Addressing the dynamic nature of reference data: a new nt database for robust metagenomic classification

Background Accurate metagenomic classification relies on comprehensive, up-to-date, and validated reference databases. While the NCBI BLAST Nucleotide (nt) database, encompassing a vast collection of sequences from all domains of life, represents an invaluable resource, its massive size —currently exceeding 10 12 nucleotides— and exponential growth pose significant challenges for researchers seeking to maintain current nt-based indices for metagenomic classification. Recognizing that no nt-based indices exist for the widely used Centrifuge classifier, and the last public version was released in 2018, we addressed this critical gap.

We present a new Centrifuge-compatible nt database, meticulously constructed using a novel pipeline incorporating different quality control measures, including reference decontamination and filtering. These measures demonstrably reduce spurious classifications, and through temporal comparisons, we reveal how this approach minimizes inconsistencies in taxonomic assignments stemming from asynchronous updates between public sequence and taxonomy databases. These discrepancies are particularly evident in taxa such as Listeria monocytogenes and Naegleria fowleri, where classification accuracy varied significantly across database versions.

This new database, made available as a pre-built Centrifuge index, responds to the need for an open, robust, nt-based pipeline for taxonomic classification in metagenomics. Applications such as environmental metagenomics, forensics, and clinical metagenomics, require comprehensive taxonomic coverage and will benefit from this resource. Our new nt-based index highlights the importance of treating reference databases as dynamic entities, subject to ongoing quality control and validation akin to software development best practices. This dynamic update approach is crucial for ensuring the accuracy and reliability of metagenomic analysis, especially as databases continue to expand in size and complexity . Key words: Metagenomics; Taxonomic Classification; Reference Database; NCBI BLAST nt; Centrifuge; Recentrifuge; Quality Control; Data Contamination; High Performance Computing

The capacity to accurately and sensitively profile microbial commu- nities is critical to a broad swath of biological applications, spanning biomedical research, healthcare, and environmental biosurveil- lance. Microbial community features are indicative of a multitude of human health-relevant outcomes [1], including cancer [2], infec- tion [3, 4, 5], trauma [6, 7], chronic injury [8, 9], and neurological disease [10, 11, 12]. Similarly, profiling microbial communities in non-human environments can provide detection and association Compiled on: June 12, 2024. Draft manuscript prepared by the author. 1 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted June 14, 2024. ; https://doi.org/10.1101/2024.06.12.598617doi: bioRxiv preprint 2 | Database Core processing Post-processing Downstream
 pipeline (e.g. AI/ML) Pipeline for Centrifuge
 database preparation Pre-processing data FASTQ metadata Recentrifuge’s interactive plot for visualization and exploration of results HTML Centrifuge C++ 
 python Validation
 analysis NCBI BLAST nt (FASTA) NCBI BLAST nt (Centrifuge) NCBI Taxonomy Figure 1. General layout of a metagenomic classification pipeline based on the NCBI BLAST nt database using Centrifuge and Recentrifuge. This flowchart is showing a typical configuration of a pipeline for taxonomic classification in metagenomics using the NCBI BLAST nt database [37, 39]. Data (and metadata, as available) is feed into a FASTQ format pre-processing block that performs quality checks using software such as fastp [40] and MultiQC [41]. Filtered FASTQ files are then processed by Centrifuge [20] using an indexed version of the nt database. Finally, Recentrifuge [36] post-process Centrifuge output, including negative control samples (if any), to further filter and prepare

for interactive exploration, downstream analysis and validation tests. Typical downstream jobs include metagenome assemblers and artificial intelligence/machine learning (AI/ML) training pipelines. The pipeline for building the Centrifuge reference database based on nt is detailed in Figure 3. information relevant to bioremediation, environmental surveil- lance [13, 14], pest control [15], and wastewater monitoring [16, 17]. In these applications spaces, the microbial community is often profiled via metagenomic sequencing, which facilitates compre- hensive detection of fastidious or difficult-to-culture microorgan- isms. Metagenomic classification transforms raw shotgun DNA sequences into microbiome composition profiles, which can then be interrogated for signatures or biomarkers of any given label of interest. A broad array of software platforms have been developed for classification of shotgun metagenomic sequence reads, including Kraken [18, 19], Centrifuge [20], MetaPhlAn [21], CLARK [22, 23], DIAMOND [ 24], Kaiju [ 25], GOTTCHA [ 26], metakallisto [ 27], KMCP [28], LMAT [29, 30], and ganon [ 31]. Each of these plat- forms relies, in some form, on a reference database containing known sequences or sequence signatures from a subset of refer- ence genomes. These databases may include a range of indexed

