Multisubstrate specificity shaped the complex evolution of the aminotransferase family across the tree of life

In core metabolism, a single or a few enzymes, with high specificity and efficiency, are typically responsible for catalyzing each biochemical step. However, ATs often retain broad substrate specificity (23), which is typically seen in secondary or specialized metabolic enzymes that underwent significant copy number expansion (33–38). We therefore investigated how many ATs are present in various extant organisms by obtaining and analyzing all putative AT sequences from the genomes of 90 organisms across the ToL. We queried the genomes of 15 species from eubacteria (39), archaea (40), archaeplastida (plants) (41, 42), holozoa (animals) (43), holomycota (fungi) (44), as well as major protist taxa, spanning multiple phyla (45) (SI Appendix, Fig. S1B and Table S2), for protein sequences containing conserved protein family (Pfam) domains known for AT enzymes (SI Appendix, Tables S1 and S2). These species were selected based on the availability of high-quality genomes that represent the diversity of phyla and clades (SI Appendix, Fig. S1B and Table S2). Inspection of Pfam by HMMER (46) corrected the Pfam annotations for 93 ATs (Dataset S1) and a length criterion (250 to 1,200 amino acids) excluded potential pseudogenes, resulting in 2,938 putative AT genes (Dataset S2). Since our ability to detect homology fades over long evolutionary timescales (47, 48), a structure-assisted multiple sequence alignment, MAFFT-DASH (49), was used to capture the deep evolutionary kinship among distantly related ATs (Materials and Methods).

ATs are found among two independently evolved PLP-dependent enzyme fold-types, I and IV (Fig. 1B) (23, 24). The fold-type I includes canonical AT families of class I, II, and IV, as well as smaller families of sugar ATs and alliinases, whereas the fold-type IV only comprises class III ATs. We therefore separately aligned sequences from each Pfam family, having a close phylogenetic proximity (50), and then merged into two master alignments for independently evolved fold-type I and IV (Fig. 1C). The phylogeny estimation using the approximately maximum likelihood method (51) captured the monophyletic origins of AT and related enzymes from different Pfam families (Fig. 1C). Overall, class I (PF00155) represented the largest clade having 1,373 out of the 2,938 putative AT enzymes, followed by 704 in class IV (PF00266), 462 in class II (PF00202), and 267 in class III (PF01063, Fig. 1C and SI Appendix, Table S2). Ninety sugar ATs (PF01041), involved in bacterial cell wall biosynthesis (52, 53), were enriched in eubacteria and archaea, while 40 alliinases (PF04864) were mainly restricted to plants and protists with secondary plastids, except for two amorphean protists without secondary plastids (Thecamonas trahens and Monosiga brevicollis, SI Appendix, Table S2).

The copy number analyses showed that archaea and eubacteria on average have ~15 and ~25 AT and related genes, respectively, though a few prokaryotes having parasitic, pathogenic, or symbiotic lifestyles (54–56) had much lower numbers of ATs (Fig. 1D and SI Appendix, Table S2). In theory, up to 22 AT activities are needed to establish core nitrogen metabolism for synthesis of essential amino acids and nucleic acids (23). Also, genome reconstruction studies of LUCA (57) suggest that ~25 AT and related genes (from classes I, II, III, IV, and sugar ATs) were present at the base of the ToL. Therefore, these analyses revealed that the copy numbers of AT enzymes remained constant in prokaryotes.

Eukaryotes had a higher total number of AT and related genes than archaea and eubacteria (Fig. 1D, SI Appendix, Fig. S2 A and B, and Dataset S3), with plants and photosynthetic protists having the highest numbers (Fig. 1D, SI Appendix, Fig. S2 A–C, and Dataset S3). This likely reflects the expansion of metabolic processes [e.g., photorespiration (58), chlorophyll biosynthesis (59)] necessary to maintain an autotrophic lifestyle. Aside from these photosynthetic organisms, the higher AT number of eukaryotes was attributed to functionally homologous AT isoforms for organelle targeting (Fig. 1D): e.g., three mammalian aspartate (Asp) AT copies with highly redundant functions (60, 61). Additionally, multicellular eukaryotes had tissue-specific AT paralogs, such as human AlaAT1 and AlaAT2 (62) and Arabidopsis tryptophan (Trp) ATs, TAA1, TAR1, and TAR2 (63, 64). However, the average number of AT candidate genes was only 1.5 and 2.4 times higher in animals and plants, respectively, than in eubacteria, whereas the average number of genes in the genomes of animals and plants are 8.1 and 6.8 times larger than those in eubacteria (Fig. 1E and SI Appendix, Fig. S2C and Table S2). This indicates that, despite the substantial expansion of eukaryotic genome sizes and metabolic networks (65), AT gene families did not experience substantial copy number expansion even within eukaryotes. This is in contrast to the evolutionary history of specialized metabolic enzymes, many of which are also promiscuous but underwent significant copy number expansion (33, 34). These data revealed that organisms can maintain a robust and complex core nitrogen metabolic network with a relatively small number of AT genes.

