our data highlight the strong potential of PacBio HiFi long-read sequencing, combined with short-read coverage for binning, to achieve more accurate reconstruction of soil microbial diversity. Although long-read sequencing o ffers superior contiguity and completeness for MAG reconstruction, the incorporation of short-read coverage signi ficantly improves binning accuracy. This approach underscores the potential of leveraging the vast amount of publicly available Illumina metagenomic datasets, where complementing existing SR resources with additional PacBio HiFi sequencing can maximise contiguity and accuracy without discarding previous e fforts. By providing robust depth signals, this hybrid strategy enhances the discrimination of closely related taxa. It facilitates the recovery of a broader range of genomes, even within high-complexity microbiomes such as soil. However, our

also emphasise the critical importance of sequencing depth in these ecosystems. Indeed, shallow coverage remains a limiting factor for both assembly quality and genome recovery especially for rare taxa. For future studies targeting new complex environments such as some soils, we recommend using multiple Revio SMRT Cells to increase sequencing depth and maximise the retrieval of genomes, including those from low-abundance species. However, the associated cost can remain prohibitive. In conclusion, the integration of HiFi long- and short-read technologies represents a robust, and scalable strategy for achieving genome-resolved insights into complex microbial ecosystems. This strategy leverages long reads for contiguity and short reads for depth, illustrating how recent LR technologies amplify rather than replace the value of previous SR e fforts. Continued improvements in sequencing technology, library preparation optimized for longer reads, and reference catalogues will be essential for unlocking the full potential of soil metagenomics towards describing the functions and interactions of these complex yet essential microbial communities. Supplementary Information Availability of data and material All custom scripts and the datasets generated and analysed during the current study are available in the Recherche Data Gouv, in Soil Metagenome Binning repository, 25 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint Competing interests CF is an Associate Editor in BMC Microbiome but she was not involved in the peer-review process of this manuscript. Funding This work was supported by the French National Research Agency (ANR) and France2030 under the MISTIC project (ANR-22-PEAE-0011) and the Initiative of Excellence Université Côte d’Azur (ANR-15-IDEX-01) through the Maison de la Modélisation, Simulation et Interaction. This research was also supported by the Plant Health and Environment Department of INRAE through the MetaNema project. Authors' contributions C.B., N.M., S.M., A.F., M.B.-B., C.L., J.-M.F., C.F., and E.G.J.D. conceived and designed the study. C.B., N.M., A.P. developed and implemented the software and methodological approaches. C.B., N.M., and A.F. carried out the analyses and formal data interpretation. C.B. curated the data. C.B., N.M., C.F., and É.G.J.D. prepared the original draft of the manuscript. C.L., R.V., C.F., and É.G.J.D. supervised the project. M.B.-B., A.F., P.A., D.J.S., C.F., and É.G.J.D. acquired funding. All authors reviewed and approved the fi nal manuscript.

We sincerely thank the Sophia Agrobiotech bioinformatics platform for its valuable support and expertise in data processing, as well as for providing optimised computing hardware and software (ISC PlantBIOs, https://doi.org/10.15454/qyey-ar89). We also thank Genotoul for providing high computing resources and fast reactivity during access openings (Bioinfo Genotoul, https://doi.org/10.15454/1.5572369328961167E12). We acknowledge the GenOuest bioinformatics core facility (https://www.genouest.org) for providing the computing infrastructure. Some experiments presented in this paper were also carried out using the PlaFRIM experimental testbed, supported by Inria, CNRS (LABRI and IMB), Université de Bordeaux, Bordeaux INP and Conseil Régional d’Aquitaine (see https://www.plafrim.fr). Part of the experiments (short read metagenomic sequencing, 16S rRNA gene sequencing) were performed at the PGTB (https://doi.org/10.15454/1.5572396583599417E12). We are grateful to the Genosol, Gentyane, Get-Plage and PGTB platforms for their signi ficant contributions, including soil DNA extraction and high-throughput sequencing, which have been essential for the progress of this work. We are grateful to all collaborators and technical 26 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint staff who have contributed their expertise and assistance throughout this research. This work benefited from the use of the GenoSol-CRB and GenoSol-LADM services of the GenoSol platform (https://doi.org/10.15454/L7QN45), hosted by UMR 1347 Agroecology at the INRAE Bourgogne–Franche-Comté center. The platform is supported by a grant from the French State through the National Research Agency (ANR) under the “Investments for the Future” program (ANR-11-INBS-0001) and “France 2030” (ANR-24-INBS-0001, AnaEE-France). Additional funding was provided by GIS-IBISA, the Bourgogne–Franche-Comté Regional Council, and INRAE. GenoSol-CRB is part of BRC4Env (https://doi.org/10.15454/TRBJTB), while GenoSol-LADM is integrated into the AnaEE-France research infrastructure. We sincerely thank Isabelle Kupin for her assistance in depositing the datasets.