genomes, a subset of discriminative k-mers (e.g., CLARK [22] and KrakenUniq [32]), or a curated set of marker genes (e.g., MetaPhlAn4 [21]). The selection of database and software plat- form has a clear and systematic impact on the resultant microbial profile, influencing downstream assessments [33]. Numerous ad- ditional benchmarking studies providing comparative assessments between platforms are available for assessment [34]. In addition to the reference source and taxonomic clades for inclusion, metage- nomic classification results are impacted by temporal distribution of build dates and filtering processes applied to the reference con- tent. These parameters are potentially impactful but have not been thoroughly assessed in previous studies. To address this knowledge gap, we undertook a comparison of four temporally distinct database builds, comparing the extent of read classification and overall compositional impact. We also assessed the impact of reference decontamination via the Conter- minator [35] and Recentrifuge [ 36] platforms and the impact of short reference sequence removal. For broad taxonomic coverage, we constructed databases from the NCBI BLAST Nucleotide (nt) database [37], the most comprehensive database for nucleotide BLAST search [38]. We selected the Centrifuge classifier platform for this comparison due to its classification speed and optimized indexing scheme. The NCBI BLAST nt database encompasses nearly all traditional GenBank divisions, representing a significant portion of available GenBank sequences and spanning all domains of life [37, 39]. There- fore, the growth rate of nt follows that of GenBank. Both the num- Figure 2. Evolution of NCBI GenBank database over time. This semi-log plot shows the number of bases and sequences for NCBI GenBank. For the last 3 years of the time series, from April 2021 to April 2024, we fit an exponential model Aϕ(t–t 0)/τ (with time constant τ = 1) to the number of bases in the database, with R2 confirming the goodness of the model. As timet (x-axis, date) is measured in years fromt0 = 1982, A is the intercept to t = t0 and ϕ is the base of the exponential with the meaning of the yearly rate of change [44]. So, on average, between the mentioned dates, GenBank database had a 56% growing rate in nucleotides from year to year. In addition, at least in the last decade, the slope of the number of bases has been increasing while the slope of sequences has been decreasing, with these opposite trends indicating a progressive rise in the average length of the sequences in the database. ber of sequences and the nucleotides in GenBank are experienc- ing an exponential growth (see Figure 2 for quantitative details). Building nt-based reference databases for taxonomic annotation requires increasing computational resources, such as processing power, memory, and I/O bandwidth. In particular, the build pro- cess for these databases is memory-intensive, requiring comput- ing architectures that are not broadly available, with the literature mentioning initial database build difficulties and subsequent build failures as the dataset sizes increase[31, 42]. The latest available Centrifuge database based on NCBI BLAST nt was built in 2018, with multiple users unsuccessfully requesting an updated version in recent years [43]. As a response to the need for an open, robust, nt-based pipeline for taxonomic classification in metagenomics, we provide an updated, stable, and optimized Centrifuge nt reference database. The updated database provides a critical resource for researchers (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted June 14, 2024. ; https://doi.org/10.1101/2024.06.12.598617doi: bioRxiv preprint Martí et al. | 3 Table 1. Size parameters for nt DB references and builds nt DB release date Build version name Masked nt DB size Centrifuge nt DB size Taxonomy DB std dec std dec size content Jun 26, 2022 Jun22 - 745 GiB - 442 GiB 203 MiB 2.43 Mtid Sep 23, 2022 Sep22 - 843 GiB - 500 GiB 213 MiB 2.45 Mtid Jan 04, 2023 Jan23 941 GiB 936 GiB 558 GiB 555 GiB 217 MiB 2.48 Mtid Apr 03, 2023 Apr23 1.09 TiB 1.08 TiB 615 GiB 611 GiB 220 MiB 2.50 Mtid About database sizes, dec is a decontaminated database while std is a standard database (non-decontaminated). The size of taxonomy database is the decompressed size in megabytes of the names.dmp file of the corresponding NCBI Taxonomy database available at the indicated nt database release date. The unit tid denotes (different) taxonomic identifiers, so Mtid is a million of taxonomic identifiers. using metagenomics, where comprehensive and reliable taxonomic classification across the tree of life is paramount. This work un- derscores the critical need for treating the databases as a dynamic entity requiring continuous quality control and validation, much like software development best practices.

Reference selection Four NCBI nt release dates spanning 2022-2023 were selected for assessment (Table 1). Characterization and decontamination of the reference database Figure 3 shows a flowchart of the main steps involved in the pipeline for building a reference database for Centrifuge based on the nt database. After obtaining the nt database in FASTA format using BLAST+ tools [45] and the current content of the NCBI Taxonomy database, the content of nt was explored using Recentrifuge [36]. Figure S1 shows some examples of insights into the database provided at this step. The nt database was decontaminated using Conterminator [35] in parallel and the output post-processed by a different script of Recentrifuge, which removes the detected contamination and gen- erated a contamination network plot (see Figure S2 for an example).

masking, filtering, and indexing Therefasplit script from Recentrifuge [36] was used at different stages of the pipeline to enable efficient I/O processing of the se- quence database in parallel, which contains nucleotide counts on the order of 1012. Low complexity regions were masked using sev- eral instances of NCBI dustmasker [45] in parallel. The decontami- nated and masked version of the nt database was then processed by Recentrifuge’srefafilt code to remove sequences shorter than 16 nucleotides (found to cause problems downstream) and demultiplex headers for redundant sequences as generated by recent versions of the BLAST+ command line toolblastdbcmd. Finally , indexing of the decontaminated filtered database was performed with an opti- mized version of the Centrifuge 1.0.4-β database building C++ code centrifuge-build [20] targeting Livermore Computing (LC) HPC cluster nodes with 128 cores and 2 TiB main memory. The entire build process required about three weeks of clock time. Sample sequencing and processing Defined DNA mixtures were