Sequence fingerprints.

Known NATs do not all inhabit the same connected components of the SSN (Fig 2), which indicates NATs are not one homogeneous family of acetyltransferases. We used MEME [74] from the MEME suite [75] to identify motifs in each of the SSN clusters. Sequence motifs of highly conserved residues were detectable for each of the clusters. Based on the similarity between motifs and on the clustering of the NATs in the SSN, we defined five different groups of NATs (Fig 2). We subsequently calculated sequence motifs for each of these groups. The motifs are shown in Fig 3 and Table 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Characteristic sequence motif fingerprints of NAT Groups 1 to 4.

Sequence motifs were calculated as described in the Material and Methods section and using sequences from the SSN clusters. Each position in the motif is represented by a colored bar and a one-letter code for the amino acid frequently found at that position in the GNAT fold. The height of colored bar is proportional to the frequency of the corresponding amino acid. The colors correspond to the type of amino acid (blue: R, H, K; red: D, E; green: C, S, G, Y, T; black: P, F, V, L, I, A; orange: M, W; purple: N, Q). Group 5 is not shown as there is no structure of NAA70 available.

https://doi.org/10.1371/journal.pcbi.1007988.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Regular expressions for key sequence motifs of NATs Groups 1 to 4.

All regular expressions were calculated using MEME from MEME Suite [75].

https://doi.org/10.1371/journal.pcbi.1007988.t001

Group 1 consists of Group 1a with NAA10 and NAA20, and Group 1b that contains NAA30 of all eukaryotic kingdoms. NAA10 and NAA20 are in the same connected component, while NAA30 is found in a single isolated cluster. Sequence motifs that are important for binding of substrate and acetylation in both subgroups of Group 1 NATs are localized on the α1-α2 loop, the β4 and β5 strands and the β6-β7 loop [51, 52, 54] (S3A Text) and this is true for groups 2 and 3 as well. Group 2 consists of NAA50 and NAA60. NAA50 and NAA60 do not cluster together in the SSN but the resemblance between their key sequence motifs (S3B Text) justifies placing them in the same group. Group 2 also contains a catalytically inactive yeast NAA50, a member of the fungal Sacharomicetaceae family (S3C Text). This inactive NAA50 forms a separate cluster, numbered 135, in the SSN (Fig 2). Sequence motifs on key secondary structure elements are not strictly conserved between cluster 135 and the catalytically active NAA50 (cluster 9). This explains why the inactive enzyme clusters separately from cluster 9. It is important to mention that NAA60 of animals and plants do not share the same cluster (S2C Text). Plant NAA60 bears some slight, but potentially important differences on the α1-α2 loop, where a negatively charged E in animals is replaced by a positively charged K in plants. The key tyrosine of β6-β7 loop is still present in plants, but unlike animal NAA60 which has thee tyrosines in this loop, plant NAA60 only has one. We define Group 3 around NAA40. Some plant NAA40 form a separate cluster and some others share the same cluster as animal NAA40 sequences. A striking characteristic of NAA40 is its long α0 helix and the position of its α1-α2 loop [55] which extends over and covers the binding site where the β6-β7 loop lies in other NATs. Group 4 is defined around NAA80 which is structurally different from the first three groups. Its surface shows a large cleft which is covered by loops in all other NAT structures available to date [56]. The need for a larger ligand binding site is explained by the fact that NAA80 has evolved to catalyze N-terminal acetylation of fully folded actin and harbors an extensive binding surface to actin [76]. Finally, Group 5 contains NAA70 which is a chloroplast NAT discovered in Arabidopsis thaliana [41]. NAA70 is closer to bacterial acetyltransferases than to the eukaryotic ones in Groups 1 to 4. A BLAST search against the NCBI non-redundant database [77] and excluding green plants, suggests that NAA70 is most similar to cyanobacterial proteins with the best hit being a protein from Gleocapsa sp (29,7%id over 62% query cover). We also found that NAA70 shares a high percentage of sequence identity with Enterococcus faecalis acetyltransferase whose structure has been solved (PDB code 1U6M). Unfortunately, there is not enough reliable structure information on NAA70 to be able to map the position of the key sequence motifs onto the secondary structure elements.