Ancestral ATs diversified into at least six different classes before LUCA and, today, are ubiquitously present across the ToL (Fig. 1B). However, it is unknown how these founder AT enzymes of LUCA evolved to catalyze essential AT reactions of core nitrogen metabolism, hindering the accurate prediction of AT functions and nitrogen metabolic networks that exist in various organisms. To address this issue, based on the large-scale AT phylogeny illustrating an overall relationship of Pfam families (Fig. 1C), we further built 40 FastTree trees using multiple starting trees for each AT class to avoid a potential issue of local optima (66). We then identified 62 distinct AT and related enzyme groups that represent monophyletic clades (bootstrap value ≥ 0.9) (SI Appendix, Figs. S3–S7 and Dataset S3). The “AT group names” were assigned based on biochemically or genetically validated activities from the Uniprot database query (Dataset S2), though it is important to note that some of them might have been further functionalized in certain taxa. We could not name 15 AT and related groups and hence noted as “uncharacterized (UC)". Overall, class I contained the largest number of 22 groups with 12 AT, 5 non-AT, and 5 UC enzyme groups (SI Appendix, Fig. S3) while 6 of 10 groups in class II, 7 of 11 groups in class III, and 2 of 11 groups in class IV as well as 3 groups in sugar AT and alliinases classes were AT groups (SI Appendix, Figs. S4–S7). The novel Asp ATs (PF12897, fold-type I) (26, 67), distinct from canonical Asp ATs of class I (PF00155), were found only in two eubacteria and an outgroup of the alliinase class (Fig. 1C). The 17 non-AT groups, which belong to AT classes but likely have non-AT activity (e.g., synthases, lyases, and other transferases), were still included in the analysis, as they improved the phylogenetic relationships of ATs (Dataset S4), and some non-AT members can contain ATs, e.g., a Trp AT group found in the alliinases class (68) (SI Appendix, Fig. S7). Mapping of AT and related enzymes from four model species—Escherichia coli, Saccharomyces cerevisiae, Homo sapiens, Arabidopsis thaliana—suggested complex evolutionary history of AT groups, many of which appear to have been differentially lost in multiple taxa (SI Appendix, Figs. S3–S7 and Dataset S4).

To further trace the deep evolutionary history of 62 AT and related groups since LUCA, we mapped and clustered different AT and related enzyme groups based on their presence and absence in the 90 species (Fig. 2 and Dataset S5) and calculated their percent conservation within each taxon (Fig. 2A). The most striking finding was that there was no AT group that is absolutely conserved across the ToL, which was unexpected for core metabolic enzymes catalyzing essential reactions. Only one AT group, class IV Ala-glyoxylate ATs (AGTs, SI Appendix, Fig. S6)—along with non-AT groups for CoA-utilizing synthases (class I, SI Appendix, Fig. S3) and Cys desulfurases (class IV, SI Appendix, Fig. S6)—was distributed across almost all taxa though still absent in some species, especially prokaryotes (cluster 1 or CT1 in Fig. 2B). Five AT groups—Asp ATs, Ala ATs, phosphoserine (PSer) ATs, kynurenine AT, and ornithine ATs (CT2, Fig. 2B)—were largely conserved among eukaryotes, but not in prokaryotes. Tyrosine (Tyr) AT and class II AGTs (CT4, Fig. 2B) were present in animals and land plants but largely absent in fungi. Although the outcome might vary with the inclusion of other species, further investigations of selected AT groups (e.g., PSer AT, Tyr AT) using the NCBI Landmark model species database, containing high-quality genome assemblies of 27 species across the ToL (including 10 from our 90 species, such as A. thaliana, H. sapiens, E. coli), also yielded consistent results with similar distribution patterns (SI Appendix, Fig. S8).