employed for comparative assess- ment of databases and classification parameters (Tables 1, 2). No- template process negative controls (ntc) were included in all analy- ses. Sequence libraries were prepared from gDNA via the Illumina DNA Prep library preparation kit and sequenced on the Illumina NextSeq 2000 via P2 300 cycle reagents and flow cell. Resultant data were processed for taxonomic classification via Centrifuge us- ing each of the database builds described in Table 1. Comparative analyses of reference controls were carried out in R (version 4.2.0). Phyloseq [46] was used to generate microbiome abundance profiles from count data and ggplot2 [47] was used to generate figures. The vegan package [48] was used to compute Bray Curtis dissimilarity reconterminator python centrifuge-build C++ NCBI nt (FASTA) NCBI nt (Centrifuge) NCBI BLAST nt BLAST+ tools C++ dustmasker C dustmasker C dustmasker C dustmasker C++ NCBI taxonomy NCBI nt (decontaminated) NCBI nt (conterminated)NCBI nt (conterminated)NCBI nt (conterminated)NCBI nt (decontaminated) taxonomic data conterminator C++ Conterminator output tsv contamination network PDF NCBI nt (dustmasked) refasplit python refafilt python length/reads log-histogram PDF NCBI nt (filtered) rcf python seqs/len per taxon (krona-style) HTML LEGEND file file format Required input (multiple files)

of different analysis through the pipeline (one single file per run) Data flow code lang/env Pieces of software input format NCBI BLAST nt database output format file Intermediate results (multiple files) Figure 3. General flowchart of the pipeline for building the Centrifuge reference database based on the NCBI BLAST nt database. The entire pipeline process takes several weeks using a HPC (high-performance computing) server with a high-memory-per-core ratio, a high-bandwidth low-latency parallel filesystem, 128 cores, and total memory of 2 TiB. The most demanding step is the final indexing of the filtered database using a version ofcentrifuge-build, which may expand beyond two weeks, followed by the decontamination step, which usually takes a few days. (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted June 14, 2024. ; https://doi.org/10.1101/2024.06.12.598617doi: bioRxiv preprint 4 | Table 2. DNA standard mixtures employed for comparative assessment Sample name Mixture Vendor Cat No Human spike Sample type zymo_gut_microbiome Zymobiomics gut microbiome Zymo D6331 no bacteria zymo_micro_comm_II Zymobiomics microbial community standard II Zymo D6311 no bacteria atcc_mycobiome Mycobiome mock community ATCC MSA-1010 no eukaryote atcc_oral_spike Oral microbiome mock community ATCC MSA-1004 18 ng bacteria atcc_skin_spike Skin microbiome mock community ATCC MSA-1005 14 ng bacteria atcc_gut_spike Gut microbiome mock community ATCC MSA-1006 40 ng bacteria protozoa_mix Acanthamoeba castellanii, Naegleria fowleri ATCC Various no eukaryote mammal_equal_mix Equal mix of 11 mammalian DNAs Various Various no eukaryote non-mammal_mix Equal mix of 10 non-mammalian eukaryotic DNAs Various Various no eukaryote mammal_log_mix Log concentration range of 7 mammalian DNAs Various Various no eukaryote atcc_20_strain ATCC 20 strain mix ATCC MSA-1003 no bacteria atcc_path_mix ATCC pathogen mix ATCC MSA-4000 no bacteria distances between samples.

and Discussion Comparative assessment of gDNA reference controls and mixtures In addition to the assessment of temporal database builds, we eval- uated other microbiome analytical parameters relevant to the as- sessment of accurate microbiome profiles. This warrants the use of standard controls with known microbial composition across dif- ferent kingdoms of life (Tables 1, 2). We systematically evaluated the impact of different values for the minimum length of partial hits in Centrifuge [ 20] or MHL (Minimum Hit Length), relative abundance thresholds, taxonomic levels, and sample type to gener- ate accurate microbiome profiles. We calculated precision, recall, balanced F1 scores and Bray Curtis dissimilarity distances to com- pare sequenced profiles of control standards with ground truth data across all database versions (Table 2). First, we investigated the impact of temporal database builds on microbiome profiling and observed temporal instability in Bray Curtis dissimilarity distances in 3 samples; ATCC_path_mix, proto- zoa_mix and zymo_micro_comm_II when compared to reference control profiles (Figures 4, S3). There was an increase in Bray Cur- tis dissimilarity distances in the Jan23 database versions for these samples. Upon inspection, we determined that delayed updates be- tween taxonomic labels in NCBI BLAST nt database and the current NCBI Taxonomy database at the time of the nt database release was a major factor. The nt database is relying on the Taxonomy database for the taxonomic information, but both databases are evolving over time independently, with no explicitly version mapping between them in the documentation. Traditionally , the nt database was us- ing a current version of the taxonomy database but that does not seem the case anymore, with the taxonomy used in NCBI BLAST nt database slightly trailing behind. In some cases, newly uploaded genomic sequences were lacking current taxonomic labels and then confounded Centrifuge’s database indexing algorithm and, eventu- ally , its classification process. For example, in the protozoa_mix sample, there were reads that were assigned and equally scored to multiple Naegleria fowleri (tid 5763) reference genomes that were present in the Jan23 NCBI BLAST nt database. However, when Centrifuge was reporting multi- ple assignments (Centrifuge default), only reads that were assigned to N. fowleri genomes whose tid was present in the NCBI Taxon- omy database used during the database compilation were assigned to tid 5763 while reads that were assigned to N. fowleri genomes whose tid was absent from the used taxonomy (but present in the NCBI BLAST nt database) were classified with a tid 0, normally re- served for unclassified reads. When Centrifuge was configured