1. Hug LA, Baker BJ, Anantharaman K, Brown CT, Probst AJ, Castelle CJ, et al. A new view of the tree of life. Nat Microbiol [Internet]. 11 avr 2016 [cité 30 janv 2025];1(5):16048. Disponible sur: https://www.nature.com/articles/nmicrobiol201648 2. Thompson LR, Sanders JG, McDonald D, Amir A, Ladau J, Locey KJ, et al. A communal catalogue reveals Earth’s multiscale microbial diversity. Nature [Internet]. 23 nov 2017 [cité 30 janv 2025];551(7681):457-63. Disponible sur: https://www.nature.com/articles/nature24621 3. Torsvik V, Øvreås L. Microbial diversity and function in soil: from genes to ecosystems. Curr Opin Microbiol [Internet]. juin 2002 [cité 8 sept 2020];5(3):240-5. Disponible sur: https://linkinghub.elsevier.com/retrieve/pii/S1369527402003247 4. Delgado-Baquerizo M, Oliverio AM, Brewer TE, Benavent-González A, Eldridge DJ, Bardgett RD, et al. A global atlas of the dominant bacteria found in soil. Science [Internet]. 19 janv 2018 [cité 21 juill 2025];359(6373):320-5. Disponible sur: https://www.science.org/doi/10.1126/science.aap9516 5. Sokol NW, Slessarev E, Marschmann GL, Nicolas A, Blazewicz SJ, Brodie EL, et al. Life and death in the soil microbiome: how ecological processes influence biogeochemistry. Nat Rev Microbiol [Internet]. juill 2022 [cité 9 févr 2025];20(7):415-30. Disponible sur: https://www.nature.com/articles/s41579-022-00695-z 6. Hartmann M, Six J. Soil structure and microbiome functions in agroecosystems. Nat Rev Earth Environ [Internet]. 22 nov 2022 [cité 9 févr 2025];4(1):4-18. Disponible sur: https://www.nature.com/articles/s43017-022-00366-w 7. Knight CG, Nicolitch O, Griffiths RI, Goodall T, Jones B, Weser C, et al. Soil microbiomes show consistent and predictable responses to extreme events. Nature [Internet]. 19 déc 27 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint 2024 [cité 9 févr 2025];636(8043):690-6. Disponible sur: https://www.nature.com/articles/s41586-024-08185-3 8. Li S, Lian WH, Han JR, Ali M, Lin ZL, Liu YH, et al. Capturing the microbial dark matter in desert soils using culturomics-based metagenomics and high-resolution analysis. Npj Biofilms Microbiomes [Internet]. 22 sept 2023 [cité 9 févr 2025];9(1):67. Disponible sur: https://www.nature.com/articles/s41522-023-00439-8 9. Vogel TM, Hirsch PR, Simonet P, Jansson JK, Tiedje JM, Van Elsas JD, et al. Advantages of the metagenomic approach for soil exploration: reply from Vogel et al. Nat Rev Microbiol [Internet]. oct 2009 [cité 9 févr 2025];7(10):756-7. Disponible sur: https://www.nature.com/articles/nrmicro2119-c3 10. Parks DH, Rinke C, Chuvochina M, Chaumeil PA, Woodcroft BJ, Evans PN, et al. Recovery of nearly 8,000 metagenome-assembled genomes substantially expands the tree of life. Nat Microbiol [Internet]. nov 2017 [cité 30 mai 2019];2(11):1533-42. Disponible sur: http://www.nature.com/articles/s41564-017-0012-7 11. Belliardo C, Koutsovoulos GD, Rancurel C, Clément M, Lipuma J, Bailly-Bechet M, et al. Improvement of eukaryotic protein predictions from soil metagenomes. Sci Data [Internet]. 