Intrinsically disordered domains: Sequence ➔ disorder ➔ function relationships

Abstract

Disordered domains are long regions of intrinsic disorder that ideally have conserved sequences, conserved disorder, and conserved functions. These domains were first noticed in protein–protein interactions that are distinct from the interactions between two structured domains and the interactions between structured domains and linear motifs or molecular recognition features (MoRFs). So far, disordered domains have not been systematically characterized. Here, we present a bioinformatics investigation of the sequence–disorder–function relationships for a set of probable disordered domains (PDDs) identified from the Pfam database. All the Pfam seed proteins from those domains with at least one PDD sequence were collected. Most often, if a set contains one PDD sequence, then all members of the set are PDDs or nearly so. However, many seed sets have sequence collections that exhibit diverse proportions of predicted disorder and structure, thus giving the completely unexpected result that conserved sequences can vary substantially in predicted disorder and structure. In addition to the induction of structure by binding to protein partners, disordered domains are also induced to form structure by disulfide bond formation, by ion binding, and by complex formation with RNA or DNA. The two new findings, (a) that conserved sequences can vary substantially in their predicted disorder content and (b) that homologues from a single domain can evolve from structure to disorder (or vice versa), enrich our understanding of the sequence ➔ disorder ensemble ➔ function paradigm.

Short abstract

PDB Code(s): 1R05, 1NKP, 3VXV, 2MOE, 1QK9, 6CCG

Abbreviations

CATH

class, architecture, topology, and homology

CDF

cumulative distribution function

CH

charge–hydropathy

IDPs or IDRs

intrinsically disordered proteins or regions

kip

kinase inhibitor protein

NMA–NIA

normalized monomeric area and normalized interface area

NMR

nuclear magnetic resonance

PDB

Protein Data Bank

PDD

probable disordered domain

SCOP

Structural Classification of Proteins

VSL2b

disorder predictor trained on V = variously characterized proteins with S = short and/or L = long IDRs, version 2b

1. INTRODUCTION

Molecular biologists and geneticists use domain for a contiguous set of amino acids having a particular function, for example, the autoinhibitory domain1, 2 and the transactivation domain.3, 4 Such domains are often located within intrinsically disordered protein regions and undergo disorder‐to‐structure transitions upon binding to their partners.4, 5 In contrast, structural biologists initially used domain to describe a protein unit that folds autonomously6 and has functional autonomy and/or evolutionary conservation.7, 8

Domain databases based on evolutionary conservation include Pfam,9, 10 SMART,11 and the Conserved Domain Database (CDD).12 Domain databases based on structural autonomy include the Structural Classification of Proteins (SCOP)13 and Class, Architecture, Topology, and Homology (CATH).14 Finally, a database based on both evolutionary conservation and structural autonomy is SUPERFAMILY.15

In contrast to structured proteins, intrinsically disorder proteins and regions (IDPs and IDRs) lack stable structures in solution, existing instead as highly dynamic ensembles with thousands of conformations.16, 17, 18, 19, 20 Many IDPs and IDRs undergo disorder‐to‐structure transitions upon binding partners.16, 17, 18, 19, 20, 21 These disorder‐to‐structure changes upon binding are often incomplete with flanking or looping IDRs22 that sometimes contribute positively or negatively to the binding constant. Such complexes are called fuzzy.23, 24 Some fuzzy complexes remain entirely disordered yet bind with high affinity.25 Structure‐to‐disorder transitions upon binding have also been observed for some proteins,26, 27, 28 with at least one protein showing simultaneous structural changes in both directions for different regions.26

IDPs and IDRs are abundant in all domains of life.29, 30, 31 IDPs, IDRs, and their various interactions with partners of all types32 are critically involved in many biological processes, such as molecular recognition, signaling, regulation, and cell cycle control20, 33, 34, 35, 36, 37 among many others.38, 39

In one study, the large majority of IDRs exhibited significantly less sequence conservation than the structured regions of the same proteins.40 Follow‐up studies have shown that conserved‐sequence IDRs are also common.41, 42 Conserved‐sequence IDRs were more recently called constrained disorder 42 and are correlated with tissue‐specific alternative splicing and cell regulation.43 One IDR is conserved in length and dynamic behavior but has negligible sequence conservation;44 such segments are called conserved (or flexible 42) disorder. Some IDR sequences are not conserved and change from being predicted‐to‐be disordered to being predicted‐to‐be structured; these regions are called non‐conserved.42 Finally, insertions and deletions (IDELs) are more often disordered than structured,45, 46, 47 and some IDR deletions are observed to occur in paralogues.48 In summary, from an evolutionary point of view, there are four types of IDRs: (a) conserved‐sequence (or constrained) disorder; (b) variable‐sequence (or flexible) disorder; (c) non‐conserved disorder, and (d) INDELs.

