Ancient dynamin segments capture early stages of host–mitochondrial integration

Functionally Diverse Dynamins Are Found Across Eukaryotic Supergroups.

Members of the dynamin superfamily are characterized by three major domains: an N-terminal GTPase domain (ND), a stalk region formed by helices of the middle domain (MD), and a GTPase effector domain (GED) that folds back across the stalk to regulate GTPase activity (Fig. 1A). Most dynamins act by forming homo-oligomeric rings around membrane necks through interactions of the stalk, driving scission via GTP hydrolysis (12). Starting with a curated list of functionally diverse dynamins, we used ND, MD, and GED alignments to query a broadly sampled protein sequence database (Methods, Construction of the Dynamin Database). We thus identified and aligned 3,827 eukaryotic dynamins from 854 species or strains representing all extant eukaryotic supergroups, as well as 14 bacterial dynamins (Datasets S1 and S2; see SI Text for dataset descriptions).

• Download figure
• Open in new tab
• Download powerpoint

Fig. 1.

Breaking dynamin into evolutionary segments. (A) The crystal structure of human dynamin 1 [Protein Data Bank (PDB) ID: 3SNH] (18). Dynamins have three conserved domains: the N-terminal GTPase domain (ND, light brown), the middle domain (MD, brown), and the GTPase effector domain (GED, dark brown). (B) We assemble a 556-residue concatenated alignment of the ND, MD, and GED. Our analysis relies on finding clusters of proteins with high sequence identity across short segments of this alignment. Two contiguous segments with similar clusters can be merged into a single longer segment with no loss of information. The upper-triangular matrix R_ij shows the properties of long segments starting at residue i and ending at residue j; high values indicate that the long segment is made up of multiple subsegments with similar clustering properties (SI Text, Evolutionary Segments). The saw tooth structure shows that dynamin breaks into eight natural evolutionary segments ranging from 22 to 118 residues in length: three in the ND, four in the MD, and one in the GED. Every dynamin is assigned an eight-letter signature, such that two proteins with the same letter at a given segment have high sequence identity across that segment (SI Text, Sequence Clustering).

To examine the diversity within this dataset we used the graphical CLANS tool, which clusters proteins by sequence similarity (SI Text, Sequence Clustering) (13). Applying CLANS to the full alignment, we found that the proteins split into seven major classes (Fig. S1D), each associated with distinct functional annotations (Table S1), whose mutual relationships were not well resolved (Fig. S2). Three of these—which we term superclasses A, B, and C—span multiple eukaryotic supergroups. Given a set of present-day proteins and a backbone tree of eukaryotic species (11, 14), we define the “root species” as the last common ancestor of the species in which those proteins are found and the “root protein” as the last common ancestral protein itself. Class A (blue) consists of 1,891 dynamins, including those involved in mitochondrial and peroxisomal division, vesicle scission, or cell plate formation, with LECA as the root species; class B (orange) consists of 1,390 myxovirus-resistance-like dynamins, with LECA as the root species; class C (green) consists of 136 ND-only dynamins, including those involved in cytokinesis or chloroplast division, with the last common ancestor of amoebozoans and archaeplastids as the root species. The remaining four classes are confined to individual lineages: 16 alveolate-specific proteins related to class C (Fig. S2), 252 fungal (Mgm1) and 142 metazoan (OPA1) mitochondrial fusion dynamins, and 14 bacterial dynamins.

Ancestral Dynamins Can Be Reconstructed by Parsimony.