While there are important differences between each of the groups in terms of sequence motifs, some similarities emerge (Fig 3). They are especially obvious between groups 1a, 1b and 2, where we can observe a well conserved tyrosine in the α1-α2 loop (Fig 3 and S3D Text) and most importantly, another conserved tyrosine in the β6-β7 loop (Fig 3 and S3D Text). This tyrosine is essential for function and is strictly conserved in all members of groups 1 and 2 [49–51, 54, 78] with the exception of the catalytically inactive fungal Sacchormycetaceae NAA50 where this tyrosine is lost [53]. The tyrosine in the α1-α2 loop is conserved in all NATs of Group 1a, Group 1b and Group 2 [49, 50, 79] except for NAA20 where it can be replaced by phenylalanine [51]. Group 3 and group 4 motifs clearly differ from those of Group 1 (Fig 3). Compared to the other groups, strands β4 and β5 stand out in groups 3 and 4 where they play a major role in substrate binding and catalysis. Interestingly their sequence motifs and key residue positions differ between the two groups [55, 56].

Structure comparison.

We compared structures of acetyltransferases to one another using the DALI server [80]. DALI computes and compares intramolecular distance from each structure, and no sequence information is used. Our dataset consists of structures of 38 catalytic subunits of acetyltransferases, all belonging to our SSN. The resulting dendrogram (Fig 4) illustrates the degree of similarity between all structures in the dataset. It shows a classification that overlaps with that of the SSN. In addition, it highlights that Group 1a and Group 2 are more similar to one another than to the other NATs, and to the rest of the entire superfamily. Structure of NAA30, around which we built Group 1b, has not been solved, but based on the sequence similarity it is safe to assume that it would be found in the same part of the dendrogram as Group 1a and Group 2. NAA40 (Group 3) is the NAT closest to NAA80 (Group 4), to the histone acetyltransferase GCN5, to the dopamine N-acetyltransferase and to the arylalkylamine N-acetyltransferase. The proximity of NAA40 and NAA80 is only observed in the structure-based classification and was not observed in the sequence similarity network, although NAA40 and NAA80 are closer to one another than to other known NATs in the SSN. NAA80 is also close to the histone acetyltransferase GCN5. Groups 1a, 2, 3 and 4 of NATs are more similar to one another than to the rest of the superfamily when structures are compared, but this is not the case when sequences are compared. In-between these 4 groups one finds non-NAT acetyltransferases, namely the histone acetyltransferase GCN5 (cluster 1) and serotonin acetyltransferases (cluster 122).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. DALI dendrogram for structural similarity between acetyltransferases.

The dendrogram is the result of hierarchical clustering of structures. The known NATs are closer to one another than to the rest of the superfamily. Note the non-NAT acetyltransferases located close to known NATs.

https://doi.org/10.1371/journal.pcbi.1007988.g004

It is important to mention that while we can see differences in structures of different acetyltransferases, they are still quite similar to one another. We verified the proximity of the structures by building a network based on a structure similarity matrix. The resulting network is random with all nodes connected to all nodes when we use a Z-score higher than 2, which is considered to be significant as a threshold for an edge between two nodes [81]. The Z-score threshold needs to be increased to at least 15 for cluster separation in the similarity network to appear (S1 Fig).

Similarities and differences between groups.

