FastAAI: efficient estimation of genome average amino acid identity and phylum-level relationships using tetramers of universal proteins

Abstract

Estimation of whole-genome relatedness and taxonomic identification are two important bioinformatics tasks in describing environmental or clinical microbiomes. The genome-aggregate Average Nucleotide Identity is routinely used to derive the relatedness of closely related (species level) microbial and viral genomes, but it is not appropriate for more divergent genomes. Average Amino-acid Identity (AAI) can be used in the latter cases, but no current AAI implementation can efficiently compare thousands of genomes. Here we present FastAAI, a tool that estimates whole-genome pairwise relatedness using shared tetramers of universal proteins in a matter of microseconds, providing a speedup of up to 5 orders of magnitude when compared with current methods for calculating AAI or alternative whole-genome metrics. Further, FastAAI resolves distantly related genomes related at the phylum level with comparable accuracy to the phylogeny of ribosomal RNA genes, substantially improving on a known limitation of current AAI implementations. Our analysis of the resulting AAI matrices also indicated that bacterial lineages predominantly evolve gradually, rather than showing bursts of diversification, and that AAI thresholds to define classes, orders, and families are generally elusive. Therefore, FastAAI uniquely expands the toolbox for microbiome analysis and allows it to scale to millions of genomes.

Graphical Abstract

Graphical Abstract.

Introduction

The global diversity of microbial and viral genomes is very large, estimated at over a billion species of (just) bacteria, and most of it remains undiscovered [1]. As genome sequencing can help characterizing this diversity and has recently become routine, most microbial scientists have been overwhelmed by the amount of available data. For many researchers, a complete and thorough analysis of the available genome data is not necessary. Instead, these users would be better served with straightforward methods that can identify their unknown DNA/RNA sequence(s) and be able to discriminate between well-understood taxa and those that are potentially novel. Thus, tools that can help researchers identify the most interesting or relevant genomes to their unknown query genome(s) among thousands of candidates are important. Furthermore, metagenomics, which is the sequencing of environmental DNA, allows the genetic characterization of the majority of these microbes that have resisted cultivation in the laboratory. However, current tools to analyze metagenomic data are clearly lagging behind the development of sequencing technologies (and data), and do not scale with the number of metagenome-assembled genomes (MAGs), currently in the order of hundreds of thousands, that have become available [2]. This is a major limitation for better understanding, studying, and communicating about the biodiversity of uncultivated microorganisms that run the life-sustaining biogeochemical cycles on the planet, form critical associations with their plant and animal hosts, or produce products of biotechnological value.

One commonly used pair of methods to determine the level of novelty of a newly described taxon relative to the taxa already available in reference databases are genome-aggregate Average Nucleotide Identity (ANI) and Average Amino-acid Identity (AAI). ANI represents the average nucleotide identity of all genomic regions shared between any two genomes and offers robust resolution between strains of the same or closely related species (i.e. showing 80%–100% ANI). However, ANI is not appropriate for estimating genome relatedness (and thus, classification) among more divergent genomes because nucleotide substitutions saturate and/or sequences cannot be reliably aligned at this level [3]. Instead, the AAI should be used for moderately divergent genomes as it effectively delineates relationships at the genus and family levels, similarly to ANI for the species level. It is also important to note that the ANI/AAI-based approach can be applied to any unknown genome sequence, even if it is incomplete, that encodes at least a few genes shared with at least a few reference genomes [4], whereas alternative approaches based on universal genes such as the small subunit ribosomal RNA (SSU rRNA) gene are limited to sequences carrying the corresponding genes. Accordingly, the ANI/AAI approach has facilitated the analysis of an increasing number of MAG, single-cell amplified genome or viral genome sequences, and the need to perform whole-genome ANI or AAI calculations has grown exponentially in recent years.

Despite these strengths, traditional ANI/AAI calculation is based on pairwise sequence alignments, e.g. BLAST [5], and thus does not scale well with the increasing number of genomes. For instance, performing all versus all ANI calculation among the ∼13 000 reference genomes (∼169 million pairwise comparisons) of all bacterial species taxonomically described to date [6] required ∼1 month of processing time on a computer cluster of 500 CPUs. Performing a similar comparison among 100 000 genomes (10 billion comparisons) is expected to take >1 year due to the quadratic increase in the number of comparisons to perform. This is prohibitively expensive, even for larger computer clusters. While faster ANI implementations based on the numbers of k-mers shared between genomes have been recently described by our team (e.g. FastANI) [4] and others (Skani) [7], and have been widely used, current AAI implementations are not fast enough, especially considering that the majority of genomes in the reference database are moderately or distantly related to each other. Further, AAI does not provide robust resolution at the inter-domain and phylum levels (that is, between distantly related genomes) compared to the phylogenetic analysis of the SSU rRNA or other universal genes. Alternative approaches such as phylogenetic placement based on universal genes are accurate [8, 9], but are also computationally expensive and do not scale well with the exponentially increasing number of microbial genomes recovered from environmental samples. Thus, there is an urgent need for a scalable tool to calculate genome relatedness, especially for deep-branching genomes with no close relatives available.

