Skip to main content
NCBI home page
Bioinformatics logoLink to Bioinformatics
. 2026 Jul 3;42(7):btag470. doi: 10.1093/bioinformatics/btag470

TEDLH: domain HMMs for sensitive detection of remote homologues

Claudia Alvarez-Carreño 1,2,✉, Anton S Petrov 3,4, Vaishali P Waman 5, Ian Sillitoe 6, Christine Orengo 7
Editor: Pier Luigi Martelli
PMCID: PMC13385337  PMID: 42398027

Abstract

Motivation

The Encyclopedia of Domains (TED) provides domain annotations for proteins in the AlphaFold Protein Structure Database (AFDB) using a consensus of three state-of-the-art structure-based methods. We used these annotations to construct profile Hidden Markov models (HMMs), collectively forming the TED Library of HMMs (TEDLH). TEDLH enables sensitive sequence and profile searches, supporting systematic exploration of protein domain families and their evolutionary relationships.

Results

TEDLH links 934,186 domain HMMs to experimentally determined CATH-PDB structures through direct (primary) and transitive (secondary and tertiary) relationships. Fewer than half of TEDLH HMMs are directly linked to a CATH-PDB domain; the remaining models are connected through transitive relationships. These transitive links extend coverage into more divergent regions of sequence space and better represent CATH superfamily diversity. HMM–HMM comparisons within CATH superfamily 3.30.70.100 illustrate how transitive relationships expand sequence coverage. In this superfamily, 5640 TEDLH HMMs are connected to 173 CATH-PDB representatives. Primary, secondary, and tertiary relationships progressively capture more divergent sequences (pairwise sequence identity <20%) that retain structural similarity (TM-score ≥0.6) and a conserved two-layer α/β sandwich core fold. All-against-all HMM–HMM comparisons across TEDLH also reveal sequence similarities across the CATH hierarchy (cross-hits). At low query coverage (<50%), cross-hits are more frequent between CATH classes, architectures and topologies, whereas at higher coverage thresholds (≥70%) they predominantly occur between superfamilies. These cross-hits are not driven by superfamily size or sequence diversity and can provide guidance for CATH curation. As an example, analysis of cross-hits between superfamilies 2.170.130.30 and 3.10.20.30 reveals evolutionary relationships between these groups.

Availability

TEDLH is compatible with HH-suite3 and is available from FigShare https://doi.org/10.6084/m9.figshare.28531754 for local use.

1 Introduction

Protein domains are conserved units of function and evolution. For the past thirty years, domain annotation has followed two complementary paths. Domains identified in experimentally determined structures from the Protein Data Bank (Berman et al. 2000) have been systematically classified and mapped onto protein sequences by resources such as CATH (Orengo et al. 1997), SCOP (Murzin et al. 1995), and ECOD (Cheng et al. 2014). Parallel approaches based on alignments of homologous sequences have enabled the identification of domains in protein sequences through patterns of sequence conservation [e.g. Pfam (Sonnhammer et al. 1997), SMART (Schultz et al. 1998), and TIGRFAMS (Haft et al. 2003)]. Together with structure-based classifications, these efforts laid the foundation for profile-based models, including Gene3D (Buchan et al. 2002) and SUPERFAMILY (Gough and Chothia 2002), which extend domain annotation to hundreds of millions of uncharacterized sequences while capturing both evolutionary and structural information. More recently, highly accurate protein structure prediction methods have provided access to predicted structures at an unprecedented scale. The availability of this expanded structural data creates new opportunities, as well as new challenges, for structure-informed domain annotation and classification.

The AlphaFold Database (AFDB) (Varadi et al. 2022) transformed access to protein structural data by supplying predicted three-dimensional models for more than 214 million UniProt (UniProt_Consortium 2019) sequences. Since its release, several resources have been developed to provide annotation and to integrate AFDB with other bioinformatics databases and resources. The Encyclopedia of Domains (TED) (Lau et al. 2024) catalogues 365 million putative domains in proteins from AFDB (version 4), of which more than 251 million are mapped to CATH (Orengo et al. 1997). CATH (Orengo et al. 1997) is one of the most widely used protein domain classification systems, organizing domains hierarchically into Class, Architecture, Topology, and Homologous superfamily to capture shared structural features and evolutionary relationships. Domain assignments in TED are derived from the consensus of three state-of-the-art structure-based methods: Merizo (Lau et al. 2023), Chainsaw (Wells et al. 2024), and UniDoc (Zhu et al. 2023). TED greatly expands both the number and the sequence diversity of domains that can be assigned to CATH with high confidence.

Here, we present TEDLH, a searchable library of profile Hidden Markov Models (HMMs), that captures the expanded coverage of sequence space enabled by structure-informed domain annotations. Data for TEDLH (TED Library of Hidden Markov Models) were derived from non-redundant high-quality TED domains, as determined by the Qscore, which integrates multiple indicators of assignment reliability. This strategy propagates the structure-based domains annotated in TED to the sequence level. By retrieving sequences directly from TED and using its pre-defined domain boundaries, we mitigate the risk of homologous over-extension (Gonzalez and Pearson 2010) that can occur in iterative sequence searches [e.g. PSI-BLAST (Altschul et al. 1997) or jackhammer (Johnson et al. 2010)]. Moreover, because TED annotates the AFDB (version 4), it provides a broad and diverse set of domain sequences. As a result, sequence diversity can be sourced directly from TED to build multiple sequence alignments (MSAs) rather than from general protein sequence databases. These features give TED-derived MSAs two major advantages: precise domain boundary definitions and high sequence diversity.