Groups 1 and 2 are highly similar. Because of the similarities between NAA10 in one hand, and NAA30 in the other hand, we placed them in the same group of NATs but defined two subgroups (1a and 1b), as the NAA30 cluster is not connected to those of NAA10 and NAA20. Sequence motifs on the α1-α2 loop, β4 and β5 strands and β6-β7 loop are characteristic to both subgroups and represent its signature. Protein N-termini starting with a small residue, which is exposed after removal of the initial methionine, and protein termini with the initial methionine can be acetylated by this group of NATs [7]. Even though these enzymes are obviously closely related, they employ different solutions to bind and acetylate substrates. Slight changes are sufficient to shift the substrate specificity of the GNAT fold. The same GNAT elements in Group 2 of NATs, which contains NAA50 and NAA60, are important for substrate binding and catalysis [49, 50]. Group 2 NATs acetylate protein N-termini starting with a methionine [49, 50]. Interestingly, Group 2 of NATs contains an inactive yeast NAA50 (found in cluster 135), which does not contain the characteristic β6-β7 tyrosine involved in substrate binding and which is present in Group 1 and Group 2 NATs [53]. The inactive yeast NAA50 does not have the Ac-CoA binding motif either [53]. It is therefore expected that its function is not the same as the function of NAA50 sequences which have all substrate interaction sites conserved. By comparing sequences and available structures we found that these important binding sequence motifs are absent only in the Saccharomycetaceae from the fungal Ascomycota phylum. Other families from this phylum and members of other fungal phyla seem to have conserved substrate binding sequence motifs when compared to NAA50 of other eukaryotic kingdoms. Large differences between Group 1a and Group 2 exist in the way the substrate binds to the enzyme and also in the position of the catalytic residue on the fold. The difference in catalytic strategy between Group 1a and Group 2 enzymes can be illustrated by drawing a horizontal line through the middle of the V-shaped splay across the β4 and β5 strands (Cf Fig 1); in Group 1a the active site would be above the line, while it would be below the line in Group 2. Interestingly, catalytic residues of Group 2 are conserved in Group 1a, but they are not catalytically active in Group 1a [54]. The two groups share two conserved tyrosines in each of the β6-β7 and α1-α2 loop. Both are involved in substrate binding. In Group 1a of NATs, residues at positions upstream of the mentioned tyrosine (positions -2 and -4 with respect to the Tyr) are also involved in substrate binding, but the same positions in Group 2 do not seem to be as important for binding the substrate, even though position -4 shows mutational sensitivity [51].

There are important differences between groups 3 and 4, and each is very different from the other groups. Group 3 contains NAA40, the NAT with the specificity towards the histone H4 and H2A N-termini (sequence: SGRG) [55]. The first major difference between NAA40 and Group 1 and 2 NATs is the fact that the β6-β7 loop in NAA40 has no role in determining substrate specificity [55] and does not carry an invariable tyrosine. While the α1-α2 loop of NAA40 plays a role in substrate binding [55], just like in groups 1 and 2 NATs, it has a different position in the 3D structure and its sequence motif does not bear any resemblance with the conserved motifs of groups 1 and 2. Our SSNs and phylogenetic tree all show a large distance between Group 3 and Groups 1 and 2.

Group 4 is defined by NAA80, the most recently discovered N-terminal acetyltransferase [42]. The β6-β7 loop of NAA80 is not conserved and does not play an important role in acetylation. As in most other NATs the α1-α2 loop plays an important role in substrate binding and is well conserved [56]. The α1-α2 loop sequence motif is different from those of other NATs. Moreover, NAA80 has a wider substrate binding groove between the α1-α2 and β6-β7 loops. This structural feature supports classifying NAA80 into a different NAT type. In addition NAA80 does not have the α1-α2 and β6-β7 tyrosines found in Groups 1 and 2. Our results confirm, over a larger set of sequences, an observation that has been reported earlier [56].

Predictions based on SSN and sequence motifs.

The initial SSN (Fig 2) revealed clusters containing uncharacterized sequences around clusters of known NATs. In what follows, we investigate how similar or different from known NATs these uncharacterized clusters are. To that goal, we checked for the presence in these sequences of the motifs we had characterized for the groups of known NATs (Table 1). Proteins belonging to those clusters and displaying sequence motifs close to the NAT motifs are likely to be NATs.