Here we present FastAAI, a bioinformatics tool aiming at solving the issues related to the speed and scalability of the AAI estimation among bacterial and archaeal genomes while providing higher resolution at the phylum and domain level. FastAAI is a standalone Python tool that leverages protein tetramer information in single-copy proteins (SCPs) broadly present in prokaryotic genomes (sometimes termed “universal” genes) to calculate genome similarities that are subsequently translated into traditional AAI values of relatedness. The current implementation of FastAAI can perform pairwise genome comparisons in a matter of microseconds and scales exceedingly well with a larger number of genomes compared.

Materials and methods

FastAAI implementation

FastAAI is divided into four steps (see Supplementary Fig. S17 for a graphical representation of the workflow). Briefly, FastAAI performs, (i) identification of marker proteins in each genome using a set of 122 universal SCP represented by hidden Markov models (HMMs) (Supplementary Table S6, and below), (ii) extraction of tetramers for each protein per genome, (iii) estimation of Jaccard indices based on the tetramers extracted from proteins shared by a genome pair, (iv) estimation of the average Jaccard index () followed by its translation to estimated AAI values (termed to distinguish it from AAI values estimated based on alignment-based methods). The proteins of a genome are predicted with Pyrodigal [10, 11], with FastAAI automatically selecting the optimal translation table between codon tables 4 and 11 based on which table gives more and longer predicted genes and are searched against a collection of pre-built HMMs with PyHMMER [12, 13] using profile trusted cutoffs (see Supplemental Methods for additional details on Pyrodigal and PyHMMER). The HMM models used by FastAAI are the Pfam [14] HMM collections for a set of 122 proteins previously reported to occur in most prokaryotic genomes in a single copy. These proteins were originally curated for use in the GToTree [15] tool and represent all Pfam models which were recovered in exactly one copy in at least 90% of appropriate domain (either bacterial or archaeal) genomes on RefSeq as of 9 December 2019. The exact model collections used by FastAAI can be found at GToTree/hmm_sets/hmm-sources-and-info.tsv under the Archaea.hmm and Bacteria.hmm collections; FastAAI uses all of the unique members of these two HMM sets. The second step consists of extracting all tetramers per marker protein. Tetramers are considered as present or absent regardless of the number of times they occur in a protein sequence. In the third step, for each pair of genomes, FastAAI identifies shared marker proteins; if a protein is only present in one genome, it is excluded from the comparison. Then, for each shared marker protein in the genome pair under comparison, a Jaccard similarity value is calculated based on the shared versus total tetramers of the corresponding protein sequences (graphical abstract, panel A, and Supplementary Fig. S17). Finally, FastAAI calculates the average of the previously estimated Jaccard values for all shared proteins (), the standard deviation, and uses values to estimate AAI values (; read below for estimation formula; see also Supplementary Material about why pooling together tetramers from all shared SCPs to calculate Jaccard similarity or using a smaller k-mer size than four were not implemented in FastAAI).

Testing datasets

To test performance and speed, we evaluated FastAAI on three primary datasets (Supplementary Table S1). The first dataset (RefSeq) comprised all complete reference genomes from the NCBI RefSeq database with available taxonomic classification as of 13 November 2019. In total, this dataset was composed of 5000 bacterial and 328 archaeal genomes. The second dataset (TypeMat) included 10 573 genomes (10 200 Bacteria and 373 Archaea) from type material with standing in nomenclature, which are available through the MiGA webserver with release number r2019-02 [6, 16]. Genomes in either dataset with completeness estimates below 50% or contamination higher than 30% based on MiGA estimates were removed from further analyses as low-quality genomes. Finally, the third dataset (labeled PhylaLite) was composed of 39 complete genomes (36 Bacteria and 3 Archaea), one genome per phylum from the TypeMat collection. In addition to these primary datasets, we included simulated datasets derived from a subset of genomes in the RefSeq dataset (n = 500) to evaluate FastAAI when genomes had different levels of completeness and contamination. For this, we created in silico genomes with varying degrees of completeness by randomly subsampling the single-copy marker proteins of these RefSeq genomes to a given level of completeness. Then, each of the resulting genomes was modified to increase contamination by randomly adding single-copy marker proteins from distantly related organisms to achieve a given level of contamination. All metrics used, including length, N50, completeness, and contamination, are shown for all datasets in Supplementary Fig. S2.

FastAAI’s method of estimating AAI

FastAAI utilizes PyHMMER to identify proteins in a genome’s proteome as being one of the 122 prokaryotic SCPs (discussed above). FastAAI considers a protein to be a particular SCP when PyHMMER identifies the protein’s best match to be the SCP in question and the protein has no better match among the remaining SCPs. Consequently, a single protein can only be classified as one SCP and each SCP can have only one protein as its representative within a single genome. No information about proteins that are not labeled as an SCP representative is retained by FastAAI; as a result, a genome can have at most 122 proteins considered in the process of estimating AAI (and typically around 80 SCPs since several SCPs are domain-specific; Supplementary Fig. S2). To estimate AAI for a pair of genomes, FastAAI determines which SCPs the two genomes share and collects the SCP representative proteins of each in matched-SCP pairs. FastAAI then calculates the Jaccard index between the tetramers of each pair and averages them to produce an initial estimator that we label as (graphical abstract, panel A, and Supplementary Fig. S17).

