Protein Structure Characters in the Light of Phylogenetic Systematics Authors: Nicholas J Matzke, Changhao Li, Ashar zz Malik Published: 2026 ABSTRACT AbstractProtein structure characters have great potential for improving phylogenetic inference, especially for deep nodes where amino acid sequences are highly diverged. The combination of AlphaFold structure predictions and Foldseek's “3Di” structural alphabet makes it relatively easy to conduct model-based phylogenetic inference that includes a partition of slow-evolving 3Di characters. However, we show that even identical amino acid sequences can produce substantially different 3Di characters, depending on the source of the structural model and whether inter-chain interactions are considered. We argue that such variability can be addressed with key concepts from traditional organism-based phylogenetic systematics: semaphoront, hypodigm, and character ascertainment method. To illustrate this, we develop an analogy between organismal development, taphonomy, and subsequent description and character coding by a systematist, and the process of protein synthesis, folding, and interaction and subsequent extraction, experimentation, and structural modeling by a biochemist. We conclude that differences in 3Di characters between semaphoronts are not intrinsically a problem, but they do require that the researcher uses the same replicable method on all proteins in the phylogenetic analysis. The guiding principle should be to maximize the chance that character differences in the data matrix are the results of underlying evolutionary changes, rather than artifacts due to differences in the methods used for obtaining semaphoronts and coding characters.Keywords: protein structure, structural phylogenetics, phylogenetic systematics, semaphoronts, foldseek, alphaFold We show how identical amino acid chains can produce different three-dimensional (3Di) protein structure characters. This is concerning because previous work suggests that 3Di characters are more conserved than amino acid sequences, and thus may be useful for deep phylogenetic inference. We address this puzzle with reference to the careful terminology that phylogenetic systematists developed to describe specimens and characters. Similar to organisms and fossils, proteins can exhibit different characters depending on developmental stage, the environment, and impacts from experimental procedures, as well as character coding methods. We suggest that researchers should focus on replicability and uniformity of character data acquisition methods used to generate a particular matrix, and be guided by the principle of maximizing the chance that character differences in the data matrix are the results of underlying evolutionary changes, rather than artifacts due to differences in methods used for obtaining structures and characters. Protein structural phylogenetics is an attempt to use information about protein structure, which is typically highly conserved in evolutionary history (Illergård and Ardell 2009), to estimate phylogenetic trees, or to improve the estimation of phylogenetic trees. After diverse early efforts with substantial practical limitations (reviewed by Puente-Lelievre et al. 2025), the field may be poised to take off due to the combination of two innovations. The first is easily accessible and reasonably accurate computationally predicted protein structure models produced by algorithms like AlphaFold (Jumper et al. 2021; Varadi et al. 2024), which (in monomer form) are precalculated for the entirety of UniProt, and which can be predicted for novel sequences using CoLabFold (Mirdita et al. 2022). The second is the invention of a 20-character “structural alphabet” of tertiary-interaction characters, termed “3Di” (for 3-dimensional), by the authors of the deep homology search program Foldseek (van Kempen et al. 2024; Kim et al. 2025). Foldseek can generate a 3Di structural character at every residue position of an amino acid (AA) chain, and the fact that the 3Di alphabet has 20 letters means that these codes can easily be used in readily available algorithms designed for AA-based homology searches, multiple sequence alignment, and phylogenetic modeling and inference (Puente-Lelievre et al. 2024). The combination of these tools raises the prospect that almost any protein phylogenetics exercise can be enhanced with a simple workflow (Matzke et al. 2026): (i) download or generate AlphaFold structural predictions for each protein, (ii) generate 3Di characters with Foldseek, (iii) combine AAs and 3Dis into a partitioned data matrix with simple R functions (Malik et al. 2024; Matzke et al. 2026), and (iv) conduct inference using standard programs such as IQ-TREE (Minh et al. 2022), including statistical model comparison to select optimal substitution models for the AA and 3Di partitions, adding special transition rate matrices for the 3Di characters using a transition matrix derived from the Foldseek 3DI cost matrix (Puente-Lelievre et al. 2024; van Kempen et al. 2024) or similar inferred rate matrices (Garg and Hochberg 2025). Initial results suggest that adding 3Di characters to an AA dataset can increase support and accuracy for the deepest nodes in a tree, particularly in the “twilight