IDPs and IDRs carry out molecular recognition,49 and disorder predictors have been used to identify specific IDP and IDR loci that bind to globular protein partners.50 IDR‐located binding sites have been identified by disorder prediction,34, 51, 52 and, alternatively, by linear sequence motifs.53, 54, 55 Globular protein partner binding by IDPs and IDRs is common across the three domains of life.56 These prediction‐based studies have focused on short segments (5–15 residues in length) located within longer IDRs or even within IDPs.

Much longer IDRs (≥20–30 residues in length) bound to globular protein partners, and the previous work on conserved‐sequence disorder41 shows that such regions have conserved functions, conserved sequences, and conserved disorder, and therefore, by analogy to structured domains, such IDPs or IDRs were called disordered domains.57 An especially interesting disordered domain example is found in p27^kip1, p21^{Waf1/Cip1/Sdi}, and p57^kip2.58, 59, 60 All three of these proteins have domains that are entirely disordered by both prediction and experiment and play similar key regulatory roles in controlling the cell cycle.

Like p27^kip1 and its homologues, many Pfam domains contain sequences with 100% predicted intrinsic disorder.57, 61, 62 It is noteworthy that the Pfam database recently added “disordered” as a specific entry type,10 suggesting that the definition of protein domains should now also include those that are intrinsically disordered.

In the current work, we analyzed the sequences, the induced structures, and the functions for a set of Pfam disordered domains to explore their sequence–disorder–function relationships. The results presented herein highlight interesting distinctions between disordered and structured domains and improve our understanding of protein sequence–disorder–function relationships.

2. RESULTS

2.1. Analysis of probable disordered domains

The workflow for analysis of probable disordered domains (PDDs) is shown in Figure 1. As indicated in Section 4, 206 PDDs were obtained based on the set of PDD seed sequences from a previous study.61 First, as indicated in Figure 1 on the left, the homologues to the PDD sequences were collected, giving a total of 19,577 sequences. These homologues were collected from the seed proteins for each Pfam domain having at least one PDD sequence. These homologues in the seed alignment were used to correlate sequence conservation with predicted disorder and to investigate the sequence–disorder relationships. Not used here are the many additional homologous sequences matching the full alignment that could have been found by application of the hidden Markov model predictors developed for each Pfam domain; if such domains were included, the number of PDDs could be greatly increased. Next, as indicated in Figure 1 on the right, 2,548 different structures from 110 domains were collected for members having available structures in PDB. Their likely mechanisms of disorder‐to‐structure transition were examined to further investigate disorder–function relationships.

Figure 1.

Workflow for analysis of probable disordered domains in Pfam database

2.2. Sequence–disorder relationships: Unexpected wide range of predicted disorder/structure among homologous sequences

Disorder prediction was applied to all the collected seed members for the PDDs. It was expected that the additional homologous sequences of the PDD seeds would also show highly predicted disorder. Such a result was indeed commonly observed (Figure 2a), with 78 (38%) of the Pfam domains having predicted disorder (defined by an average vsl2b score >0.6 for all of the associated seed proteins, and with an additional 69 (33%) of the Pfam, having predicted disorder PDD for more than half of the seed proteins and with substantial predicted disorder in most of the remaining seed proteins (see Section 4 and Figure S1 for more details). Thus, a total of 147 (71%) Pfam domains were found to be all or mostly predicted‐to‐be‐disordered as expected. Totally unexpected was that finding that four Pfam domains (2%) were found to contain equal numbers of predicted to be disordered and predicted to be structured members and that 55 Pfam domains (27%) were found to contain mostly predicted‐to‐be‐structured members with only a few predicted disorder members (Figure 2a).

Figure 2.