Here we use similarity to intermediate sequences to map experimentally determined structures in the PDB (Berman et al. 2000) classified by CATH (CATH-PDB) to a library of HMMs built from TED sequences (TEDLH HMMs). Intermediate sequence searches have long been employed to detect remote homologous relationships because these methods substantially increase the sensitivity of sequence comparisons (Park et al. 1997, Salamov et al. 1999, Li et al. 2000). HMM identifiers in TEDLH encode the associated CATH-PDB domain and superfamily, as well as whether the connection is direct (primary) or through an intermediate HMM (secondary or tertiary).

To demonstrate the functionality of the TEDLH library, we selected two examples from the all-against-all scanning of HMMs in TEDLH. The first example illustrates sequence relationships within superfamily 3.30.70.100. We also show examples of ancestral relationships across CATH superfamilies 2.170.130.30 and 3.10.20.30. TEDLH facilitates efficient exploration of protein domain diversity at the sequence level. TEDLH is publicly available via figshare at https://doi.org/10.6084/m9.figshare.28531754.

2 Methods

2.1 Domain selection and initial library preparation

Sequences of proteins with non-redundant domains in the TED database (Lau et al. 2024) were retrieved from UniProt and trimmed according to the domain boundaries defined in TED. Each domain sequence was labelled with its corresponding CATH superfamily assignment in TED and subsequently filtered by Qscore. The Qscore assesses the reliability of a TED domain assignment by integrating structural coverage, domain assignment consensus, compactness, and pLDDT. It ranges from 0 (no confidence) to 100 (high confidence). Qscores above 75 are considered high confidence. High-confidence TED domains were grouped according to their structure-based CATH superfamily assignments provided by TED. TED sequences within each CATH superfamily group were then clustered by sequence similarity with MMseqs2 (Mirdita et al. 2019) using a maximum E-value threshold of 1 × 10−3 and a bi-directional coverage threshold of 80. This strategy takes advantage of the high-quality domain boundaries and the inclusion of remote relatives defined by TED. Clustering 111,970,613 TED sequences with MMseqs2 version 18-8cc5c yielded 1,160,453 clusters with two or more sequences and 990,748 singletons.

Because the full potential of an HMM comes from the variation captured across multiple homologous sequences, singletons were excluded, and the remaining clusters were each aligned with MAFFT v7.526 (Katoh and Standley 2013). For clusters with more than 300 members, redundancy was reduced by selecting representative sequences at a maximum of 70% pairwise identity using MMseqs2 clustering.

Multiple sequence alignments (MSAs) were converted into individual HMMs using hhmake from HH-suite3 version 3.3.0 (Steinegger et al. 2019). To make all-against-all HMM–HMM comparisons computationally tractable, the 1,160,453 HMMs were partitioned into 39 libraries, one per CATH-Architecture. Libraries were built following the HH-suite3 user guide, omitting secondary structure scoring as recommended. The size of each library is summarized in Supplementary File 1. Each HMM was queried against the library of its corresponding architecture using a single iteration of hhblits. All-against-all HMM–HMM comparison results within each Architecture were then analysed to identify transitive relationships between HMMs.

2.2 TEDLH library curation

2.2.1 Recovering TED-HMMs related to CATH

To ensure that the final library contained only TED-HMMs with unambiguous links to experimentally determined CATH-PDB domains, we performed a cross-referencing step. Non-redundant CATH-PDB domain sequences (CATH v4_4_0) were scanned against all TED-HMMs using hhblits (Steinegger et al. 2019). A TED-HMM was considered to have a primary relationship to a CATH-PDB domain only if two criteria were met: (i) the HMM–sequence match exceeded predefined thresholds (E-value ≤ 1 × 10−1, query coverage ≥ 70%, template coverage ≥ 70%), and (ii) the TED domains used to build the HMM were all assigned by TED to the same CATH superfamily as the matched CATH-PDB domain. We adopted a relaxed E-value threshold because the agreement between independent structure-based TED assignments and CATH-PDB classifications provides additional support for homology. For each primary TED-HMM, the best-matching CATH-PDB domain was retained and encoded in the HMM header together with the corresponding CATH superfamily (Fig. 1b).

Figure 1.

Three-panel figure describing the TEDLH (TED Library of HMMs) pipeline. Panel a shows a funnel diagram where 324 million TED-100 domains are categorized as Not-assigned (85M, blue), CAT-assigned (46M, light green), or CATH-assigned (194M, dark green). Below, a workflow shows CATH-assigned domains grouped into 1 million clusters, converted into 1 million multiple sequence alignments, then built into HMMs, which are filtered down to 934,186 HMMs spanning 4,689 superfamilies (labeled TEDLH). A separate diagram illustrates the filtering strategy: a CATH-PDB domain query matches a primary hit (P, dark green), which links to secondary hits (S, light green) and tertiary hits (T, grey). Panel b shows an example HHblits search output: a query protein (G7CLC7_MYCT3) searched against the tedlh_class3 database, returning a ranked hit table. The top hit (3.50.50.60|1d4cA02|P878178, highlighted) is annotated in an inset showing how hit identifiers encode a CATH superfamily code, related CATH domain, and unique identifier, with "P" marking it as a primary hit. Panel c shows three cartoon protein structure models illustrating the structural similarity underlying the HMM match.