zone” where AA sequence similarity has decayed to very low levels (Puente-Lelievre et al. 2024). Another preliminary study suggests that adding 3Di characters statistically significantly improves average resolution of nodes in a sample of 10 protein superfamilies, although some of these superfamilies show little improvement, particularly if sequences are short and dominated by alpha-helices (Fullmer et al. 2025). Garg and Hochberg (2025) show that 3Di characters have much slower rates of change than amino acids, potentially substantially improving our confidence in estimates of classic puzzles such as rooting the Tree of Life using ancient gene duplications. However, they also note that 3Di characters can inherit information from structures that might bias phylogenetic inference. For example, in a clade of proteins where some members oscillate between two functional folds, while other members specialize on only one of the two folds, the 3Di-based phylogenetic inference will group a bistable protein based on which conformation was used to generate the 3Di characters, perhaps contradicting the phylogenetic history inferred from AAs (Garg and Hochberg 2025). Such issues suggest that, as structural phylogenetics comes into view, we need to do some careful thinking about fundamental concepts such as homology, and how they relate to new datasets such as 3Di characters, and likely additional alphabets and encodings of structure that may emerge in the future. Homology Puzzles: Same Sequence, Different 3Di Characters Homology puzzles are not difficult to come across while working with 3Di characters. We give an example from our own work. The bacterial flagellum has long been a model system in biochemistry, and the question of how such a complex system might evolve is a longstanding research topic (Pallen and Matzke 2006). As the system is very ancient, likely dating back to the last bacterial common ancestor, the relationships between flagellar and nonflagellar proteins, or between pairs of flagellar proteins in the case of ancient duplications, are often in the twilight zone. It is therefore appealing to attempt to use structure predictions and 3Di characters to improve resolution. We examined the flagellar export gate proteins, consisting of FliP, FliQ, FliR, and FlhB, which form a complex at the bottom of the Type 3 Secretion System export apparatus, which serves as both the rotating base of the flagellum and the system that selects and exports the protein subunits to progressively construct the flagellar rod, hook, and filament. When an experimental structure was resolved with cryo-electron microscopy (Kuhlen et al. 2018, 2020), it was revealed that FliPQR are structural homologs, and FlhB may share a region of homology as well (Fig. 1). Ribbon model of cryoelectron microscopy structure of the FliP5Q4R1-FlhB1 complex. FliQ consists of two bent alpha helices, a structure which is closely superimposable on two regions each of FliP and FliR. Colors and rotations done in ChimeraX 1.10.1. Structure source: https://www.rcsb.org/3d-view/6S3L, model created by Kuhlen et al. (2020). To test the behavior of Foldseek's structureto3didescriptor function, which converts a protein structure backbone chain into a series of 3Di characters equal to the number of AA residues, we generated 3Di characters on a number of input files: Kuhlen et al. (2020)'s FliPQR-FlhB structure, 6S3L, as downloaded from RCSB unmodified. Kuhlen et al. (2020)'s FliPQR-FlhB structure, 6S3L, as downloaded from RCSB unmodified. The same structure, but with each chain split out into a separate PDB file. These produced identical 3Di characters to #1, so were not examined further. The same structure, but with each chain split out into a separate PDB file. These produced identical 3Di characters to #1, so were not examined further. The same structure, but with the PDB files modified so that all chains were given the same label, “Z.” PDB structure files typically label distinct subunits/chains alphabetically. For example, in 6S3L, the 11 proteins of the FliP5Q4R1-FlhB1 structure are sequentially labeled as chains A-K in the PDB file. By default, Foldseek's structureto3didescriptor encodes 3Di characters by encoding the closest within-chain interaction for each amino acid. However, after relabeling all chains as “Z” (or any arbitrary single letter), Foldseek treats all the sequences as one large chain, allowing the coding of 3Di characters for between-chain interactions when an amino acid's closest neighbor is in another chain. A researcher might do this if they hypothesize that important and conserved structural interactions occur at the quaternary as well as tertiary level. The same structure, but with the PDB files modified so that all chains were given the same label, “Z.” PDB structure files typically label distinct subunits/chains alphabetically. For example, in 6S3L, the 11 proteins of the FliP5Q4R1-FlhB1 structure are sequentially labeled as chains A-K in the PDB file. By default, Foldseek's structureto3didescriptor encodes 3Di characters by encoding the closest within-chain interaction for each amino acid. However, after relabeling all chains as “Z” (or any arbitrary single letter), Foldseek treats all the sequences as one large chain, allowing