Histidinol-phosphate (HisP) ATs, 7,8-diaminopelargonic acid (DAPA) ATs, and acetylornithine ATs (CT6, Fig. 2B) were conserved across the ToL except in animals and nonphotosynthetic protists, owing to the lack of histidine (His), biotin, arginine (Arg) biosynthesis in most of these taxa (69–71). Interestingly, HisP ATs were found in choanoflagellates (e.g., M. brevicollis, Fig. 2B and Dataset S5), suggesting that de novo His biosynthesis was likely lost after the divergence of animals and choanoflagellates ~600 Mya (72). Ala/glutamine (Gln)/lysine (Lys)/Arg ATs and prephenate ATs, along with glutamate (Glu) semialdehyde aminomutase involved in chlorophyll biosynthesis (73), were highly conserved in bacteria, plants, and photosynthetic protists (CT8, Fig. 2B). Some AT groups show highly lineage-specific distribution, which includes animal- and fungi-specific branched-chain amino acid ATs, and land plant-specific Trp ATs that likely functionalized from alliinases (CT5, 11, and 9, respectively, in Fig. 2B). Interestingly, ISS1 (or also known as VAS1, class I, SI Appendix, Fig. S3) (74) and d-amino acid AT-related enzyme (class II, SI Appendix, Fig. S4) groups, whose functions are not fully understood, were highly conserved among all plant species and photosynthetic protists, except diatom Phaeodactylum tricornutum (CT9, Fig. 2B).

The above AT mapping across the ToL showed that most AT groups are not ubiquitous, unlike other core enzymes carrying out essential metabolic reactions (11, 13, 15, 65, 75–78). Also, closely related ATs (e.g., class II ATs) did not cluster together at all based on taxonomic distribution (Fig. 2). This is likely because AT enzymes from a distant AT group were frequently recruited and replaced existing AT enzymes in certain taxa to carry out essential AT reactions—a phenomenon known as nonorthologous gene displacement (32). One clear example was the enzymes responsible for Ala AT activity, which is essential in all organisms (23). In eukaryotes, Ala AT activity is provided mainly by class I Ala ATs (97% conserved among eukaryotes, Fig. 2A), and to a lesser extent by the side activities of class II and IV AGTs, and PSer ATs (Fig. 2C and SI Appendix, Fig. S9A). In contrast, in some prokaryotes like E. coli, Ala AT activity can be redundantly provided by a number of ATs, e.g., avtA, aspC, tyrB, alaA, serC, and ilvE (79) that are poorly conserved among prokaryotes (Fig. 2 and SI Appendix, Fig. S9A). Since at least one AT enzyme with Ala AT activity was present for all species analyzed, Ala AT activity is conserved across the ToL but mediated by taxon-specific AT enzymes (SI Appendix, Fig. S9A). Additionally, the Tyr AT group (class I) is widespread among eukaryotes but not in many fungi (Fig. 2, SI Appendix, Fig. S3, and Dataset S4), whose Tyr AT genes were likely lost and replaced with aminoadipate ATs that belong to a distant clade of class I [e.g., yeast Aro8 and Aro9 (80), SI Appendix, Figs. S3 and S9B]. Asp AT activity is provided by canonical Asp AT enzymes (class I) in most organisms and additionally by prephenate AT enzymes in many photosynthetic organisms and some prokaryotes (SI Appendix, Fig. S9C). However, a few eubacteria have noncanonical Asp ATs (PF12897), (26) and we could not identify any known Asp AT enzymes in many archaea species (SI Appendix, Fig. S9C). Similarly, no clear orthologues were found in land plant genomes for aminoadipate AT enzymes required for Lys metabolism (Fig. 2B and SI Appendix, Fig. S3). While some of these missing genes could be due to sequencing and/or annotation errors (81, 82), these results overall suggest that there are still unknown ATs responsible for catalyzing these essential AT reactions.

Overall, the global AT mapping across the ToL revealed that essential AT activities were not necessarily carried out by conserved, orthologous enzymes belonging to the same AT class and group in different taxonomic lineages. This likely explains the difficulty in homology-based assignment of AT functions. Instead, we observed many instances of widespread AT enzyme replacement from distantly related, nonorthologous AT groups, where ATs from unrelated groups and even different classes were recruited to redundantly or alternatively carry out specific AT reactions. Therefore, in contrast to other core metabolic enzymes, many ATs underwent a complex evolutionary history characterized by the extensive recruitment and replacement of distantly related ATs.

