Physicochemical, functional, and evolutionary characteristics of protein loop regions in human and Escherichia coli proteomes

Abstract

Protein loops play crucial roles in the formation of binding and enzyme active sites. However, general structural and biological characteristics of these loops remain unclear. In this study, we investigated loops from structural and evolutionary perspectives using the entire protein data bank (PDB), Homo sapiens, and Escherichia coli proteins. We found that loop sequences tended to be unique among species. However, loop properties exhibited high similarity or conservation. Class, Architecture, Topology, and Homologous superfamily (CATH) classification analysis, which clusters domains within protein chains into superfamilies indicating an evolutionary relationship, suggested that the terminal residues of most loops connected to the same superfamily. The functions of conserved loops were not consistently conserved. The amino acid composition profiles showed different preferences. Collectively, this study provides an overview of loops from structural, functional, and evolutionary perspectives and a vast natural loop repertoire for future investigations.

Significance

This investigation into protein loops, essential for molecular binding and enzymatic activity, enhances our understanding of their diverse roles across various organisms. By exploring extensive datasets, the study reveals distinctive properties and functions of these loops, offering insights into structures, evolutionary patterns, and potential functional capacities. These insights open avenues for novel applications, such as research into their implications in disease and therapeutic development. Additionally, the huge natural loop repertoire presented in this study serves as a valuable resource for future research.

Introduction

Protein loops play a critical role in protein function. In particular, loops on protein surfaces often participate in enzyme reactions and ligand/protein binding as they are exposed to solvents. Although the loop regions of homologous proteins may display higher variability in terms of sequence composition, they also exhibit conserved architectures to some extent [1]. Exchanging loop sequences for proteins can enable gain of function [2–4]. Notably, reconstruction of humanized antibodies (non-human antibodies modified to reduce the risk of inducing immune responses in humans) or creation of antibody-mimetic proteins (smaller proteins with antigen-binding properties) supports loop sequence functions [5–7].

Loops also contribute to the evolution of protein function. Loop regions generate variability through diverse sequence compositions, leading to new or altered functions [8]. Sequence changes in loop regions are common during the evolution of enzyme functions [9]. In fact, considering structural and sequence (dis)similarity of loop regions is a more efficient means of classifying proteins or characterizing protein evolution compared with conventional approaches (e.g., measuring conserved protein core regions) [10]. Accordingly, many studies have sought to elucidate the role of protein loop regions in protein function and evolution. Typically, these studies have focused on loop structural and physicochemical properties: loop sequence, number of residues in a loop, local loop structure, and distance and stretch (normalized distance) between the endpoints of a loop [2,6,11–16]. For instance, in contrast to the loop stretch, Choi et al. demonstrated that loop distance distribution is independent of the number of residues, while highly stretched loops are easier to predict than contracted loops [17].

Amino acid composition of protein loops has also been investigated from a functional characteristic perspective. For example, glycine is enriched in interface loop regions; phenylalanine is overabundant in hotspot loops, exhibiting a high average binding energy; thus, hot loops are commonly used to recognize protein targets [18]. Omega loops, often associated with regulatory functions and biomolecular recognition [19–21], are biased toward specific amino acids, such as glycine, proline, tyrosine, aspartate, serine, and asparagine [19,22]. Moreover, eukaryotic proteins have evolved myriad notable features compared with prokaryotic proteins, including longer sequences [23–26] and abundant multi-domain proteins [27–29]. Amino acid frequencies have been studied in-depth within intrinsically disordered regions, where the difference in disordered regions between eukaryotic and prokaryotic proteins is caused by a shift in serine, proline, and isoleucine frequencies [30]. However, to the best of our knowledge, the analysis of loop properties on a larger scale and comparative analysis between eukaryotes and prokaryotes from an evolutionary perspective has not been conducted. Moreover, the amino acid frequencies of large-scale loops from eukaryotic and prokaryotic proteins are less understood.

In this study, we aimed to describe and compare the properties of protein loops acquired from the entire protein data bank (PDB), Homo sapiens, and Escherichia coli and assess the loop structures and functions from an evolutionary perspective. By characterizing the three loop properties, i.e., length, distance, and stretch, we demonstrate similarities among human and E. coli datasets. In particular, similar loop distances and stretch distributions, including longer loops, were observed in human proteins compared with those in E. coli proteins. Loop functions identified based on Class, Architecture, Topology, and Homologous superfamily (CATH) and Pfam annotations show that protein loops are highly capable of gaining or losing unique functions in humans or E. coli. Preferences for specific amino acids vary between eukaryotes and prokaryotes. For instance, differences in serine, glycine, and alanine frequencies were observed between humans and E. coli.