Around Groups 1a and 1b, no neighboring cluster showed sequence motifs close to those found in clusters 2 (NAA30), 8 (NAA20) and 10 (NAA10). Therefore, none of the connected clusters to either NAA10 or NAA20 were considered as potential new NATs. In group 2, the connected components around clusters 9 (NAA50) and 24 (NAA60) contain uncharacterized clusters numbered 49, 120, 128, 135, 136, 212 and 83 and 209, respectively (Cf Fig 3). All clusters, in addition to an isolated cluster numbered 207, have sequence motifs resembling that of group 2 (S6A Text) but also displaying significant differences. In the 6 clusters around cluster 9 (NAA50) (Fig 2), NAA50 is the only confirmed acetyltransferase and we found sequences annotated as NAA50 both in clusters 9 and 135. There are X-ray structures available for both of these clusters, and proteomics and biochemical data show that they may differ in their substrate specificity [53, 88]. We observe a difference in sequence motifs; a mutation-sensitive phenylalanine [49] in the α1-α2 loop of NAA50 (first Phe in the motif of group 2 shown on Fig 3) is replaced by a less bulky leucine in sequences from cluster 135 (Fig 6A and 6C). We observe the same differences in the α1-α2 loop between cluster 9 (NAA50) and uncharacterized clusters 49, 120, 128 and 212 (Fig 6A). Two residues downstream from the leucine/phenylalanine substitution, we observe a conserved isoleucine in cluster 120 instead of a highly conserved valine in NAA50. This valine forms van der Waals contacts with the substrate in NAA50 [49] and is, thus, important for substrate binding. Moreover, the α1-α2 loop in cluster 120 sequences contains two conserved prolines (Fig 6A) unlike NAA50 that contains only one (Fig 3). The characteristic β6-β7 motif of NAA50 (Cf Table 1 and Fig 3) is not present in cluster 135, which doesn’t have the conserved tyrosine in this loop (second tyrosine of the sequence motif on Fig 3). Structurally, the differences between cluster 9 (NAA50) and cluster 135 enzymes are precisely in the β6-β7 loop, which is longer in cluster 135 structure (S6B Text). Sequences from all other clusters around cluster 9, carry the same β6-β7 loop motif as NAA50 (Fig 6A). Finally, there are differences in sequence motifs carried by the β4 strand; the methionine responsible for interacting with the substrate in NAA50 (third position in β4 motif on Fig 4) is substituted by a glutamine in clusters 49 and 128 and by a glutamate in clusters 135 and 136, while sequences in cluster 212 retain the conserved methionine. Based on the presented differences we predict that clusters 49, 120, 128, 136, and 212 have substrate specificities different from that of NAA50. Sequences from clusters 49 and 120 were selected for experimental verification (See section 4.2 below).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Variations of sequence motifs in key positions on the GNAT fold suggest novel NATs with different ligand specificities.

When we compare the sequence motifs of NAA50 (cluster 9) and NAA60 (cluster 24) to the corresponding motifs of their surrounding clusters, we notice a number of small but meaningful differences (A). These differences occur on key positions of the GNAT fold and are illustrated here on the X-ray structure of NAA50 (PDB 3TFY) (B) The sequence differences located on the α1-α2 loop, β4 and β5 strands and β6-β7 loop residues are likely to result in altered specificity. The structure superimposition between human NAA50 from cluster 9 (orange, PDB 3TFY) and yeast NAA50 from cluster 135 (green, PDB 4XNH) highlights the small differences between residues involved in substrate binding in these two proteins with reportedly different specificities [88] (C).

https://doi.org/10.1371/journal.pcbi.1007988.g006

Sequences in clusters 83 and 209 around cluster 24 (NAA60) (Fig 2) could display substrate specificities distinct from that of NAA60. The main difference is in the α1-α2 loop (Fig 6A); a conserved acidic residue in NAA60 is substituted by a conserved positive residue four residues downstream of the mutation-sensitive phenylalanine. Given the importance of the α1-α2 loop [49–51, 54] this is likely to result in a change in ligand specificity for clusters 83 and 209. Cluster 83 contains only plant sequences while cluster 24 contains only animal sequences. The differences in sequence motifs we observe could indicate differences in specificities between plant and animal NAA60.

In group 3, the component around cluster 20 (NAA40) also includes clusters 13, 47, 104 and 222 (Fig 2) which share NatD motifs to some extent. Crucial differences exist though. Sequences in clusters 13, 104 and 222 lack an aspartic acid conserved in cluster 20 and located on the β3 strand of NAA40. Clusters 13, 47 and 222 lack a tryptophan conserved in the α1-α2 loop of NAA40 (S6C Text). Since both residues are crucial for substrate binding, the four clusters around NAA40 might have different substrate similarities. In particular D127 in β3, together with Y136 and Y138 in β4, and E129 located between β3 and β4, is involved in interactions with the first 4 residues of the NatD substrate (H4 and H2A histones) and their mutation greatly reduces the catalysis rate [55]. Interestingly cluster 20 contains not only the well characterized animal NAA40 but also sequences from the plant phylum Chlorophyta. Other plant sequences found in this group form cluster 104 and belong to the phylum Streptophyta. Fungal sequences form cluster 47 and 222 and do not mix with plant and animal sequences. The observed differences between cluster 20 and 104 indicate that fungal, plant and animal NAA40 might have different specificities.