Comparison of FastAAI to DIAMOND-based AAI using the RefSeq dataset revealed a linear correspondence between and AAI values [R²: 0.96, root mean squared error (RMSE): 2.22%; Supplementary Fig. S12A]. However, despite the low RMSE and the high R-squared of the linear regression, AAI values predicted using the regression coefficients were consistently lower compared to the actual AAI values, especially in the inter-phylum and inter-class levels, as evidenced by the curvature over the regression line at these regions. To solve this and improve the correspondence between FastAAI’s estimate and AAI values, we transformed by applying several transformations to linearize the relationship. The linearization coefficients were optimized using a gradient descent method aiming at decreasing the RMSE, excluding values with > 0.9 (see below). Finally, to have a 1:1 correspondence, we applied a linear model on top of the initial transformation. The final transformation defined as the estimated AAI () is described by the following formula:

The linear regression of AAI versus values in the 0%–100% range shows a slightly lower R-squared of 0.95 and a higher RMSE (22.02%; Supplementary Fig. S12B). This underperformance is due to data points with values higher than 0.9 (and thus, the corresponding values of ) collapsing to the AAI range of ∼95%–100%, increasing the regression error and generating curvature in these portions of the plot (Supplementary Fig. S12B). The low correspondence between FastAAI’s and AAI values in the upper range is presumably due to the high sequence conservation of the universal genes (thus, low resolving power) at this (the species) level. Therefore, FastAAI summarizes these high values as “>90% AAI” and we recommend using instead ANI (e.g. FastANI) or traditional AAI to compare genomes related at this level, as also suggested previously [3, 4]. As AAI values tend to remain >30% even across domains, FastAAI is also insensitive below 30% AAI and so summarizes estimates at or below 30% as “<30% AAI”.

Estimation of 16S rRNA gene identities, AAI, and GTDBtk-RED values

We used INFERNAL (INFERence of RNA ALignment) v1.1.2 [17] to identify and extract 16S rRNA gene sequence identities for all genomes in the RefSeq dataset. In cases where more than one 16S rRNA gene sequence was found in a single genome, the longest sequence was selected as the representative for the genome, or if multiple sequences had the same length, the representative was randomly chosen among them. We then performed pairwise alignments and determined the identities of each pair of sequences using a custom python script (available at https://github.com/cruizperez/Bioinformatic_Tools), which uses the Needleman–Wunsch algorithm with the same parameters as implemented in EMBOSS Needle [18]. We used the aai.rb tool from the Enveomics collection [19] with DIAMOND as the searching tool to calculate pairwise AAI values. Finally, we performed classifications and extracted Relative Evolutionary Diverge (RED) values based on the Genome Taxonomy Database (GTDB) using GTDBtk with release #202 of the database.

FastAAI taxonomic classification accuracy

We evaluated the ability of FastAAI values to discriminate genomes at different taxonomic levels by using the TypeMat dataset as the ground truth and reference taxonomy. For this, we determined threshold values delimiting taxonomic ranks by finding valleys in the distribution of values between inter- versus intra-rank comparisons for the TypeMat genome dataset. For example, genomes assigned to the same domain typically show values above 35.3% and those assigned to different domain typically have values below 35.3%. Genomes in the RefSeq dataset (and their associated taxonomy, assumed to be correct) were treated as queries, and their taxonomic classification based on the estimated values against TypeMat genomes was compared to their available taxonomy (from RefSeq). For each rank, we built a confusion matrix where we labeled the classifications obtained based on value thresholds and compared them to the available taxonomic classification of the RefSeq query genome from NCBI as follows: True positive (TP) for pairs with > taxonomic rank threshold (or just “threshold” for simplicity) and where query and reference shared the same taxon, true negative (TN) for pairs with < threshold and where the pair belonged to a different taxon in the rank, false positive (FP) for pairs with > threshold and where the pair belonged to a different taxon in the rank, and false negatives (FN) for pairs with < threshold and where the pair belonged to the same taxon. From this confusion matrix, we calculated the accuracy, recall, precision, and F1 score for each taxonomic rank as follows:

See also supplementary online material that includes additional details on the FastAAI versus GTDB comparisons, the code implementation of the FastAAI algorithm, and FastAAI’s querying options, parallelization, and results formatting.

Results

We focus our analysis below primarily on bacterial genomes because of their large numbers in the public datasets. FastAAI applies similarly to archaeal genomes but for genomes of other domains (e.g. eukaryotic) adjustments in the collection of universal SCPs will be required for best results. For our accuracy evaluations, we primarily focus on the comparison of FastAAI to 16S rRNA gene identities since the latter gene represents the backbone of bacterial taxonomy and has been most commonly used to assess genetic relatedness among taxa/genomes. We provide computational and performance comparisons to traditional alignment-based methods such as DIAMOND [20] and EzAAI [21] as well as alternative methods such as GTDB-Tk [22].