Open in a new tab

TED Library of HMMs. (a) Construction of TEDLH from TED domain annotations. TED domain sequences were grouped by CATH superfamily and clustered by sequence similarity using MMseqs2. Sequences within each cluster were aligned to generate multiple sequence alignments (MSAs), which were subsequently converted into profile hidden Markov models (HMMs). TEDLH comprises 934,186 TED-HMMs with primary, secondary, or tertiary relationships to CATH-PDB domains. (b) Example output from an HHblits search against TEDLH. Inset: information encoded in the HMM header, including the associated CATH superfamily (sf). (c) Regions of structural similarity between the query example (UniProt ID: G7CLC7) and the corresponding CATH-PDB template (CATH domain: 1d4cA02) are highlighted in colour.

To identify TED-HMMs related to CATH-PDB beyond primary links, we analysed all-against-all HMM–HMM comparisons within each Architecture. Secondary relationships were defined as TED-HMMs linked to a primary HMM that matches a CATH-PDB domain, and tertiary relationships were defined as TED-HMMs linked only to secondary HMMs. The same thresholds used for primary hits (E-value ≤ 1 × 10−1, query and template coverage ≥70%) were applied for secondary and tertiary relationships, and all sequences in the TED-HMM MSA were required to share the same CATH classification as the CATH-PDB domain to which they are ultimately linked. As a final filtering step, structural similarity was verified by superposing a representative of each primary TED-HMM onto its closest CATH-PDB hit (Supplementary File 1). TED-HMMs whose representative structure did not meet the structural similarity threshold (i.e. having SSAP score < 70) were excluded from the final database. TED-HMMs with primary, secondary, or tertiary relations to CATH-PDB were compiled into the final TED HMM Library (TEDHL). Using this strategy, we recovered 934,186 TED-HMMs that could be associated with a CATH-PDB domain through transitive homologous relationships. Thus, primary, secondary and tertiary relationships reflect both sequence similarity and concordant structure-based domain classification.

The final library (TEDLH) was built with the HH-suite3 using these 934,186 TED-HMMs, divided into three sets of (i) Class 1, Mainly Alpha (212,345 HMMs spanning 1,592 CATH superfamilies); (ii) Class 2, Mainly Beta (254,974 HMMs spanning 1,077 CATH superfamilies); and (iii) Class 3, Alpha Beta (466,867 HMMs spanning 2,019 CATH superfamilies). Overall, TEDHL HMMs map to 50,167 CATH-PDB sequences from 4,688 CATH superfamilies, representing 83% of superfamilies in CATH v4.4. The 971 superfamilies that are not represented in TEDLH tend to be small helical domains with few representatives in CATH (Supplementary File 1).

2.3 Sequence and structure analyses

2.3.1 All-against-all comparisons of HMMs in TEDLH

The 934,186 TED-HMMs constituting the TEDLH library were searched against the Class 1, Class 2, and Class 3 libraries using hhblits from the HHsuite (Steinegger et al. 2019). All hits above threshold (E-value < 1 × 10−1, no minimal query or template coverage) were recorded and analysed.

2.3.2 Cluster map

Cluster map visualizations of relationships within TEDLH HMMs were done with CLANS (Frickey and Lupas 2004) using E-values from precomputed all-against-all scans.

2.3.3 Sequence and structure analysis

Pairwise comparisons at the sequence level (HHM-HHM) were performed with hhalign from the HHsuite (Steinegger et al. 2019), and at the structure level with TM-align (Zhang and Skolnick 2005). We used CATH-SSAP (Orengo and Taylor 1996) for superpositions of multiple structures and structure-derived MSAs. Topology diagrams were downloaded from ProteoVision (Penev et al. 2021).

3 Results

3.1 Contribution of primary and transitive links to TEDLH coverage of CATH superfamilies

The TEDLH library comprises HMMs derived from TED domain sequences that can be associated with CATH-PDB domains either directly or through intermediate HMMs. We first identified TEDLH HMMs exhibiting direct sequence similarity with CATH-PDB domains (primary links). TEDLH HMMs that display significant similarity to these primary HMMs were then collected as secondary links. Finally, we retrieved HMMs related to the set of secondary relationships (tertiary links). This series of transitive connections enable the detection of more remote homologs (Park et al. 1997, Salamov et al. 1999, Li et al. 2000). The HMM identifier indicates the level of the transitive relationship to a CATH-PDB domain and the corresponding CATH superfamily (Fig. 1b).

For each CATH superfamily, we assessed how well TEDLH models capture the sequence space associated with experimentally determined structures. Primary relationships are detected through direct searches of CATH-PDB domains against the TEDLH HMM library, whereas secondary and tertiary relations encompass more divergent homologues, thus, extending coverage into more remote regions of the sequence space. We quantified the proportion of TEDLH HMMs that are directly linked to a CATH-PDB domain (primary relationships) across CATH structural classes. In CATH Class 1, 58% of TEDLH HMMs are associated with a primary relationship to a CATH-PDB domain. In Class 2, primary relationships account for 33.4% of TEDLH HMMs. In Class 3, 42.9% of TEDLH HMMs are directly linked to a CATH-PDB domain. The remaining coverage is contributed by secondary and tertiary relationships. Secondary and tertiary links broaden the mapping of sequence space so that the diversity of domains in CATH can be explored by transitive relationships.

