## Main

The human gut microbiome has been implicated in important phenotypes related to human health and disease1,2. However, incomplete reference data that lack sufficient microbial diversity3 hamper understanding of the roles of individual microbiome species and their functions and interactions. Hence, establishing a comprehensive collection of microbial reference genomes and genes is an important step for accurate characterization of the taxonomic and functional repertoire of the intestinal microbial ecosystem.

The Human Microbiome Project (HMP)4 was a pioneering initiative to enrich knowledge of human-associated microbiota diversity. Hundreds of genomes from bacterial species with no sequenced representatives were obtained as part of this project, allowing their use for the first time in reference-based metagenomic studies. The Integrated Gene Catalog (IGC)5 was subsequently created, combining the sequence data available from the HMP and the Metagenomics of the Human Intestinal Tract (MetaHIT)6 consortium. This gene catalog has been applied successfully to the study of microbiome associations in different clinical contexts7, revealing microbial composition signatures linked to type 2 diabetes8, obesity9 and other diseases10. But, as the IGC comprises genes with no direct link to their genome of origin, it lacks contextual data to perform high-resolution taxonomic classification, establish genetic linkage and deduce complete functional pathways on a genomic basis.

Culturing studies have continued to unveil new insights into the biology of human gut communities11,12 and are essential for applications in research and biotechnology. However, the advent of high-throughput sequencing and new metagenomic analysis methods—namely, involving genome assembly and binning—has transformed understanding of the microbiome composition in both humans and other environments13,14,15. Metagenomic analyses are able to capture substantial microbial diversity not easily accessible by cultivation by directly analyzing the sample genetic material without the need for culturing, although biases do exist16. This can be achieved by binning de novo-assembled contigs into putative genomes, referred to as metagenome-assembled genomes (MAGs). However, current challenges associated with metagenome assembly and binning can result in incorrectly binned contigs, which substantially affects further taxonomic and functional inferences. Therefore, the use of MAGs requires careful considerations17, but they provide important insights into the uncultured microbial diversity in the absence of isolate genomes.

Recent studies have massively expanded the known species repertoire of the human gut, making available unprecedented numbers of new cultured and uncultured genomes16,18,19,20,21. Two culturing efforts isolated and sequenced over 500 human-gut-associated bacterial genomes each19,21, while three independent studies16,18,20 reconstructed 60,000–150,000 MAGs from public human microbiome data, most of which belong to species lacking cultured representatives. Combining these individual efforts and establishing a unified nonredundant dataset of human gut genomes is essential for driving future microbiome studies. To accomplish this, we compiled and analyzed 204,938 genomes and 170,602,708 genes from human gut microbiome datasets to generate the Unified Human Gastrointestinal Genome (UHGG) and Protein (UHGP) catalogs, the most comprehensive sequence resources of the human gut microbiome established thus far.

## Results

### More than 200,000 human gut microbial genomes in the UHGG catalog

We first gathered all prokaryotic isolate genomes and MAGs from the human gut microbiome (publicly available as of March 2019). We compiled the isolate genomes from the Human Gastrointestinal Bacteria Culture Collection (HBC)19 and the Culturable Genome Reference (CGR)21, as well as cultured human gut genomes available in the NCBI22, PATRIC23 and IMG24 repositories, which include genomes from several other large studies11,12,25. In addition, we included all of the gut MAGs generated in Pasolli et al.20 (CIBIO), Almeida et al.18 (EBI) and Nayfach et al.16 (HGM). To standardize the genome quality across all sets, we used thresholds of >50% genome completeness and <5% contamination, combined with an estimated quality score (completeness –5 × contamination) > 50. The final numbers of genomes matching these criteria were 734 (HBC), 1,519 (CGR), 651 (NCBI), 7,744 (PATRIC/IMG), 137,474 (CIBIO), 87,386 (EBI) and 51,489 (HGM), resulting in a total of 286,997 genome sequences (Fig. 1a and Supplementary Table 1). These represented 204,938 nonredundant genomes on the basis of a Mash26 distance threshold of 0.001 (99.9% nucleotide identity) and only considering one genome per species per sample to account for the fact that the three large MAG studies analyzed many samples in common. Genomes were recovered in samples from a total of 31 countries across six continents (Africa, Asia, Europe, North America, South America and Oceania), but the majority originated from samples collected in China, Denmark, Spain and the United States (Fig. 1b).