the coding of 3Di characters for between-chain interactions when an amino acid's closest neighbor is in another chain. A researcher might do this if they hypothesize that important and conserved structural interactions occur at the quaternary as well as tertiary level. Using the ChimeraX Fetch command, we downloaded the databased AlphaFold structure models for the most similar sequences to the FliP, FliQ, FliR, and FlhB sequences used in Kuhlen et al. (2020)'s structure. These were all 100% identical in shared sequence; however, some sequences had small differences in length compared to the same protein in the structure, perhaps due to posttranslational or other modifications to the chains used in the experimental 6S3L structure. Using the ChimeraX Fetch command, we downloaded the databased AlphaFold structure models for the most similar sequences to the FliP, FliQ, FliR, and FlhB sequences used in Kuhlen et al. (2020)'s structure. These were all 100% identical in shared sequence; however, some sequences had small differences in length compared to the same protein in the structure, perhaps due to posttranslational or other modifications to the chains used in the experimental 6S3L structure. All of the resulting 3Di characters were manually aligned by eye, taking care to attempt to align the strongest structural homology region (the last 2 alpha helices in the FliP and FliR structures, which are approximately superimposable on the FliQ structure). While 3Di alignment programs are available (Matzke et al. 2026), our purpose was illustrative rather than phylogenetic; thus, we did not build the large dataset of hundreds of homologous sequences required for automated alignment of these very distant relatives. Instead, our goal is simply to display similarities and differences in the 3Di characters produced by these different input files. The resulting 3Di characters are shown in Fig. 2. Identical peptides can produce different 3Di characters in different circumstances, as shown by 3Di characters derived from the FliPQR-FlhB structure 6S3L. See text for details. We are initially confronted by a paradox: although 3Di characters have been shown to be more conserved than AAs in several studies, here we have identical amino acid sequences producing similar but decidedly nonidentical 3Di characters. This puzzle can be resolved by consideration of the Foldseek algorithm and the input structures. First, in an experimentally resolved structure, identical peptides can have somewhat different structures based on their position in the overall quaternary structure and how they pack with their neighbors. Second, Foldseek's structureto3didescriptor algorithm works by assigning each residue in a chain a geometric “theoretical center,” and then identifying which neighboring amino acid in the backbone has the closest theoretical center. This pair of amino acids then has a series of measurements taken, describing the distance and geometric relationship of the backbones containing the pair of residues. These measurements are then fed into the 3Di alphabet classifier, which was trained on thousands of aligned proteins to produce an alphabet that maximized sensitivity for retrieving deep homologs in database searches. Each residue is thus classified into a 3Di structural character. If we relabel all of the chains as “Z,” Foldseek then treats the whole structure as one large chain, and it is allowed to find closest-neighbor residues from other chains, which is quite a frequent occurrence in an integrated structural complex such as FliPQR. This could be expected to change some 3Di characters. Third, the AlphaFold-predicted structures, while done for identical or near-identical sequences, were statistical predictions of monomer structures, with no accounting for the influence of neighbors. It can be seen in Fig. 1 that the AlphaFold 3Di sequences share some features with the 3Dis from the original FliPQR chain labeling, and some with the all-labeled-Z chains. Concepts From Phylogenetic Systematics to the Rescue: Semaphoronts and Hypodigms At this point, some researchers might be tempted to throw up their hands and declare that 3Di characters have been debunked, seeing as minor differences in source files can produce different 3Dis from identical AAs. We argue that this would be an incorrect conclusion. While these kinds of ambiguities are quite rare in sequence-based phylogenetics, where the character states are determined by the genetic code, they are quite common in more traditional fields of biology such as morphological systematics and paleontology. These fields have traditionally relied on extensive comparative analyses of bones and other morphology to assign homology and then to semi-subjectively code observed variations as discrete characters. This is a large area, but a few basic concepts are useful to resolve the 3Di puzzle: semaphoront and hypodigm. The term semaphoront was coined by Hennig in 1966 (Hennig 1999 [1966 original]) to denote a “signal bearer”, ie a specimen of an organism, from which characters may be derived. In the best case, it is the museum-preserved and catalogued specimen of a whole organism, but quite often only a partial specimen is available: the dropped tail of a lizard, the jawbone