Materials and methods

Loop definition

The ‘ss_dis.txt’ file distributed by PDB (December 2019), which utilized DSSP-based secondary structure assignments, was used to investigate loop structure. Specifically, a protein loop structure was defined as any region classified as “B,” “S,” “T,” or space in the secondary structure assignment. The following additional criteria were applied to curate the data: (1) number of residues of a loop ranging between 4–50; (2) terminal residues not located in disordered regions; (3) loop stretches no more than 1.2; (4) loop sequences without unknown residues (shown as “X” in the PDB FASTA file); and (5) protein chain termini skipped when extracting loop regions. For nuclear magnetic resonance (NMR) structures, the first model was used for further analysis. Additionally, to validate our findings with a newer dataset, we later repeated the analysis using the PDB data as of October 2024; these validation results and updated methods are presented in Supplementary Figure S11-14, Supplementary Table S9-10, and Supplementary Text S1 “Validation Analysis with Updated Dataset (2024)” in the supplementary material.

Dataset construction

Three-dimensional structures and secondary structure assignments with missing protein region information were downloaded from the PDB (December 2019). A non-redundant chain dataset (the redundancy among chains was removed with >90% sequence identity (S90)) served as the primary dataset, and, redundant (the entire PDB) and S100 (removing redundant sequences based on the 100% identity threshold) datasets were also assessed (Figure 1). Note that the ‘bc_(sequence identity).txt’ files (e.g., bc_100.txt), provided by the PDB (January 2020), were used to remove redundant sequences. Protein chains were further divided into different datasets according to the source organisms: the entire PDB, Homo sapiens, and E. coli. It is worth noting that protein chains with multiple source organisms were removed. For the non-redundant S90 datasets, 55,654, 12,212, and 2,334 protein chains from the entire PDB, Homo sapiens, and E. coli, respectively, were used separately for loop extraction. The number of protein chains was 417,925, 103,984, and 21,958 for the redundant datasets and 85,029, 20,283, and 4,143 for the non-redundant S100 datasets.

Figure 1.

Workflow of this study.

Loop properties

Three loop properties were used: loop length, distance, and stretch. Loop length is defined as the number of residues in the loop, loop distance (LD) as the distance between the first and last C alpha atoms of the loop, and loop stretch (LS) as the normalized loop distance, LS=LD/LD_max. The maximum distance (LD_max) is associated with the number of residues (n) and calculated using the formulas g(n/2–1)+d and g(n–1)/2 for even and odd n, respectively, where g=6.046 and d=3.46 Å [17,31,32]. Loop stretch indicates how stretched a loop is. Ideally, loop stretch for a fully stretched loop should be 1. However, because the values of g and d are theoretical approximations, loop stretch values can be >1 [31].

Structural and functional domain annotations for loops

Structural domains of loops were assigned using the CATH classification [33] (Figure 1). Specifically, combined with the Structure Integration with Function, Taxonomy, and Sequence (SIFTS) database [34], CATH v4.2 and CATH daily update file (latest classifications, 2020/01/13) were used to annotate superfamilies of both terminal residues of a loop.

From an evolutionary perspective, conserved loop sequences between humans and E. coli were identified. To ensure the accuracy of organism assignments, we performed manual curation of all long conserved loops (>=10 residues) in our dataset. This curation process involved verification of source organism annotations and protein sequence origins using UniProt cross-references. The results of this verification process are provided in Supplementary Table S2. These were then used to analyze the structures and functions using CATH classification and the Pfam database [35]. CATH classification was similar to that used in the above analysis. Regarding Pfam analysis, the corresponding human and E. coli loops were obtained according to the conserved loop sequences. The Pfam IDs of the terminal residues of a loop were extracted from the SIFTS files. Eventually, Pfam clans, which consist of groups of Pfam families organized by sequence, structural, or functional similarities, were subsequently identified and extracted based on the corresponding Pfam IDs.

Superfamily context repertoire

CATH protein structure classification database was applied to provide information on the evolutionary relationships of the protein domains; those sharing evolutionary similarities were assigned to the same CATH superfamily [33]. Multiple protein domains within a single protein chain can be clustered into either the same or different superfamilies. Considering the CATH superfamily, loops can be fully or partially located within or outside protein domains, thus allowing them to link a common or diverse superfamily. It is worth noting that only superfamilies of the terminal residues of a loop were considered. As loop sequences may be duplicated, each redundant loop sequence possesses a superfamily context repertoire composed of all superfamilies associated with different loops that share the same sequence, regardless of whether they are located in other regions of the same protein chain or in different molecules. Superfamily context repertoires can be homogeneous or heterogeneous, depending on whether the loops share the same superfamily context. For example, many loops, whether found in the same or different protein chains, share an identical sequence, referred to as sequence-No.1. If all these loops also share the same pair of connected superfamilies (a loop’s two ends), the superfamily context of sequence-No.1 is considered homogeneous.

Amino acid composition profiles

Amino acid composition of the protein loop sequences from different datasets, i.e., different organisms, sub-datasets of loops connecting common or diverse superfamily contexts, and conserved loops, were calculated. Clustered heat maps were generated using Ward’s variance minimization algorithm, and the distance was computed using Euclidean.