to report a unique tid per read using the lowest-common-ancestor (LCA) strategy (pipeline default), those reads were given a final root-only assignment (tid 1). As another example, the results for the zymo_micro_comm_II sample were impacted when using the Apr23 database due to varied taxonomic classification of Listeria monocytogenes genomes (Figures 4B, S3). In that databse, the addi- tion of a Listeria monocytogenes genome (CP046449.1) annotated as Listeria sp. LM90SB2 (tid 2678528 in the NCBI Taxonomy database, with parent being unclassified Listeria, tid 2642072) resulted in a final LCA classification of reads at only the genus level ( Listeria). Overall, our observations highlight the impact on the classification of specific taxa due to the discrepancy of taxonomic information between the nt database and the version of the NCBI Taxonomy database used for indexing the Centrifuge database. Next, we tested the impact of the following MHL values: 15 (minimum value), 22 (Centrifuge’s default), 30, 40, and 50 on the number of unique NCBI tax ids assigned to sequenced reads. Here, we observed that the number of unique NCBI tax id assignments and Bray Curtis dissimilarity distances decreases at increasing MHL values (Figure 4A,B). An inflection point was observed at MHL 30 (Figure 4A), followed by an increase in F1 and precision scores com- pared to MHL 15 and 22 (Figure 4D). This suggests that a MHL of 30 should be close to the optimal value in reducing false positive as- signments across these particular samples (Figure S3B) by reducing the noise of spurious assignments without a significant increase in false negatives. As most microbiome analysis protocols remove low-abundance features to account for potential sequencing errors and false pos- itives, we assessed the impacts of different thresholds (0, 0.001, 0.01, 0.1, 0.5, 1) on classification performance. Overall, we observed that a relative abundance threshold of 0.1% resulted in the highest F1 scores across samples (Figure 4D). As expected and previously observed, our data demonstrated that classification performance decreases at lower taxonomic ranks with an increasing number of detected false positives (Fig. S3B). Furthermore, samples consisting mainly of bacterial content had a better classification performance than eukaryotic samples (Figure 4C,D). Finally, we compared the number of annotated sequences be- tween the different versions of the Jan23 database and observed a decrease along the implemented improvements (Fig 4C). The standard version (Jan23_std) had an average number of 26,309,720 annotated sequences, followed by the decontaminated version (Jan23_dec) with 26,298,931 annotated sequences and lastly , the de- contaminated and filtered version (Jan23_filt) had 26,219,236 anno- tated sequences. However, Jan23_dec and Jan23_filt databases had higher non-root to root-only assignment ratios, suggesting that additional annotated reads with the standard version (Jan23_std) were assigned to the root-only level. (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted June 14, 2024. ; https://doi.org/10.1101/2024.06.12.598617doi: bioRxiv preprint Martí et al. | 5 0 K 10 K 20 K 30 K 20 30 40 50 MHL no. of unique NCBI tax ids sample ATCC_20_strain atcc_gut_spike atcc_mycobiome atcc_oral_spike ATCC_path_mix atcc_skin_spike mammal_equal_mix mammal_log_mix non−mammal_mix protozoa_mix zymo_gut_microbiome zymo_micro_comm_II 0.1 0.2 0.3 0.4 0.5 mhl15 mhl22 mhl30 mhl40 mhl50 MHL BC distance 0.1 0.2 0.3 0.4 Jun22_dec Sep22_dec Jan23_dec Jan23_std Jan23_filt Apr23_decon Apr23_std database BC distance 0.0 0.1 0.2 0.3 0.4 0.5 phylum class order family genus species taxonomic rank BC distance 0.1 0.2 0.3 0.4 bacteria eukaryote sample type BC distance 26,309,720 26,298,931 26,291,236 1e+05 1e+06 1e+07 1e+08 Jan23_std Jan23_dec Jan23_filt database no. of classified sequences 118.98 124.768 124.421 2 3 5 Jan23_std Jan23_dec Jan23_filt database log10 (non−root/root only assignments) F1 precision recall 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 mhl mhl15 mhl22 mhl30 mhl40 mhl50 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 0.25 0.50 0.75 1.00 database Jun22_dec Sep22_dec Jan23_dec Jan23_std Jan23_filt Apr23_dec Apr23_std 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 tax_rank phylum class order family genus species 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 0.25 0.50 0.75 1.00 type bacteria eukaryote A B C D Figure 4. Comparative assessment of gDNA reference controls. A) The number of unique NCBItid captured at different MHL values across samples and averaged across all temporal database versions. Data points are colored according to sample. B) Comparison of Bray Curtis dissimilarity distances at an abundance threshold of 0.1% between MHL values, database versions, taxonomic ranks, and sample types. For every variable that is being compared, Bray Curits dissimilarity distances were averaged across other variables. C) The number of annotated sequences and ratio of non-root to root-only assigned reads across different versions of the Jan23 database. Mean values are annotated at the top of each boxplot. D) Comparison of classification accuracy metrics (F1 balanced accuracy , precision, and recall) between MHL values, database versions, taxonomic ranks, and sample types across abundance thresholds. Potential implications While this work primarily focuses on providing a readily usable and improved nt-based database for metagenomic classification with Centrifuge and Recentrifuge, the underlying methodology and findings have implications that extend beyond this immediate application. Firstly , the availability of a comprehensive, decontaminated nt database opens new avenues for developing and training machine learning models for metagenomic analysis. Such models, trained on cleaner data, have the potential to achieve higher accuracy and better generalization capabilities, particularly for challenging tasks such as identifying novel or rare taxa. This could lead to improved tools for metagenome-based