The frequent recruitment and displacement of AT enzymes from distant AT groups may be facilitated by the promiscuity and functional redundancy of AT enzymes among different groups. However, AT substrate specificity has not been extensively studied (23). We therefore experimentally examined the extent of overlap in substrate specificities within the same and among different AT groups. Here, we employed aromatic (Aro) AT activity as a testbed, as it utilizes structurally distinct substrates but is widely detected across distantly related ATs within class I (SI Appendix, Table S3), which include the AT groups of ISS1 (83), canonical Asp AT (84), aminoadipate AT (80), HisP AT (85), Tyr AT (86), kynurenine AT (87), and prephenate AT (88) (SI Appendix, Fig. S3). As their substrate specificity has not been fully defined, we expressed, purified, and characterized recombinant enzymes from i) canonical Aro AT groups (as positive controls), ii) non-Aro AT groups with no reported canonical Aro ATs, and iii) non-AT group (i.e., A. thaliana C-S lyase, At SUR1, as a negative control, SI Appendix, Table S3). Their substrate specificities were determined by a multisubstrate enzyme assay using 15 keto acceptors with Gln or Glu as an amino donor depending on the enzymes. AT reactions were initially carried out for an extended reaction time with high substrate concentrations. Then, multiple reaction products were analyzed simultaneously using liquid chromatography-mass spectroscopy (LC–MS) (SI Appendix, Fig. S10A) to calculate their percent conversion of the keto acid substrates to the corresponding amino acid products (Fig. 3A).

No Aro AT activity was detected in the Ala AT of A. thaliana (At AlaAT1) and E. coli (Ec alaA), nor in At SUR1 (Fig. 3A), consistent with prior reports (23, 89, 90). In contrast, Aro AT activity was detectable, though at varied degrees, in all of the analyzed non-Aro AT group enzymes (Fig. 3A and SI Appendix, Table S3), which include At HisN6 and kynurenine AT (At KAT) that are previously not known to have Aro AT activity in plants (91). This result revealed that Aro AT activity is widespread in many ATs beyond those currently designated as canonical Aro ATs. To our surprise, HisP ATs from A. thaliana and E. coli (At HisN6A, At HisN6B, and Ec hisC, respectively), having only 32% sequence identity (Fig. 3A and SI Appendix, Table S4), showed a very similar substrate preference, as reflected by the relative abundance of the produced amino acids: His > Trp > Tyr > phenylalanine (Phe, Fig. 3A and SI Appendix, Table S5). Similarly, At KAT and Ec ybdL, having 40% sequence identity, produced the same eight amino acids with a similar substrate preference, although Ec ybdL showed stronger activity with Phe than At KAT (Fig. 3 and SI Appendix, Table S5). We also detected the conservation of similar substrate promiscuity across different groups. All Aro ATs including ones from ISS1, His AT, Tyr AT, and kynurenine AT groups, except prephenate AT, showed activity with Met, His, Leu, and homo-Ala (Fig. 3A and SI Appendix, Table S5). Therefore, although many ATs show very broad substrate promiscuity, we identified the presence of common signatures of multisubstrate specificity among different ATs even after being separated for billions of years of evolution.

For quantitative comparison, we further conducted kinetic characterization of Ec ybdL and At KAT in their linear ranges of activity (SI Appendix, Fig. S10B). Both enzymes had the highest kinetic efficiency with 4-methylthio-2-oxobutanoic acid (4MTOB), followed by imidazol-5-yl pyruvate (I5P) and phenylpyruvate (PPY, keto acids of Met, His, and Phe, respectively, Fig. 3B and SI Appendix, Fig. S10 C–H), which largely agreed with the results of the initial substrate specificity screening (Fig. 3A). Interestingly, Ec ybdL and At KAT showed very similar response curves to these three substrates, except for overall higher turnover rates (k_cat) for Ec ybdL than At KAT (SI Appendix, Fig. S10H). Additionally, both enzymes showed similar substrate inhibition with 4MTOB and I5P, but not with PPY (Fig. 3B), as was also reported for human KAT (92). Importantly, unlike the animal KATs (87) and Ec ybdL having kynurenine AT activity, At KAT was not capable of using kynurenine as an amino donor (SI Appendix, Fig. S11). This result confirmed the lack of Ec ybdL contamination in the recombinant At KAT preparation and showed functional specialization among kynurenine AT orthologs in different organisms.