FastAAI accurately estimates AAI

FastAAI is intended to replace traditional AAI in resolving relationships above the species level; at and within the species level, ANI is sensitive and efficient enough (e.g. FastANI implementation) and thus, already serves as a complementary method to AAI. Therefore, we focused our performance evaluation of FastAAI on genomes with a relatedness between 30% and 90% AAI. Using the RefSeq dataset (see the ‘Materials and methods’ section), we calculated a reference AAI value for each pair of genomes using DIAMOND and compared predictions made with FastAAI against these values. FastAAI’s estimator (see the ‘Materials and methods’ section for details on the calculation of achieves a highly accurate estimation of DIAMOND-based AAI between 30% and 90% , with an average prediction error of only 1.3% within this range (R-squared: 0.98, RMSE: 1.3%; Fig. 1; see Supplementary Fig. S12 for AAI <30% and >90%). The strong linear correlation between and DIAMOND-based AAI values demonstrates that FastAAI can provide a similarly robust estimate of genome relatedness above the species level (Fig. 2B and C). For the most distantly related genomes (e.g. related at around 30% AAI), the and DIAMOND-based AAI values deviated the most from each other, which we investigated below with comparisons against the 16S rRNA gene and showed that this was due to higher accuracy of at the domain and phylum levels.

Figure 1.

Identity correspondence between AAI and derived from the FastAAI value transformation using the RefSeq dataset (5328 genomes, 28.39 million pairwise comparisons). Each dot represents a pairwise comparison colored by the lowest taxonomic rank (figure key) that the two genomes in the pair share. Note the linear correspondence between both values and the low error associated with the linear regression.

Figure 2.

Distribution of 16S rRNA gene identities, AAI, and values, according to the lowest taxonomic level shared by the pairs of genomes analyzed. Underlying data are based on all versus all comparisons of the TypeMat genomes. Note the large overlaps between taxonomic ranks using all methods, but also the virtually no overlap between the inter- (red distribution) and intra-domain (yellow distribution) groups for , which is consistent to the picture based on 16S rRNA gene identities. Solid lines represent the means of the distribution while dotted lines represent one standard deviation around the mean.

FastAAI is faster than alternative AAI implementations

To demonstrate the gains in speed of FastAAI compared to alignment-based methods for calculating AAI, we performed an all-versus-all comparison of 100 randomly chosen RefSeq genomes (10 000 pairwise comparisons) using FastAAI, DIAMOND-based AAI [19], and the recently published EzAAI [21]. EzAAI uses MMseqs2 [23] as a substitute for either BLAST or DIAMOND as its alignment engine but calculates AAI in a fundamentally similar manner; that is, finding reciprocal best-match alignments between the proteins of each pair of genomes it processes and calculating AAI based on these best matches. As DIAMOND does not independently function as a pipeline for both protein prediction and alignment in the manner that FastAAI and EzAAI do, all three tools were supplied with the same set of predicted proteins generated from the original 100 RefSeq genomes using Prodigal [11]. Timing covered all of the remaining comparable steps of each tool, including data formatting, database creation as needed, the calculation of AAI, and writing results. All tools were installed via Anaconda and compared on Georgia Tech’s PACE Phoenix cluster utilizing small compute nodes equipped with two Intel Xeon Gold 6226 CPUs.

The average time per pairwise genome comparison was 9.6 s [standard deviation (SD) = 4.16] for DIAMOND (version 2.0.1), 32.0 s (SD = 12.1) for EzAAI (version 1.2), and 0.019 s (SD nonapplicable) for FastAAI. Although this test already demonstrates more than two orders of magnitude speedup over both DIAMOND and EzAAI, it is in fact a nearly worst-case scenario for demonstrating the speed gains of FastAAI relative to these tools. As can be seen in Supplementary Fig. S16, the relative speed gain per comparison of FastAAI over DIAMOND and EzAAI increases as the datasets processed grow larger, with FastAAI becoming >5 orders of magnitude faster per comparison than either DIAMOND or EzAAI in our largest test dataset (∼2.29 billion pairwise comparisons) when a run started from predicted proteins. Protein prediction is, in fact, the most rate limiting among all steps of FastAAI (Supplementary Fig. S15).