Identification of orthologous-like sequences for Archaeoglobus fulgidus dataset analysis

BLASTp [36] version 2.13.0 was used to identify orthologous-like sequences in the human and E. coli. datasets. Selected amino acid sequences of Archaeoglobus fulgidus were used as a query sequence and searched against PDB. To obtain the best matched sequence, an e-value threshold of 1e-2 and a minimum query coverage of 50% were applied.

Results

Protein loop extraction and statistics

We downloaded protein secondary and 3D structure data from the PDB in December 2019. The number of protein chains used in this study are listed in Table 1. Combined with the SIFTS database, we identified protein loops and calculated loop features, namely length, distance, stretch, amino acid sequence, and connected secondary structure. For a protein with multiple NMR structures, the first model was selected for further analyses. Multiple loops on a single protein molecule were identified and analyzed separately. Subsequently, we assigned structural and functional annotations to loops by detecting CATH superfamily of loop terminal residues. The number of loops, unique loop sequences, loops annotated by the CATH database, and percentages of non-redundant loop sequences are shown in Table 1. Some loops from the entire PDB, humans, and E. coli were not annotated by CATH after removing redundant sequences in the PDB with 90% sequence identity (S90), with the corresponding percentages of 28.98%, 31.37%, and 24.41%, respectively. In this study, a “loop” indicates a single loop region on a protein structure, while a “loop sequence” indicates the amino acid sequence of the loop regions. Therefore, two different loops in different protein structures can have the same loop amino acid sequence. Compared with the total number of loops, the number of unique loop sequences decreased by 24.21%, 21.93%, and 10.77% in the PDB, human, and E. coli datasets, respectively. Intriguingly, the non-redundant loop sequences were extremely high for the S90 datasets, up to 91.75% in E. coli and 85.13% in the human dataset. The higher percentage of non-redundant loop sequences in E. coli may be due to preference for specific loops in human protein or biases of data depositions in the PDB. Analysis of the 2024 dataset corroborated these statistical trends (Supplementary Table S9).

Table 1.

Protein loop statistics

	Redundancya	Protein chains	# of Loops	Annotated Loopsb	Unique Loop Seqc	Non-redundant Loop Seqd
Entire PDB	All	417,925	4,512,768	3,447,038	733,928	40.49%
	S100	85,029	905,728	678,737	487,130	73.42%
	S90	55,654	555,516	394,535	421,047	85.72%
Homo sapiens	All	103,984	1,027,593	828,506	157,799	43.16%
	S100	20,283	192,496	145,620	99,171	71.55%
	S90	12,212	102,901	70,626	80,338	85.13%
Escherichia coli	All	21,958	279,503	221,870	45,202	40.25%
	S100	4,143	49,331	40,027	27,052	71.59%
	S90	2,334	24,818	18,761	22,146	91.75%

^a Redundant or non-redundant sequence in the PDB. All: Redundant PDB data; S100 or S90 dataset: similar sequences were removed by 100% or 90% sequence identities.

^b Annotated Loops: superfamilies of loops were annotated by the CATH database.

^c Unique Loop Seq indicates that only one sequence remained among the duplications in the same chain or other molecules.

^d Non-redundant Loop Seq indicates that such sequence cannot be detected in the same chain or other molecules (i.e., sequence is unduplicated).

Comparisons of protein loop properties among the entire PDB, Homo sapiens, and Escherichia coli datasets

We first investigated three main loop structural properties: loop length, distance, and stretch (Figure 2). Loop length frequency distribution ranged from 4–50. Predictably, we found that the frequencies decreased with increasing loop length and that the percentage of short loops (four and five residues) was high in E. coli proteins compared to human proteins. The frequencies of loops on protein structures are shown in Figure 2A and the frequencies of unique loop sequences are shown in Figure 2B. A notable difference in the frequencies of 4 and 5 residue loops was observed between human and E. coli proteins, while the frequencies of other loop lengths were similar. Loop distance increased as the number of loop residues increased, while loop stretch exhibited an opposite trend (Figure 2C and 2D). We also observed loop distance and stretch distributions (Figure 2E and 2F). Two peaks in the loop distance distribution (5–6 Å and 8–10 Å) and one in the loop stretch distribution (approximately 0.6) were observed. Analysis of the 2024 dataset further reinforced these characteristic distribution patterns (Supplementary Figure S11).

Figure 2.

Comparison of protein loop characteristics among the entire PDB, Homo sapiens, and Escherichia coli. (A)–(B) Frequency distribution of protein loops with the residue number ranging between 4–50. The frequency was calculated according to individual loops and unique loop sequences. Distance (C) and stretch (D) of loops containing different residues. The Y axis in (C) and (D) represents average distance and stretch values in terms of length, respectively. Loop distance (E) and stretch (F) frequency distribution.