Even though there are four clusters around NAA80 (cluster 63, group 4) (Fig 2), we did not find any variations in their key sequence motifs (S6D Text). The clustering in this case was likely based on taxonomical differences.

Experimental verification of clusters 49 and 120.

To evaluate the accuracy of our predictions, we recombinantly expressed two candidates from the clusters 49 and 120, purified them and tested their ability to acetylate N-termini from a selection of 24 amino acids-long synthetic peptides (Fig 7). One of the candidate enzymes was N1Q410 from the fungus Dothistroma septosporum. After expression and subsequent purification of N1Q410 (S7A Text), we tested its ability to acetylate N-termini of different sequences in a DTNB-based spectrophotometric assay (Fig 7A). The first seven peptides represent typical substrates for the seven known NATs in higher eukaryotes (NatA, SESS; NatB, MDEL; NatC, MLPG; NatD, SGRG; NatE, MLGP; NatF, MLGP; NatH, DDDI) [7, 42]. The subsequent six peptides have been selected dependent on the initial results for both proteins, resembling amino acid combinations that are potential substrates. Although the overall activity of N1Q410 was relatively low, there was a clear preference for methionine starting peptides, especially MDEL (21.09 ± 4.03 μM) and MEEE (15.10 ± 0.25 μM) (Fig 7A). The putative NAT A0A194XTA9 from the fungus Phialocephala scopiformis (S7A Text) showed a higher activity in general as well as a broader substrate specificity (Fig 7B). Similar to N1Q410 only peptides starting with a methionine were Nt-acetylated by A0A194XTA9, with the peptides MAPL (50.92 ± 1.89 μM), MFGP (48.94 ± 2.50 μM) and MVEL (148.74 ± 2.25 μM) showing the highest activities.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Purification and DTNB-based activity assays of the putative NATs N1Q410 and A0A194XTA9.

Putative NAT activities were tested by DTNB-based assays. 3μM of purified N1Q410 (A) and A0A194XTA9 (B) were incubated with a selection of 24 amino acids-long synthetic peptides (300 μM), and Ac-CoA (300 μM) for 1 hour at 37°C. The formation of Nt-acetylated product was spectrophotometrically determined. Shown is the mean ± SD (n = 3).

https://doi.org/10.1371/journal.pcbi.1007988.g007

Group 2 also harbors clusters 9 and 24 containing known NATs NAA50 and NAA60, respectively. Thus, we would expect that proteins from other Group 2 clusters would express NAT-activity and further that these display a substrate specificity similar to what is observed for NAA50 and NAA60. NAA50 and NAA60 have overlapping substrate specificities in vitro, but in vivo substrates are not likely to overlap since NAA50 is nuclear/cytosolic and partly anchored to the ribosome via NAA15-NAA10 [89, 90] while NAA60 acetylates transmembrane proteins via anchoring to the cytosolic side of the Golgi-membrane and other cellular membranes [33, 91]. Both enzymes may acetylate a variety of Met-starting N-termini, in particular Met-Leu, Met-Ala, Met-Val, Met-Lys, Met-Met [38, 92].

