Introduction

Intrinsically disordered proteins (IDPs) and intrinsically disordered regions (IDRs) of proteins have been challenging the traditional structure–function paradigm over the last two decades. They exhibit a range of conformations, from molten globules to random coils1, with their net charge correlated to their conformational preference2,3. The absence of distinct structure in IDPs/IDRs can be attributed to their high net charge and low hydrophobicity4. The flexible nature of IDPs/IDRs allows them to be easily regulated by post-translational modifications (PTMs)5. Phosphorylation is a post-translational modification enriched in IDPs/IDRs6, that alters the charge of serine, threonine and tyrosine amino acids by substituting a hydroxy functional group with a negatively charged phosphate group. Phosphorylation plays a crucial role in many biological processes, where it can modify charge and hydrophobicity, and modulate interactions with partners7.

According to the polyelectrostatic model, interconverting conformations of charged IDPs and/or IDRs bind to their partners through an average electrostatic field caused by long-range electrostatic interactions8,9,10. Phosphorylation can modify the electrostatic interactions of disordered regions by altering their charge which can either enhance or reduce their binding affinity for binding partners10. Conformational ensembles of IDPs/IDRs depend on the balance of all charge interactions in the protein, rather than individual short-range interactions between the binding site of the protein and a binding partner. Therefore, they can interact with partners through several distinct binding motifs and/or conformations, and various functional elements will switch between availability for interaction and burial10.

It is becoming clear that phase-separating proteins and other molecules, such as mRNA, mediate a variety of biological functions11. Proteins containing IDRs are commonly found in membraneless organelles, formed by liquid–liquid phase separation (LLPS)12. Messenger RNAs are also commonly sequestered during LLPS, typically alongside proteins with IDRs and specific RNA-binding motifs12, so specific and non-specific charge-charge interactions are likely to be involved. Phosphorylation (and therefore charge variation) is a known control mechanism in LLPS formation13. Databases of proteins observed to associate with membraneless organelles are being collated. Notably, the DrLLPS resource14 classifies proteins undergoing LLPS as core (localised proteins that have been verified as essential for granule assembly/maintenance), client (localised but non-essential), and regulator (contribute to modulation of LLPS but are not located in the condensate).

PhosIDP is an addition to the protein-sol software suite15, created for visualizing the effects of phosphorylation on protein sequence charge profiles, and thereby facilitating functional hypotheses. In order to assist in understanding the behaviour of IDPs/IDRs, the web tool visualizes changes resulting from phosphorylation across a sequence, rather than focusing on the consequences of individual phosphorylation sites. Further, offline calculations for datasets of proteins are presented that demonstrate the role of phosphorylation in modulating charge for IDRs in core LLPS proteins, but not client proteins.

Results

Phosphorylation mediates substantial charge alteration in the IDR of CIRBP

As an exemplar of proteins of interest, human cold-inducible RNA binding protein (CIRBP, UniProt16 ID Q14011) is used, in particular as a representative of a core set of proteins that are crucial for the formation of stress granules (SGs), examples of liquid–liquid phase separated membraneless organelles17. CIRBP possesses a structured N-terminal RNA recognition motif (RRM) that mediates specific RNA interactions, and a C-terminal disordered arginine/glycine-rich (RGG) region that is involved in weak multivalent RNA interactions17. Both the positive charge and disorder of the unphosphorylated RGG region are apparent in Fig. 1 (panels A and B). It is known that phosphorylation can regulate the phase transitions of stress granules13. Here, the incorporation of phosphorylation (as recorded in UniProt) shows that the C-terminal end of the disordered region of CIRBP becomes substantially more negatively charged, noting that the N-terminal end of the disordered region remains positively charged.

Figure 1
figure 1

Results of phosIDP server for CIRBP (UniprotKB: Q14011), a stress granule-associated protein. (A) and (B) Changes to charge and fold propensity upon phosphorylation are depicted in lighter shades. Additional delineation of the profiles is given with a green line for unphosphorylated values and a pink line for phosphorylated values. The sawtooth window for smoothing disordered and ordered regions is shown in blue. (C) Phosphorylation sites from UniProt are displayed. (D) Location of known Pfam domains.

Full size image

Nucleophosmin, a protein with order–disorder transitions coupled to phosphorylation