Significant diversity of predicted disorder among the homolog sequences within PDDs. (a) The pie chart showing the numbers and percentage of domains having different compositions of seed members. In each group, examples were presented as the Pfam phylogenetic trees mapped with average VSL2b disordered score into each sequence. (b) A set of 56 PDDs having both 100% predicted‐to‐be‐disordered and 100% predicted‐to‐be‐structured members. The distribution of predicted disorder percentage for each domain is shown as boxplots. (c) Distribution of predicted disorder percentage of all the 19,577 seeds. (d) Distribution of average predicted disorder percentage for 206 domains. PDD, probable disordered domain; VSL2B, disorder predictor trained on V = variously characterized proteins with S = short and/or L = long IDRs, version 2b

Thus, four groups having distinct seed member compositions were observed as shown in the pie chart of Figure 2a. In each group, one domain example is presented by mapping the average VSL2b score into Pfam phylogenetic trees, where red and blue indicate predicted‐to‐be‐disordered and predicted‐to‐be‐structured seeds, respectively. In the latter three groups, there were 56 domains that contained both fully predicted‐to‐be‐disordered (100% residues with VSL2b score >0.5) and fully predicted‐to‐be‐structured seeds (100% residues with VSL2b scores <0.5, Figure 2b).

Overall, the disorder prediction showed significant diversity, as shown in Figure 2c for all the seeds (predicted disorder percentage ranged from 0 to 100%) and in Figure 2d across the domains (average predicted disorder percentage ranged from 13 to 100%), respectively. Among the 19,577 seeds, only 30% of them (5,833) were PDD (highlighted in the red bar in Figure 2c), and 14% (2,755) of them were predicted to be fully structured (0% residues predicted to be disordered, highlighted in the blue bar in Figure 2c). Also, only 74 of the 206 domains (36%) were predicted to be >95% disordered (highlighted in red bar in Figure 2d). The wide range of predicted disorder or structure for the homologous sequences within a domain is completely unexpected. This observation was validated on a larger set of representative proteomes reduced to a sequence identity of 55%. The disorder contents in both sets showed close agreement across all domains studied (Figure S2), which suggests the seed sequence set does not contain a biased order–disorder content. These striking differences of predicted disorder or structure suggested that, for these domains, their sequences are likely more conserved than their structures. Thus, the widely held view that structure is more strongly conserved than sequence63, 64, 65, 66, 67 is clearly not followed by all IDP or IDR domains, which instead often show the unexpected contrary behavior that sequence is more strongly conserved than structure.

Further investigation of sequence conservation indeed shows that conserved positions have higher disorder prediction scores on average than non‐conserved positions in PDDs (Figure S3a). The structured domain set showed similar pattern (Figure S3b). Therefore, many PDDs contain conserved sequences and conserved disorder, which is similar to structured domains having conserved sequences and conserved structure.

The evolutionary origins of these structural differences were investigated by calculating the disorder content of the last common ancestor of each domain (Figure 3). In many cases, the estimated disorder content of the domain ancestor is consistent with overall disorder content of domain examples, which indicates that different structural differences may be an ancient feature of these domains. However, in a few cases, disorder is significantly different in the domain ancestor than in the domain examples, either higher in three cases (Figure 3: PF00560, PF01821, PF00096) or lower in one case (Figure 3: PF00641). This seems to indicate a structural shift from disorder to order, or order to disorder, in these domains across evolutionary history.

Figure 3.

Estimated disorder content of the last common ancestor of each of the 56 PDDs having both 100% predicted‐to‐be‐disordered and 100% predicted‐to‐be‐structured members. The estimated percentage disordered residues of the ancestor (circles) and the 95% confidence interval of the estimate (error bars) are plotted along with the median percentage disordered residues (X) for comparison. PDD, probable disordered domain

2.3. Unusual disorder–function relationships: Disorder‐to‐structure transition depends upon various mechanisms

Given the wide range of disorder and structure within a domain, we first examined the distribution of disorder prediction for the identified PDB structures. This was done to exclude the possibility that domains could form structures simply because they are predicted‐to‐be structured members. Histograms in Figure 4a confirmed that majority of identified structures were predicted to be disordered on their sequences, with only a small set (109, 11.7%) predicted to be fully structured. An obvious question is why do predicted‐to‐be‐disordered proteins have structures in PDB?

Figure 4.