To determine how many species were included in this gut reference collection, we clustered all 286,997 genomes using a multistep distance-based approach (Methods) with an average nucleotide identity (ANI) threshold of 95% over at least a 30% alignment fraction (AF)27. The clustering procedure resulted in a total of 4,644 inferred prokaryotic species (4,616 bacterial and 28 archaeal; Supplementary Table 2). We found the species clustering results to be highly consistent with those previously obtained16,18,20 (Supplementary Table 3). The best quality genome from each species cluster was selected as its representative on the basis of genome completeness, minimal contamination and assembly N50 (with isolate genomes always given preference over MAGs), and the final set was used to generate the UHGG catalog (Fig. 1c). Of the 4,644 species-level genomes, 3,207 were >90% complete (interquartile range, IQR = 87.2–98.8%) and <5% contaminated (IQR = 0.0–1.34%), with 573 of these having the 5S, 16S and 23S rRNA genes together with at least 18 of the standard tRNAs (Extended Data Fig. 1). These 573 genomes (535 from isolates and 38 from MAGs) satisfy the ‘high quality’ criteria set for MAGs by the Genomic Standards Consortium28. The rRNA operon has previously been shown to be a problematic region to assemble from short-read metagenomic datasets13,16,18,20, which might explain the low number of high-quality MAGs. Thereafter, we classified each species representative using the Genome Taxonomy Database Toolkit29,30 (GTDB-Tk; Extended Data Fig. 2), a standardized taxonomic framework based on a concatenated protein phylogeny representing >140,000 public prokaryote genomes, fully resolved to the species level (see Methods for details on the taxonomy nomenclature used). However, over 60% of the gut genomes could not be assigned to an existing species, confirming that the majority of the UHGG species lack representation in current reference databases.

To obtain further insights into the quality of the UHGG genomes, we inferred the level of strain heterogeneity within each MAG with CMseq20. The median strain heterogeneity (proportion of polymorphic positions) of the UHGG MAGs was 0.06% (IQR = 0.01–0.25%; Extended Data Fig. 1c and Supplementary Table 1), which is below the 0.5% threshold defined previously20 to distinguish medium- from high-quality MAGs. We believe that this additional metric on strain heterogeneity is a useful complement to the standard completeness and contamination estimates, providing further evidence of the overall high quality of the genomes included here.

### Comparison of species recovered in individual studies

We investigated how many of the 4,644 gut species were found in the different study collections to determine their level of overlap and reproducibility, as well as the ratio between cultured and uncultured species (Fig. 2a). Each of the large MAG studies used a different assembly and binning approach: the CIBIO study used metaSPAdes31 and MetaBAT 2 (ref. 32) for assembling and binning sequencing runs previously merged by sample; the HGM study used MEGAHIT33 to assemble runs merged by sample and applied a combination of MaxBin 2 (ref. 34), MetaBAT 2 (ref. 32), CONCOCT35 and DAS Tool36 for binning and refinement; and the EBI study used metaSPAdes31 and MetaBAT 2 (ref. 32) for assembling and binning individual runs without merging by sample. Despite these methodological differences, the largest intersection found was between these collections of MAGs, with the same 1,081 species detected independently in the CIBIO, EBI and HGM datasets, but not in any of the cultured genome studies. By restricting the analysis to genomes recovered from 1,554 samples common to all three MAG studies, we found that 93–97% of the species from each set were detected in at least one other MAG collection and 79–86% were detected across all three (Extended Data Fig. 3a). A similar level of species overlap was observed when comparing studies on a per-sample basis (Extended Data Fig. 3b). Furthermore, conspecific genomes recovered from the same samples across different studies had a median ANI and AF of 99.9% and 92.1%, respectively (94.5% AF with ≥90% complete genomes and 86.6% AF with medium-quality genomes; Extended Data Fig. 3c). These results suggest that the large-scale studies of human gut MAGs16,18,20 generally recovered highly similar genomes. However, the smaller AF values detected among genomes that were <90% complete suggest that caution is needed when using medium-quality genomes in downstream analyses.

Rarefaction analysis indicated that the number of uncultured species detected has not reached a saturation point, meaning that additional species remain to be discovered (Fig. 2b). However, these most likely represent rarer members of the human gut microbiome, as the number of species is closer to saturating when only considering those with at least two conspecific genomes.

We also investigated the intersection between the three large culture-based datasets: the HBC, CGR and NCBI (which contains gut genomes from the HMP4). Unlike the MAGs, the majority of cultured species were unique within a single collection (486/698; 70%), with only 70 (10%) common to all three collections (Extended Data Fig. 3d). This may be due to varied geographic sampling between the collections (Asia, Europe and North America) or might highlight the stochastic nature of culture-based studies.

### Most gut microbial species lack isolate genomes

We found that 3,750 (81%) of the species in the UHGG catalog did not have a representative in any of the human gut culture databases. To extend the search to isolate genomes from other environments or lacking information on the isolation source, we compared the UHGG catalog to all NCBI RefSeq isolate genomes. We identified an additional set of 438 species closely matching cultured genomes (88 from human body sites, 29 from other animals, 3 from plants and the remainder (318) from unknown sources), leaving 3,312 (71%) UHGG species as uncultured (Supplementary Table 2).