We are interested in inferring sequence and function of the root proteins of present-day dynamin superclasses. Although evolutionary rates can vary over time and across the protein sequence at billion-year timescales, maximum-likelihood methods incorporating rate heterogeneity can accurately reconstruct protein phylogenies. In effect, conserved regions of a protein similar to the ancestral sequence provide a reliable phylogenetic signal (Fig. S3). However, outside these regions, the reconstructed sequences are often inaccurate. Here we present a sequence reconstruction method based on parsimony, which can outperform maximum likelihood when evolution is heterogeneous (15). This relies on a compact representation of protein sequence that makes clear which regions of an ancestral protein have been reliably reconstructed. We break up the alignment into a small number of contiguous “evolutionary segments” (SI Text, Evolutionary Segments). For a given protein at a given segment, we compress the entire amino acid sequence into a single letter, such that two proteins with the same letter have similar sequences, beyond some threshold identity X% (SI Text, Sequence Clustering). Each present-day protein is thus assigned a short signature consisting of a string of letters: Each position corresponds to a segment, and each letter labels a cluster of proteins with highly similar sequences across the relevant segment. Dollo parsimony then implements the following rule (16): If two present-day proteins have the same letter at a given segment, then their root protein must also have that letter. Ancestral proteins are thus assigned a signature, but with gaps (?) wherever the sequence cannot be reconstructed. If an ancestral signature has a certain letter, it implies the ancestral sequence is >X% identical to present-day proteins with the same letter across the relevant segment.

Crucially, the results of Dollo parsimony are guaranteed to be valid independent of the choice of segments and thresholds, as long as convergence can be ruled out. To prevent convergence, we must use some combination of long segments and high thresholds, so the chance of two unrelated proteins accidentally having >X% identity across the entire segment is negligible. However, if the segments are too long or the threshold is too high, it is likely that only recently diverged protein pairs have >X% identity across whole segments. We would then be able to reconstruct recent ancestral proteins, but the signatures of ancient proteins would consist mostly of gaps. Also, if the threshold were so low that any pair of homologous proteins would have >X% identity across all segments, all proteins both present day and ancestral would trivially have the same signature. This would prevent us from inferring the function of ancestral variants. The details of how segment boundaries are chosen and how sequence stretches at each segment are grouped into clusters corresponding to letter labels are given in SI Text, Evolutionary Segments and SI Text, Sequence Clustering. To understand our results, only the following properties of segments and letter clusters are relevant. First, we find that dynamins split into eight well-defined evolutionary segments ranging in length from 22 to 118 residues, smaller than domains but significantly correlated with secondary-structural elements (Fig. 1B) (17, 18). In practice, segments 20 residues or longer are sufficient to prevent accidental convergence and can retain nearly as much phylogenetic information as an entire protein (Fig. S3D). Our segments are therefore sufficiently long for parsimony to be valid. Second, when we examine the clusters of sequences corresponding to each segment-wise letter label, almost all clusters (223 of 238 clusters across segments; Dataset S1, sheet 2) have a median pairwise identity >X = 35%. This is comparable to the level of conservation between pairs of proteins with similar function (19). Therefore, if we do find present-day proteins with signatures nearly identical to those of some reconstructed ancestor, we can infer the function of the ancestor based on the function of its extant descendants.

Ancient Segments, Unchanged for Billions of Years, Are “Living Fossils.”

It is illuminating to benchmark our segment-based approach against a maximum-likelihood phylogenetic analysis; further validation steps are described in Fig. S3 and SI Text, Benchmarking Segment-Based Parsimony. Each letter label at each segment is associated with a cluster of proteins. For each such cluster, we define three attributes (Fig. 2 and Dataset S1, sheet 2):

• Attribute 1: The pairwise identity between sequences within a cluster. This allows us to estimate the sequence of root proteins: If convergence is ruled out, the root protein of a pair of present-day proteins is at least as similar to each of them as they are to one another. Of 238 clusters, 223 have a median intracluster pairwise identity >35%. If the phylogeny within such clusters is balanced, this implies that the root protein is >35% identical to at least one present-day protein over the relevant segment.

