(C) PLOS One
This story was originally published by PLOS One and is unaltered.
 . . . . . . . . . .



A predicted structural interactome reveals binding interference from intrinsically disordered regions [1]

['Junhui Peng', 'Laboratory Of Evolutionary Genetics', 'Genomics', 'The Rockefeller University', 'New York', 'United States Of America', 'Li Zhao']

Date: 2026-01 

Proteins function through dynamic interactions with other proteins in cells, forming complex networks fundamental to cellular processes. While high-resolution and high-throughput methods have significantly advanced our understanding of how proteins interact with each other, the molecular details of many important protein-protein interactions are still poorly characterized, especially in non-mammalian species, including Drosophila. Recent advancements in deep learning techniques have enabled the prediction of molecular details in various cellular pathways at the network level. In this study, we used AlphaFold2 multimer to examine and predict protein-protein interactions from both physical and functional datasets in Drosophila. We found that functional associations contribute significantly to high-confidence predictions. Through detailed structural analysis, we also found the importance of intrinsically disordered regions in the predicted high-confidence interactions. Our study highlights the importance of disordered regions in protein-protein interactions and demonstrates the importance of incorporating functional interactions in predicting physical interactions between proteins. We further compiled an interactive web interface to present the predictions, facilitating functional exploration, comparative analysis, and the generation of mechanistic hypotheses for future studies.

Understanding which proteins interact with each other and how they interact is essential for uncovering fundamental biology and for identifying new pathways involved in health and disease. However, identifying protein-protein interactions experimentally is often challenging and error-prone, and many organisms still lack comprehensive interaction maps. In this study, we use AlphaFold2-multimer, a powerful artificial intelligence tool, to generate high-confidence predictions of PPIs in Drosophila melanogaster, a widely used model organism. We highlight the importance of incorporating often neglected functional associations when predicting protein-protein interactions at a genomic scale. The predictions enable us to examine how intrinsically disordered regions can mediate binding across large interaction networks, revealing widespread, structurally plausible interactions in vivo. Overall, our work demonstrates how AlphaFold predictions can greatly expand our understanding of the structural forces that shape protein interaction networks and help reveal hidden layers of cellular complexes and pathways.

Funding: This work was supported by National Institutes of Health (NIH) MIRA R35GM133780 and the Allen Distinguished Investigator Award from Paul G. Allen Family Foundation to L.Z. The work was also partly supported by the Shapiro-Silverberg Fund for the Advancement of Translational Research to J.P. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Copyright: © 2026 Peng, Zhao. This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

In this study, we combine AlphaFold2 multimer and the STRING database [ 14 ] to examine and predict the structural details of protein-protein interactions in D. melanogaster. We aim to systematically and comprehensively evaluate different sources of protein-protein interactions within the Drosophila genome, including physical interactions inferred from experiments or homology and functional associations derived from co-expression, co-existence in the same pathways, text-mining, genomic context, etc. Our study provides a resource for the structural details of possible physical interactions in the Drosophila genome, and highlights the importance of incorporating functional PPIs in future high-throughput computational studies. We also reveal the important and prevalent roles of intrinsically disordered regions in the Drosophila PPI network. We further compiled our analysis and results into an interactive web interface for users to explore, validate, and expand protein interaction hypotheses.

AlphaFold2 Multimer [ 24 , 25 ] has provided a new tool for evaluating direct and indirect protein interactions, potentially distinguishing true, stable interactions from experimental artifacts [ 18 , 26 – 28 ], and enhancing our understanding of the structure and function of interaction networks. From an evolutionary biological perspective, AlphaFold2 Multimer is also a powerful tool for studying the interactions and changes in interactions of homologous proteins across different lineages [ 29 ].

Drosophila is an important model species for biological and biomedical research. Despite significant efforts to study Drosophila protein-protein interactions both in vivo and in vitro, many uncertainties remain in existing PPI databases, such as false positives [ 11 ], inconsistency between databases [ 12 ], and underrepresentation of transient or weak interactions [ 5 ]. The combination of complex tissues and small body sizes poses unique challenges in biochemistry compared to other model organisms such as rodents, particularly due to the difficulty in obtaining sufficient amounts of protein samples from specific tissues for high-quality experiments such as affinity purification mass spectrometry. Additionally, methods such as yeast two-hybrid may introduce errors due to system differences [ 13 ]. Therefore, it is essential to evaluate these databases to gain a deeper understanding of the reliability and biases of previous methods [ 11 , 14 ]. Furthermore, while much is known about the interactions of ordered structural components, characterizing the interaction details involving disordered proteins or disordered regions remains challenging [ 1 , 15 – 21 ]. It has been shown that disordered regions can adopt diverse binding modes, ranging from transient fuzzy complexes to disorder-to-order transitions upon binding [ 15 – 17 , 19 – 22 ]. Recent experimental and computational studies further highlight conditionally folding as an important disorder-to-order binding mechanism [ 19 , 23 ]. Despite these studies, systematic characterization of different disorder binding modes and their contributions to protein-protein interactions at the proteome scale remains limited.