16 juin 2022 [cité 29 janv 2025];9(1):311. Disponible sur: https://www.nature.com/articles/s41597-022-01420-4 12. Ji P, Zhang Y, Wang J, Zhao F. MetaSort untangles metagenome assembly by reducing microbial community complexity. Nat Commun [Internet]. 23 janv 2017 [cité 9 févr 2025];8(1):14306. Disponible sur: https://www.nature.com/articles/ncomms14306 13. Treangen TJ, Salzberg SL. Repetitive DNA and next-generation sequencing: computational challenges and solutions. Nat Rev Genet [Internet]. janv 2012 [cité 9 févr 2025];13(1):36-46. Disponible sur: https://www.nature.com/articles/nrg3117 14. Bertrand D, Shaw J, Kalathiyappan M, Ng AHQ, Kumar MS, Li C, et al. Hybrid metagenomic assembly enables high-resolution analysis of resistance determinants and mobile elements in human microbiomes. Nat Biotechnol [Internet]. août 2019 [cité 9 févr 2025];37(8):937-44. Disponible sur: https://www.nature.com/articles/s41587-019-0191-2 15. Wang Y, Zhao Y, Bollas A, Wang Y, Au KF. Nanopore sequencing technology, bioinformatics and applications. Nat Biotechnol [Internet]. nov 2021 [cité 9 févr 2025];39(11):1348-65. Disponible sur: https://www.nature.com/articles/s41587-021-01108-x 16. Plaza Oñate F, Roume H, Almeida M. Recovery of Metagenome-Assembled Genomes from a Human Fecal Sample with Pacific Biosciences High-Fidelity Sequencing. Gill SR, éditeur. Microbiol Resour Announc [Internet]. 16 juin 2022 [cité 30 janv 2025];11(6):e00250-22. Disponible sur: https://journals.asm.org/doi/10.1128/mra.00250-22 28 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint 17. Koren S, Walenz BP, Berlin K, Miller JR, Bergman NH, Phillippy AM. Canu: scalable and accurate long-read assembly via adaptivek-mer weighting and repeat separation. Genome Res [Internet]. mai 2017 [cité 18 juill 2025];27(5):722-36. Disponible sur: http://genome.cshlp.org/lookup/doi/10.1101/gr.215087.116 18. Wenger AM, Peluso P, Rowell WJ, Chang PC, Hall RJ, Concepcion GT, et al. Accurate circular consensus long-read sequencing improves variant detection and assembly of a human genome. Nat Biotechnol [Internet]. oct 2019 [cité 18 juill 2025];37(10):1155-62. Disponible sur: https://www.nature.com/articles/s41587-019-0217-9 19. Plassart P, Terrat S, Thomson B, Griffiths R, Dequiedt S, Lelievre M, et al. Evaluation of the ISO standard 11063 DNA extraction procedure for assessing soil microbial abundance and community structure. PloS One. 2012;7(9):e44279. 20. Chen S. Ultrafast one-pass FASTQ data preprocessing, quality control, and deduplication using fastp. iMeta [Internet]. mai 2023 [cité 17 juill 2025];2(2). Disponible sur: https://onlinelibrary.wiley.com/doi/10.1002/imt2.107 21. Chen S, Zhou Y, Chen Y, Gu J. fastp: an ultra-fast all-in-one FASTQ preprocessor. Bioinformatics [Internet]. 1 sept 2018 [cité 17 juill 2025];34(17):i884-90. Disponible sur: https://academic.oup.com/bioinformatics/article/34/17/i884/5093234 22. Abellan-Schneyder I, Matchado MS, Reitmeier S, Sommer A, Sewald Z, Baumbach J, et al. Primer, Pipelines, Parameters: Issues in 16S rRNA Gene Sequencing. Tringe SG, éditeur. mSphere [Internet]. 