• Attribute 2: The root species of a cluster. This allows us to estimate the age of root proteins: Assuming no horizontal transfer, if a cluster contains proteins from multiple eukaryotic supergroups, its root species and root protein are necessarily ancient, over 1.5 billion y old (11). In contrast, the root proteins of supergroup-specific clusters are potentially more recent. We find few ancient clusters: Class A dynamins have an ancient cluster at each of segments 1–8, class B dynamins have an ancient cluster at each of segments 1–4 and 6, and the ND-only class C dynamins have an ancient cluster at each of segments 1–3.

• Attribute 3: Whether the cluster is monophyletic. This establishes that the root protein of one cluster is not ancestral to that of any other cluster. To be monophyletic, member sequences of a cluster must repeatedly form a well-defined clade over 1,000 bootstrap phylogenetic trees calculated for a given segment (Methods, Maximum-Likelihood Analysis). The majority of supergroup-specific clusters were strongly monophyletic (median monophyly support: class A, 800/1,000; class B, 567/1,000; and class C, 759/1,000; Fig. 2C, Left). However, we found that the monophyly support levels of ancient clusters were significantly different from those of supergroup-specific clusters (P = 5E-6, Kolmogorov–Smirnov test; Fig. 2C, Right). This is striking, given that both ancient and supergroup-specific clusters had sequence identities in the same range of 50–90%. Whereas ancient class C clusters were monophyletic (median monophyly support: 356/1,000; Fig. 2C, Center), ancient class A and B clusters were not (median monophyly support: class A, 0/1,000; class B, 0/1,000; Fig. 2C, Center).

• Download figure
• Open in new tab
• Download powerpoint

Fig. 2.

Living fossils: tight, ancient, paraphyletic clusters. (A) (Left) Schematic protein tree. Proteins are grouped into clusters of high pairwise identity corresponding to letter labels (colored ovals and circles). For each cluster we show extant proteins (solid dots) and their last common ancestor (root proteins: open dots). If descendants of a root protein are confined to the cluster itself, the cluster is monophyletic; otherwise it is paraphyletic, with other clusters branching out of it. (Right) Extant proteins are mapped to the leaves of a species tree. If a cluster contains proteins from multiple species, their last common ancestor is its root species. Three superclasses (blue, orange, and green) emerge from three root proteins (“A,” “B,” and “C”). Proteins similar to A and B survive to the present day, and no proteins similar to C survive. (B) Maximum-likelihood trees for one trial of segment 2. We show a consensus split network (48) representing 100 bootstrap replicates; nodes with strong bootstrap support appear as thin stems. Ancient clusters (2A, 2B, and 2M) are shown with colored overlays and the rest with letters; not all clusters are labeled. Paraphyletic clusters 2A and 2B form “galls” from which other class A (blue) and class B (orange) clusters emerge; monophyletic clusters like 2M have no outward branches. Clusters 2U and 2W map to the center of the tree due to long-branch attraction (Fig. S3D). We map ancestral proteins to a species tree (Mya: millions of years ago); species group labels are as in Fig. 3. Ancient clusters contain proteins from multiple supergroups (labels in boxes). (C) To find the phylogenetic attributes of clusters corresponding to each letter label, the analysis shown in Fig. 2B for segment 2 is repeated for each segment separately. We show the attributes of all 110 clusters in our dataset with more than 10 members. Attribute 1: pairwise identity (y axis, Left and Center; dashed line, 35% identity). Attribute 2: root species (open circles, ancient, root species >1.5 billion y old; solid circles, supergroup-specific, root species <1.5 billion y old). Attribute 3: monophyly (x axis, Left, Center, and Right, monophyly support out of 1,000 trees; Right, cumulative distribution of monophyly support). If a cluster is tight (pairwise identity >35%), ancient (root species >1.5 billion y old), and paraphyletic (monophyly support <∼100/1,000), its members are living fossils: extant proteins similar in sequence to the ancient root proteins of functional superclasses. All ancient class A and class B clusters (e.g., 2A and 2B, Center) are paraphyletic and therefore ancestral. The ancient class C clusters (1L, 2M, and 3N) are monophyletic.