By calculating the number of genomes contained within each cultured and uncultured human gut species, we found that species containing isolate genomes represented the largest clusters, while those exclusively encompassing MAGs tended to be the rarest, as discussed previously16,18,20. For example, only 2 of the 25 largest bacterial clusters were exclusively represented by MAGs (Fig. 2c), with 1,212 uncultured species represented by a single genome (80% of which originated from samples only analyzed in one of the MAG studies; Extended Data Fig. 4). The bacterial species most represented in our collection were Agathobacter rectalis (recently reclassified from Eubacterium rectale37), Escherichia coli D and Bacteroides uniformis (Fig. 2c, Extended Data Fig. 5a and Supplementary Table 2), whereas the most frequently recovered archaeal species was Methanobrevibacter A smithii, with 608 genomes found across all six continents (Extended Data Fig. 6). We inferred the level of geographic diversity of each species by calculating the Shannon diversity index on the proportion of samples in which each species was found per continent. The largest species clusters displayed similarly high levels of geographic distribution, indicating that the most highly represented species were not restricted to individual locations (Fig. 2c and Extended Data Fig. 5b).

We determined how representative the UHGG catalog is of the human gut microbial diversity by mapping 1,005 independent metagenomic datasets against the 4,644 UHGG species (Fig. 2d and Supplementary Table 4). Using Kraken 2 (ref. 38), we obtained a median classification rate of 85.9% (IQR = 83.5–88.1%). Notably, this corresponded to a median improvement of 155% over the standard RefSeq database. The increase in classification rate was most pronounced in non-Western samples from Cameroon, Ethiopia, Ghana and Tanzania, highlighting the potential of the UHGG catalog to improve the study of microbiome diversity from these understudied populations.

The phylogenetic distribution of the 4,616 bacterial (Fig. 3a) and 28 archaeal (Extended Data Fig. 6) species revealed that uncultured species exclusively represented 66% and 31% of the phylogenetic diversity of Bacteria and Archaea, respectively, with several phyla lacking cultured representatives (Fig. 3b). The four largest monophyletic groups lacking cultured genomes were the 4C28d-15 order (167 species, recently proposed as the novel order Comantemales ord. nov.39; Fig. 3c), order RF39 (139 species), family CAG-272 (88 species) and order Gastranaerophilales (67 species). While none have been successfully cultured, several have been described in the literature, including for RF39 (ref. 16) and Gastranaerophilales (previously classified as a lineage in the Melainabacteria40), which are characterized by highly reduced genomes with numerous auxotrophies. This analysis suggests that, despite recent culture-based studies11,12,19,21, much of the diversity in the gut microbiome remains uncultured, including several large and prevalent clades.

### Expanding the set of proteins in the human gut microbiome

Metagenomic approaches have the ability to leverage gene content information not only for more precise taxonomic analysis but also to predict the functional capacity of individual species of interest in comparison to marker-gene-based methods (for example, relying solely on the 16S rRNA gene or a limited number of diagnostic genes). We built the UHGP catalog with a total of 625,255,473 full-length protein sequences predicted from the 286,997 analyzed genomes herein. These were clustered at 50% (UHGP-50), 90% (UHGP-90), 95% (UHGP-90) and 100% (UHGP-100) amino acid identity, generating between 5 to 171 million protein clusters (Fig. 1c and Extended Data Fig. 7a). While the number of UHGP-95 and UHGP-90 clusters showed a steady increase as a function of the number of genomes considered, those from UHGP-50 are reaching a saturation point (Fig. 4a), in line with previous estimates6.

To determine how comprehensive the UHGP is when compared to existing human gut gene catalogs, we combined the UHGP-90 (n = 13,910,025 protein clusters) with the IGC5, a collection of 9.9 million genes from 1,267 gut metagenome assemblies, which we grouped into 7,063,981 protein clusters at 90% protein identity (referred to as IGC-90). Nearly all samples used to generate the IGC were also included in the UHGP catalog (except for 59 transcriptome datasets), but the latter was generated from a larger and more geographically diverse metagenomic dataset (including samples from Africa, South America and Oceania). Combining the UHGP-90 and IGC-90 resulted in a set of 15.2 million protein clusters, with an overlap of 5.8 million sequences (Fig. 4b). This revealed that 81% of the IGC is represented in the UHGP catalog, with the missing 19% likely representing fragments of prokaryotic genomes that are <50% complete and viral or eukaryotic sequences, plasmids or other sequences not binned into MAGs. In fact, only 0.2% (n = 34,070 clusters) of the UHGP-90 was predicted to be of viral origin (on the basis of eggNOG annotations), as compared to the 5% estimate obtained in a previous human gut gene catalog6 included in the IGC. Most notably, though, the UHGP provided an increase of 115% in coverage of the gut microbiome protein space over the IGC (from 7,063,981 to a total of 15,217,595 protein clusters).