24 févr 2021 [cité 5 mai 2025];6(1):e01202-20. Disponible sur: https://journals.asm.org/doi/10.1128/mSphere.01202-20 23. Magoč T, Salzberg SL. FLASH: fast length adjustment of short reads to improve genome assemblies. Bioinformatics [Internet]. 1 nov 2011 [cité 17 juill 2025];27(21):2957-63. Disponible sur: https://academic.oup.com/bioinformatics/article/27/21/2957/217265 24. Martin M. Cutadapt removes adapter sequences from high-throughput sequencing reads. EMBnet.journal [Internet]. 2 mai 2011 [cité 18 juill 2025];17(1):10. Disponible sur: http://journal.embnet.org/index.php/embnetjournal/article/view/200 25. Chen S. Ultrafast one-pass FASTQ data preprocessing, quality control, and deduplication using fastp. iMeta [Internet]. mai 2023 [cité 18 juill 2025];2(2). Disponible sur: https://onlinelibrary.wiley.com/doi/10.1002/imt2.107 26. Rognes T, Flouri T, Nichols B, Quince C, Mahé F. VSEARCH: a versatile open source tool for metagenomics. PeerJ [Internet]. 18 oct 2016 [cité 18 juill 2025];4:e2584. Disponible sur: https://peerj.com/articles/2584 27. Andrews S. FastQC: A quality control tool for high throughput sequence data. 2010; Disponible sur: http://www.bioinformatics.babraham.ac.uk/projects/fastqc 29 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint 28. Ewels P, Magnusson M, Lundin S, Käller M. MultiQC: summarize analysis results for multiple tools and samples in a single report. Bioinformatics [Internet]. 1 oct 2016 [cité 19 juin 2025];32(19):3047-8. Disponible sur: https://academic.oup.com/bioinformatics/article/32/19/3047/2196507 29. Li D, Liu CM, Luo R, Sadakane K, Lam TW. MEGAHIT: an ultra-fast single-node solution for large and complex metagenomics assembly via succinct de Bruijn graph. Bioinformatics [Internet]. 15 mai 2015 [cité 30 janv 2025];31(10):1674-6. Disponible sur: https://academic.oup.com/bioinformatics/article/31/10/1674/177884 30. Benoit G, Raguideau S, James R, Phillippy AM, Chikhi R, Quince C. High-quality metagenome assembly from long accurate reads with metaMDBG. Nat Biotechnol [Internet]. sept 2024 [cité 30 janv 2025];42(9):1378-83. Disponible sur: https://www.nature.com/articles/s41587-023-01983-6 31. Li H. Minimap2: pairwise alignment for nucleotide sequences. Birol I, éditeur. Bioinformatics [Internet]. 15 sept 2018 [cité 9 févr 2025];34(18):3094-100. Disponible sur: https://academic.oup.com/bioinformatics/article/34/18/3094/4994778 32. Vasimuddin Md, Misra S, Li H, Aluru S. Efficient Architecture-Aware Acceleration of BWA-MEM for Multicore Systems. In: 2019 IEEE International Parallel and Distributed Processing Symposium (IPDPS) [Internet]. Rio de Janeiro, Brazil: IEEE; 2019 [cité 9 févr 2025]. p. 314-24. Disponible sur: https://ieeexplore.ieee.org/document/8820962/ 33. Danecek P, Bonfield JK, Liddle J, Marshall J, Ohan V, Pollard MO, et al. Twelve years of SAMtools and BCFtools. GigaScience [Internet]. 29 janv 2021 [cité 9 févr 2025];10(2):giab008. Disponible sur: https://academic.oup.com/gigascience/article/doi/10.1093/gigascience/giab008/6137722 34. Kang DD, Li F, Kirton E, Thomas A, Egan R, An H, et al. MetaBAT 2: an adaptive binning algorithm for robust and efficient genome reconstruction from metagenome assemblies. PeerJ [Internet]. 