Identified structures were mostly predicted to be disordered on their sequences. (a) Prediction of disorder on matched structures. (b) Distribution of average predicted disorder for the four groups. MBD: macromolecular‐binding domains. Ion: ion‐binding. SS: disulfide bonds. Non‐consistent: domains without consistent mechanism of structural formation

Regarding the existence of structures for the predicted‐to‐be‐disordered proteins, three mechanisms were found that could potentially bring about disorder‐to‐structure transitions: (a) disulfide bond formation; (b) ion binding; and (c) macromolecular binding. In addition, there is a fourth group for which the stabilization mechanisms are uncertain or inconsistent. For each group, the numbers of domains and structures and the range of the average predicted disorder are summarized in Table 1. Percentage of disordered residues for each structure was averaged over all structures for each domain type. The distribution of average percentage of predicted disorder in the four different groups is shown in Figure 4b. The first three groups had significantly higher predicted disorder than the fourth, and many of domains in the first three groups have experimental evidence supporting disorder‐to‐structure transitions that depend on the indicated stabilizing factors. Particularly, the intrinsic disorder for the macromolecular‐binding domains was verified by searching for experimental evidence in literature. The stabilization factors for the fourth group are discussed in more detail in the Supporting Information (Table S2 and the text below the table). Here, we discuss the first three groups.

Table 1.

Likely mechanism for predicted‐to‐be‐disordered domains to form PDB structures

Groups of stabilization factors	Numbers of domains	Numbers of PDB chains	Range of average predicted disorder% (median)
Disulfide bonds	25	267	26–100% (72%)
Ion‐binding	16	419	19–100% (74%)
Macromolecular‐binding	60	1,084	12–100% (100%)
Uncertain stabilization mechanisms	7	16	52–100% (100%)
	11	357	17–74% (34%)

Abbreviation: PDB, Protein Data Bank.

2.3.1. Disorder‐to‐structure transitions upon disulfide bond formation or ion binding

Disulfide bonds are formed by oxidation of thiol groups between cysteines in a relatively oxidizing environment. Because the interior of cells is a reducing environment, disulfide bonds are usually not found in intracellular proteins but instead are found in extracellular, secreted, and periplasmic proteins,68 although disulfide bonds can also be formed in cytoplasmic proteins under certain conditions,69 such as oxidative stress.70 Among the 25 disulfide‐bond‐containing domains, most of them (22 domains) are secreted proteins (annotation from UniProt), two (PF06747, PF05051) are found in the intermembrane space of mitochondria, and the remaining one (PF09256) is an extracellular domain found in membrane receptors; therefore, the environment of this example is similar to that of a secreted protein. Some secreted proteins in gram negative bacteria have been shown to be unfolded inside the cell because structure‐inducing disulfide bonds remain unconnected in the reducing conditions and then adopt structures outside the periplasm as a result of disulfide bond formation in the oxidative environment.71 In eukaryotes, disulfide bond formation in secreted proteins involves the endoplasmic reticulum and is more complicated but still depends on movement from a reducing environment to a more oxidative environment.72 Furthermore, recent work in bacteria shows that their secreted proteins need to be at least partially unfolded inside the cell to interact with the secretion apparatus and become folded outside the cell, where the extra‐cellular folding is brought about not just by disulfide bond formation but by a variety of mechanisms.73 Thus, secreted proteins represent a new cohort that are transiently disordered and have been called “delayed folding proteins.”74 The observation that the delayed folding of many secreted proteins is regulated by disulfide bond formation confirms our previous conjecture62 that disordered domains that become folded by disulfide bond formation are likely secreted proteins that remain unstructured inside the cell and then gain structure outside the cell following secretion.69, 72, 73, 74

Proteins with ion binding functions are significantly enriched in IDPs.75 Various types of ions, such as Ca²⁺, Zn²⁺, Cd²⁺, Co²⁺, Mg²⁺, Cu²⁺, and Na⁺, have been identified in the current study. Among these, Zn²⁺ binding is the most commonly found type. This is consistent with our previous findings that zinc fingers are the second among the top 20 domains that are strongly correlated with predicted disorder,35 and zinc is the third among the top 20 ligand keywords that are strongly correlated with predicted disorder.38 The folding of a typical zinc finger depends on the binding of Zn²⁺—the conformation changes from an unfolded state into a highly folded structure in the presence of zinc.35, 76 In addition, Gla domains (PF00594) are mostly disordered in the absence of calcium ions and undergo a disorder‐to‐structure transition upon calcium binding, suggesting that calcium ions are required for proper Gla‐domain folding.77