Taken together, these results uncovered remarkable similarities of multisubstrate specificity and kinetic properties between ATs from the same groups but taxonomically distant species (e.g., E. coli vs. Arabidopsis). This conserved promiscuity may also represent their ancestral activity that may date back to LUCA. On the other hand, we detected the overlapping multisubstrate specificity among distantly related, nonorthologous AT groups within class I ATs (Fig. 3A), which likely provides functional redundancy and facilitates the AT displacement between distant AT groups (Fig. 2).

To elucidate underlying mechanisms of the overlapping substrate specificity, or functional redundancy, among different AT groups (Fig. 3), we examined whether the presence of certain active site residues or motifs correlates with their substrate specificity. We first obtained experimental or homology-based structures of representative enzymes from each AT group (108 enzymes in total, Dataset S6) and determined the active site residues manually by referring to two crystal structures bound with a hydrophilic Asp [PDB ID: 1ARG (93)] or a hydrophobic Phe [PDB ID: 1W7M (94)] ligand, respectively. The original multiple sequence alignment was then used to calculate the consensus of the structurally conserved (≥10%) active site residues for each AT group (Fig. 4A).

As expected (23, 24), the Asp and Lys residues involved in cofactor PLP binding (red arrows in Fig. 4A) were highly conserved across all AT groups belonging to PLP-fold type I. The C-terminal Arg residue, which recognizes substrate’s carboxylate (yellow arrows in Fig. 4A), was mostly conserved with some exceptions, such as POP2 and DAPA AT (class II) that use non-α-amino acid substrates having an increased molecular distance between the transferable amino and carboxyl groups (96, 97). Besides, several other residues were generally well conserved for most PLP-fold type I groups (black arrows in Fig. 4A), while each AT class had certain motifs that are conserved among most or only within a subset of AT groups (orange and blue boxes in Fig. 4A, respectively). Class IV ATs do not share a ubiquitously conserved motif with other class ATs, consistent with its membership in a different fold-type of PLP enzymes (24, 98) (Fig. 1).

Next, we analyzed whether distantly related AT groups with similar multisubstrate specificity have shared residues or motifs among each other, by comparing the active site residues of class I AT enzymes having Aro AT activity: ISS1, Tyr AT, kynurenine AT groups (Fig. 3), as well as canonical Asp AT and aminoadipate AT groups (80, 83–88). However, the active site sequences of these five AT groups were not more similar to each other than other class I ATs (Fig. 4A) and did not share common residues or motifs in their active sites. Therefore, we then employed a structure-based comparison to identify ligand interacting residues by docking Phe-external aldimine complex, as a generic aromatic substrate, to the active sites of their homodimeric complex structures (Fig. 4B and SI Appendix, Fig. S12). Interestingly, the unique binding pose of Phe in each AT group requires the involvement of several active site residues that were not necessarily conserved at the level of the peptide sequence (Fig. 4A and SI Appendix, Table S6). For example, nonconserved N-terminal aliphatic amino acids interact with the aromatic ring of Phe in Arabidopsis ISS1 (Met 19), Arabidopsis TAT1 (Ile 13), yeast Aro8 (Leu 25), E. coli tyrB (Leu 39), and Arabidopsis HisN6A (Ile 34) (red stars on Fig. 4B and SI Appendix, Fig. S13). Furthermore, a corresponding residue is absent for human KAT1 and Streptomyces bingchenggensis prephenate AT (Sb PphAT), but present for non-Aro ATs such as human and Arabidopsis Ala ATs (Val 64 and Val 82, respectively, SI Appendix, Table S6). Similarly, the conserved Tyr near the phosphate group of PLP (SI Appendix, Table S6) interacts with both the aromatic ring of Phe and the phosphate group of PLP in Arabidopsis ISS1 (Tyr 64) and Arabidopsis TAT1 (Tyr 71), but only with PLP in the human KAT1 (Tyr 63), yeast Aro8 (Tyr 105), and Arabidopsis HisN6A (Tyr 80) (purple stars on Fig. 4B). Instead, the Phe 278, Tyr 251, and Tyr 256 residues interact with the substrate Phe in KAT1, Aro8, and HisN6A, respectively (orange stars on Fig. 4B).