We also compared the read mapping rate using the same 1,005 metagenomic samples tested against the UHGG catalog (Supplementary Table 4). Even though the classification rate was substantially higher when using the UHGG catalog than with RefSeq, the increase with the UHGP-90 over IGC-90 was more modest (median of 5%; Extended Data Fig. 7b). These results suggest that, although the UHGP collectively encompasses a much larger number of protein clusters, most of the newly added proteins are at lower abundance/prevalence within individual samples. However, as the UHGP was generated from individual genomes and not from their original unbinned metagenome assemblies, our catalog also has the advantage of providing a direct link between each gene cluster and its genome of origin. To this end, we have also generated high-quality subsets of the UHGP-95, UHGP-90 and UHGP-50 consisting of protein clusters where at least two proteins from different genomes belonging to the same species were retrieved (UHGP-95-HQ, n = 10,798,224; UHGP-90-HQ, n = 8,082,122; UHGP-50-HQ, n = 3,088,278). This clustering criterion was used to control for the presence of contaminating sequences within each MAG and for the possibility that multiple copies of the same protein-coding sequence may be present in one genome. The UHGP ultimately allows the combination of individual genes with their genomic context for an integrated study of the gut microbiome.

### Functional capacity of the human gut microbiota

We used the eggNOG41, InterPro42, COG43 and KEGG44 annotation schemes to capture the full breadth of functions within the UHGP. However, we found that 41.5% of UHGP-100 was poorly characterized, as 27.3% lacked a match to any database and a further 14.2% only had a match to a COG with no known function (Fig. 4c). On the basis of the distribution of COG functions, the most highly represented categories were related to amino acid transport and metabolism, cell wall/membrane/envelope biogenesis and transcription.

We further leveraged the set of 171 million proteins derived from the human gut genomes to explore the functional diversity within each of the UHGG species. Protein sequences from all conspecific genomes were clustered at 90% amino acid identity to generate a pan-genome for each species. Analysis of the functional capacity of the UHGG species pan-genomes identified a total of 363 KEGG modules encoded by at least one species (Extended Data Fig. 8a and Supplementary Table 5). Most conserved modules were related to ribosomal structure, glycolysis, inosine monophosphate biosynthesis, gluconeogenesis and the shikimate pathway—all representing essential bacterial functions. However, we found that, for certain phyla such as Myxococcota, Bdellovibrionota, Thermoplasmatota, Patescibacteria and Verrucomicrobiota, a substantial proportion of the species pan-genomes remained poorly characterized (Extended Data Fig. 8b). At the same time, species belonging to the clades Fibrobacterota, Bacteroidota, Firmicutes I, Verrucomicrobiota and Patescibacteria had the highest proportion of genes encoding carbohydrate-active enzymes (CAZy; Extended Data Fig. 8b). As most of these lineages are largely represented by uncultured species (Fig. 3b), this suggests that the gut microbiota may harbor many species with important metabolic activities yet to be cultured and functionally characterized under laboratory conditions.

### Patterns of intraspecies genomic diversity

With the protein annotations and pan-genomes inferred for each of the UHGG species, we explored their intraspecies core and accessory gene repertoire. Only near-complete genomes (≥90% completeness) and species with at least ten independent conspecific genomes were analyzed. The overall pattern of gene frequency within each of the 781 species considered here showed a distinctive bimodal distribution (Extended Data Fig. 9), with most genes classified as either core or rare (that is, present in ≥90% or <10% of conspecific genomes, respectively). We analyzed the pan-genome size for each species in relation to the number of conspecific genomes to look for differences in intraspecies gene richness. We observed distinct patterns across different gut phyla, with species from various Firmicutes clades showing the highest rates of gene gain (Fig. 5a). There was wide variation in the proportion of core genes between species even among clades with more than 1,000 genomes (Fig. 5b), with a median core genome proportion (percentage of core genes among all genes in the representative genome) estimated at 66% (IQR = 59.6–73.9%).

To distinguish the functions encoded in the core and accessory genes, we analyzed their associated annotations. Core genes were well covered, with a median of 96%, 94%, 92% and 69% of the genes assigned with an eggNOG, InterPro, COG and KEGG annotation, respectively (Fig. 5c). In contrast, the accessory genes had a significantly higher proportion of unknown functions (P < 0.001), with a median of 21% of the genes (IQR = 16.7–27.3%) lacking a match in any of the databases considered. Thereafter, we investigated the functions encoded by the core and accessory genes on the basis of the COG functional categories. Genes classified as core were significantly associated (adjusted P < 0.001) with key metabolic functions involved in nucleotide, amino acid and lipid metabolism, as well as other housekeeping functions (for example, related to translation and ribosomal structure; Fig. 5d). In contrast, accessory genes had a much greater proportion of COGs without a known function and of genes involved in replication and recombination, which are typically found in mobile genetic elements (MGEs; Fig. 5d). A significant number of accessory genes were related to defense mechanisms, which encompass not only general mechanisms of antimicrobial resistance (AMR) such as ABC transporter efflux pumps but also systems targeted toward invading MGEs (for example, CRISPR–Cas and restriction modification systems against bacteriophages). These results highlight the potential of this resource to provide better understanding of the dynamics of chromosomally encoded AMR within the gut and allow deciphering of the extent to which the microbiome may be a source of both known and novel resistance mechanisms.