The potential of our server to highlight the role phosphorylation could play in structural transitions is demonstrated with human nucleophosmin, which (in common with CIRBP) is classified as a core protein in regard to formation of membraneless organelles14. Nucleophosmin shuttles between the nucleus and cytoplasm, undertaking multiple roles18. The current focus is on how prediction of structured and unstructured regions in the phosIDP server, and the presence of phosphorylation sites, correlates with segments for which structural flexibility is known to correlate with function. The core domain is structurally polymorphic, with pentamer formation dependent on phosphorylation status19. Phosphorylation biases towards monomer over pentamer, and this balance is also influenced by ligand binding, contributing to the rich functional properties of nucleophosmin20. Only a part of the folded core domain is predicted as structured, using sequence-based fold propensity (Fig. 2). Colour-coding is common between the predicted structured/unstructured plot (Fig. 2B) and the core domain structure (Fig. 2E, 4n8m19), with the predicted structured and unstructured regions forming the two parts of the monomer. Extensive interactions of the monomer within the pentamer (Fig. 2E) are consistent with the 3D structural stability being dependent upon oligomerisation. Furthermore, the localisation of phosphorylation sites at the monomer interface likely reflects their role in the monomer—pentamer equilibrium19. Therefore, if a user were studying the phosIDP server results, in the absence of known 3D structure, a reasonable prediction would be that conformation predicts only as weakly folded, and that phosphorylation alters the structured/unstructured balance, thereby potentially mediating function. Such a hypothesis could then be subject to the types of experimental analysis that have been applied to investigate the nucleophosmin core domain.

Figure 2
figure 2

Sequence and structure of human nucleophosmin (UniprotKB: P06748). Windowed charge scores (A), predicted fold propensities (B), phosphorylation sites (C), and Pfam domains (D) are shown, for unmodified and phosphorylated nucleophosmin. The Pfam domains correspond with predicted structured domains, for which solved structure is available, and labelled core domain and C-terminal (C-term) domain. (E) Using the same colour-coding as fold propensity prediction (yellow/folded and blue/unfolded), a cartoon of a core domain monomer is shown against a surface of the remaining 4 monomers in a pentamer unit (PDB42 id 4nm819). Known phosphorylation sites are indicated with green spacefill and residue numbers. (F) Also with the yellow and blue colour-coding for predicted structural order from the phosIDP server, the C-terminal domain is drawn (first model from 2vxd21). Known phosphorylation sites for this sequence are shown, along with an indication of the N- and C-terminal ends of the domain.

Full size image

The other domain known to be structured for nucleophosmin, and also detected with the phosIDP sequence analysis, lies at the C-terminus (CTD, Fig. 2 panels B, D, F). The CTD is predicted as relatively weakly folded, with colour-coding transferred from sequence prediction to the NMR structure (2vxd21), consistent with observation that its folding stability is significantly lower than that of the pentamer unit formed by the core domain, as measured by thermal and chemical denaturation22. Indeed, mutation of the CTD has been associated with loss of structure and amyloid formation, possibly related to pathogenesis23,24. These results again indicate the utility of the server for identifying regions that are predicted to be of relatively low folded state stability. With regard to phosphorylation, notable amongst CTD sites (Fig. 2F) is the location of Ser243 at the start of the first helix in the CTD, and lying in the portion predicted (from sequence, coloured blue) to have only marginal folded state stability. Since favourable interactions arising from a phosphorylated helix N-cap are greater than any amino acid N-cap25, it is possible that phosphorylation at Ser243 is coupled with CTD folding stability and thereby with biological activity.

Charge distributions and phosphorylation in stress granule proteins

Whilst proteins termed core, such as CIRBP and nucleophosmin, are central to SG formation, client proteins can be found in SGs but do not themselves cause formation of SGs14. Having found in the two examples studied that SG core protein phosphorylation is extensive, we wondered how core protein charge compares with that of client proteins. In order to study protein regions rather than the overall charge of the protein, and examine local effects, charge in overlapping 21 amino acid windows was summed. Further this was divided into regions predicted to be structured or intrinsically disordered, according to the fold propensity results of the server. Distributions of the windowed net charge values are shown for IDRs of core and client proteins, and their phosphorylated counterparts are compared (Fig. 3). Interestingly, unphosphorylated core proteins tend towards overall charge neutrality more than client proteins, but with a distinct enrichment for positive charge. Upon phosphorylation the enrichment for positive charge is largely removed, perhaps indicative (on average) of reduced RNA interaction and a tendency towards SG dissolution more so than formation, although both behaviors have been observed experimentally, dependent on the system13. Although the calculations presented in Fig. 3 are offline, they demonstrate the relevance of studying charge distributions and how they change with phosphorylation.

Figure 3
figure 3

Core (panel A) and client (panel B) net charge distributions, and phosphorylated counterparts (denoted P), are calculated from overlapping 21 amino acid windows within predicted IDRs. For each of core and client proteins, the difference (phosphorylated—unphosphorylated) is also shown, as a line plot. For simplicity, the net charge in each window is shown on the x-axis. In terms of net charge per amino acid, the displayed values range from a minimum of − 0.476 to a maximum of 0.476.