2.3.2. Disorder‐to‐structure transition upon DNA/RNA/protein binding

A set of domains that function via induced folding upon DNA or RNA binding was collected. Evidence of disorder‐to‐structure transition upon DNA/RNA binding is known for all the examples in this study (summarized in Table S3). For instance, DNA‐induced folding of the basic regions of HLH domains (Figure 5) is well documented. HLH domains are developmental regulators of transcription and generally function as dimers, with each monomer containing a basic region (~18 residues) necessary for DNA binding, a helix–loop–helix motif (~34 residues), and a leucine‐zipper region (~30 residues) (Figure 5). The basic region was predicted to be disordered (Figure 5a), which is supported by NMR data demonstrating that this region of the Max protein is poorly folded in the absence of DNA78 (Figure 5b), but acquires an α‐helical conformation upon binding to DNA (Figure 5c). This activity has also been reported in MyoD79 and USF proteins,80 supporting our predictions of HLH disorder (Figure 5a) and suggesting that the disordered basic region is required for DNA recognition.

Figure 5.

DNA‐induced folding of the disordered basic region from HLH domain of human Max protein. (a) VSL2b disorder prediction for the sequences. Red: disorder; blue: order; white: ambiguous. (b) NMR structure of the HLH domain in the absence of DNA (PDB id: 1R05). Basic region: 1–18 residues. Helix–loop–helix: 19–53 residues. Leucine zipper: 54–87 residues. (c) X‐ray structure of the HLH domain bound with DNA (PDB id: 1NKP). The two structures represent the same regions (23–102 residues) from human Max protein (UniProt ID: P61244). DNA chains are colored in light blue and green. NMR, nuclear magnetic resonance; PDB, PDB, Protein Data Bank; VSL2B, disorder predictor trained on V = variously characterized proteins with S = short and/or L = long IDRs, version 2b

Predicted disorder or structure was examined in more detail by sequence alignment and structural superposition for Methly‐CpG binding domain (MBD, PF01429). MBD includes members predicted to be structured and fully disordered (from 1 to 100% disorder). The multiple sequence alignment suggests 15 highly conserved residues, several of which are directly involved in methyl‐CgG binding based on available PDB structures (Figure 6). Obvious conformational changes were observed when comparing the least and most disordered members (Figure 6a,d).

Figure 6.

Diversity of structure/disorder in the MBD domain (PF01429). Top: multiple sequence alignment was carried out by T‐coffee and displayed by ESPript 3.0. Predicted disorder for the sequences increases from top (structure) to bottom (full disorder). The asterisk (*) at the bottom of the alignment indicates highly conserved DNA‐binding sites four available structures were shown below. (a) A structural member (PDB id: 3VXV, chain A). (b) A mediate disordered member (PDB id: 2MOE, chain A). (c) A higher disordered member (PDB id: 1QK9, chain A). (d) A fully disordered member in its DNA bound state (PDB id: 6CCG, chain A). (e) Superposition of the least and fully disordered member (sequence identity 42%) in their bound state. The binding partners are not shown here. MBD, methly‐CpG binding domain; PDB, PDB, Protein Data Bank

Protein‐binding PDDs contained two groups: disordered protein binding and structured protein binding. All of the PDDs found here were predicted to be disordered by structure‐based normalized monomeric area and normalized interface area (NMA–NIA) analysis (see Supporting Information for more detail), and most of them were also predicted to be highly disordered (>75% of the residues) by the sequence‐based VSL2b predictor, suggesting a strong consistency between these two methods (Figure S4a). Most PDDs were bound to structured partners. Selected examples were labeled by their PDB ids, and their structures in the binding complexes were shown in Figure S4b–g. All the disordered domain structures lack globularity, and experimental evidence of intrinsic disorder in the absence of their binding interactions was found for 56 examples (summarized in Table S3).

Overall, the apparent contradiction between the prediction of disorder and existence of PDB structures can be reconciled for the majority of the examples because of the various stabilization factors described above that bring about disorder‐to‐structure transitions. These transitions make disordered domains distinct from well‐structured domains with regard to the mechanisms underlying of protein–protein and protein–nucleic acid interactions.