We next investigated intraspecies single-nucleotide variants (SNVs) within the UHGG species. We generated a catalog consisting of 249,435,699 SNVs from 2,489 species with three or more conspecific genomes (Fig. 6a). For context, a previously published catalog contained 10.3 million single-nucleotide polymorphisms from 101 gut microbiome species45. Of note, more than 85% of these SNVs were exclusively detected in MAGs, whereas only 2.2% were exclusive to isolate genomes (Fig. 6b). We found the overall pairwise SNV density between MAGs to be higher than that observed between isolate genomes (Fig. 6c). This was irrespective of the level of strain heterogeneity of the MAGs, as there was no correlation between SNV density and the degree of strain heterogeneity estimated with CMseq (Extended Data Fig. 10). Next, we assigned the detected SNVs to the continent of origin of each genome and observed that 36% of the SNVs were continent specific. Notably, genomes with a European origin contributed to the most exclusive SNVs (Fig. 6d). However, genomes from Africa contributed over three times more variation on average than European or North American genomes. Pairwise SNV analysis also supported a higher cross-continent SNV density, especially between genomes from Africa and Europe (Fig. 6e). Our results suggest that there is high strain variability between continents and that a considerable level of diversity remains to be discovered, especially from under-represented regions such as Africa, South America and Oceania.

### Resource implementation

Both the UHGG and UHGP catalogs are available as part of a new genome layer within the MGnify46 website, where summary statistics of each species cluster and their functional annotations can be interactively explored and downloaded (see ‘Data availability’ for more details). We have generated a Bitsliced Genomic Signature Index (BIGSI)47 of the UHGG catalog, which allows users to interactively query sequence fragments <5 kb in length to search for similar sequences in this collection.

We plan to periodically update the resource (approximately every 6–12 months) as new genomes are generated and made publicly available. MAGs will be retrieved from the European Nucleotide Archive (ENA), where a new MAG analysis class was recently implemented48. Genomes (MAGs or isolates) will be incorporated in the resource either as new species or by replacing uncultured reference genomes with better quality versions. We will adopt a versioning system whereby previous iterations of the catalog will still be accessible after subsequent updates to ensure reproducibility.

## Discussion

We have generated a unified sequence catalog representing over 200,000 genomes and 171 million protein sequences of the human gut microbiome. Of the 4,644 species contained in the UHGG catalog, 71% lack a cultured representative, meaning that the majority of microbial diversity in the catalog remains to be experimentally characterized. During preparation of our manuscript, a new collection of almost 4,000 cultured genomes from 106 gut species was released49, which will be incorporated in future versions of the resource. As 96% of these genomes were reported to have a species representative in the culture collections included here, we do not anticipate that this dataset will provide a substantial increase in the number of species discovered. Nevertheless, our analyses suggest that additional uncultured species from the human gut microbiome are yet to be discovered, highlighting the importance and need for culture-based studies. Furthermore, given the sampling bias toward populations from China, Europe and the United States, we expect that many under-represented regions still contain substantial uncultured diversity.

By comparing recently published large datasets of uncultured genomes16,18,20, we were able to assess the reproducibility of the results from each study. We show that, despite the different assembly, binning and refinement procedures used in the three studies, almost all of the same species and similar strains were recovered independently when using a consistent sample set. Although these results increase confidence in the use of MAGs, new methods for metagenome assembly, binning and quality control continue to be developed to overcome existing limitations, meaning that improved versions of the MAGs included here will likely be generated in the future.

With the establishment of this massive sequence catalog, it is evident that a large portion of the species and functional diversity within the human gut microbiome remains uncharacterized. Moreover, knowledge of the intraspecies diversity of many species is still limited owing to the presence of a small number of conspecific genomes. Having this combined resource can help guide future studies and prioritize targets for further experimental validation. Using the UHGG or UHGP catalogs, the community can now screen for the prevalence and abundance of species or genes in a large panel of intestinal samples and in specific clinical contexts. By pinpointing particular taxonomic groups with biomedical relevance, more targeted approaches could be developed to improve understanding of their role in the human gut. The functional predictions generated for the species pan-genomes could also be leveraged to develop new culturing strategies for isolation of candidate species. Target-enrichment methods such as single-cell50 and/or bait-capture hybridization51 approaches could also be applied. Given the large uncultured diversity still remaining in the human gut microbiome, having a high-quality catalog of all currently known species substantially enhances the resolution and accuracy of metagenome-based studies. Therefore, the presented genome and protein catalogs represent a key step toward a hypothesis-driven, mechanistic understanding of the human gut microbiome.