To investigate whether loop distance and stretch are related to loop length, we analyzed their distributions using loops containing 8, 10, 12, 14, and 16 residues from the entire PDB, human, and E. coli datasets (Supplementary Figure S1A–S1F). The distance distributions were not altered for each dataset when the number of residues increased, whereas the loop stretch distributions shifted toward shorter stretches as the number of residues increased. In addition, quantile-quantile plots (Q–Q plots), a method for evaluating the similarity between two probability distributions by comparing their quantiles, showed that the human and E. coli loop distance distributions were from the same probability distribution (Supplementary Figure S1G and S1H). Altogether, these observations suggest that loop distance may be independent of residue number, whereas loop stretch may not. Moreover, these properties are not unique to certain organisms, indicating that they are evolutionarily conserved.

To explain the two peaks in the loop distance distributions, we examined the linked secondary structure types and calculated the amino acid composition of the corresponding loops. A tendency for the linked secondary structure types was observed. That is, compared to the peak at approximately 8–10 Å, that at 5–6 Å exhibited a preference for linking two strands (strand-strand) compared to the other connection types (helix-helix, helix-strand, and strand-helix). In total, 71.90%, 77.80%, and 67.88% loops with distances of 5–6 Å from the entire PDB, human, and E. coli datasets were found to link strand-strand, respectively. Therefore, these shorter distance loops are used to connect antiparallel β strands. Regarding the amino acid composition of the loops from all datasets, the 5–6 Å loops preferred glycine, while the 8–10 Å peak contained more leucine residues. The relationships between amino acid composition and loop stretch were also investigated; the loops from the entire PDB, human, and E. coli datasets were divided into 12 stretch bins with a 0.1 difference for the neighboring bins (Supplementary Figure S2). For each amino acid, we further calculated the variance of the 12 stretch bins. Findings indicated that amino acids, such as cysteine, methionine, histidine, tryptophan, and arginine, contributed less to loop stretching than aspartate, glycine, and proline. See Supplementary Figure S2 for full details.

Structural and functional analysis based on CATH classification

Next, we analyzed the protein loop structures and functions based on the evolutionary relationships of the protein domains using the CATH database. The four main levels analyzed for in the CATH classification hierarchy are class (C), architecture (A), topology/fold (T), and the homologous superfamily (H). In addition, we detected the superfamilies of the terminal residues of a loop as the “context” of the loop. Therefore, the superfamily context of a loop was described as a combination of CATH superfamily annotations (or superfamily IDs) for the loop’s two ends. The annotations at the two ends can be the same, different, only one end may be annotated, or neither end may be annotated. The S90 datasets of the entire PDB, human, and E. coli were used. If the two superfamily IDs assigned to the loop’s two ends are A and B, then all extracted loops can be divided into four groups: A=B, A!=B, A or B, and null (Figure 3A). A=B indicates that loops connect to an identical superfamily, whereas A!=B represents loops connected to different superfamilies. In the A or B group, only one of the terminal residues is annotated with a superfamily, whereas the null group contains no terminal residues found in a superfamily. Most loops were successfully annotated by CATH as a superfamily for both terminal residues. To confirm whether loops linking common or diverse superfamilies were related to loop length, we assessed the loop length distributions of the A=B and A!=B groups using loops from the entire PDB, human, and E. coli S90 datasets (Figure 3B–D). In comparing loops among different organisms, the loops linking identical superfamilies showed similar distribution patterns of loop length, however, the distributions of loops linking diverse superfamilies did not, and differed from the A=B group. The robustness of these findings was validated using the new 2024 dataset, which showed consistent patterns across all analyses (Supplementary Figure S12).

Figure 3.

Superfamily context of loops using the CATH database. (A) CATH classification of loops. A and B represent the superfamilies of terminal residues. Four groups were generated based on a loop connecting common, diverse, or no superfamilies. (B)–(D) Loop length distributions for loops connecting common and diverse superfamilies in the entire PDB, Homo sapiens, and E. coli datasets.

Although the context analyses were performed for individual loops, different loops may share the same loop amino acid sequence. Therefore, to investigate the effect of amino acid sequences on the superfamily context of loops, we investigated and compared the superfamily context of each loop sequence from different loops. Since the same loop sequence may originate from different proteins, the superfamily context repertoire of a loop sequence can be homogeneous (a loop sequence from different loops shares a common superfamily context) or heterogeneous (multiple superfamily contexts were observed). Loops used in this analysis were filtered using the following criteria: (1) the loop shares its sequence with other loops, and (2) both terminal residues of the loop have the CATH annotation. A total of 41,668, 7,688, and 1, 251 loop sequences were obtained from the entire PDB, Homo sapiens, and E. coli, respectively. The filtered loop sequences were then assigned to either homogeneous or heterogeneous repertoires by gradually reducing the CATH levels from 4 to 1 (from top to bottom; the four levels were C, A, T, and H). The number of loop sequences divided into homogeneous or heterogeneous repertoires is shown in Table 2. Many heterogeneous repertoires were found, indicating that such loop sequences can be involved in multiple superfamily contexts. When we increased the CATH levels, the number of heterogeneous repertoires was predicted to decrease as fewer structural or functional constraints were considered. However, the number of heterogeneous repertoires did not decrease considerably with higher CATH levels, suggesting that the heterogeneity was not explained at the topology or homologous superfamily level of CATH classifications. These distinctive patterns of repertoire distribution and their implications for structural diversity were successfully reproduced in our analysis of the 2024 dataset (Supplementary Table S10). We also examined loop length distributions of the homogeneous and heterogeneous repertoire loop sequences (Supplementary Figure S3). Consequently, the number of loops with four residues was extremely high for heterogeneous repertoires. This might be due to the short length of loops containing four residues, making them highly likely to be observed in multiple domains or superfamilies.