3. DISCUSSION

3.1. A surprisingly wide range of predicted disorder was observed among the homologues within many individual domains

Prediction of disorder or structure is strikingly different for homologous sequences within many of the different Pfam domains. Sets of seed proteins varied from 13% (PF00023, Ank) to 100% disorder on average among the domains. Of these, only ~36% (75 of 206 domains) were predicted to be >95% disordered on average. When examining the disorder prediction of individual members for a given domain, the percentage range was even wider, ranging from fully ordered to fully disordered. We did not expect that proteins suggested to be homologues by hidden Markov models in Pfam would show such a wide range of predicted disorder. These results are certainly contrary to the general view that structure is more highly conserved than sequence.63, 64, 65, 66, 67

3.2. Disordered domains form structures as a result of various mechanisms

Various mechanisms cause disordered domains to form structures, including disulfide bonds, ion coordination, macromolecular partner binding, and many others. These major three factors are known to assist protein folding and were used to divide disordered domains in previous work.62 The examination of structural characteristics provides reasonable explanations of prediction of disorder and formation of structure for PDDs. This investigation suggests that future disorder predictors should consider these specific types of intrinsic disorder, which are likely to be distinct from IDPs or IDRs that do not undergo disorder‐to‐structure transitions.

3.3. Different disorder amount is likely related to crucial functional differences

Given the significantly different amount of predicted disorder among members in the same domain, it raises the question whether this correlates with their distinct functions. Experiment studies that correlate the disorder amount with functional differences remain largely un‐investigated, although distinct functions between different domain members are known for many examples. For instance, the G_γ domain (PF00631) has two members (G_γ1 and G_γ2) with significantly different amounts of predicted disorder (100 and 33%, respectively). Evidence shows that both G_γ1 and G_γ2 bind to the same partner (G_β1), forming β1γ1 and β1γ2 dimers having significant differences in their functions.81, 82 An experimental investigation of whether there is truly a correlation between functional differences and intrinsic disorder in G_γ would be of great interest.

This work raises important questions about homologous domains that contain both predicted‐to‐be‐disordered and predicted‐to‐be‐structured members. For example, what are the underlying selective advantages that influence a predicted‐to‐be‐disordered domain to evolve into a predicted‐to‐be‐structured domain or vice versa? These conserved‐sequence‐associated changes in structure from disorder‐to‐structure or vice versa might be associated factors such as gain‐or‐loss‐of‐function arising from the altered structural tendency or such as increased resistance or increased sensitivity to regulatory protease digestion. Also, what are the detailed changes in sequence that enable a sequence to change from predicted‐to‐be‐structured into predicted‐to‐be‐disordered (or vice versa) while still being indicated to be conserved by hidden Markov models? Work on these questions is in progress.

In summary, here, we provide further evidence for the existence of intrinsically disordered domains. We show that many disordered Pfam domains have conserved sequences, conserved disorder, and conserved functions, which is analogous to the definition of structured domains, which have conserved sequences, conserved structures, and conserved functions. Our findings strongly support disordered domains as a new type of protein functional element that is distinct from classic structured domains. Our findings also suggest that disordered domains have the unexpected capacity to evolve from disorder to structure (or vice versa) over time.

4. MATERIALS AND METHODS

4.1. Datasets

4.1.1. Structured protein dataset

Of note, 664 structured proteins were derived from non‐redundant (sequence identity <25%) X‐ray crystallography structures from PDB. These are single‐chain monomers having no missing density, no disulfide bonds/ions/small molecules, and not including secreted proteins and coiled coils (because these proteins often predicted to be disordered).

4.1.2. IDP and IDR dataset

One hundred and thirteen IDPs and 692 IDRs were obtained from DisProt.39 Only regions longer than 20 consecutive residues were included for further analysis.

4.1.3. PDD dataset

The initial disordered dataset contained a set of domain seed members for which all of their residues were predicted by VSL2b to have scores >0.50.61 The term “domain” here refers to a Pfam‐A entry. Each Pfam entry includes a set of representative members of the family, hidden Markov models built from the seed alignment, and the automatically generated full alignment containing the proteins belonging to the entry. The term “seed” here refers to an individual protein sequence (in part or in whole) from the seed alignment. The term “member” here refers to an individual protein from the seed alignment, or full alignment, and the difference is spelled out.

The previously used requirement that all residues have VSL2b scores >0.5061 is too restrictive and misses many experimentally confirmed IDPs. Two methods have been used to estimate whole protein disorder, the charge–hydropathy plot,83, 84 and the cumulative distribution function.34, 85 However, neither of these methods has been extended for the problem of distinguishing structured protein regions from IDRs, and many disordered Pfam domains are IDRs rather than IDPs. Thus, we developed a new approach using the average value of per‐residue VSL2b predictions of disorder that we tested for both IDPs and IDRs. In plots of the histograms of the VSL2b average values for the 664 structured proteins, the 113 IDPs and the 692 IDRs in the dataset described above, values in the range of 0.4–0.6 gave good separations of the structured proteins from both the IDPs and IDRs as shown in Figure S1. Not one of the structured protein exhibited an average VSL2b score >0.60; therefore, we used this value here.