3.2 Cross-hits across CATH hierarchy levels

All-against-all HMM–HMM comparisons in TEDLH also recover sequence similarity across-superfamilies (cross-hits) (Fig. 2c). These cross-hits can occur between superfamilies within a different CATH Class, Architecture, Topology or Homology. At an E-value threshold of 1 × 10−3, the number of superfamilies that retrieve cross-hits varies depending on the minimum query coverage threshold. The distribution of cross-hits follows a bimodal distribution, with one peak in the 20%–50% query coverage range and a second above 70%. Between 20% and 50% query coverage thresholds, cross-hits between Classes, Architectures, and Topologies are more frequent, consistent with recurring local sequence similarities that are well-documented in the literature (Alva et al. 2010, Nepomnyachiy et al. 2014, Ferruz et al. 2020) and reflect shared evolutionary history (Alva et al. 2015, Kolodny et al. 2021, Alvarez-Carreño 2025). Cross-hits above 70% query coverage are more frequent between superfamilies and may indicate previously undetectable relationships, offering valuable guidance for the manual curation of CATH. The top 25 superfamilies with the most cross-hits are shown in Fig. 2d and Table 1. The number of cross-hits does not appear to be driven by superfamily size or sequence diversity, as no clear pattern is seen with the number of HMMs or the number of S95/S35 clusters in CATH (Table 1).

Figure 2.

For image description, please refer to the figure legend and surrounding text.

Open in a new tab

TED-HMM relationships to CATH-S100 and cross-hits across TEDHL superfamilies. (a) Proportions of TED-HMMs exhibiting primary, secondary, or tertiary connections to CATH-PDB domains, shown separately for class 1, class 2, and class 3. (b) Distribution of primary, secondary, and tertiary TED-HMM relationships to CATH-S100 domains for the 25 most represented superfamilies in TEDLH. (c) Histograms of TEDLH cross-hits across different levels of the CATH hierarchy at varying query coverage thresholds, with a fixed window of ≥ 70% template coverage. (d) The 25 CATH superfamilies with the highest number of TEDLH cross-hits at a 70% bi-directional coverage threshold.

Table 1.

Diversity of the top 25 CATH-PDB superfamilies with the most cross-hits at 70% coverage.

Superfamily (name) HMMs in TEDLH Domains in CATH S95 in CATH S35 in CATH
2.60.40.10 (Immunoglobulins) 40,105 31,905 4,547 873
1.10.10.10 (Winged helix-like DNA-binding domain superfamily/Winged helix DNA-binding domain) 36,848 3,444 768 524
2.60.40.1930 (Macroglobulin [MG2] domain) 4,542 181 19 13
3.30.70.100 5,640 723 168 101
3.40.50.2300 (Response regulator) 12,541 2,543 616 402
2.40.50.140 (Nucleic acid-binding proteins) 10,631 2,879 415 227
3.40.50.720 (NAD(P)-binding Rossmann-like Domain) 12,953 11,728 1,430 647
3.30.70.260 (ACT domain) 2,581 335 87 50
1.10.260.40 (lambda repressor-like DNA-binding domains) 6,099 674 112 77
1.25.40.10 (Tetratricopeptide repeat domain) 33,812 759 200 155
2.60.40.1220 2,242 64 15 10
2.60.40.1120 (Carboxypeptidase-like, regulatory domain) 7,242 46 15 10
2.60.40.1080 2,097 33 8 6
2.60.40.2550 858 2 1 1
3.10.20.90 (Phosphatidylinositol 3-kinase Catalytic Subunit; Chain A, domain 1) 4,831 1,882 310 171
3.10.20.30 (Beta-grasp domain) 1,528 520 108 48
2.60.40.650 539 43 6 4
3.30.70.330 (RRM domain) 4,216 1,527 351 193
3.30.70.1070 1,005 3 3 3
2.60.40.3880 165 40 2 2
3.30.450.20 (PAS domain) 14,256 841 157 106
2.60.120.260 (Galactose-binding domain-like) 9,297 1,427 236 144
2.30.30.140 1,788 641 130 77
1.10.10.60 (Homeodomain-like) 3,289 1,221 282 176
3.30.70.920 (Lrp/AsnC effector binding domain/regulation of amino acid metabolism [RAM] domain) 603 129 23 14
Open in a new tab

S95 and S35: representative sequences at a maximum of 95% and 95% sequence identity, respectively. RRM: RNA recognition motif.

The distribution of global cross-hits across superfamilies follows a power law, with most superfamilies having less than 50 cross-hits. The superfamily with the largest number of cross-hits to other superfamilies in TEDLH at a 70% bi-directional coverage threshold is 2.60.40.10, Immunoglobulins (Fig. 2d). HMMs of superfamily 2.60.40.10 cross hit 225 other superfamilies, of these, 158 cross-hits are between CATH Homology levels; 12 are between CATH Topologies levels; 29 are between CATH Architectures; and 26 between CATH Classes.

3.3 Homology relationships within superfamily 3.30.70.100

To illustrate the ability of TEDLH-HMMs to extend sequence coverage, we analysed 5,640 models within superfamily 3.30.70.100. These HMMs are linked, directly or via transitive connections, to 173 non-redundant CATH-PDB domains. A cluster map representation was generated from precomputed pairwise HMM–HMM comparisons and between CATH-PDB sequences and HMMs (Fig. 2a). By design, at a relaxed E-value threshold of 1 × 10−1, all TEDLH-HMMs are linked to at least one CATH-PDB representative (Fig. 3a).

Figure 3.