26 juill 2019 [cité 9 févr 2025];7:e7359. Disponible sur: https://peerj.com/articles/7359 35. Pan S, Zhao XM, Coelho LP. SemiBin2: self-supervised contrastive learning leads to better MAGs for short- and long-read sequencing. Bioinformatics [Internet]. 30 juin 2023 [cité 9 févr 2025];39(Supplement_1):i21-9. Disponible sur: https://academic.oup.com/bioinformatics/article/39/Supplement_1/i21/7210480 36. Sieber CMK, Probst AJ, Sharrar A, Thomas BC, Hess M, Tringe SG, et al. Recovery of genomes from metagenomes via a dereplication, aggregation and scoring strategy. Nat Microbiol [Internet]. juill 2018 [cité 19 nov 2021];3(7):836-43. Disponible sur: http://www.nature.com/articles/s41564-018-0171-1 37. Hyatt D, Chen GL, LoCascio PF, Land ML, Larimer FW, Hauser LJ. Prodigal: 30 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint prokaryotic gene recognition and translation initiation site identification. BMC Bioinformatics [Internet]. déc 2010 [cité 21 juill 2021];11(1):119. Disponible sur: https://bmcbioinformatics.biomedcentral.com/articles/10.1186/1471-2105-11-119 38. Chklovski A, Parks DH, Woodcroft BJ, Tyson GW. CheckM2: a rapid, scalable and accurate tool for assessing microbial genome quality using machine learning. Nat

[Internet]. août 2023 [cité 9 févr 2025];20(8):1203-12. Disponible sur: https://www.nature.com/articles/s41592-023-01940-w 39. Portik DM, Feng X, Benoit G, Nasko DJ, Auch B, Bryson SJ, et al. Highly accurate metagenome-assembled genomes from human gut microbiota using long-read assembly, binning, and consolidation methods [Internet]. Microbiology; 2024 [cité 9 févr 2025]. Disponible sur: http://biorxiv.org/lookup/doi/10.1101/2024.05.10.593587 40. Bowers RM, Kyrpides NC, Stepanauskas R, Harmon-Smith M, Doud D, Reddy TBK, et al. Minimum information about a single amplified genome (MISAG) and a metagenome-assembled genome (MIMAG) of bacteria and archaea. Nat Biotechnol [Internet]. août 2017 [cité 30 janv 2025];35(8):725-31. Disponible sur: https://www.nature.com/articles/nbt.3893 41. Maurice N, Lemaitre C, Vicedomini R, Frioux C. Mapler: a pipeline for assessing assembly quality in taxonomically rich metagenomes sequenced with HiFi reads. Alkan C, éditeur. Bioinformatics [Internet]. 2 juin 2025 [cité 22 juill 2025];41(6). Disponible sur: https://academic.oup.com/bioinformatics/article/doi/10.1093/bioinformatics/btaf334/8157 874 42. Chaumeil PA, Mussig AJ, Hugenholtz P, Parks DH. GTDB-Tk: a toolkit to classify genomes with the Genome Taxonomy Database. Bioinformatics [Internet]. 1 mars 2020 [cité 27 juill 2020];36(6):1925-7. Disponible sur: https://academic.oup.com/bioinformatics/article/36/6/1925/5626182 43. Matsen FA, Kodner RB, Armbrust EV. pplacer: linear time maximum-likelihood and Bayesian phylogenetic placement of sequences onto a fixed reference tree. BMC Bioinformatics [Internet]. déc 2010 [cité 19 juin 2025];11(1):538. Disponible sur: https://bmcbioinformatics.biomedcentral.com/articles/10.1186/1471-2105-11-538 44. Steinegger M, Söding J. MMseqs2 enables sensitive protein sequence searching for the analysis of massive data sets. Nat Biotechnol [Internet]. nov 2017 [cité 30 janv 2025];35(11):1026-8. Disponible sur: https://www.nature.com/articles/nbt.3988 45. Quinlan AR, Hall IM. BEDTools: a flexible suite of utilities for comparing genomic features. Bioinformatics [Internet]. 