of a fossil, etc. Sometimes the semaphoronts can be some other item, such as a photograph of an endangered species that cannot be collected, or a recording of a bird song. As any biology student knows, organisms are not static. Instead, especially for multicellular organisms, they go through an elaborate process of development, through which many morphologies may be displayed during growth (Fig. 3, black lines), or due to sexual differentiation or environmental effects. All of these stages might be potential sources of phylogenetic characters to be sampled by researchers (Fig. 3, gray circles). Furthermore, even at the point of death or sampling, the “development” of the semaphoronts may not cease. Taphonomy (Martin 1999) is the study of how an organism changes after death, including the processes of decay, fossilization, and degradation of a fossil, as well as potential impacts by human researchers during the process of fossil preparation and study. This can be important for phylogenetics, as there are known cases where taphonomic processes can bias phylogenetic inference (Sansom and Wills 2013). Schematic overview of how phylogenetic systematists sample characters from across the lifespan of organisms. Top: As organisms develop (thick black lines for female vs. male developmental trajectories), different characters are available for sampling. After death, taphonomic processes might further modify the specimen. Gray circles: Sources of characters that researchers might collect at different life stages. Bottom: Often, characters for a species are not all available from a single specimen (technically a “semaphoront,” a signal-bearing item). Therefore, systematists propose a hypodigm. Coded characters are pooled from across the specimens for phylogenetic inference. Abbreviations: aDNA = ancient DNA; aAAs = ancient amino acids; OTU = Operational Taxonomic Unit); ML = Maximum Likelihood inference. The final concept necessary here is the hypodigm (Fig. 3), a concept invented by the paleontologist George Simpson (Simpson 1940) to describe a hypothesis about which fossil specimens belong to which species. A hypodigm, in modern terms, is a collection of semaphoronts that researchers have presumed can be grouped in order to provide characters describing a species or, more agnostically, an “Operational Taxonomic Unit” (OTU) proposed for a particular phylogenetic analysis without necessarily being a formal taxonomic species description. These semaphoronts might all come from a single very well-studied specimen, but very often, systematists will describe characters from females, males, juveniles, embryos, as well as characters obtained using advanced technology on small parts of specimens (eg sequencing, chromosome structure, sperm ultrastructure) or using visual or auditory observation of live organisms in the wild (behavior, song recordings). How does this very brief review of classic concepts from systematics help to resolve the puzzles of 3Di characters? We propose that a distant, but helpful, analogy can be made between the processes of organismal growth, development, and taxonomy, and the subsequent character sampling process of systematists (Fig. 3), and the processes of protein synthesis, folding, and subsequent sampling and experimental description of proteins (Fig. 4). Proposal for how protein phylogeneticists might make use of concepts from organismal phylogenetics (Figure 3) in order to help interpret different sources of characters. Colors as in Fig. 3. AlphaFold and Foldseek's 3Di classifier (solid black boxes) are, from the point of view of character collection, black boxes that provide another source of characters from the amino acid sequence input. While the analogies between proteins and developing organisms are only approximate, there are some relevant similarities. A mature, folded protein is the result of a series of processes: gene expression, transcription, translation, possibly post-transcriptional and post-translational modifications, and folding as influenced by a variety of interactions with the environment, chaperones, and cooperative partners. The mature protein is not itself static and may oscillate between multiple conformations for functional or other reasons. Any of the stages from gene expression to folding to function might be a source of phylogenetic characters. Finally, in an analogy to organismal death and taphonomy, proteins in an experimental setting are extracted, purified, and subjected to elaborate preparations such as crystallization, which hopefully enables an approximation of one of the biologically relevant conformations. AlphaFold (AF) predictions provide a shortcut to structural models, but the models may differ depending on (eg) whether a monomer or multimer prediction is used. From the point of view of phylogenetic systematics, all of the resulting experimental and computational models constitute possible semaphoronts, signal bearers, from which characters for a particular protein might be coded. Foldseek provides one method for coding characters from these 3D structures The key point is that in protein structural phylogenetics, as with organismal phylogenetics, we should move away from the assumption, derived from sequences, that there is one and only