Table 2.

Superfamily context repertoire of loops from S90 non-redundant datasets

	# of unique loop sequences	# of redundant loop sequences	CATH_levelsa	Homogeneous	Heterogeneous
Entire PDB	301,462	41,668	(C, A, T, H)	22,102	19,566
			(C, A, T)	22,860	18,808
			(C, A)	24,069	17,599
			(C)	29,496	12,172
Homo sapiens	55,277	7,688	(C, A, T, H)	6,455	1,233
			(C, A, T)	6,514	1,174
			(C, A)	6,608	1,080
			(C)	6,915	773
Escherichia coli	16,948	1,251	(C, A, T, H)	922	329
			(C, A, T)	961	290
			(C, A)	1,001	250
			(C)	1,125	126

^a This analysis was performed considering different CATH levels. Consider the superfamily ID 1.10.510.10 as an example. 1.10.510.10: (C, A, T, H); 1.10.510: (C, A, T); 1.10: (C, A); 1: (C).

In the previous analyses, we separately analyzed human and E. coli datasets and compared the results. Here, we detected the conserved loop sequences between humans and E. coli according to 100% sequence identity to investigate the structures and functions of protein loops from an evolutionary perspective. In total, 3,350, 2,037, and 1,633 conserved loop sequences were identified in the redundant (all) and non-redundant datasets (S100 and S90), respectively. The conserved loop sequences from the S90 datasets were further analyzed. Compared to the total loop sequences of human and E. coli datasets, 2.03% of human and 7.37% of E. coli loop sequences were conserved between the two species. In terms of individual loops, 3.27% of human and 10.54% of E. coli loops were related to conserved loop sequences. The results suggested that the conserved loop sequences were overrepresented in protein chains from both humans and E. coli but even more overrepresented in E. coli, as the percentage differences between unique loop sequences and individual loops of E. coli (3.17%) was much larger than that of humans (1.24%) (i.e., compared to humans, the conserved sequence tended to be duplicated in E. coli). The top 30 conserved loop sequences from humans and E. coli are listed in Supplementary Table S1 according to the ranking of the corresponding loop numbers of each sequence (the full list of conserved loops is provided as Supplementary Table S11). The abundance of each conserved loop sequence differed between humans and E. coli. While we found that a small fraction of long conserved loops originated from PDB misannotations (Supplementary Table S2), this did not affect the overall trends observed in our analysis.

After characterizing the conserved loops, we examined their functional annotations using CATH and Pfam databases. Length distributions of individual loops associated with the conserved loop sequences are shown in Supplementary Figure S4A. Regarding the CATH classifications, similar results were found in that many loops failed to be annotated by CATH, and most successfully annotated loops were found to connect an identical superfamily (Supplementary Figure S4B). However, the conserved loop may be the result of (1) evolutionary conservation following the divergence of a protein loop from a common ancestral sequence; (2) convergence on the same loop sequence in unrelated proteins; (3) recombination or convergence if the same sequence is present at different locations in the same superfamily [1,37–40]. Considering we found that most conserved loops were assigned to the A=B group, this may indicate common ancestry of the proteins as they originate from the same superfamily, while the loops could be either (1) or (3) of the above, suggesting a mix of homologous and convergent cases. Note that only a few conserved loops were detected in the A!=B group (Supplementary Figure S4B, 22 from humans and 26 from E. coli), which might clarify the case of obvious convergence. The CATH classification analysis on the A!=B group suggested that the conserved loop sequence tended to show a similar superfamily context in unrelated proteins within or between species (Supplementary Table S3). Supplementary Table S3 provides insights regarding whether the same structural/functional solution has been discovered by evolution from independent starting points or if these are merely coincidences.

In addition, the top ten conserved loops of the CATH superfamily and Pfam clan are displayed in Supplementary Figure S5, and their functional descriptions are listed in Supplementary Table S4. Comparing humans and E. coli, several fundamental functions were conserved. For the CATH database, the conserved functions were immunoglobulins, P-loop-containing nucleotide triphosphate hydrolases, nicotinamide adenine dinucleotide phosphate (NADP)-binding Rossmann-like domain, periplasmic binding protein-like II, and glutaredoxin. For the Pfam database, the conserved functions were P-loop containing nucleoside triphosphate hydrolase (P-loop_NTPase), periplasmic binding proteins (PBPs), NADP_Rossmann, beta _propeller, and thioredoxin. However, many functions were not conserved. These observations indicate that although the loop sequences were conserved, it was highly possible to gain or lose unique functions in humans and E. coli.