FastAAI is faster and has a lower RAM requirement than phylogeny-based methods

There are alternative methods to measure relatedness among genomes and/or taxonomically classify them aside from the traditional ANI/AAI; most notably, GTDB-Tk uses a combination of approximate maximum-likelihood-based phylogenetic placement and FastANI to classify query genomes. Because GTDB-Tk can use only its own prebuilt database, we created an equivalent FastAAI database from the same set of genomes that are used in the latest GTDB database (rel. 202, last updated Jul. 9^th, 2021), totaling 45 555 bacterial and 2 339 archaeal genomes. We randomly sampled three sets of 100 genomes from these 47 894 genomes and used these sets to test both GTDB-Tk and FastAAI by querying each set of 100 selected genomes against the full complement of 47 894 target genomes. To ensure that the comparison between FastAAI and GTDB-Tk was fair, both tools were run with 10 threads, supplied with 256 GB of RAM, used their respective premade databases as search targets, and were supplied with query sets composed of genomes (not proteins or HMM results), as this is the starting point for a GTDB-Tk search. In addition to memory usage, we measured both the wall time and CPU time for each tool, reflecting the user’s waiting time from program start to program end and the sum of computer time across all 10 threads, respectively. On average, FastAAI completed in 2.9% of the wall time, 6.5% of the CPU time, and required only 0.69% of the RAM required by GTDB-tk. All comparison results can be seen in Table 1.

Table 1.

FastAAI versus GTDB-Tk runtime and resource usage comparisons

Test	Wall Time	CPU Time	Max RAM use (GB)
FastAAI, Test 1	00:02:24	00:18:36	1.58
GTDBtk, Test 1	01:26:38	05:18:55	232.08
FastAAI, Test 2	00:02:20	00:19:09	1.56
GTDBtk, Test 2	01:17:41	04:12:42	231.67
FastAAI, Test 3	00:02:31	00:20:41	1.68
GTDBtk, Test 3	01:22:38	05:24:31	232.10

Computational resource comparison between FastAAI and GTDB-Tk were performed based on three parameters (columns) for 100 randomly selected GTDB representative genomes against the GTDB release 202. All times are reported as HH:MM:SS.

To further illustrate the minimal RAM required by FastAAI, we ran additional all-versus-all tests with FastAAI on varying sizes of random samples of the GTDB genomes, ranging in size from 100 versus 100 to the entirety of GTDB release 202 (47 894 versus 47 894 genomes) using 14 processors (one complete node on our cluster). The whole GTDB all-versus-all test was the longest-running and most memory-intensive of these tests (∼2.29 billion pairwise comparisons), which FastAAI was able to complete in a wall time of 43 min, 15 s, a CPU time of 07:25:23, and using only 3.27 GB of RAM. Provided a preprocessed database (i.e. universal proteins identified and tetramers extracted), even a comparison of this enormous scale could be accomplished on a typical personal laptop with FastAAI within a day. Performing the same task with GTDB-Tk would take well in excess of 500 CPU hours and require hundreds of gigabytes of RAM that are available only on supercomputers. The results for the other tests can be seen in Supplementary Tables S5, S9, and S10.

FastAAI provides the same taxonomic assignments as phylogeny-based methods

In addition to speed, we evaluated FastAAI’s ability to find the same best-match database genome as that identified by phylogeny-based approaches for query genomes of varied novelty relative to the database. For this, we first compared the AAI values estimated by FastAAI among 10 575 genomes selected from NCBI’s nonredundant database to the tree distances in a phylogeny of these same genomes constructed with ASTRAL [24], a state-of-the-art tool that reconciles multiple individual gene trees through maximizing shared tree topology, explicitly reconstructing the species tree instead of the average gene tree. We noted strong correspondence between estimated AAI and tree distances (R-squared = 0.70; Supplementary Fig. S11, and Supplementary Results). Similar results were also obtained with the RED values of GTDB-Tk assignments (Supplementary Figs S8 and S9). Furthermore, we classified a large collection of MAGs (n = 52 515) recently reported (GEMs collection [2]) against the GTDB database using GTDB-Tk and FastAAI. In 95.3% of the cases, GTDB-Tk and FastAAI agreed on taxonomic assignment down to the lowest level provided by GTDB-Tk and down to at least the genus level in 97% of cases. The remaining ∼3% of the cases represented mostly genomes that were novel at higher taxonomic levels (e.g. above family), and we were unable to fully assess the quality of these assignments due to the difficulty on confidently aligning the SCP sequences at this level and/or the putatively chimeric sequences of several of the corresponding MAGs (e.g. conflicting signal between different SCP). At this level of taxonomic novelty, we believe that our comparison to the 16S rRNA gene based on a highly curated dataset (RefSeq genomes) is more appropriate (see below). Nonetheless, our evaluation demonstrated that FastAAI maintains high agreement in its calls (∼97% for our test datasets) relative to slower phylogenetic classification approaches (Supplementary Tables S3, S4, and S8).

We also tested FastAAI’s ability to recover larger groupings of genomes by examining revisions to the genus Clostridium suggested by phylogenomic approaches [8]. Specifically, these phylogenomic results indicate that Clostridium should be divided into three new genera to resolve polyphyly within the canonical (i.e. NCBI-taxonomy) genus. We performed an all-versus-all AAI comparison of genomes in the phylum Bacillota (formerly Firmicutes) collected from RefSeq, then examined the subset of genomes overlapping with those in the revised genera. FastAAI’s results agree heavily with the reclassifications, consistently finding higher and nonoverlapping AAI values for genome pairs within the suggested reclassification genera than to other genomes in the canonical Clostridium genus. FastAAI also consistently identifies those within-genus relationships as the closest above-species matches for the entire collection of genomes in the Bacillota genera, even in instances where outside-genus genomes were still very similar. These results are expanded on in the supplementary results and are summarized in Supplementary Fig. S10.