diagnostics, discovery of novel micro- bial enzymes, and understanding the complex interactions within microbial communities. Secondly , our observation of temporal inconsistencies in taxo- nomic assignments stemming from asynchronous updates between NCBI databases highlights a crucial challenge in metagenomic anal- ysis. It underscores the need for a paradigm shift in how we ap- proach reference data, moving from static resources to dynamically maintained and versioned collections. Drawing inspiration from software development best practices, continuous integration and continuous delivery (CI/CD) pipelines could be adapted for refer- ence databases, ensuring timely updates, robust validation, and improved reproducibility across studies. Additionally , the pipeline developed for database construction, encompassing decontamination, filtering, and validation steps, can be generalized to other large reference databases commonly em- ployed in metagenomic analysis. Applying similar quality control measures to resources like the NCBI nr (non-redundant protein) database or specialized databases for viral or fungal identification could significantly enhance the accuracy and reliability of classifi- cation in those domains. This has the potential to improve research outcomes in fields as diverse as environmental monitoring, disease diagnostics, and evolutionary biology . Furthermore, the availability of a comprehensive, up-to-date, and validated database encompassing all domains of life—such as the one presented here—offers a unique advantage for researchers in fields that require analysis of complex microbial communities across diverse environments. For example, studies investigating host-microbe interactions, exploring the microbiome of underex- plored ecosystems, or tracing the origins of emerging pathogens can benefit greatly from such a resource. However, the exponential growth of sequencing data and the associated computational burden highlight the need for innovative strategies to ensure the long-term sustainability of such compre- hensive resources. This might involve exploring distributed com- puting approaches, developing more efficient indexing and search algorithms, splitting the database, or even dividing the classifica- tion problem, e.g., converting the classification goal in a hierarchi- cal problem by using a "forest" of specific classifiers to decide the annotation at different taxonomic levels instead of a single classifier covering the entire tree of life. Data availability As of May 2024, the Jan23 decontaminated and filtered Centrifuge nt database is the updated database version. It can be downloaded from LLNL Green Data Oasis (GDO) public ftp server at ftp:// gdo-bioinformatics.ucllnl.org/centrifuge/nt_wntr23 using any software supporting anonymous ftp download. To ease the down- load process, the database is split in 71 ultra-compressed7z files of (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted June 14, 2024. ; https://doi.org/10.1101/2024.06.12.598617doi: bioRxiv preprint 6 | 4 GiB or less with name formatnt_wntr23_filt.cf.7z.*. The cor- responding taxonomy files for that version of the database, useful for Recentrifuge runs, are available atftp://gdo-bioinformatics. ucllnl.org/centrifuge/nt_wntr23/taxonomy. Shotgun metagenomic sequences of standard controls are up- loaded to NCBI SRA with the Bioproject ID PRJNA1118346. Practical notes Despite the most demanding step is the indexing of the nt database as described above, even running the Centrifuge classifier with such a large indexed database has some important computational requirements, especially regarding free disk space and available main memory. The size of the ultra-compressed database files amount to 284 GiB. Once decompressed from this format, the size of the Centrifuge (compressed) database files is 422GiB. Given that size, a computer with at least half TiB of main memory is recom- mended for running the Centrifuge classifier, especially when using a high number of parallel threads. In addition, Centrifuge may take more than half an hour just for loading the nt database from disk into memory . Given this overhead, it is recommended to use a sam- ple sheet file (–sample-sheetargument in Centrifuge) to process multiple samples without requiring additional database loads. Also, when replicating our pipeline, it is strongly recommended to use the LCA strategy with Centrifuge, by using the-k 1 argument in the call to the classifier. Finally, it is straightforward to use Re- centrifuge to post-process Centrifuge’s results, with detailed doc- umentation in Recentrifuge’s wiki athttps://github.com/khyox/ recentrifuge/wiki/Running-recentrifuge-for-Centrifuge. Declarations List of abbreviations DB: Data Base; HPC: High Performance Computing; I/O: In- put/Output; LCA: Lowest Common Ancestor; MHL: Minimum Hit Length; WGS: Whole-Genome Shotgun; Competing Interests The authors declare that they have no competing interests. Funding This study was supported by the Lawrence Livermore National Lab- oratory’s Laboratory Directed Research and Development Program (LDRD). This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344. LLNL Disclaimer: Neither the United States government nor Lawrence Livermore National Security , LLC, nor any of their employees make any warranty, ex- pressed or implied, or assume any legal liability or responsibility for the accuracy, completeness, or usefulness of any information, apparatus, product, or process disclosed, or represent that its use would not infringe privately owned rights. Reference herein to any specific commercial product, process, or service by trade name, trademark, manufacturer, or otherwise does not necessarily consti- tute or imply its endorsement, recommendation, or favoring by the United States government or Lawrence Livermore National Security , LLC. The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States government or Lawrence Livermore National Security , LLC, and shall not be used for advertising or product endorsement purposes.