## Methods

### Genome collection

We compiled all the prokaryotic genomes publicly available as of March 2019 that were sampled from the human gut. To retrieve isolate genomes, we surveyed the IMG24, NCBI22 and PATRIC23 databases for genome sequences annotated as having been isolated from the human gastrointestinal tract. We complemented this set with bacterial genomes belonging to two recent culture collections: the HBC19 and CGR21. To avoid including duplicate entries due to redundancy between reference databases, we combined genomes obtained from the PATRIC and IMG repositories and added only those without an identical genome in the sets extracted from NCBI, HBC and CGR. This was determined by comparing isolate genomes between different databases using Mash v2.1 (ref. 26; ‘mash dist’ function) and only selecting one genome among those estimated to be identical (Mash distance of 0). MAGs (that is, uncultured genomes) were obtained from Pasolli et al.20 (CIBIO), Almeida et al.18 (EBI) and Nayfach et al.16 (HGM). For the CIBIO set, only genomes retrieved from samples collected from the intestinal tract were used.

Metadata for each genome were first retrieved from the five large human gut studies16,18,19,20,21. These were further enriched with data obtained using the ENA API (https://www.ebi.ac.uk/ena/portal/api) and the NCBI E-utilities (http://eutils.ncbi.nlm.nih.gov/). Metadata on the isolate genomes from IMG and PATRIC were retrieved using the GOLD52 system and the PATRIC FTP website (ftp://ftp.patricbrc.org/patric2/current_release/RELEASE_NOTES/genome_metadata), respectively. We only extracted metadata on the geographic origin of each genome, as other factors such as disease status and demographic information were missing from most of the samples.

### Assessing genome quality

Assembly statistics were calculated with the ‘stats.sh’ script from BBMap v38.75 (https://sourceforge.net/projects/bbmap/). Genome quality (completeness and contamination) was estimated with CheckM v1.0.11 (ref. 53) using the ‘lineage_wf’ workflow to select only genomes that passed the following criteria: >50% genome completeness, <5% contamination and an estimated quality score (completeness – 5 × contamination) > 50. We also searched for the presence of rRNAs in each genome with the ‘cmsearch’ function of INFERNAL v1.1.2 (ref. 54; options ‘-Z 1000 --hmmonly --cut_ga --noali --tblout’) against the Rfam55 covariance models for the 5S, 16S and 23S rRNAs. tRNAs of the standard 20 amino acids were identified with tRNAScan-SE v2.0 (ref. 56) with options ‘-A -Q’ for archaeal species and ‘-B -Q’ for species belonging to bacterial lineages.

To investigate the level of strain heterogeneity represented within each MAG, we used the CMseq tool (https://github.com/SegataLab/cmseq) as previously described20. Briefly, metagenomic reads from the sample used to generate the MAG were aligned to the respective MAG using bowtie v2.2.3 (ref. 57), with the resulting alignment file indexed and sorted with samtools v1.5 (ref. 58). The level of strain heterogeneity was estimated with the ‘polymut.py’ script from the CMseq package by calculating the number of nonsynonymous substitutions detected out of all positions mapped with a depth of coverage of at least 10 reads and base quality of at least 30 (a minimum of 100 positions were needed to estimate strain heterogeneity).

### Species clustering

We clustered the total set of 286,997 genomes at an estimated species level (ANI ≥ 95%; ref. 27) using dRep v2.2.4 (ref. 59) with the following options: ‘-pa 0.9 -sa 0.95 -nc 0.30 -cm larger’. Because of the computational burden of clustering the entire genome set, we used an iterative approach where random chunks of 50,000 genomes were clustered independently. The selected representatives from each chunk were combined and subsequently clustered, reducing the final computational load. To ensure that the best quality genome was selected as the species representative in each iteration, a score was calculated for each genome on the basis of the following formula:

$${\mathrm{Score}} = {\mathrm{CMP}} - 5 \times {\mathrm{CNT}} + 0.5 \times {\mathrm{log}}\left( {{\mathrm{N}}50} \right)$$

where CMP represents the completeness level, CNT is the estimated contamination and N50 is the assembly contiguity characterized by the minimum contig size in which half of the total genome sequence is contained. The genome with the highest score was chosen as the species representative, with cultured genomes prioritized over uncultured genomes (that is, if a MAG had a higher score than an isolate genome, the latter would still be chosen as the representative).