Improved resolution at the phylum level and thresholds for describing other taxonomic ranks

Our comparison of DIAMOND-based AAI and 16S rRNA gene identities using the bacterial RefSeq genomes demonstrated a strong correlation (Supplementary Fig. S9A), with overlaps between adjacent taxonomic ranks in terms of the AAI values. For example, about 30% of the genomes grouped at the genus level as their lowest shared rank show similar AAI values to genomes grouped at the family level, consistent with the patterns observed in previous studies based on smaller genome datasets [8]. In these comparisons, we observed that the lowest values for 16S rRNA gene identities (70%–80%) corresponded to (traditional) AAI values between 30% and 40%. Importantly, the same range in AAI values (30%–40%) also included genome pairs with higher 16S rRNA gene identities (80%–85%), revealing substantial overlaps in AAI values at higher taxonomic levels (i.e. low resolution), as noted previously [25]. However, in the case of FastAAI, 16S rRNA gene identities between 70% and 80% corresponded to FastAAI’s values between 32.28 and 33.57 with almost no overlap with same-versus-different-domain level comparisons (Fig. 2 and Supplementary Fig. S9B), revealing higher resolution of FastAAI at the phylum-to-domain level compared to those of AAI and 16S rRNA gene identities (also see Supplementary Fig. S9 for discussion of a few outlier genome pairs observed).

More specifically, for 16S rRNA gene identities, we identified the inter-domain and inter-phylum valleys at 70.4% and 80.2%, respectively (Fig. 2A). We observed that only 1.1% of same-domain comparisons were below 70.4%, while only 0.54% of different-domain comparisons were above this threshold, indicating limited overlap and good discrimination capacity at the domain level based on 16S rRNA gene identities. However, the same values for phylum-level comparisons revealed a lower discriminatory capacity; 34.7% of the comparisons were below the 80.2% threshold and 10.9% were above. Note that phyla are typically designated based on their branching patterns in the 16S rRNA gene phylogeny, not necessarily the 16S rRNA gene identity; hence, the results reported here do not necessarily reflect low performance of the 16S rRNA gene phylogeny approach. Our goal was to examine if convenient sequence identity thresholds for discriminating taxonomic ranks can be determined that could guide future taxonomic studies and description and to evaluate FastAAI’s domain- and phylum-level resolution. Compared to 16S rRNA gene identity values, we found that the thresholds discriminating domain- and phylum-level comparisons using FastAAI were 35.3% and 40.5% . Interestingly, we found that only 0.07% of different-domain and 0.03% of same-domain comparisons were below and above the domain-threshold (35.3%), indicating similar—if not higher—discriminatory power at this level for relative to the 16S rRNA gene. Phylum-level comparisons were nonetheless similar, if not slightly better, to those observed for 16S rRNA with 20.4% same-phylum and 6.47% different-phylum comparisons below and above the identified threshold (40.5 ), respectively. These results indicate that FastAAI offers a lower frequency of misclassifications (or higher resolving power) than (traditional) AAI and 16S rRNA gene identities at the domain and phylum levels for deep-branching genomes. It should be also noted that the large number of genome pairs sharing the same phylum rank but showing identities below the discrimination threshold in all three metrics indicates that at least some of the corresponding genomes may deserve a distinct (novel) phylum designation based on genomic relatedness. We evaluated the remaining taxonomic ranks (i.e. family and genus) in a similar fashion, and showed that reliable FastAAI thresholds can be established for each rank. These thresholds are available in Supplementary Table S7 which can be used to guide (future) taxonomic descriptions and classifications of new taxa.

We further interrogated the distributions of FastAAI-based AAI values within phyla to determine whether or not reliable AAI thresholds can be identified that distinguish different ranks, similar to the 95% ANI that distinguishes (different) species [4] or the thresholds that distinguish domain and phylum ranks mentioned above. We found that well-sampled phyla (see Supplementary Table S2 for which phyla were used) might show such AAI thresholds for separating (different) classes (of the same phylum), orders, and families at roughly 45%, 50%, and 55% AAI (Table 2), respectively. Notably, the distribution of AAI values within these ranks for the well-sampled phyla were relatively narrow/tight (Supplementary Figs S3, S4, and S6), further supporting the practical usefulness of these thresholds. In contrast, more poorly-sampled phyla displayed less consistency in the AAI thresholds to separate these ranks. This is especially true for Mycoplasmatota, which lacks any apparent distinction in AAI values at taxonomic distances above the genus level (Supplementary Fig. S5), suggesting inconsistent standards for the taxonomic classification of genomes of this phylum in terms of overall genomic relatedness compared to other phyla. In general, there was even more variability in the AAI value distributions to distinguish different genera of the same family or different species of the same genus across all phyla, with same-genus comparisons in particular exhibiting extremely wide ranges in AAI values between 50% and >90% AAI (e.g. Supplementary Fig. S7). Collectively these results reinforce our above finding that consistent AAI boundaries may be established for the higher taxonomic ranks (order, class), but may prove more challenging to establish and less effective at lower ranks (family, genus).