For the set of PDD sequences using VSL2b scores >0.60, 206 domain entries were identified and considered as PDDs. Each chosen domain had at least one seed that was identified as a PDD by this criterion.

4.2. Disorder prediction

Disorder prediction was used to examine whether or not additional homologues of the PDD sequences were also highly predicted to be disordered. All the seed sequences of the 206 domains were collected and predicted for disorder on the whole parent Uniprot sequences using the PONDR VSL2b predictor. Seed sequences were chosen because these are manually selected by Pfam for significant sequence identity with each other and more likely to contain additional disordered homolog within a domain having PDD seeds. VSL2b was used because it showed the best overall performance for long IDRs (≥30 consecutive residues) in a comprehensive comparison of 19 predictors.86 Because the length of disordered domains was proposed to be over 20–30 residues,57 it is expected that VSL2b would show good accuracy. We also chose this predictor for the reason of consistency (it was used in our previous work61 for the statistical analysis).

4.3. Classify PDDs by distinct seed compositions

To examine the seed member composition for each PDD, average disorder scores by VSL2b on seed sequences were used to partition the domains into groups. The sequences with average disorder score >0.6 were defined as predicted‐to‐be‐disordered seed members, and those with average disorder score ≤0.6 were defined as predicted to be ordered seed members. In each domain, the numbers of the two types of seeds were calculated, and four groups were classified: (a) domains containing all predicted‐to‐be‐disordered seeds; (b) domains containing mostly predicted‐to‐be‐disordered seeds; (c) domains containing equal numbers of seed types (number difference <2); (d) domains containing mostly predicted‐to‐be‐structured seeds.

Evolutionary relationships between ordered and disordered seeds were examined with domain sequence‐based phylogenies. Phylogenies were built from Pfam model domain alignments using FastTree87 and rooted by the midpoint method.88 The maximum likelihood disorder content and 95% confidence intervals of the last common ancestor of each domain family, according to phylogeny midpoint, was determined.89 Plots of domain families were branch length‐scaled by node depth for display, but ancestors were colored according to their estimated disorder content using the original branch lengths. Phylogenetic analysis and plotting was performed with the R package phytools.90

4.4. Examination of structure/disorder–function relationship

To examine the structure/disorder–function relationship for the PDDs, X‐ray and NMR structures in PDB were identified and possible mechanisms of disorder‐to‐structure transition were provided. Matched PDB structures for the domains were from members in Pfam full alignment. As the focus was on the PDDs, structures with multiple domains on the target chains were removed for further analysis. Short matched peptides (<20 residues) were also not included. Sequence‐based disordered predictions were applied to the whole parent UniProt sequences of the structural members, and the average percentage of predicted disorder was calculated for each matched domain region.

Next, likely mechanisms of structural formation for these predicted‐to‐be‐disordered domains were used to partition them into four groups. These are: (a) domains form structures because of disulfide bonds; (b) domains that form structures because of ion binding; (c) domains that form structures because of macromolecular binding; and (d) domains form structures with no consistent mechanism for structure formation. The method for partitioning into these groups is the same as that used in our previous work.62

CONFLICT OF INTEREST

The authors declare that they have no conflict of interest with the contents of this article.

Supporting information

Appendix S1 Supplemental Methods and Materials

Figure S1 Using average VSL2b score to determine the threshold for IDPs, IDRs and structured proteins.

Figure S2 Consistence of disorder prediction between RP55 and Seed sets for the 206 domains.

Figure S3 Sequence Conservation in PDDs and structured domains

Figure S4 NMA–NIA boundary distances divided protein‐binding interactions into disordered or structured protein binding

Table S1 Percentage of IDPs, IDRs and structured proteins included within three different thresholds of average VSL2b score

Table S2 Likely stabilization factors for highly predicted‐to‐be‐disordered monomer structures

Table S3 Evidence of intrinsic disorder for macromolecular‐binding disordered domains

Zhou J, Oldfield CJ, Yan W, Shen B, Dunker AK. Intrinsically disordered domains: Sequence ➔ disorder ➔ function relationships. Protein Science. 2019;28:1652–1663. 10.1002/pro.3680