Six-panel figure illustrating hit relationships for the CATH superfamily 3.30.70.100, centred on domain 1opzA00. Panel a (top left): A large cluster plot showing several separate point clouds connected by faint grey lines, representing sequence similarity relationships. Points are coloured by category: dark blue (CATH domain), dark green (primary), light green (secondary), and grey (tertiary). A dashed box highlights one central cluster, which is magnified in three smaller panels to its right, each isolating one link type: primary links (62 HMMs), secondary links (268 HMMs), and tertiary links (165 HMMs). Panels b–d (top right): Three ensembles of superposed protein structures: teal (primary), light green (secondary), and grey (tertiary), illustrating how structurally similar the hits in each category are (primary hits overlay tightly; tertiary hits are more divergent). Bottom left: Four circular insets show individual protein structures for domain 1opzA00 and three example hits (1opzA00_P584884, 1opzA00_S670065, 1opzA00_T689290), connecting back to points in the dashed box in panel a. Below each pair, similarity scores are listed: for the primary link, E-value 5.3×10⁻⁵, TMscore 0.67155, identity 19.7%; for the secondary link, E-value 6.1×10⁻¹¹, TMscore 0.67744, identity 14.5%; for the tertiary link, E-value 1.0×10⁻⁷, TMscore 0.73496, identity 23.8%. Panel e: A scatter plot of TMscore (y-axis, 0.4–1.0) versus sequence identity percentage (x-axis, 0–100%), with points coloured by primary/secondary/tertiary category. A dashed purple vertical line marks the "Twilight Zone" around 30% identity. Marginal histograms above and to the right show the count distributions for identity and TMscore respectively. The plot shows that even at low sequence identity many hits still achieve high TMscores (structural similarity). Panel f (bottom): A multiple sequence alignment showing 1opzA00 aligned against three AlphaFold model sequences (AF-A0A7Y8JK28-F1-model_v4_TED02, AF-A0A6L5FDN9-F1-model_v4_TED02, AF-A0A6J4NZK7-F1-model_v4_TED02). Above the alignment, a secondary structure cartoon (arrows for beta strands, coils for alpha helices) marks the corresponding structural elements. Aligned residues are highlighted blue (beta strand regions) or yellow-green (helix regions), showing conserved structural regions despite sequence divergence.

Open in a new tab

Representation of primary secondary and tertiary relationships between HMMs and CATH S100 domains of superfamily 3.30.70.100. (a) Cluster map of HMM-HMM similarities for superfamily 3.30.70.100 showing connections at an E-value threshold of 1 × 10−1. Nodes representing domain 1opzA00 and three associated HMMs are highlighted: 1opzA00_P584884 (primary HMM), 1opzA00_S670065 (secondary HMM), and 1opzA00_T689290 (tertiary HMM). A representative structure for each HMM is shown. Insets: Same map, highlighting primary, secondary, and tertiary connections to CATH domain 1opzA00. (b) Structural superposition of CATH domain 1opzA00 with one representative from each of its 65 primary HMMs. (c) Structural superposition of CATH domain 1opzA00 with representatives from its 268 secondary HMMs. (d) Structural superposition of CATH domain 1opzA00 with representatives from its 165 tertiary HMMs. (e) Relationship between structural similarity (TM-score) and sequence identity for 1opzA00 compared with representatives of its primary, secondary, and tertiary HMMs. (f) Structure-informed multiple sequence alignment of domains 1opzA00, AF-A0A7Y8JK28-F1-model_v4_TED02, AF-A0A6L5FDN9-F1-model_v4_TED02, and AF-A0A6J4NZK7-F1-model_v4_TED02 with secondary structural elements.

We examined the set of TEDLH relationships associated with the CATH-PDB domain 1opzA00 (superfamily 3.30.70.100), a soluble domain of the P-type copper ATPase from Bacillus subtilis (Banci et al. 2003). In total, 62 primary, 268 secondary, and 165 tertiary relations were identified (Fig. 3a–d). A superposition of 1opzA00 to representatives of primary, secondary, and tertiary connections shows high structure similarity (TMscores > 0.6) despite low pairwise sequence identities (Sequence identity < 20%). Primary, secondary, and tertiary HMMs progressively sample more distant homologues, capturing sequences with lower pairwise identity and greater structural variation. Despite this divergence, structural superpositions across multiple representatives reveal a conserved core fold, consisting of a β-α-β-β-α-β arrangement that forms a two-layer α/β sandwich. The analysis of superfamily 3.30.70.100 illustrates the ability of TED-HMMs to detect remote homology beyond regions covered by experimentally determined structures.

3.4 Insights into protein fold divergence

3.4.1 Superfamilies sharing common ancestry

To investigate potential evolutionary relationships between distinct folds, we examined cross-superfamily HMM–HMM similarities. Comparisons between 3u7zA00_P233639 (CATH superfamily 2.170.130.30) and 2m19A00_P481943 (CATH superfamily 3.10.20.30) reveal significant sequence similarity (HHalign Probability 93.94%, E-value 1.5 × 10−7), suggesting a shared evolutionary origin (Fig. 4). Although domains in these superfamilies show detectable sequence similarity, their differences emerge at a higher level of structural organization.

Figure 4.