Both N1Q410 and A0A194XTA9 display clear N-terminal acetyltransferase activity confirming that these are true NATs (Fig 7). Furthermore, both enzymes prefer Met-starting N-termini among the peptides tested. A0A194XTA9 has a clear preference for the NatE/NatF (NAA50/NAA60) type of substrates strongly suggesting that this NAT is either a NAA50 or NAA60 type of enzyme in P. scopiformis. For N1Q410, we observe a preference for N-terminal peptides where Met is followed by an acidic residue at the second position, very similar to NAA20/NatB activity [39] despite the fact that it harbours sequence motifs highly similar to those of Group 2 NATs. This is an example of how sensitive N-acetyltransferase ligand specificity can be to subtle sequence changes. In this case they mainly consist in: (1) two substitutions in the α1-α2 loop where the highly conserved F and V in the NAA50 motif are replaced by an L and an I, respectively, (2) NAA50 has only one proline in the α1-α2 loop while cluster 120 has 2 and (3) the highly conserved M in the β4 strand of NAA50 is at a different position in cluster 120 sequences. Thus, N1Q410 might be a NAA20 type enzyme which is clustered among NAA50/NAA60 type enzymes in Group 2, or there might be other factors skewing the substrate preference in vitro.

Collection of sequence dataset.

All members of GCN5-related acetyltransferase superfamily contain the GNAT fold. As there is no finer classification to aid dataset creation, we retrieved all UniProt sequences that match the GNAT fold signature as defined by PROSITE [64, 98]. According to PROSITE there are four types of GNAT fold–GNAT (code: PS51186), GNAT_ATAT (code: PS51730), GNAT_NAGS (code: PS51731) and GNAT_YJDJ (code: PS51729). These PROSITE signatures match sequences from all domains of life (around 900 000 sequences in UniProt). We restricted our dataset to only eukaryotic entries (more than 50000 sequences) in agreement with the focus of this work. We kept all SwissProt (manually curated) sequences and Homo sapiens TrEMBL (not reviewed) sequences in the dataset. The remaining TrEMBL sequences were filtered to reduce the size of the dataset. Filtering of TrEMBL sequences was performed using h-cd-hit [70, 99] in three steps–a first run performed at 90% identity, a second at 80% and a third at 70% identity. The threshold was set to be 70% sequence identity as this usually indicates shared function [100–102]. We created two datasets using this strategy: the full-sequence dataset (S1 Dataset) and the GNAT-domain dataset (S2 Dataset). We used the pfamscan tool from Pfam [103] together with HMMER3.2.1 [104] to locate the GNAT fold boundaries in the full-sequence dataset in order to generate the GNAT-domain dataset.

Generating the SSNs.

The final, filtered, dataset (14396 sequences) was used to generate the SSN using EFI-EST [71] with the following parameters: E-value of 10⁻¹⁵ and alignment score of 30 (S7 File). The chosen values ensured that sequences clustering together were closely related (Cf S1A Text) with a minimal sequence identity equal to 40% yielding isofunctional clusters. The shortest sequence kept in the network was 34 amino acids long. It is not known yet what the minimal functional part of the GNAT fold is. The resulting network was analyzed using Cytoscape [73]. To visualize the network in Cytoscape we used γfiles organic algorithm by γWorks (https://www.yworks.com/). In addition to the network made from E-value thresholds equal to 10⁻¹⁵ and alignment score equal to 30 we created several other networks, mainly for the purpose of finding the best dataset for phylogenetic analyses (see Phylogeny section below for more details). Parameters for these networks were: for E-value of 10⁻⁵, alignment scores of 15, 20, 25, 30, 35 or 40; for E-value of 10⁻¹⁰, alignment scores of 15, 20, 25, 30, 35 or 40; for E-value of 10⁻¹⁵, alignment score 16, 25, 30, 35 or 40; for E-value equal to 10⁻²⁰, alignment score equal to 20, 30, 35 or 40.

Identification of isofunctional clusters and their neighbours.

In the resulting SSN (E-value 10⁻¹⁵ and alignment score 30) there were no clear boundaries between different clusters. In order to identify separate clusters, we applied the clusterONE algorithm [72] which is designed to recognize dense and overlapping regions in a graph. The search for dense regions in a network (clusters) was performed with the following parameters: minimum size of 10 sequences for a cluster to be considered, minimum density: auto, edge weights: percentage identity, and the remaining settings were taken at their default values. Next, we identified known NATs, and other non-NAT acetyltransferases, in their corresponding clusters (using annotation details added to the network) and we let these clusters be defined by experimentally confirmed enzymes (based on the assumption of cluster isofunctionality). Given the high percentage identity inside the identified clusters, we assumed cluster isofunctionality (i.e. similar ligand specificity) and transferred annotation from experimentally confirmed proteins to unknown ones within the same cluster. We also created a simplified network using the clusterONE results as input. We represented each cluster defined by ClusterONE as a single node. Nodes in the simplified network are connected by an edge if at least one edge exists between nodes of two given clusters in the original network. After adding all nodes and edges to the simplified network, we applied γFiles (https://www.yworks.com) orthogonal algorithm to get the final view.