To further investigate the within-species population diversity, we calculated pairwise distances for all conspecific genomes using Mash v2.1 (ref. 26; default sketch size). From these results, we generated individual distance trees for each species using the ‘complete’ hierarchical clustering method implemented in the Fastcluster R package60. We calculated the number of clusters recovered using a distance cutoff of 0.03 (97% ANI) and 0.01 (99% ANI).

### Evaluating reproducibility of the methods

The species clusters inferred here were compared with those previously generated in human gut MAG studies16,18,20 from a common set of genomes. Similarity between species clusterings was estimated using the adjusted Rand index (ARI) computed in the Scikit-learn Python package61. This metric considers both the number of clusters and cluster membership to compute a similarity score ranging from 0 to 1.

Conspecific genomes recovered in the same metagenomic samples but in different studies were compared with FastANI v1.1 (ref. 27) with default parameters to obtain both the maximum AF and ANI for each pairwise comparison.

### Inferring cultured status

To determine cultured status, the UHGG species representatives were searched against NCBI RefSeq release 93 after excluding uncultured genomes (that is, metagenome-assembled or single-cell amplified genomes). Genome alignments were performed in two stages: (1) Mash v2.1 (ref. 26) was used as an initial screen (using the function ‘mash dist’) to identify the most similar RefSeq genome to each of the UHGG species and (2) ‘dnadiff’ from MUMmer v4.0.0beta2 (ref. 62) was subsequently used to compute whole-genome ANI for the genome pairs. A species was considered to have been cultured if (1) it contained a cultured gut genome from the UHGG catalog or (2) it matched an isolate RefSeq genome with at least 95% ANI over at least 30% of the genome length. Available metadata related to each RefSeq genome were retrieved from the ENA API (https://www.ebi.ac.uk/ena/portal/api/) using the corresponding BioSample accession.

### Calculating the number of conspecific genomes

For an accurate assessment of the number of nonredundant genomes belonging to each species, we de-replicated all conspecific genomes at a 99.9% ANI threshold using dRep with options ‘-pa 0.999 --SkipSecondary’. Furthermore, the frequency of each species was only counted once per sample to avoid cases where the same genome was recovered multiple times because of overlapping samples between the three MAG studies.

### Estimating geographic diversity

A geographic diversity index was estimated to assess how widely distributed each species was. We calculated the Shannon diversity index on the proportion of samples in which each species was found per continent. This metric combines both richness and evenness, such that the level of estimated diversity is highest in species found across all continents at a similar proportion.

### Phylogenetic analyses

Taxonomic annotation of each species representative was performed with GTDB-Tk v0.3.1 (refs. 29,30; database release 04-RS89) using the ‘classify_wf’ function and default parameters. To use consistent species boundaries between the genome clustering and taxonomic classification procedures, genomes were assigned at the species level if the ANI to the closest GTDB-Tk species representative genome was ≥95% and the AF was ≥30%. In this taxonomy scheme, genera and species names with an alphabetic suffix indicate taxa that are polyphyletic or needed to be subdivided on the basis of taxonomic rank normalization according to the current GTDB reference tree. The lineage containing the type strain retains the unsuffixed (valid) name, and all other lineages are given alphabetic suffixes, indicating that they are placeholder names that need to be replaced in due course. Taxon names above the rank of genus appended with an alphabetic suffix indicate groups that are not monophyletic in the GTDB reference tree but for which there exists alternative evidence that they are monophyletic groups. We also generated NCBI taxonomy annotations for each species-level genome on the basis of its placement in the GTDB tree, using the ‘gtdb_to_ncbi_majority_vote.py’ script available in the GTDB-Tk repository (https://github.com/Ecogenomics/GTDBTk/).

Maximum-likelihood trees were generated de novo using the protein sequence alignments produced by GTDB-Tk: we used IQ-TREE v1.6.11 (ref. 66) to build a phylogenetic tree of the 4,616 bacterial and 28 archaeal species. The best fit model was automatically selected by ‘ModelFinder’ on the basis of the Bayesian information criterion (BIC) score. The LG+F+R10 model was chosen for building the bacterial tree, while the LG+F+R4 model was used for the archaeal phylogeny. Trees were visualized and annotated with Interactive Tree Of Life (iTOL) v4.4.2 (ref. 67). Phylogenetic diversity (PD) was estimated by the sum of branch lengths, with the amount that was exclusive to uncultured species calculated as PDtotal – PDcultured. Uncultured monophyletic groups were defined as nodes in the tree containing child leaves exclusively comprising uncultured genomes.