There are two reasons a cluster might not be monophyletic: Either its phylogenetic support is weak (e.g., cluster 2E in Fig. 2B) or it is paraphyletic, meaning its root protein is ancestral to members of multiple clusters. As evident in Fig. 2B (and also Fig. S3A) the ancient class A and class B clusters at each segment are truly paraphyletic: They are not only ancient, but also ancestral to other supergroup-specific class A and class B clusters. To reflect their important status, we label the unique ancestral cluster at each segment after the class itself (e.g., 2A or 2B at segment 2; for ancestral clusters, the same letter indicates distinct sequences at each segment; e.g., 1A and 2A are distinct). These clusters are living fossils: sets of extant proteins similar in sequence (over the relevant segment) to the ancient root proteins of functional superclasses. Although sequence remnants of the root class C dynamin are lost, those of the root class A and class B dynamins persist to the present day.

Entire Ancestral Dynamin Variants Are Preserved in Extant Species.

Every protein in our dataset is assigned an eight-letter signature (Table S1 and Dataset S1, sheet 1; the segment value is redundant and therefore omitted in signatures; e.g., 1A, 2A, … , 8A becomes AAAAAAAA; class C dynamins have three-letter signatures). By carrying out the parsimony analysis independently for each segment, we can reconstruct the full signatures of ancestral protein variants. We find that LECA had a class A dynamin identical in signature to that of mitochondrial and peroxisomal division dynamins of many extant species (AAAAAAAA; Fig. 3A). Vesicle scission dynamins in fungi and metazoans retain similarity to the ancestral variant only in their N-terminal GTPase domains; those of land plants and alveolates are even more diverged. A class A variant of unknown function in green algae (A:A::XYZ) appears to have duplicated in land plants, into a vesicle dynamin (:W:-:XYZ) and a phragmoplastin (AVAA:XYZ) responsible for cell plate formation (20) (Fig. S3). LECA also had a class B dynamin whose closest extant variants are found in the metazoans (BBBB:B::; Fig. 3B). These dynamins have a patchy present-day distribution and appear to have undergone a recent radiation in fungi; in vertebrates these myxovirus-resistance-like proteins have antiviral functions (21).

• Download figure
• Open in new tab
• Download powerpoint

Fig. 3.

Punctuated diversification of dynamins across 1.8 billion y. The heterogeneous nature of dynamin evolution is highlighted, over time and across the protein sequence. The evolutions of different dynamin superclasses (colored) and of mitochondrial FtsZ (black) are overlaid on the eukaryotic tree (gray). The timing of speciation events, from 2,000 million years ago (Mya) to the present day, is approximately as reported in Parfrey et al. (11); the FtsZ loss data are derived from Fig. S6. Certain events (mitochondrial genome loss, FtsZ loss, other gene loss and duplication events, and horizontal transfers) are shown on the correct branches but not necessarily at the correct times. Species group labels: Al, alveolates; St, stramenopiles; Ar, archaeplastids; GA, green algae (including land plants); RA, red algae; Ex, excavates; Am, amoebozoans; Op, opisthokonts; Me, metazoans; Fu, fungi. We represent dynamins with 8-letter signatures (Dataset S1, sheets 1 and 2; class C dynamins span only segments 1–3). Each segment-wise letter labels a cluster of similar sequences; two proteins with the same letter at a given segment thus have high sequence identity across that segment. The key (Right) shows the full set of 238 letter clusters across segments and superclasses, sorted as in Dataset S1, sheet 2. The letters A and B represent distinct clusters at each segment, corresponding to the reconstructed root proteins of classes A and B; the root protein of class C cannot be reconstructed. Most supergroup-specific clusters are labeled with the shorthand symbol “:” but are assigned full cluster labels in Dataset S1; no two instances of : in this figure correspond to the same full label. A subset of clusters, including all those spanning multiple supergroups, is labeled by segment-wise letters for ease of reference. Gaps are labeled “-”; segments of ancestral proteins that cannot be reconstructed are labeled “?.” We show functional annotations [mitochondrial division (MID), vesicle scission (VES), etc.] where they have been experimentally verified (Table S1). (A) (Left) Class A dynamins specialized for mitochondrial division (MID; solid blue), or lone/bifunctional (1/BIF; dashed blue and lavender; “lone” indicates a single copy per genome). (Right) Derived class A dynamins specialized for vesicle scission (VES; lavender) or phragmoplast/cell plate formation [phragmoplastin (PHR); purple]. (B) Class B dynamins involved in antiviral activity (AVA) (orange). (C) Class C dynamins specialized for cytokinesis (CYT) (green) or chloroplast division (CHD) (dark green). The cytokinetic (LMN) and plastid division dynamins (PQR) are sister groups, descended from an ancestral dynamin of unknown sequence (???) (Fig. S2). The vertical arrow shows horizontal gene transfer by secondary endosymbiosis of red algae by diatoms (Bacillariophyta). Some rootings of the eukaryotic tree would place the ancestral class C dynamin at LECA itself (Fig. S4).