To better understand and compare the functions of conserved loop sequences, we acquired ligands/chemical components of related PDB entries and analyzed their ligand-binding sites in humans and E. coli. We found that not all conserved loops participated in ligand binding. Moreover, the functions of the conserved loop sequence can also be conserved; that is, for a given conserved loop sequence, the loop regions from human and E. coli PDB entries were observed to potentially bind not only to the same ligand but also to completely different ligands or only one of the loop regions was identified as a ligand. To visualize these three conditions, the conserved loop sequences, VAEH, ADNSG, and GPNMGG, were chosen as examples (Figure 4). Each loop sequence of the human proteins (top column) was identical to that of the corresponding E. coli protein (bottom). For VAEH, the phosphate ion (PO₄) interacted with residue E44 of 2Y1H (chain A) of humans, and pyridoxal-5'-phosphate (PLP) was found to bind to residue H119 of 4LW2 (chain A) of E. coli. For ADNSG, guanosine-5'-diphosphate (GDP) bound to residues A48 and S51 of 5G53 (chain C) in humans, but no ligand was observed in 6S0K (chain L) of E. coli. For GPNMGG, the same ligand, adenosine-5'-diphosphate (ADP), was identified, which binds to M1137 and G1139 of 2O8B (chain B) in humans, compared to M617 and G619 of 6I5F (chain A) in E. coli. Specifically, we assume that this loop (GPNMGG) is a P-loop, as it showed the same feature as that reported previously [41]; that is, the structural P-loop contains a four-residue backbone fragment (i.e., the GXXX pattern) that primarily connects a strand and helix. Additionally, these loops (2O8B and 6I5F) were classified into the 3.40.50.300 CATH superfamily, which is described as a P-loop containing nucleotide triphosphate hydrolases.

Figure 4.

Examples of conserved loop sequences between humans and E. coli. Visualization of the ligand-binding sites of the proteins from (A) human and (B) E. coli were generated using PyMol. Three conserved loop sequences are shown. The top (human) sequence was the same as that from the corresponding bottom (E. coli). Protein loop regions are marked in yellow, and the remaining regions are represented in dark green. The binding sites are shown as sticks, and each residue is labeled with one letter. Ligands are shown as balls.

Amino acid composition and comparison among various groups with different resource organisms

To examine the contribution of amino acids to individual loops and their properties, we calculated and compared amino acid composition profiles among different species, superfamily combinations, and repertoire groups. The loops in the PDB, human, and E. coli datasets were divided into six categories: all loops, all unique loop sequences, A=B, A!=B, homogeneous, and heterogeneous groups. In addition, the conserved unique loop sequences and their related loops within 3D structures of proteins from humans and E. coli were analyzed. As shown in Figure 5, a clustered heatmap was used to visualize the similarities and differences in amino acid composition profiles in different groups (the amino acid composition ratios for each group are shown in Supplementary Table S5). As a result, various loop categories from different organisms were separated, suggesting that the amino acid differences among organisms were larger than those within an organism. Interestingly, A!=B, heterogeneous and conserved loops exhibited opposite trends, as each category among organisms was closely related. Specifically, because the loop number of the A!=B category was low (Figure 3A), we assumed that only a few loops could connect different CATH superfamily members, and these loops may tend to be conserved between humans and E. coli. Likewise, the number of loops with a heterogeneous superfamily context repertoire was small (Table 2), and they may also be conserved. Collectively, we hypothesized that few loops that showed more “complex” functions (e.g., connecting various superfamilies or involved in heterogeneous superfamily context) might tend to be conserved. These characteristic patterns of amino acid compositions were consistently observed in the 2024 dataset (Supplementary Figure S13).

Figure 5.

Clustered heatmap of amino acid composition profiles of loops from various groups with different resource organisms. Protein loops from three resource organisms are included: Entire_PDB, humans, and E. coli. The Loop (all) and Unique loop sequence categories represent loop sequences acquired based on individual loops and unique loop sequences, respectively. The A=B and A!=B categories represent loops linking common and diverse superfamilies. Homogeneous and heterogeneous categories were the heterogeneity of the superfamily context of loops.

To elucidate factors affecting the unique amino acid composition distributions of the different groups, we compared each amino acid composition profile among groups (among categories and resource organisms, Figure 6). Each amino acid composition that differed by >1.5% was considered to be the main contributor to the differences [30]. In terms of categories (Figure 6A and 6B), the amino acid composition distributions of all loops from humans and E. coli were highly similar to those of all and unique sequences. Compared to similar distributions between A=B and A!=B in humans, the A=B category in E. coli preferred glycine, while the A!=B category preferred leucine and proline. For both humans and E. coli, the homogeneous category showed preference for proline over glycine in the heterogeneous category. Notably, the conserved loops from both humans and E. coli were biased toward glycine and aspartic acid. In terms of resource organisms (Figure 6C), we observed that loops (all loops and unique loop sequences) in humans displayed preferences for serine compared to that of glycine and alanine in E. coli. This may be due to serine serving as a target of serine/threonine kinases prevalent in eukaryotes [42]. One potential reason for the abundance of glycine and alanine in E. coli is that the cost of producing alanine and glycine (also serine) is the lowest [43]. We further examined the frequencies of glycine, alanine, and serine in humans and E. coli (Supplementary Figure S6) and found that glycine and alanine decreased in all datasets in humans compared to E. coli and that serine frequency exhibited an opposite trend. Moreover, the frequency differences in glycine, alanine, and serine were widespread and not unique to particular datasets. Analysis of the 2024 dataset reaffirmed these characteristic amino acid preferences and their organism-specific distribution patterns (Supplementary Figure S14).