Table 2.

Average AAI and standard deviations for pairs of RefSeq Genomes calculated with FastAAI

Taxonomic distance	Average AAI	AAI standard deviation
Same phylum	43.29	2.26
Same class	49.38	4.15
Same order	58.13	9.09
Same family	81.34	8.32
Same genus	77.69	13.62
Same species	94.98	0.37

Phyla do not display evidence for punctuated evolution based on AAI patterns

Using our computed AAI matrices, we also tested if lineage evolution is characterized by “bursts” of diversification as suggested previously based on the progressive coarse graining of 16S rRNA gene phylogenetic trees (punctuated evolution; [26]) or instead evolution is more gradual, with a roughly constant rate of (new) lineage appearance. For this, we constructed neighbor joining phylogenies from the all-versus-all AAI matrices of each bacterial phylum and applied the coarse graining methodology and other related approaches (see supplementary material) to determine whether the same pattern of punctuation could be identified within each bacterial phylum based on AAI values. Our approaches produced findings inconsistent with the ”punctuated” evolution scenario, and instead supported slow and steady diversification over deep evolutionary time with significantly more action near the tips of the trees, presumably due to selection not having enough time to remove deleterious mutations at this timescale [27] (Supplementary Figs S3 and S14). One explanation for this discrepancy compared to the results of the previous study may be that evolutionary signal simply differs between the 16S rRNA gene and the whole-genome, including the limited resolution offered by the 16S rRNA gene among closely related genomes.

FastAAI is robust to varying levels of genome completeness and contamination

Most studies using culture-independent genomic techniques filter recovered prokaryotic genomes according to their quality level based on estimated completeness and contamination [28]. Gene phylogenies or classification methods that rely on a single or few genes are often affected by genome incompleteness due to the absence of the required gene(s) for the analysis. In addition, contaminating gene sequences (i.e. chimeric genomes) can be sources of error that can affect homology-based classification methods such as MEGAN [29] or MyTaxa [30]. In such cases, AAI is generally more robust because the number of proteins or genome fragments used in the calculations may be adequate even for relatively incomplete genomes (e.g. ∼200 shared proteins are typically adequate [4]). Given that FastAAI relies on a reduced set of universal proteins, it is important to assess the consistency of the results at variable levels of genome quality. For this, we selected 500 genomes from the RefSeq dataset that were estimated to be 100% complete with 0% contamination. We predicted the proteins for each selected genome and used FastAAI to calculate their genome relatedness against the smaller RefSeq dataset to create a baseline of reference values. Each set of proteins per genome was then randomly subsampled and in-silico contaminated to create a set of genomes with varying levels of contamination and completeness (see the ‘Materials and methods’ section and Supplementary Fig. S1). Each of these incomplete and contaminated genomes was subsequently searched against the RefSeq dataset and the resulting values of these searches were compared to the complete-genome results (baseline) to determine the degree of deviation from the baseline. Our results demonstrated that in the absence of contamination, FastAAI could recover almost identical values for genomes with varying levels of completeness, even down to 10% complete genomes, only deviating by, at most, 2.26% from the complete-genome-based values in the genomes of the lowest completeness (Supplementary Fig. S1). This result showed that FastAAI is robust even for incomplete genomes, as long as about eight of the universal proteins are present (roughly corresponding to 10% completeness). As expected, contamination resulted in an increased deviation of values from the baseline. In the 100% completeness genomes, the deviation increased from 0% (100% completeness, 0% contamination) to 0.95% (100% completeness, ∼37% contamination), a relatively small value considering the high level of contamination. This trend was also observed in lower completeness levels, where we observed the highest values of deviation, ranging from 2.26% to 3.98% in genomes with 10% completeness and contamination of up to 37% (a very low-quality genome; Supplementary Fig. S2). Therefore, considering that proposed classification standards for the quality of assembled genomes establish a low quality-draft as a genome with <50% completeness and <10% contamination [28], we expect FastAAI to perform well, with <1% deviation from the true estimate for such low-quality genomes.

Discussion

This study introduces FastAAI, a tool for the fast and accurate estimation of the whole-genome genetic relatedness and classification of microbial genomes that provides results that are similar, if not better, with methods such as AAI, SSU rRNA gene identities, and phylogenetic trees, while dramatically improving speed. The high accuracy in taxonomic assignments of query genomes that show different degrees of novelty relative to the database genome is apparently due to the robust phylogenetic signal that (shared) tetramers of universal genes carry. We considered alternative amino acid k-mer sizes, but discarded smaller values of k due to the high likelihood of k-mer overlaps occurring by chance (low specificity) and discarded larger values of k due to the insufficient sensitivity over the small sequence space of single proteins, an issue previously identified in other k-mer based approaches [31]. We also demonstrated substantial advantages over newer alternative tools for the classification of microorganisms in terms of speed and computational resources. The high accuracy and speed, especially in comparisons among deep-branching genomes (i.e. higher taxonomic levels), will be important for environmental and clinical samples that are expected to harbor substantial novel diversity [9].