Two-panel figure illustrating a "chain of links" connecting two different CATH topologies (3.10.20.30 and 2.170.130.30) through a series of intermediate structural hits. Panel a (left, large network plot): A large network diagram, with clusters of dots connected by faint grey lines. Points are coloured by category: dark blue (CATH domain), dark green (primary connection), light green (secondary connection), and grey (tertiary connection), with a subset outlined in black marking members of topology 2.170.130.30. Four pink/magenta points, numbered 1–4, are highlighted along a path stretching from the 3.10.20.30 cluster toward the 2.170.130.30 cluster, representing a chain of intermediate structural hits linking the two topologies. To the right of the network, two topology cartoon diagrams represent the fold architectures at each end of the chain. Four boxed cartoon protein structures show the actual structures corresponding to points 1–4: Box 1: 2l52A00 Box 2: 2m19A00_P481943 Box: 3u7zA00_P233639 Box 4: 4zrpA02 Curved arrows connect box 1 to box 2 (TM-score = 0.75582), box 2 to box 3 (TM-score of 0.59300), and box 3 to box 4 (TM-score = 0.79450), illustrating how structural similarity gradually changes along the chain from one fold topology to the other. Panel b (bottom): Three stacked pairwise sequence alignment blocks (in HHsearch/HHblits output style), each comparing two of the four structures in sequence: - 2l52A00 vs. 2m19A00_P481943 (Probab=97.57, E-value=4.1e-12, Score=63.37, Aligned_cols=83, Identities=36%, Similarity=0.586) - 2m19A00_P481943 vs. 3u7zA00_P233639 (Probab=93.94, E-value=1.5e-07, Score=42.53, Aligned_cols=86, Identities=30%, Similarity=0.391) - 3u7zA00_P233639 vs. 4zrpA02 (Probab=64.91, E-value=6.1e-05, Score=29.80, Aligned_cols=91, Identities=26%, Similarity=0.370) Each alignment block shows the secondary structure elements (SSE track with blue arrows for strands and yellow-green coils for helices).

Open in a new tab

Distinct folds with a common ancestor. (a) Cluster map of HMM-HMM similarities between superfamilies 2.170.130.30 and 3.10.20.30 showing connections at an E-value threshold of 1 × 10−3. Structure of 2l52A00 (representative of superfamily 2.170.130.30) and 4zrpA02 (representative of superfamily 3.10.20.30). Three-dimensional structure and topology diagrams of domains 2l52A00 and 4zrpA02 coloured by secondary structural element. (b) HMM-HMM alignments.

Pairwise structural comparisons between 2l52A00 (representative of superfamily 3.10.20.30) and 4zrpA02 (representative of superfamily 2.170.130.30) show overall structural similarity (TM-score 0.59). The observed differences arise mainly from changes in the composition and arrangement of secondary-structure elements (SSEs). In domain 2l52A00, the SSEs are organised as ββ–[αα]–ββ, whereas in domain 4zrpA02 they adopt a ββ–[αββ]–ββ arrangement. In both cases, the N- and C-terminal β-hairpins form a comparable β-sheet that superimposes well, while the regions highlighted in brackets represent the variable structural segment that distinguishes the two domains (Fig. 4b and c). This pattern is consistent with a possible interconversion between disorder and secondary structural elements during evolution. CATH superfamily 2.170.130.30 includes several cobalamin-binding proteins, such as transcobalamin-1 (4kkiA02), transcobalamin-2 (2bb6A02, 4zrpA02, 5nsaA00), and intrinsic factor (2pmvA02). These domains share a characteristic β-hairpin, which is absent from domains in superfamily 3.10.20.30. The β-hairpin provides a surface involved in the coordination of cobalamin (Fig. 4a, inset 4).

The results of pairwise comparisons of HMMs in TEDLH suggests that these superfamilies, currently placed in different classes, could be considered for merging in CATH. The class distinction is based on the proportion of residues in helix versus strand conformations: superfamily 2.170.130.30 is assigned to class 2 (mainly β) due to a higher proportion of β-strand residues and a lower proportion of α-helices compared to superfamily 3.10.20.30.

4 Discussion

The identification of domains at the structure level in TED brings in a wealth of annotation for domains that would have been very difficult to identify by sequence alone (Lau et al. 2024). Domains in TED are identified based on three complementary domain-chopping methods [Merizo (Lau et al. 2023), Chainsaw (Wells et al. 2024) and UniDoc (Zhu et al. 2023)], providing precise and reliable domain boundaries, which we used to generate MSAs that retain structural information and serve as a robust source of evolutionary data. Using these annotations, we generated high-quality sequence HMMs, creating TEDLH, a library derived from MSAs of sequence clusters covering non-redundant TED domains assigned to CATH.

TEDLH currently contains high-confidence domains mapped to CATH and linked, directly or via transitive relationships, to a CATH-PDB structure. Each TEDLH HMM can be traced to its closest CATH-PDB representative, facilitating evolutionary and functional interpretation and downstream annotation. By including remote homologues through transitive links, TEDLH captures structural similarity (TM-score > 0.50) even when sequence identity is low (<20%). In contrast to purely structure-based comparisons and embeddings-based comparisons, which require additional evidence to infer evolutionary relationships, sequence matches against TEDLH are based on conserved amino-acid positions and insertion/deletion patterns, providing directly interpretable signals of shared ancestry.

The patterns of partial cross-hits (query coverage < 50%) observed for a large proportion of superfamilies represented in TEDLH is consistent with previous observations that protein folds often share localized regions of similarity (Friedberg and Godzik 2005, Alva et al. 2015, Ferruz et al. 2020, Kolodny et al. 2021). These local regions of similarity likely reflect remnants of shared evolutionary history between structurally distinct folds (Fetrow and Godzik 1998, Lupas et al. 2001, Alva et al. 2015, Kolodny et al. 2021). Global cross-hits (query and template coverage ≥ 70%) likely reflect remote homology relationships that become detectable with the expanded sequence coverage offered by TEDLH profiles.