Figure 6.

Amino acid frequency differences among groups and source organisms. Comparisons among groups from (A) Homo sapiens and (B) E. coli. Positive values represent amino acids showing a higher frequency in the former categories and negative values represent a higher frequency in the latter categories. E.g., for loop (all) vs. UniqueSeq (all), frequency difference=loop (all)–UniqueSeq (all). (C) Comparison of amino acid composition between humans and E. coli. The E. coli frequencies are used as a baseline; that is, positive values indicate that the amino acids are more abundant in humans. Domains: Domain regions of protein chains that form a loop. Chains: entire protein chains with a loop. Compared to Chains, Chain_remove_loop indicates that the loop regions were removed, and only the remaining segments of protein chains were used.

Discussion

Loop properties of various membranes and soluble proteins have been studied previously. In particular, Choi et al. [17] reported that loop distance distribution patterns appear to be independent of loop length, whereas loop stretch is not. We observed similar results in a large-scale loop dataset comprising different organisms, suggesting that these findings were not unique to a specific species and that such loops tend to be conserved during evolution. Additionally, compared to Choi’s studies, we found two similar peaks (5–6 Å and 8–10 Å) in the loop distributions and determined that the 5–6 Å distances strongly preferred antiparallel beta loops. In this study, we further analyzed the amino acid composition to elucidate the underlying mechanisms of such preferences. Loops with distances of 5–6 Å from the PDB, human, and E. coli datasets exhibited preference for glycine. One possible reason for this is that glycine is the smallest amino acid and might contribute to fitness of short distances. In particular, we observed that loops with distances of 8–10 Å did not show any preference for the linked secondary structure types but had a bias toward leucine. However, reasons for these preferences remain unknown. Furthermore, the distances of long loops containing 31–50 residues were considerably larger in the human dataset than in the E. coli dataset (Mann–Whitney U-test, p=0.0429). Their corresponding stretches were also larger in humans, but not significantly (Mann–Whitney U-test, p=0.1366). These observations indicate that long loops from E. coli (>30 residues) might tend to contract.

In general, loop stretch is assumed to be <1. However, as certain parameters used for calculating the loop stretch are empirical values, loop stretches can be >1. Therefore, we also examined loops where stretches were >1.2. Limited loops with a stretch of more than 1.2 were detected; specifically 20 loops from humans and one from E. coli (Supplementary Table S6). Therefore, discarding loops with stretches >1.2 would not affect the performance of our analyses. Meanwhile, given the limited number of loops, estimating the contributions of amino acids to the loops with a larger stretch was difficult (Supplementary Table S7). Therefore, we hypothesized that the higher stretches were due to improper 3D structures deposited in the PDB.

Loop sequence compositions exhibit high variability in homologous proteins with conserved architectures [1]. In fact, loop sequences often change during enzyme evolution [9]. In this study, we found that more than 85% of the loop sequences from the entire PDB, human, and E. coli were observed only once, indicating that most of the loops were unique at a protein chain level. Moreover, several loop sequences were detected multiple times and were involved in homogeneous or heterogeneous superfamily context repertoires. These loop sequences may contribute to the variability in the sequence compositions of proteins and chains.

Generally, our observations suggest that certain loop features are robust during the evolutionary process. However, eukaryotes and prokaryotes have their own loop biases. For example, abundance of conserved loop sequences differs between humans and E. coli. The top sequence for humans is QPGGS, which was detected in 102 loops compared to NKHMNADTD in humans, which was found in 38 loops (Supplementary Table S1). In addition, the VAEH and ADNSG loops formed different three-dimensional conformations and bound (or do not bind) different ligands in human and E. coli proteins (Figure 4). Combined with previous studies of the conformational ensemble of loops and the associated dynamics [44,45], our examples provide important insights regarding the mechanisms of conformational loop selection, which might inform the prediction of accurate positions for various loop fragments [46].