We also examined why Ec ybdL (99) and At KAT had different specificity toward Phe and kynurenine substrates (Fig. 3 and SI Appendix, Figs. S10 and S11). Docking of these substrates to these two enzymes showed that both enzymes had nearly identical active site residues (SI Appendix, Table S6); however, there were five residues that are distant to the substrate and located around the active site entrance, which appear to make the Ec ybdL active site wider with higher solvent accessibility than At KAT, likely aiding the access of bulky aromatic substrates into the active site of Ec ybdL (SI Appendix, Fig. S14 A and B). Additionally, the N-terminal helix that plays a critical role in the substrate binding of ATs by conformational change (94, 100, 101) had more hydrophobic surface in At KAT (I74 and V77) than Ec ybdL (A25 and Q28) (SI Appendix, Fig. S14 C and D), potentially restricting the mobility of the N-terminal helix in At KAT.

Overall, these sequence and structural analyses of distantly related Aro ATs suggest that they achieve similar multisubstrate specificity using different active residues that are not necessarily conserved at peptide sequences. Additionally, residues that do not directly participate in substrate recognition appear to contribute to AT substrate specificity by influencing overall active site conformation (100).

Species used in this study were selected based on the availability of high-quality proteomes (designated as “reference proteome” by Uniport, except some plant species queried from Phytozome). Taxon sampling was performed for each kingdom and protists to include species that would best represent the diversity of the investigated taxonomic group. It is important to note that the selection of the number of species from each taxon is not impartial, as it overrepresents eukaryotic lineages, particularly archaeplastida, holozoa, and holomycota. However, since species from these groups are also disproportionately represented in scientific research (115), we concluded that their overrepresentation would benefit a broader audience.

To identify putative AT sequences, we performed Pfam analysis. All protein sequences were downloaded in August 2021 for 90 species from three public databases: Phytozome (42) and Phycocosm (44) for plants and UniProt (116) for the others. To avoid redundant protein sequences, we used primary transcripts and nonredundant proteome data for plants and the others, respectively (Dataset S7). The proteins shorter than 250 amino acids were excluded owing to the potential for nonfunctional genes. The sequences after length filtration proceed to Pfam annotation by HMMER v3.3.1 (hmmscan with a default setting; E-value cutoff = 0.01) (117) using Pfam profile hidden Markov models obtained from Pfam v35.0 (46). In the case of multiple hits for the same region in the protein sequence, we chose the Pfam with a lower E-value. Then, we obtained 2,938 putative AT sequences (Dataset S2) by searching Pfam IDs in which proteins display AT activity: PF00155, PF00202, PF00266, PF04864, PF01041, PF12897 (PLP fold type I), and PF01063 (PLP fold type IV). For the extremely long proteins (>1,200 amino acids), we extracted the only AT Pfam domain region. In the case that multiple AT Pfam domains exist, the longer AT region was selected.

For structure-guided multiple sequence alignment, amino acid sequences were aligned by a MAFFT-DASH (49) using a BLOSUM62 scoring matrix (gap opening penalty = 1.53). We assumed closer associations among the proteins belonging to the same Pfam. Thus, all putative AT sequences were divided into seven groups according to Pfam IDs, followed by multiple sequence alignment for each group with iterative refinement methods (L-INS-i for PF01041, PF04864, and PF12897; FFT-INS-i for PF00155, PF00202, PF0266, and PF01063). Then, considering the evolutionary divergence of PLP fold type I, we merged the alignments for six Pfam (i.e., PF00155, PF00202, PF00266, PF04864, PF01041, and PF12897) into a single multiple sequence alignment for PLP fold type I with a progressive method (FFT-NS-2). The alignment for PF01063 was regarded as the one for PLP fold type IV. For further analysis, we deleted DASH sequences from the alignments to avoid misinterpretation.

To construct phylogenetic trees for large alignments, we employed an approximately maximum likelihood method. The alignments for PLP fold types I and IV were analyzed by FastTree v2.1.11 (51) with LG+CAT model and the option “-mlacc 2 -slownni -spr 4 -pseudo”. To avoid a potential issue of a local optima described in the previous study (66), we generated 40 FastTree trees using 20 random and 20 parsimony starting trees prepared by RAxML-NG (v1.2.0) with a JTT+G model, and then selected the most likelihood trees. For the PLP-fold type I, we additionally built 40 FastTree trees for each class (e.g., class I ATs, Alliinases, Sugar ATs) and then selected the most likelihood trees. To evaluate the reliability of each split in the tree, we used local support values (118) computed in FastTree from 1,000 resamples. We analyzed the tree and modified the tree topology using an ETE 3 Toolkit (119) in a Python environment (v3.6.13). To identify the highly supported phylogenetic branches, the nodes with low supporting values (<90%) were deleted and then integrated into parental nodes. The condensed phylogenetic trees were manually drawn using Microsoft PowerPoint with local support values as branch support.