TEDLH models are searchable with HHblits and HHsearch from HH-suite3 (Steinegger et al. 2019), allowing both single-sequence queries and HMM–HMM comparisons. TEDLH complements established HMM-based resources for sequence similarity detection at the domain level such as the Pfam protein family database (Paysan-Lafosse et al. 2025) [harmonized with the ECOD structural classification (Cheng et al. 2014)] and Gene3D (based on the CATH classification of protein structures). TEDLH extends access to AFDB-derived structural diversity at the sequence level and provides a valuable resource for investigating remote homology and informing the ongoing curation of domain classifications. TEDLH will be expanded in the future to incorporate additional superfamilies not yet classified in structural databases, further increasing its coverage and utility for detecting remote homology.

Supplementary Material

btag470_Supplementary_Data
btag470_supplementary_data.docx (793.9KB, docx)

Acknowledgement

The authors thank the CATH group for insightful discussions.

Contributor Information

Claudia Alvarez-Carreño, Department of Structural and Molecular Biology, University College London, London, WC1E 6BT, United Kingdom; NASA Center for the Origin of Life, Georgia Institute of Technology, Atlanta, GA 30332, United States.

Anton S Petrov, NASA Center for the Origin of Life, Georgia Institute of Technology, Atlanta, GA 30332, United States; School of Chemistry and Biochemistry, Georgia Institute of Technology, Atlanta, GA 30332, United States.

Vaishali P Waman, Department of Structural and Molecular Biology, University College London, London, WC1E 6BT, United Kingdom.

Ian Sillitoe, Department of Structural and Molecular Biology, University College London, London, WC1E 6BT, United Kingdom.

Christine Orengo, Department of Structural and Molecular Biology, University College London, London, WC1E 6BT, United Kingdom.

Author contributions

Claudia Alvarez-Carreño (Conceptualization [equal], Data Curation [equal], Formal Analysis [equal], Writing Original Draft [lead]), Anton S. Petrov (Supervision [equal], Writing Review Editing [equal]), Vaishali P. Waman (Writing Original Draft [equal], Writing Review Editing [equal]), Ian Sillitoe (Writing Review Editing [equal]), and Christine Orengo (Writing Original Draft [equal])

Supplementary material

Supplementary material is available at Bioinformatics online.

Conflict of interests

None declared.

Funding

CAC is funded by the Royal Society Newton International Fellowship [NIF\R1\231037]. ASP acknowledges funding from NASA Grant 80NSSC24K0344. VPW is funded by Wellcome Trust [221327/Z/20/Z]. IS is funded by BBSRC [BB/W018802/1].

Data availability

The data underlying this article are available in FigShare, at https://doi.org/10.6084/m9.figshare.28531754.