one set of “correct” characters that can be coded from a protein. What we actually have is a number of possible semaphoronts that give us various windows into a protein's structural conformations during its life cycle, and then we have a number of ways of encoding conserved structural characters for phylogenetic inference—and the number is sure to increase as researchers invent additional structural alphabets beyond 3Di. This point of view suggests that researchers should not focus on whether there is a “correct” set of structural characters for a particular protein, and instead should put effort into ensuring that their methodology for producing structural characters is replicable, and is the same for all of the proteins under study (Puente-Lelievre et al. 2024). For example, due to the variability in 3Di coding documented above, it would likely be hazardous to run a phylogenetic analysis where half the proteins had 3Dis derived from experimental structures, and half from AlphaFold predictions. Similarly, it could be hazardous to combine monomer and multimer structural predictions, or within-chain 3Di characters with within-complex 3Di characters (where all chains are given the same chain label, allowing Foldseek to recognize between-chain interactions in its coding). Additional cases could arise from structures derived from different experimental techniques (eg X-ray crystallography vs. cryo-EM), or from NMR structures which can provide an ensemble of structures rather than a single estimate. In both cases, identical AA sequences might produce different 3Di characters. While the relative rarity of experimental structures often leaves researchers in a “beggars can’t be choosers” situation when seeking data for a phylogeny of a protein clade, some possibilities for accounting for 3Di variability can be considered. First, for each position in each AA sequence with multiple structures, a researcher could randomly sample each 3Di character from the available collection of 3Di strings at that position, creating a pseudobootstrap sequence roughly analogous to the structural variants sampled with MD in Malik et al. (2020)'s structural phylogenetics method. Repeating Maximum Likelihood many times for a collection of these pseudobootstrapped sequences would measure the variability in results due to possible variation in characters. However, sampling characters from different structures might well create pseudobootstrap 3Di sequences that represent no observed structural estimate, with uncertain results. A probably more valid alternative would be to simply sample the available structures for each OTU at random and measure the resulting variability in phylogenetic results. These randomization approaches are, however, computationally expensive and thus only feasible for smaller phylogenetic analyses. A more efficient alternative could be to calculate pairwise structural distances between all available structures, and then for each OTU, select the single structure closest to the overall cluster. This would presumably tend to select the structures with conformations closest to the overall average conformation, which arguably approximates the goal of obtaining characters with methods/data that are the “same for all of the proteins under study.” On the other hand, this method has the risk of favoring a low-frequency conformation for a protein if it is closer to the overall average. This paper focuses on the effects of differences in ascertainment procedure that a phylogeneticist might use to obtain 3Di characters from available estimated structures. However, we do not mean to suggest that the issue of error in experimental or AlphaFold structures should be ignored. In the case of AlphaFold, estimates of structural uncertainty are readily available, or can potentially be approximated by collecting a sample of estimated structures from different starting conditions. Such sources of uncertainty might be incorporated in phylogenetic analyses by sampling procedures such as those discussed above, or perhaps by weighting the likelihood calculation of each site by some function of the local structural uncertainty. However, both theory and our early experience suggest that such procedures may have diminishing returns. The low-confidence regions of AlphaFold structures tend to be loops and other less-stable regions outside the core stable folds, and these regions tend to be more evolutionarily more labile in sequence and structure, and thus are less likely to produce the conserved characters that are highly influential in supporting the existence of inferred clades deep in the tree where it is hoped 3Di will help resolve relationships where AA similarity has decayed. All of the above suggestions are worthy of further testing against reference datasets. Nevertheless, it may be impossible to avoid judgment calls in some situations. However, as we have argued in this paper, this situation is not radically different from morphological characters in traditional systematics. We suggest that the guiding principle should be to maximize the chance that character differences in the data matrix are the results of underlying evolutionary changes, rather than artefactual differences in the methods used for character ascertainment.