Intriguingly, we found that protein loops exhibited preferences for specific amino acids, such as glycine, proline, aspartic acid, and leucine. The abundances were widespread among various loop sets (i.e., different categories in this study) and organisms, excluding serine and alanine, which were particularly enriched in humans and E. coli, respectively (Supplementary Table S5). Although human and E. coli protein loops similarly favored certain amino acids, their preferences for loop sequences were extremely divergent (Supplementary Table S1). In addition, many loops from categories or organisms have different biases toward particular amino acids. We assumed that this might be partially associated with post-translational modifications of loops. For example, a potential reason for serine enrichment in humans is that serine is a phosphorylation target, whereas between 1/3 and 2/3 of the eukaryotic proteome are phosphorylated [47,48]. Conserved loops showed preference for glycine and aspartic acid (both are relatively small), suggesting that these two amino acids may be essential for the flexibility and solvent exposure of loops.

In addition, to gain a comprehensive understanding of properties between eukaryotic and prokaryotic proteomes, we characterized the loop properties in Saccharomyces cerevisiae and Archaeoglobus fulgidus as additional species for comparison. Compared to entire PDB, human, and E. coli datasets, similar findings were observed (Supplementary Figure S7 and S8). Amino acid compositions of yeast loops exhibit higher similarity with human sequences compared to E. coli (Supplementary Figure S9), as expected. Additionally, archaea displayed unique features, including preference for short loops and particularly long distance loops containing 25 or 28 residues (Supplementary Figure S8). We speculated that these observations might be due to species-specific features or the small dataset size (only 151 protein chains in archaea; Supplementary Figure S8). To explore this, we ran BLAST with the 151 sequences from archaea to identify orthologous-like sequences in human and E. coli and characterize the loop properties. We found 56 and 42 matched chains in human and E. coli, respectively, and the corresponding loops displayed similarity and dissimilarity to human and E. coli, including similar preference for short loops and different amino acid compositions (Supplementary Figure S10). Unfortunately, due to the small dataset size, the cause of dissimilarity observed in archaea is difficult to explain.

Furthermore, we examined missing regions in a loop to investigate the effect of intrinsically disordered regions (IDRs). We calculated the percentages of loops that include at least one disorder site or fully disordered loops and found that the relative proportions were small among the species, suggesting that the IDRs might not have a considerable effect on our results (Supplementary Table S8).

Moreover, recent advances in protein structure prediction, particularly the development of AlphaFold, offer exciting new avenues for extending this analysis. While this study focused on experimentally determined structures to ensure high confidence in our findings, integrating AlphaFold models could offer several benefits. However, careful validation is essential when using predicted structures, especially for loop regions, as they tend to be more flexible and may be predicted with lower accuracy than core structural elements. Future studies that combine experimental structures with carefully validated predicted models could provide even more comprehensive insights into the properties and evolution of loop regions.

Conclusion

In conclusion, this study provides an in-depth understanding of loop structural properties, functions, and evolution on a large scale. It also provides a vast natural loop repertoire for further investigations, such as the mechanisms for controlling functions and protein folding. We performed analyses using all PDB entries, human, and E. coli proteins, which is considered sufficient to describe loop properties and functions, as well as basic structural analyses. Additionally, to assess the robustness of our findings against potential database biases, we performed a validation analysis using the newer PDB database (October 2024), which contains approximately 1.4-fold more entries than our initial dataset from 2019. This validation analysis demonstrated that our key findings regarding loop properties, including loop length and stretch distributions and amino acid composition between different species, remain consistent even with the substantially larger dataset. This consistency across different temporal snapshots of the PDB suggests that our observations reflect genuine physicochemical and biological properties. However, since this study is dependent on database information, certain biases may have occurred due to the data deposition. For example, if the depth and width of protein analysis differ considerably between humans and E. coli, limitations might arise when assessing the differences between humans and E. coli. To improve the performance, we removed redundant protein sequences; however, certain biases may be difficult to remove. Therefore, a more complete analysis of loop structures and their conformational changes is needed. These analyses could contribute to the discovery of further characteristics of naturally emerging and designed proteins and finer predictions of robust protein structures.

Conflict of interest

The authors declare no conflicts of interest.

Author contributions

H.N. designed the research. L.Z. performed the experiments. L.Z. and H.N. analyzed the data. L.Z. and H.N. prepared the manuscript.

Data availability

Any additional information required to reanalyze the data reported in this paper is available upon reasonable request.

A preliminary version of this work was deposited in bioRxiv (https://doi.org/10.1101/2023.06.09.544332) on June 11, 2023).

Acknowledgments

The authors would like to extend gratitude to Prof. Kengo Kinoshita and Prof. Matsuyuki Shirota for valuable discussions. This study was financially supported by the Development of Key Technologies for the Next-Generation Artificial Intelligence/Robots of the New Energy and Industrial Technology Development Organization (NEDO), Research Support Project for Life Science and Drug Discovery (Basis for Supporting Innovative Drug Discovery and Life Science Research (BINDS)) from AMED under Grant Number JP24ama121019 and JSPS KAKENHI Grant Numbers JP18K11520, JP22K12240. During the first revision, OpenAI o1-preview model was used for code generation and Anthropic Claude 3.5 Sonnet (New) model was used for English composition. During the second and third revisions, Anthropic Claude Opus 4, Sonnet 4, and Sonnet 4.5 models were used to compose the response to reviewers.