References

  1. Altschul SF, Madden TL, Schäffer AA  et al.  Gapped BLAST and PSI-BLAST: a new generation of protein database search programs. Nucleic Acids Res  1997;25:3389–402. [DOI] [PMC free article] [PubMed] [Google Scholar]
  2. Alva V, Remmert M, Biegert A  et al.  A galaxy of folds. Protein Sci  2010;19:124–30. [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Alva V., Söding J., Lupas A.N.  A vocabulary of ancient peptides at the origin of folded proteins. eLife  2015; 4: e09410. [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Alvarez-Carreño C.  Piecing together the history of protein folds from a fragmented evolutionary record. Genome Biol Evol  2025;17:evaf148. [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Banci L, Bertini I, Ciofi-Baffoni S  et al.  A core mutation affecting the folding properties of a soluble domain of the ATPase protein CopA from Bacillus subtilis. J Mol Biol  2003;331:473–84. [DOI] [PubMed] [Google Scholar]
  6. Berman HM, Westbrook J, Feng Z  et al.  The protein data bank. Nucleic Acids Res  2000;28:235–42. [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Buchan DWA, Shepherd AJ, Lee D  et al.  Gene3D: structural assignment for whole genes and genomes using the CATH domain structure database. Genome Res  2002;12:503–14. [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Cheng H, Schaeffer RD, Liao Y  et al.  ECOD: an evolutionary classification of protein domains. PLoS Comput Biol  2014;10:e1003926. [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. Ferruz N, Lobos F, Lemm D  et al.  Identification and analysis of natural building blocks for Evolution-Guided Fragment-Based protein design. J Mol Biol  2020;432:3898–914. [DOI] [PMC free article] [PubMed] [Google Scholar]
  10. Fetrow JS, Godzik A.  Function driven protein evolution. A possible proto-protein for the RNA-binding proteins. Pac Symp Biocomput  1998:485–96. [PubMed] [Google Scholar]
  11. Frickey T, Lupas A.  CLANS: a java application for visualizing protein families based on pairwise similarity. Bioinformatics  2004;20:3702–4. [DOI] [PubMed] [Google Scholar]
  12. Friedberg I, Godzik A.  Connecting the protein structure universe by using sparse recurring fragments. Structure  2005;13:1213–24. [DOI] [PubMed] [Google Scholar]
  13. Gonzalez MW, Pearson WR.  Homologous over-extension: a challenge for iterative similarity searches. Nucleic Acids Res  2010;38:2177–89. [DOI] [PMC free article] [PubMed] [Google Scholar]
  14. Gough J, Chothia C.  SUPERFAMILY: HMMs representing all proteins of known structure. SCOP sequence searches, alignments and genome assignments. Nucleic Acids Res  2002;30:268–72. [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Haft DH, Selengut JD, White O.  The TIGRFAMs database of protein families. Nucleic Acids Res  2003;31:371–3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Johnson LS, Eddy SR, Portugaly E.  Hidden Markov model speed heuristic and iterative HMM search procedure. BMC Bioinformatics  2010;11:431. [DOI] [PMC free article] [PubMed] [Google Scholar]
  17. Katoh K, Standley DM.  MAFFT multiple sequence alignment software version 7: improvements in performance and usability. Mol Biol Evol  2013;30:772–80. [DOI] [PMC free article] [PubMed] [Google Scholar]
  18. Kolodny R, Nepomnyachiy S, Tawfik DS  et al.  Bridging themes: short protein segments found in different architectures. Mol Biol Evol  2021;38:3030. [DOI] [PMC free article] [PubMed] [Google Scholar]
  19. Lau AM, Bordin N, Kandathil SM  et al.  Exploring structural diversity across the protein universe with the encyclopedia of domains. Science  2024;386:eadq4946. [DOI] [PMC free article] [PubMed] [Google Scholar]
  20. Lau AM, Kandathil SM, Jones DT.  Merizo: a rapid and accurate protein domain segmentation method using invariant point attention. Nat Commun  2023;14:8445. [DOI] [PMC free article] [PubMed] [Google Scholar]
  21. Li W, Pio F, Pawłowski K  et al.  Saturated BLAST: an automated multiple intermediate sequence search used to detect distant homology. Bioinformatics  2000;16:1105–10. [DOI] [PubMed] [Google Scholar]
  22. Lupas AN, Ponting CP, Russell RB.  On the evolution of protein folds: are similar motifs in different protein folds the result of convergence, insertion, or relics of an ancient peptide world?  J Struct Biol  2001;134:191–203. [DOI] [PubMed] [Google Scholar]
  23. Mirdita M, Steinegger M, Söding J.  MMseqs2 desktop and local web server app for fast, interactive sequence searches. Bioinformatics  2019;35:2856–8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Murzin AG, Brenner SE, Hubbard T  et al.  SCOP: a structural classification of proteins database for the investigation of sequences and structures. J Mol Biol  1995;247:536–40. [DOI] [PubMed] [Google Scholar]
  25. Nepomnyachiy S, Ben-Tal N, Kolodny R.  Global view of the protein universe. Proc Natl Acad Sci USA  2014;111:11691–6. [DOI] [PMC free article] [PubMed] [Google Scholar]
  26. Orengo CA, Michie AD, Jones S  et al.  CATH—a hierarchic classification of protein domain structures. Structure  1997;5:1093–108. [DOI] [PubMed] [Google Scholar]
  27. Orengo CA, Taylor WR.  SSAP: sequential structure alignment program for protein structure comparison. Methods Enzymol  1996;266:617–35. [DOI] [PubMed] [Google Scholar]
  28. Park J, Teichmann SA, Hubbard T  et al.  Intermediate sequences increase the detection of homology between sequences. J Mol Biol  1997;273:349–54. [DOI] [PubMed] [Google Scholar]
  29. Paysan-Lafosse T, Andreeva A, Blum M  et al.  The pfam protein families database: embracing AI/ML. Nucleic Acids Res  2025;53:D523–d534. [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Penev PI, McCann HM, Meade CD  et al.  ProteoVision: web server for advanced visualization of ribosomal proteins. Nucleic Acids Res  2021;49:W578–w588. [DOI] [PMC free article] [PubMed] [Google Scholar]
  31. Salamov AA, Suwa M, Orengo CA  et al.  Combining sensitive database searches with multiple intermediates to detect distant homologues. Protein Eng  1999;12:95–100. [DOI] [PubMed] [Google Scholar]
  32. Schultz J, Milpetz F, Bork P  et al.  SMART, a simple modular architecture research tool: identification of signaling domains. Proc Natl Acad Sci USA  1998;95:5857–64. [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Sonnhammer EL, Eddy SR, Durbin R.  Pfam: a comprehensive database of protein domain families based on seed alignments. Proteins  1997;28:405–20. [DOI] [PubMed] [Google Scholar]
  34. Steinegger M, Meier M, Mirdita M  et al.  HH-suite3 for fast remote homology detection and deep protein annotation. BMC Bioinformatics  2019;20:473. [DOI] [PMC free article] [PubMed] [Google Scholar]
  35. UniProt_Consortium. 2019. UniProt: a worldwide hub of protein knowledge. Release 2018/11/06. [DOI] [PMC free article] [PubMed]
  36. Varadi M, Anyango S, Deshpande M  et al.  AlphaFold protein structure database: massively expanding the structural coverage of protein-sequence space with high-accuracy models. Nucleic Acids Res  2022;50:D439–d444. [DOI] [PMC free article] [PubMed] [Google Scholar]
  37. Wells J, Hawkins-Hooker A, Bordin N  et al.  Chainsaw: protein domain segmentation with fully convolutional neural networks. Bioinformatics  2024;40:btae296. 10.1093/bioinformatics/btae296 [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Zhang Y, Skolnick J.  TM-align: a protein structure alignment algorithm based on the TM-score. Nucleic Acids Res  2005;33:2302–9. [DOI] [PMC free article] [PubMed] [Google Scholar]
  39. Zhu K, Su H, Peng Z  et al.  A unified approach to protein domain parsing with inter-residue distance matrix. Bioinformatics  2023;39:btad070. 10.1093/bioinformatics/btad070 [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

btag470_Supplementary_Data
btag470_supplementary_data.docx (793.9KB, docx)

Data Availability Statement

The data underlying this article are available in FigShare, at https://doi.org/10.6084/m9.figshare.28531754.

Articles from Bioinformatics are provided here courtesy of Oxford University Press

RESOURCES