1. Martí JM, Martínez-Martínez D, Rubio T, Gracia C, Peña M, Latorre A, et al. Health and Disease Imprinted in the Time Variability of the Human Microbiome. mSystems 2017;2(2):10.1128/msystems.00144–16. https://doi.org/10. 1128/msystems.00144-16. 2. Sepich-Poore GD, et al. The microbiome and human cancer. Science 2021;371(6536). 3. Johnson RC, et al. Correlation between nasal microbiome com- position and remote purulent skin and soft tissue infections. Infection and Immunity 2015;83(2):802–811. 4. Karl JP . Gut Microbiota-targeted Interventions for Reducing the Incidence, Duration, and Severity of Respiratory Tract In- fections in Healthy Non-elderly Adults. Military Medicine 2021;186(3-4):e310–e318. 5. van der Lelie D, Taghavi S. COVID-19 and the Gut Microbiome: More than a Gut Feeling. mSystems 2020;5(4). 6. Avila-Herrera A, et al. Metagenomic features of bioburden serve as outcome indicators in combat extremity wounds. Sci- entific Reports 2022;12(1):13816. 7. Kok CR, et al. Targeted metagenomic assessment reflects criti- cal colonization in battlefield injuries. Microbiology Spectrum 2023;11(6):e0252023. 8. Kalan L, et al. Redefining the Chronic-Wound Microbiome: Fungal Communities Are Prevalent, Dynamic, and Associated with Delayed Healing. mBio 2016;7(5). 9. Misic AM, Gardner SE, Grice EA. The Wound Microbiome: Modern Approaches to Examining the Role of Microorganisms in Impaired Chronic Wound Healing. Advances in Wound Care 2014;3(7):502–510. 10. Alhasson F, et al. Altered gut microbiome in a mouse model of Gulf War Illness causes neuroinflammation and intesti- nal injury via leaky gut and TLR4 activation. PLoS ONE 2017;12(3):e0172914. 11. McGrattan AM, et al. Diet and Inflammation in Cognitive Ageing and Alzheimer’s Disease. Current Nutrition Reports 2019;8(2):53–65. 12. Tremlett H, et al. The gut microbiome in human neurological disease: A review. Annals of Neurology 2017;81(3):369–382. 13. Brodie EL, et al. Urban aerosols harbor diverse and dynamic bacterial populations. Proceedings of the National Academy of Sciences of the United States of America 2007;104(1):299–304. 14. Jaing C, et al. Sierra Nevada sweep: metagenomic measure- ments of bioaerosols vertically distributed across the tropo- sphere. Scientific Reports 2020;10(1):12399. 15. Martí JM, Arias-Giraldo LF, Díaz-Villanueva W, Arnau V, Rodríguez-Franco A, Garay CP . Metatranscriptomic dynamics after Verticillium dahliae infection and root damage in Olea europaea. BMC plant biology 2020;20:1–15. https://doi.org/ 10.1186/s12870-019-2185-0. 16. Garner E, et al. Next generation sequencing approaches to evaluate water and wastewater quality. Water Research 2021;194:116907. 17. Keshaviah A, et al. Wastewater monitoring can anchor global disease surveillance systems. Lancet Global Health 2023;11(6):e976–e981. 18. Wood DE, Salzberg SL. Kraken: ultrafast metagenomic se- quence classification using exact alignments. Genome biology 2014;15:1–12. 19. Wood DE, Lu J, Langmead B. Improved metagenomic analysis with Kraken 2. Genome biology 2019;20:1–13. 20. Kim D, et al. Centrifuge: rapid and sensitive classification of metagenomic sequences. Genome Research 2016;26(12):1721– 1729. 21. Blanco-Miguez A, et al. Extending and improving metage- nomic taxonomic profiling with uncharacterized species using MetaPhlAn 4. Nature Biotechnology 2023;41(11):1633–1644. (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted June 14, 2024. ; https://doi.org/10.1101/2024.06.12.598617doi: bioRxiv preprint Martí et al. | 7 22. Ounit R, Wanamaker S, Close TJ, Lonardi S. CLARK: fast and accurate classification of metagenomic and genomic sequences using discriminative k-mers. BMC genomics 2015;16:1–13. 23. Ounit R, Lonardi S. Higher classification sensitivity of short metagenomic reads with CLARK-S. Bioinformatics 2016;32(24):3823–3825. 24. Buchfink B, Xie C, Huson DH. Fast and sensitive protein align- ment using DIAMOND. Nature methods 2015;12(1):59–60. 25. Menzel P , Ng KL, Krogh A. Fast and sensitive taxonomic classi- fication for metagenomics with Kaiju. Nature Communications 2016;7:11257. 26. Freitas TA, et al. Accurate read-based metagenome characteri- zation using a hierarchical suite of unique signatures. Nucleic Acids Research 2015;43(10):e69. 27. Schaeffer L, et al. Pseudoalignment for metagenomic read assignment. Bioinformatics 2017;33(14):2082–2088. 28. Shen W, et al. KMCP: accurate metagenomic profiling of both prokaryotic and viral populations by pseudo-mapping. Bioin- formatics 2023;39(1). 29. Ames SK, et al. Scalable metagenomic taxonomy classifi- cation using a reference genome database. Bioinformatics 2013;29(18):2253–2260. 30. Ames SK, Gardner SN, Marti JM, Slezak TR, Gokhale MB, Allen JE. Using populations of human and microbial genomes for organism detection in metagenomes. Genome research 2015;25(7):1056–1067. https://doi.org/10.1101/gr.184879. 114. 31. Piro VC, Dadi TH, Seiler E, Reinert K, Renard BY. ganon: precise metagenomics classification against large and up- to-date sets of reference sequences. Bioinformatics 2020 07;36(Supplement_1):i12–i20. https://doi.org/10.1093/ bioinformatics/btaa458. 