Because the ancestral class C dynamin appears lost and therefore cannot be reconstructed by parsimony, we carried out a phylogenetic analysis of GTPase domains (concatenated segments 1, 2, and 3) for every class C dynamin in our dataset (Fig. S2). This revealed that the cytokinetic dynamins were monophyletic (LMN: 88% bootstrap support), and represented a sister group to the chloroplast division dynamins (PQR: 43% bootstrap support). The last common ancestor of amoebozoans and archaeplastids therefore had a class C cytokinetic dynamin (LMN) (22) that persists essentially unchanged in many lineages, as well as another class C dynamin of unknown sequence (???) (Fig. 3C). Descendants of this unknown ancestor are specialized for chloroplast division in archaeplastids and plastid-bearing stramenopiles such as diatoms (PQR); other diverged descendants are present but uncharacterized in alveolates, amoebozoans, one rhizarian (Bigelowiella natans), and one diatom (Thalassiosira oceanica). Because diatom plastids are derived from the secondary endosymbiosis of red algae (23) and the chloroplast dynamins of these lineages branch close to one another (Fig. S2), we attribute the occurrence of the PQR variant in stramenopiles to horizontal transfer.

Extant Eukaryotes Have Separate Mitochondrial and Vesicle Scission Dynamins or a Potentially Bifunctional Dynamin.

Alveolates, land plants, metazoans, and fungi each have distinct class A dynamins specialized for mitochondrial division (the ancestral variant; Fig. 3A, Left) and vesicle scission (a derived variant; Fig. 3A, Right). In amitochondriate eukaryotes (24) such as microsporidia and the Entamoebidae, the ancestral class A variant appears diverged or deleted (although microsporidia retain the diverged fungal vesicle dynamin). Surprisingly, several eukaryotic lineages encode just a single class A dynamin; these include both the oomycete and diatom lineages of stramenopiles (25/30 species or strains, median protein repertoire 17,368), red algae (6/7 species or strains, median protein repertoire 7,176), and excavates (30/36 species or strains, median protein repertoire 9,858). Assuming that dynamins are essential for mitochondrial division as well as for clathrin-coated vesicle scission, we predict these lone dynamins to drive both processes. There is partial although not conclusive experimental support for this idea. The class A dynamin of the excavate Trypanosoma brucei localizes to mitochondria (25) but appears to be required for mitochondrial fission as well as endocytosis (26); that of the amitochondriate excavate Giardia lamblia colocalizes with clathrin-coated vesicles (27). The class A dynamin of the red alga Cyanidioschyzon merolae has a punctate distribution on the cell surface between mitochondrial division events, suggesting it might also participate in vesicle scission (28, 29). The amoebozoan Dictyostelium discoideum has two class A variants, of which DymA is ancestral; this protein is not essential for mitochondrial fission but influences mitochondrial morphology (30) and participates in both mitochondrial and vesicle activities (31).