15 mars 2010 [cité 18 juill 2025];26(6):841-2. Disponible sur: https://academic.oup.com/bioinformatics/article/26/6/841/244688 46. Quast C, Pruesse E, Yilmaz P, Gerken J, Schweer T, Yarza P, et al. The SILVA ribosomal 31 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint RNA gene database project: improved data processing and web-based tools. Nucleic Acids Res [Internet]. 27 nov 2012 [cité 18 juill 2025];41(D1):D590-6. Disponible sur: http://academic.oup.com/nar/article/41/D1/D590/1069277/The-SILVA-ribosomal-RNA-ge ne-database-project 47. Letunic I, Bork P. Interactive Tree of Life (iTOL) v6: recent updates to the phylogenetic tree display and annotation tool. Nucleic Acids Res [Internet]. 5 juill 2024 [cité 24 juill 2025];52(W1):W78-82. Disponible sur: https://academic.oup.com/nar/article/52/W1/W78/7645242 48. Bickhart DM, Kolmogorov M, Tseng E, Portik DM, Korobeynikov A, Tolstoganov I, et al. Generating lineage-resolved, complete metagenome-assembled genomes from complex microbial communities. Nat Biotechnol [Internet]. mai 2022 [cité 26 avr 2023];40(5):711-9. Disponible sur: https://www.nature.com/articles/s41587-021-01130-z 49. Eisenhofer R, Nesme J, Santos-Bay L, Koziol A, Sørensen SJ, Alberdi A, et al. A comparison of short-read, HiFi long-read, and hybrid strategies for genome-resolved metagenomics. Chen WH, éditeur. Microbiol Spectr [Internet]. 2 avr 2024 [cité 24 juill 2025];12(4). Disponible sur: https://journals.asm.org/doi/10.1128/spectrum.03590-23 50. Deng F, Han Y, Li M, Peng Y, Chai J, Yang G, et al. HiFi based metagenomic assembly strategy provides accuracy near isolated genome resolution in MAG assembly. iMetaOmics [Internet]. 8 juill 2025 [cité 24 juill 2025]; Disponible sur: https://onlinelibrary.wiley.com/doi/10.1002/imo2.70041 51. Goussarov G, Mysara M, Cleenwerck I, Claesen J, Leys N, Vandamme P, et al. Benchmarking short-, long- and hybrid-read assemblers for metagenome sequencing of complex microbial communities. Microbiology [Internet]. 25 juin 2024 [cité 24 juill 2025];170(6). Disponible sur: https://www.microbiologyresearch.org/content/journal/micro/10.1099/mic.0.001469 52. Knight CG, Nicolitch O, Griffiths RI, Goodall T, Jones B, Weser C, et al. Soil microbiomes show consistent and predictable responses to extreme events. Nature [Internet]. 19 déc 2024 [cité 16 juin 2025];636(8043):690-6. Disponible sur: https://www.nature.com/articles/s41586-024-08185-3 53. Kim CY, Ma J, Lee I. HiFi metagenomic sequencing enables assembly of accurate and complete genomes from human gut microbiota. Nat Commun [Internet]. 26 oct 2022 [cité 24 mars 2023];13(1):6367. Disponible sur: https://www.nature.com/articles/s41467-022-34149-0 32 .CC-BY-ND 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 12, 2025. ; https://doi.org/10.1101/2025.09.12.675765doi: bioRxiv preprint