### BIGSI construction

A BIGSI47 was generated for all species-level genomes with BIGSI v0.3.8. First, k-mers of size 31 were extracted from each genome with McCortex v1.0.1 (ref. 68; ‘mccortex31 build -k 31’). Thereafter, Bloom filters were built for each k-mer set using ‘bigsi bloom’ and inserted into the BIGSI index with ‘bigsi build’. BIGSI config parameters h (number of hash functions applied to each k-mer) and m (Bloom filter’s length in bits) were set at 1 and 28,000,000, respectively. A final API layer for querying the index was built using hug (http://www.hug.rest/) and hosted on the MGnify46 website at https://www.ebi.ac.uk/metagenomics/genomes.

### Pan-genome analysis and functional annotation

Protein-coding sequences (CDS) for each of the 286,997 genomes were predicted and annotated with Prokka v1.13.3 (ref. 69), using Prodigal v2.6.3 (ref. 70) with options ‘-c’ (predict proteins with closed ends only), ‘-m’ (prevent genes from being built across stretches of sequence marked as Ns) and ‘-p single’ (single mode for genome assemblies containing a single species). Pan-genome analyses were carried out using Roary v3.12.0 (ref. 71). We set a minimum amino acid identity for a positive match at 90% (‘-i 90’), a core gene defined at 90% presence (‘-cd 90’) and no paralog splitting (‘-s’). Normalized pan-genome size was estimated by dividing the total number of core and accessory genes by the number of genes contained in the species representative genome.

The UHGP catalog was generated from the combined set of 625,255,473 CDS predicted. Protein clustering of the UHGP and IGC5 was performed with the ‘linclust’ function of MMseqs2 v6-f5a1c72 with options ‘--cov-mode 1 -c 0.8’ (minimum coverage threshold of 80% the length of the shortest sequence) and ‘--kmer-per-seq 80’ (number of k-mers selected per sequence, increased from the default of 21 to improve clustering sensitivity). The ‘--min-seq-id’ option was set at 1, 0.95, 0.9 and 0.5 to generate the catalogs at 100%, 95%, 90% and 50% protein identity, respectively. We clustered the IGC only at 90% and 50% protein identity, as it was originally de-replicated at 95% nucleotide identity5. Functional characterization of all protein sequences was performed with eggNOG-mapper v2 (ref. 73; database v5.0 (ref. 41)) and InterProScan v5.35-74.0 (ref. 42). COG43, KEGG44, CAZy74 and viral annotations were derived from the eggNOG-mapper results. Differences in annotation coverage and COG functional categories between the core and accessory genes were evaluated with two-tailed Wilcoxon rank-sum tests in R v3.6.0 (function ‘wilcox.test’). Expected P values were corrected for multiple testing with the Benjamini–Hochberg method. Cohen’s d effect sizes were estimated with the function ‘cohen.d’ from the Effsize75 R package. To accurately estimate the proportion of each KEGG module in the species pan-genome, we used the compositional data analysis R package CoDaSeq76. Pseudocounts for zero-count data were first imputed using a ﻿Bayesian–multiplicative simple replacement procedure implemented in the ‘cmultRepl’ function (method ‘CZM’). Final counts were thereby converted to centered log ratios using the ‘codaSeq.clr’ function to account for the compositional nature of the data and for differences in pan-genome size.

### SNV analyses

A total of 2,489 species with at least three conspecific genomes were used to generate a catalog of SNVs. For each species, we mapped all conspecific genomes to the representative genome using the ‘nucmer’ program from MUMmer v4.0.0.beta2 (ref. 62) and filtered alignments using the ‘delta-filter’ program with options ‘-q -r’ to exclude chance- and repeat-induced alignments. Thereafter, we identified SNVs using the ‘show-snps’ program. Single-base insertions and deletions were not counted as SNVs. Each SNV locus was included in the catalog only when the alternate allele was detected in at least two conspecific genomes. The final SNV catalog was generated by unifying the SNV coordinates on the basis of their position in the species representative genome. The SNV entries in the catalog were characterized as genome type or continent specific on the basis of whether the alternate allele could be found solely in genomes from a specific genome type or continent. The number of continent-specific SNVs was normalized by the number of genomes from the corresponding continent to estimate the contribution per genome to the continent-specific SNV discoveries.

Similar programs and parameters were used for the pairwise genome alignment, but in this case only near-complete genomes (≥90% completeness) and species with at least ten independent conspecific genomes were considered. Because of the high computational demand, pairwise alignments of species encompassing more than 1,000 genomes were limited to the 1,000 best quality genomes. A total of 29,283,684 pairwise genome alignments were performed between almost 113,000 genomes from 909 species. For each pairwise comparison, we estimated the total number of SNVs and the overall density as the number of SNVs per kilobase. In addition, the pairwise comparisons were organized on the basis of the type and continent origin of the genomes in the pair for further downstream analyses. A two-tailed Wilcoxon rank-sum test was used to evaluate differences in SNV distribution. Resulting P values were corrected for multiple testing with the Benjamini–Hochberg method.

### Reporting Summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.