FastAAI accelerates the calculation of AAI because it does not require the alignment of proteins between genomes, and instead estimates AAI between two genomes from the shared fraction of tetramers over a set of universal SCPs shared by the two genomes. This allows FastAAI to rapidly calculate AAI from massive datasets: even in all-versus-all comparisons of tens of thousands of genomes (i.e. hundreds of millions or billions of pairwise comparisons), FastAAI requires only a few hours of runtime and very modest computational resources. Indeed, the overwhelming majority of runtime for FastAAI in all of our applicable tests is consumed in the preprocessing stage, i.e. protein prediction and HMM searches to identify SCPs (Supplementary Fig. S15), so the average time per comparison between a pair of genomes decreases as the number of comparisons performed increases (Supplementary Fig. S16). This relative speedup is due to the fact that even though the total number of pairwise genome comparisons increases with the square of the number of genomes, the slow preprocessing step grows linearly and only the number of fast tetramer pairwise comparisons increases with the square of the genome number. Similarly, comparison of a single novel genome to a large, preprocessed reference data set requires additional preprocessing of only the novel genome, which dominates comparison times against even our large data sets. Notably, FastAAI retains its performance even in the presence of highly incomplete or contaminated genomes, which are commonly recovered by metagenomic surveys. Finally, for database searches, FastAAI reliably retrieves the closest relative in most cases even using a highly incomplete database, which we expect that will only improve as databases become more comprehensive with genome representatives from uncultured microorganisms.

Our results show that, in most cases, FastAAI alone is a sufficiently accurate estimator of genomic relatedness and other tools have modest additional benefits (see Supplemental Tables S3 and S4). However, no tool is without its limitations. FastAAI’s estimates of genomic relatedness are generally highly accurate and consistently able to place genomes among their most similar relatives for all datasets tested here. Despite this generally excellent performance in classification, FastAAI does not share the nearly perfect accuracy that analogous ANI estimators such as FastANI have achieved for closely related genomes, which have allowed these ANI estimators to simply replace traditional ANI estimation. Indeed, for a vanishingly small fraction of genome pairs (e.g. a few hundred of the total ∼2.2 billion comparisons within the GTDB rel. 207 dataset), FastAAI can produce inaccurate estimates of AAI that deviate from a traditionally calculated AAI by as much as 20%. These outliers did not display any unusual features in the number of SCPs recovered, the standard deviation of Jaccard values across their shared SCPs or in genome lengths, meaning the underlying reason for these errors may be related to chimeric genomes and/or missing true orthologous for a few of the SCP genes. Further, FastAAI shares a limitation with traditional AAI in the form of limited resolution among extremely distantly related genomes (see Fig. 2). For genomes at this level of novelty, other approaches such as phylogenetic reconstruction may still be more appropriate. FastAAI can therefore effectively complement, if not replace, the abovementioned methods for genome comparisons at the genus level and above, while reverting to the traditional AAI, or even better ANI, for closely related genomes with high FastAAI identities (species level). Thus, species-level comparisons should instead be resolved with FastANI or regular ANI comparisons, and particularly novel genomes with no apparent close relatives in the reference databases at the (same) genus level can be processed with FastAAI and/or using phylogenetic methods when the number of genomes is tractable. Finally, FastAAI is a database-dependent tool; in cases where a highly novel genome has no relative in the database such as a highly novel pathogen, FastAAI may be unable to do more than quantify its extreme novelty (by the relative low AAI value to the best match reported).

In conclusion, our results demonstrate that FastAAI will be a valuable tool for the scientific community aimed at the fast and accurate genome relatedness estimation and classification of microorganisms from clinical, industrial, or environmental sources.

Supplementary data

Supplementary data are available at NAR online.

Conflict of interest

None declared.

Funding

Funding to pay the Open Access publication charges for this article was provided by U.S. NSF grants DBI1356288NSF, DBI1356380NSF, DBI1759831NSF, and DBI1759892.

Data availability

The TypeMat and PhylaLite genomes are available at http://microbial-genomes.org/projects/20. Archives containing the versions used in this paper are available upon author request. RefSeq genomes are available at https://www.ncbi.nlm.nih.gov/refseq/. The GTDB genomes used in this paper can be found at https://data.gtdb.ecogenomic.org/releases/release202/. Accessions for the RefSeq genomes and lists of genomes used in subsamples of the GTDB dataset are available in supplemental materials. Databases for GTDB release 202, ASTRAL, and GEMS are available on FigShare at 10.6084/m9.figshare.25746876, 10.6084/m9.figshare.20522040, and 10.6084/m9.figshare.20522088, respectively.

FastAAI is an open-source tool freely available under the MIT license at https://github.com/cruizperez/FastAAI, https://doi.org/10.6084/m9.figshare.28774625, and through a PyPi installation at https://pypi.org/project/FastAAI-release.