Full size image

Discussion

Previous work has demonstrated the importance of net charge and also the balance between positive and negative charge in the analysis of IDR conformation and function26. Further, the role played by sequence distribution of charged residues3,27 has been studied, as well as incorporation of non-charged residues with a hydropathy component28, and the analysis of both charge and hydrophobic patterning in the formation of collapsed or expanded states29. Here, the crucial role that phosphorylation plays in many systems, not only through site-specific interactions, but also with global switching of charge profiles, can be quickly and conveniently studied with phosIDP, subject to data collation in the UniProt database. The visualisation omits potential coupling of phosphorylation with sequence-specific effects, such as those associated with proline30,31, leaving scope for improvement as these become better characterised. Two examples were taken from a dataset of ‘core’ proteins that are integral for membraneless organelle formation. For CIRBP (Fig. 1), phosphorylation of a disordered RGG domain, summed over several sites, reduces the net positive charge close to neutrality, similar to the effect seen overall in IDR regions of the core protein set (Fig. 3). Given the mRNA localisation in membraneless organelles, it is apparent that charge modulation could fine tune stability. The charge change due to phosphorylation for predicted structured and disordered regions can be used to obtain the change in net charge per residue (NCPR) for an entire set of proteins. Average changes in NCPR upon phosphorylation are − 0.034 for the core set, and − 0.015 for client proteins. It has been observed that distributions of NCPR values, within sets of homologues for two proteins that enable condensate formation with mRNAs, are relatively narrow around a mean NCPR of approximately + 0.0232. Thus potential charge change on phosphorylation is comparable in magnitude with values that modulate condensate formation. It is recognised that phosphorylation sites collated in UniProt may under-represent the complete set of sites, and (acting in the opposite sense), the collated sites may not all be modified at the same time. Interestingly though, coupled modification within phospho-islands has been well-documented33. In the nucleophosmin example (Fig. 2), phosphorylation is enriched close to the boundaries of predicted structured and disordered regions, which for the oligomerisation domain relates to structural transitions that are known to be modulated by phosphorylation. We suggest that the phosIDP web tool will be particularly useful when looking for these regions of predicted weak disorder or structure, and the effect of phosphorylation, which can in turn be matched to available experimental data or lead to experiment design.

Methods

Using the PhosIDP web tool

As part of the protein-sol software suite15, PhosIDP is freely available at https://protein-sol.manchester.ac.uk/phosidp without requiring a license or registration. In order for the software to execute the user needs to provide a valid UniProt ID16. Results are displayed graphically with additional information provided below the plots. Data is also available to download as text. The custom URL is available for 7 days from the day of creation.

The phosIDP algorithm

Upon receiving a UniProt protein ID16, phosIDP calculates charge at neutral pH as a sequence-based profile, averaged over a sliding window of 21 amino acids15, with a second charge profile added that corresponds to the addition of a double negative charge for each phosphorylation site recorded in UniProt (Fig. 1A). A hydropathy scale34 is combined with net charge to predict IDRs4,35, in a scheme repurposed from the existing implementation on protein-sol15. The scale for fold propensity (Fig. 1B) extends from positive/folded to negative/unfolded. For both the sequence and fold propensity plots, lines tracking the calculated values (and aiding visualisation) are added to the column plots, green for values prior to phosphorylation and pink when phosphorylated. To extend predictions for the amino-terminal and carboxy-terminal 10 amino acids, that are omitted when fitting a sliding 21 amino acid window, predicted values at the eleventh residue in from either of the termini are extrapolated out. An additional feature of the folded state prediction is a sawtooth window that smooths structured and intrinsically disordered regions. Short stretches of IDR within ordered regions may legitimately describe the behavior of loops, but could also obscure the overall domain structure. The sawtooth envelope is created from the predicted profile with the caveat that disordered regions with fewer than 10 amino acids in length take on the average prediction for a window of up to 41 amino acids centered on that region. Additionally, predicted IDRs separated by structured regions of fewer than 10 amino acids are combined into one longer IDR. For reference, phosphorylation sites, as curated by UniProt (Fig. 1C), and Pfam domains36 (Fig. 1D) are displayed.

Supplementary Fig. 1 illustrates the effect of varying the sliding window between 5 and 40, for both charge and fold propensity profiles (using CIRBP, Q14011). It is concluded that a window length of 21 is a compromise between more rapid profile variation at lower lengths (5, 10), and a particularly smoothed profile at longer length (40). A comparison between protein disorder prediction schemes is given in Supplementary Fig. 2. Three additional prediction schemes were used, from those that score highly in a recent review37, IUPred2-Long38, MetaDisorderMD39, and PDisorder (http://www.softberry.com). Accounting for the sense of the predictions, two schemes predict disorder as positive, and two predict disorder as negative, agreement is reasonable for CIRBP and nucleophosmin. An advantage of the fold propensity algorithm in the current application is that it can be easily adjusted according to the charge modulation upon phosphorylation.

Datasets of proteins in membraneless organelles

Three publically available databases were studied in order to generate a dataset of proteins that have been located in membraneless organelles. MSGP40 is a database of proteins in mammalian stress granules, and PhaSepDB41 contains proteins that undergo LLPS in various organelles. Each of these have overlap with DrLLPS14, with the advantage of DrLLPS recording proteins as core, client, or regulator. Since our aim was to compare core and client proteins, DrLLPS was used to generate sets of 56 core proteins and 728 client proteins. The two proteins studied in detail were chosen to be high ranking in terms of percentage of predicted disordered residues that are recorded as phosphorylation sites (nucleophosmin ranks 1 and CIRBP 6 of the 56 core proteins), and with reports of biophysical characterisation.