The functions of AT and related groups were inferred from the publicly available experimental data. First, we picked four species whose enzymes have been well defined (i.e., H. sapiens, A. thaliana, S. cerevisiae, and E. coli), and then inquired about the functions of AT candidates from literature search. Also, we collected additional 68 enzymes whose experimental data are available in the UniProtKB (116), the BRENDA (120), or the SABIO-RK (121) databases. This information was utilized to annotate names for the AT and related groups. We labeled the group with UC when any group member was unavailable in the collection. Only enzyme clades that have more than five sequences from at least four species or three kingdoms with a bootstrap value higher than 90% were assigned AT groups names. Sequences that did not belong to any groups were indicated as “unassigned”.

We computed how many AT-coding genes each species carries for 62 AT and related groups. This dataset was used for two different analyses to examine taxonomic conservation and distribution. For the taxonomic conservation, the sum of the gene counts was calculated in each kingdom for each AT and related group. The values were used to generate a heatmap using Microsoft Excel.

To investigate the taxonomic distribution, we first transformed the gene counts into the possession (i.e., 0: absence, >1: presence) of individual AT and related groups. A heatmap was generated with columns being clustered (distance metric: “euclidean”, linkage method: “ward”) by the seaborn package v0.11.2 (122) in a Python environment, whereas rows were arranged phylogenetically by referring to the previous studies (39–41, 43–45) to be able to trace the evolutionary history of individual ATs.

The taxonomic distribution of selected AT groups was separately examined using the NCBI Landmark model species database, as described in SI Appendix, SI Method.

All AT candidates determined by Pfam analysis were analyzed for their predicted subcellular localization. The presence of N-terminal presequences for subcellular localization was analyzed using TargetP-2.0 server (123). We focused on the mitochondrial and chloroplast transit peptide and ignored the presence of signal peptides for secretory pathways. Based on this prediction, the AT candidates were classified into 3 groups according to subcellular location: mitochondria, chloroplast, and others.

AT enzymes were expressed and purified as described previously (114) and in SI Appendix, SI Method. Multisubstrate specificity screenings as well as the enzyme kinetic analyses of ATs were conducted as described previously (114) and in SI Appendix, SI Method.

We initially picked 111 model proteins over 62 AT-like groups (1 to 4 enzymes from each group) for structural analysis. The crystal structures of 57 proteins were obtained from the Research Collaboratory for Structural Bioinformatics Protein Data Bank (RCSB PDB) (124). For the other proteins, homodimeric structure models were generated by AI-based structure prediction using AlphaFold v2.1.0 (125, 126). The predicted structures were manually checked whether the dimeric interfaces and the active site formed properly. In the case of a distorted structure or collapsed active site, we constructed another structure model by conventional homology modeling using SWISS-MODEL (127). When the models were still poor, they were excluded from further analyses. Finally, we obtained the structures for 107 proteins over 56 AT groups (Dataset S6). Then, the structures were aligned with each other using the SALIGN module of MODELLER v10.2 (128) and then used for active site analysis. Active site residues were determined manually referring to the protein–ligand complex structures determined previously with Asp and Phe (PDB ID: 1ARG and 1W7M, respectively) (93, 94). The corresponding positions of the active site residues were examined in multiple sequence alignments and minor residues (i.e., <10% conservation among model proteins) were removed. As a result, we obtained 50 and 32 residues for PLP fold types I and IV, respectively. Next, the putative active site residues were extracted from multiple sequence alignments for 2,938 ATs. The lists of active site residues were divided according to AT-like groups, followed by generation of consensus sequences using WebLogo (95).

The docking simulations of external aldimine intermediates were performed against 9 ATs (Dataset S8) using AutoDock v4.2.6 (129) as described in SI Appendix, SI Method.

K.K., S.-W.H., Y.Y., and H.A.M. designed research; K.K., S.-W.H., R.K., H.S., and H.A.M. performed research; H.S. contributed new reagents/analytic tools; K.K., S.-W.H., and H.A.M. analyzed data; and K.K., S.-W.H., Y.Y., and H.A.M. wrote the paper.

The authors declare no competing interest.