32. Breitwieser FP , Baker DN, Salzberg SL. KrakenUniq: confi- dent and fast metagenomics classification using unique k-mer counts. Genome Biology 2018;19(1):198. 33. Xu R, Rajeev S, Salvador LCM. The selection of software and database for metagenomics sequence analysis impacts the out- come of microbial profiling and pathogen detection. PLoS ONE 2023;18(4):e0284031. 34. McIntyre ABR, Ounit R, Afshinnekoo E, et al. Comprehen- sive benchmarking and ensemble approaches for metagenomic classifiers. Genome Biology 2017;18(1):182. 35. Steinegger M, Salzberg SL. Terminating contamination: large- scale search identifies more than 2,000,000 contaminated en- tries in GenBank. Genome Biology 2020;21(1):115. 36. Marti JM. Recentrifuge: Robust comparative analysis and contamination removal for metagenomics. PLoS Compu- tational Biology 2019;15(4):e1006967. https://doi.org/10. 1371/journal.pcbi.1006967. 37. Coordinators NR. Database resources of the National Center for Biotechnology Information. Nucleic Acids Research 2012 11;41(D1):D8–D20. https://doi.org/10.1093/nar/gks1189. 38. Altschul SF, Gish W, Miller W, Myers EW, Lipman DJ. Ba- sic local alignment search tool. Journal of molecular biology 1990;215(3):403–410. 39. Ye J, McGinnis S, Madden TL. BLAST: improvements for better sequence analysis. Nucleic acids research 2006;34(suppl_2):W6–W9. 40. Chen S. Ultrafast one-pass FASTQ data preprocessing, quality control, and deduplication using fastp. iMeta 2023;2(2):e107. https://onlinelibrary.wiley.com/doi/abs/ 10.1002/imt2.107. 41. Ewels P , Magnusson M, Lundin S, Käller M. MultiQC: sum- marize analysis results for multiple tools and samples in a sin- gle report. Bioinformatics 2016 06;32(19):3047–3048. https: //doi.org/10.1093/bioinformatics/btw354. 42. Plyusnin I, Vapalahti O, Sironen T, Kant R, Smura T. Enhanced Viral Metagenomics with Lazypipe 2. Viruses 2023;15(2).https: //www.mdpi.com/1999-4915/15/2/431. 43. Chorlton SD. Ten common issues with reference sequence databases and how to mitigate them. Frontiers in Bioinformat- ics 2024;4:1278228. 44. Martí JM, Garay CP . Not just BLAST nt: WGS database joins the party . BioRxiv 2019;p. 653592. 45. Camacho C, et al. BLAST+: architecture and applications. BMC Bioinformatics 2009;10:421. 46. McMurdie PJ, Holmes SP . An R Package for Reproducible Inter- active Analysis and Graphics of Microbiome Census Data. PLoS ONE 2013;8(5):e61217. 47. Wickham H. ggplot2: Elegant Graphics for Data Analysis. Springer-Verlag New York; 2016. 48. Oksanen J, Simpson GL, Blanchet FG, Kindt R, Legendre P , Minchin PR, et al. vegan: Community Ecology Package; 2024, r package version 2.6-7. (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted June 14, 2024. ; https://doi.org/10.1101/2024.06.12.598617doi: bioRxiv preprint 8 | BacteriaBacteria Pseudomonadota Gamma...teria Enterobacterales    1% P...s Pseudomonadaceae Pseudomonas    1% Xanthomonadales Xanthomonadaceae    1% M...s Moraxellaceae    0.8% 7 more Arenicellales    0.8% Alphaproteobacteria Hyphomicrobiales Rhizo...aceaeRhizobi...m groupAgrobacteriumAgrobacteri...ens complex Agrobacterium radiobacter    7% 4 more St...aeR...men...es uncultured Roseibium sp.    1% 3 more Rickettsiales A...eW...eWo...iaunclassi...olbachia Wolbachia endosymbiont of Cotesia melitaearum    0.9%8 more Rhod...ales Paracoccaceae   2% Roseobacteraceae   2% Sphin...dales Sphingomonadaceae   1%1 more Rhodospirillales   0.9% uncla...teria Alphaproteobacteria bacterium CMAR1   0.9% Betaproteobacteria Burkholderiales   1% Terrabacteria groupBacillota BacilliBac...lesB...e Bacillus   0.9% 7 more P...e Paenibacillus    0.7% ClostridiaE...s 2 more E...sC...a unclassi "ed Candidatus Limiplasma Candidatus Limiplasma sp.    0.8% L...sV...eu...e Vallitaleaceae bacterium    1% 1 more Ch...esChr...eaeuncla...aceae Christensenellaceae bacterium    3% Negat...cutes Selenomonadales    0.7% Acti...tota Acti...etes...St...ae Streptomyces    0.9% ... Microbacteriaceae    0.9% 5 more7 more Aci...iia A...s u...s Acidimicrobiales bacterium    1% Cyan...roup Cy...ta C...e ...NostocaceaeDesmonostoc Desmonostoc sp.    0.6% 3 more Myco...tota Mo...es ... ... Phytoplasma    0.7% Chlo...xota An...ae ... ... 0.9%    Aggregatilineales bacterium Ca...ae Cal...les Cal...eae unclassif...lineaceae 4%   Caldilineaceae bacterium FCB group Bacteroid...ota group Bacteroidota Fla...iia Fl...es 1%   Flavobacteriaceae Bact...idia Ba...es 1 more ...... 1%   Bacteroidaceae bacterium ...... 1%   uncultured Microbacter sp. Mar...les...... 1%   uncultured Saccharicrinis sp. ...... 0.8%   Marini "laceae bacterium 4 more Ignavi...eriota IgnavibacteriaIgn...lesMeli...ceaeunclas...raceae1%   Melioribacteraceae bacterium Candidatus ...cibacterotaunclassi "...ibacterota 1%   Candidatus Latescibacterota bacterium3 more PVC groupPla...ota Planct...ycetiaPir...les Pi...aeMa...us unclassi...iblastus 1%   Mariniblastus sp. T...eun...ae3%   Thermoguttaceae bacterium Ph...aeS...sun...es2%   Sedimentisphaerales bacterium ... 