Network analyses.

The topology of the simplified network was analyzed using Network Analyzer through Cytoscape. Mainly, we used node degree and betweenness centrality, where node degree tells how many neighbors a node has and betweenness centrality describes how important is a given node for interactions between different parts of a network. Network analyzer calculates betweenness centrality using the algorithm by Brandes [105]. Betweenness centrality was calculated on the representative, “pivot” network which is not weighted and not directed.

Motif discovery.

We used the MEME tool [74] to find characteristic sequence motifs within clusters. Each motif search was performed on all sequences of a given cluster. Enriched motifs were discovered relative to a random model based on frequencies of letters in the supplied set of sequences. As we work with protein sequences, zero to one occurrence of each motif per sequence was expected and searched for. A maximum of 25 unique motifs were searched for per sequence set, with 5 to 10 amino acid width. Only motifs with e-value below 1 were taken onto account. All motifs were visualized using SequenceLogoVis [106].

Choice of sequence dataset for phylogeny.

Since there are no clear boundaries between different acetyltransferases, due to lack of detailed classification, we based our phylogeny analysis on our SSNs. We used the more stringent SSN (E-value = 10⁻¹⁵, alignment score = 30) and selected three representative sequences for each cluster. In order to create the dataset for phylogenetic analyses, we created several networks that allowed for more connections between nodes (and clusters) (see Table A in S4 Text) and looked for the SSN with the largest single connected component (the largest group of clusters) exhibiting smallworld properties [109]. We calculated smallworldness for each of the largest connected components using NetworkX Python library [110].

Sequence alignment for phylogeny.

We selected three sequences per cluster to generate the multiple sequence alignment (S2 File). If a cluster contained sequences from SwissProt, those sequences were used in the alignment. Otherwise, TrEMBL sequences were randomly selected as cluster representatives. As sequence divergence within the acetyltransferase superfamily is extremely high, we used only the highly conserved A and B motifs of the GNAT fold. The alignment was generated using Clustal Omega [111] and the full alignment was constructed step by step. Sequences from closely related clusters were aligned first and different alignments were then merged using MAFFT [112]. Merging two alignments using MAFFT was always performed using “anchor” sequences and ensuring that both alignments had one set of five sequences (i.e. one cluster) in common (“anchor” sequences). That also ensured that corresponding secondary structure elements was kept intact after merging. Alignments generated for merging were manually edited, using acetyltransferases with known structures used as reference to increase the alignment precision.

Model of evolution.

To select the right amino acid replacement model, which describes the probabilities of amino acid change in the sequence, we used ProtTest3 [113]. As input, we used the previously generated multiple sequence alignment. Tested substitution model matrices were JTT [114], LG [115], DCMut [116], Dayhoff [117], WAG [118] and VT [119]. All rate variations were included in the calculation (allowing proportion of invariable sites or +I [120], discrete gamma model or +G [121] (with 4 rate categories) and a combination of invariable sites and discrete gamma model or +I+G [122]. Empirical amino acid frequencies were used. We calculated a maximum likelihood tree to be used as starting topology.

Construction and evaluation of the phylogenetic tree.

Finally, a maximum likelihood tree was calculated using RAxML [123] based on the generated alignment (S1 File). We used LG+G+F model of evolution since it provided the best fit according to prottest3 [113] calculation (with AIC, AICc and BIC models selection strategies). Ten searches for the best tree were conducted. Once the best tree was calculated, its robustness was assessed using bootstrap. As stop criterion we used a frequency-based criterion, by calculating the Pearson’s correlation coefficient [124]. After bootstrapping was complete, we used transfer bootstrap expectation (TBE) [125] which has been shown to be more informative than Felsenstein’s bootstrap method for larger trees built with less similar sequences.