1%   Phycisphaerae bacterium Ver...otaOpi...tae...... 0.8%   Opitutales bacteriumLen...otaLe...ia...... 0.7%   Victivallales bacterium SpirochaetotaSpi...tiaSpi...lesT...eu...e1%   Treponemataceae bacterium 1 more S...eu...e2%   Sphaerochaetaceae bacteriumuncl...etia 1%   Spirochaetia bacterium Thermo...eriotaDesu...adia Desu...alesG...eu...e 2%   Geopsychrobacteraceae bacterium Desulf...rioniaDesu...alesD...eP...o environmental samples0.8%   uncultured Pseudodesulfovibrio sp. 2 more 5 more Aci...ota Terr...obia uncla...lobia 1%   Terriglobia bacterium Vic...ria Vic...les un...es 4%   Vicinamibacterales bacterium Bacteria incertae sedisB...a C...a ... 0.7%    Candidatus Poribacteria bacterium Can...ota... ... 1%   Candidatus Binatia bacterium Synergistota Syn...tia S...s ... ... 0.9%    Aminobacteriaceae bacterium Myx...ota u...a 1%   Myxococcota bacterium Length (log10) 1.0 2.4 3.8 5.1 6.5 7.9 rootroot Eukaryota Opisthokonta MetazoaEumetazoa Caniformia    2% Felidae    0.9% Odontoceti    0.7% Bovidae    0.7% Microchiroptera    0.9% 3 moreHomo sapiens    1% Cercopithecinae    0.7%Mus    0.6% Cricetidae    0.8% 5 more Passeriformes    2%Galliformes    0.6%28 more Durocryptodira    0.7%Toxicofera   0.8% Anura   0.8% Perciformes   1% 17 more Cyprinodontoidei   0.6% 4 more Syngnathinae   0.6% Carangaria   0.7% 5 more Oncorhynchus    0.6% Cyprininae    0.7% Characiphysae    0.8% melanogaster group    0.6% Culicidae    0.8% Anthophila    0.7% Formicidae    0.6% 7 more Papilionoidea    1% 15 morePolyphaga    0.8% Timema californicum    0.6% 14 moreHemiptera    0.7% Eumalacostraca    0.7% 0.9%    Arachnida 0.9%    Autobranchia 0.7%    Gastropoda 1%   Platyhelminthes 3 more Fungi Dikarya 0.9%    Hypocreales6 more 0.7%    Aspergillus20 more 0.6%    Pleosporomycetidae 0.9%    Agaricomycetes 0.7%    Fungi incertae sedis ViridiplantaeStreptophytaStreptophytina 1%   NPAAA clade 0.8%   Rosales 0.6%   Malpighiales5 more 0.9%   Brassicaceae 7 more 0.7%   Solanaceae 0.6%   Lamiales 0.6%   campanulids 0.6%   Triticeae 0.6%   PACMAD clade 1%   Sar 16 more Viruses Riboviria Ortho...viraePisuviricota 8%   Severe acute respiratory syndrome coronavirus 2N...a 6 more Revtra...icetesO...s 1%   Human immunode "ciency virus 1 Bacteria envi...ples 5%   uncultured bacterium Pseudomonadota 0.9%    Gammaproteobacteria 8 more Terr...roup 0.8%    Bacillota 10 more26 more Length (log10) 1.0 2.4 3.8 5.1 6.5 7.9 A B C Figure S1. Recentrifuge’s scripts allow the characterization and interactive exploration of the content of the nt database during the building process of the nt database for Centrifuge. (A) Recentrifuge’srcf interactive plot showing the content of the nt database for the number of sequences and the logarithm of their length in the color scale. (B) Recentrifuge’srefafilt output showing the distribution of the sequence lengths in the nt database. (C) Recentrifuge’srcf interactive plot displaying new bacterial taxa not present in the previous seasonally compiled database showing the number of sequences and the logarithm of their length in the color scale. (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted June 14, 2024. ; https://doi.org/10.1101/2024.06.12.598617doi: bioRxiv preprint Martí et al. | 9 in millions (106) Figure S2. Example of inter-kingdom contamination network plot generated during the process of building the nt database for Centrifuge. The contamination was detected between the five default "kingdoms" as defined in Conterminator software, with the width and color scale of the arrows related to the number of sequences involved. The rate of contamination detected in the database (non-redundant sequences) before contamination removal was significant: almost half a percentage point. mammal_equal_mix mammal_log_mix non−mammal_mix protozoa_mix zymo_gut_microbiome zymo_micro_comm_II ATCC_20_strain atcc_gut_spike atcc_mycobiome atcc_oral_spike ATCC_path_mix atcc_skin_spike Jun22_decSep22_decJan23_decJan23_stdJan23_filt Apr23_decon Apr23_std Jun22_decSep22_decJan23_decJan23_stdJan23_filt Apr23_decon Apr23_std Jun22_decSep22_decJan23_decJan23_stdJan23_filt Apr23_decon Apr23_std Jun22_decSep22_decJan23_decJan23_stdJan23_filt Apr23_decon Apr23_std Jun22_decSep22_decJan23_decJan23_stdJan23_filt Apr23_decon Apr23_std Jun22_decSep22_decJan23_decJan23_stdJan23_filt Apr23_decon Apr23_std 0.0 0.2 0.4 0.6 0.0 0.2 0.4 0.6 database BC distance sample ATCC_20_strain atcc_gut_spike atcc_mycobiome atcc_oral_spike ATCC_path_mix atcc_skin_spike mammal_equal_mix mammal_log_mix non−mammal_mix protozoa_mix zymo_gut_microbiome zymo_micro_comm_II A B 0.00 0.25 0.50 0.75 1.00 1 10 100 1000 abundance threshold no. of false positives mhl mhl15 mhl22 mhl30 mhl40 mhl50 0.00 0.25 0.50 0.75 1.00 1 10 100 abundance threshold no. of false positives tax_rank phylum class order family genus species Figure S3. Additional comparative assessment of gDNA reference controls. A) Bray Curtis dissimilarity distance between database versions across all gDNA reference control samples. B) Number of false positive assignments between MHL values and taxonomic rank across abundance thresholds. (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted June 14, 2024. ; https://doi.org/10.1101/2024.06.12.598617doi: bioRxiv preprint