Cells are dynamic environments where thousands of proteins interact simultaneously at any given time, and these protein-protein interactions (PPIs) are essential for various biological functions. PPIs can range from stable, permanent associations – such as those within stable protein complexes and molecular machinery – to transient, short-lived interactions with a wide range of affinities and lifetimes, as seen in signal transduction pathways [ 1 – 4 ]. Understanding the organization and dynamics of the cellular PPI network is crucial for a comprehensive picture of cellular functions. Over the past decades, diverse high-throughput methods have been developed and used to study these interactions in different organisms, presenting a systematic overview of how proteins work together in cells [ 5 ]. For instance, affinity purification mass spectrometry [ 6 ] has been instrumental in identifying protein complexes. Yeast two-hybrid systems [ 7 ] and other in vitro techniques such as cross-linking mass spectrometry [ 8 ] have also been used to map PPIs at high throughput. More recently, in vivo proximity labeling methods such as BioID [ 9 ] and TurboID [ 10 ] have emerged as powerful tools for capturing dynamic protein interactions within living cells.

Results

Predicted high-confidence D. melanogaster physical protein-protein interaction network We downloaded all the protein-protein interactions (PPIs) with the highest STRING confidence scores (≥0.9) from the STRING database, version 11.5 [14], which include both direct physical interactions and indirect functional associations. The STRING database provides five main, non-exclusive categories of supporting evidence for the PPIs, including experimental, co-expression, curated databases, genomic context, and text-mining of published literature. In addition to these five categories, the STRING database has also provided a physical subnetwork by incorporating evidence from the curated database organized in IMEx [30] and BioGRID database [31]. Since it has been indicated that functionally associated proteins, such as co-expressed proteins, are more likely to interact with each other than random protein pairs [32–34], we did not restrict our analysis to the physical subnetwork, as our goal is to systematically predict and examine all possible protein-protein interactions. We used AlphaFold2 multimer to predict the structures of 27,711 protein pairs from the STRING database with the highest STRING confidence scores (≥ 0.9), involving approximately 5000 protein-coding genes clustered at 50% protein-sequence identity (Methods). The 27,711 protein pairs can be divided into five mutually exclusive categories according to STRING confidence scores in the five different supporting evidence categories, including 7838 in experimental, 7088 in co-expression, 6764 in curated databases, 5995 in text-mining, and 6 in genomic context (S1 Fig). We used pDockQ to assess AlphaFold2 multimer prediction confidence [26,35]. Among the 27,711 protein pairs, 8,101 pairs (29.2%) were predicted to have direct physical interactions at acceptable confidence with pDockQ ≥ 0.23, and 3,621 pairs (13.1%) at high confidence with pDockQ ≥ 0.5. A summary of the predictions can be found in S1 Appendix. This suggests that a relatively small percentage of the previously reported functional associations may have direct physical interactions. We further examined whether the 3,621 predicted high-confidence PPIs (pDockQ ≥ 0.5) are supported by the STRING-curated physical subnetwork. Among these predicted high-confidence PPIs, 52% (1867 pairs) are supported by highest STRING confidence scores for physical interaction (≥ 0.9, S2 Fig), ~ 70% (2541 pairs, which include the 1867 pairs with scores ≥ 0.9) by high STRING confidence scores (≥ 0.7, S2 Fig), and ~81% (2929, which include the 2541 pairs with scores ≥ 0.7) by median STRING confidence scores (≥ 0.4, S2 Fig). When further categorizing the 3,621 predicted high-confidence PPIs into five STRING database categories according to their highest STRING confidence scores, we found that 31.6% (1144) are best supported by experimental data, 13.7% (496) by co-expression, 19.7% (715) by curated database, 34.9% (1263) by text-mining, and 0.1% (3) by genomic context, specifically gene fusions. These results suggest the robustness of our AlphaFold2 Multimer predictions, with ~81% of the predicted high-confidence PPIs supported by reasonable physical interaction evidence in the STRING database. Interestingly, many of the remaining ~19% predicted interactions that lack physical evidence in STRING are supported by indirect functional associations, such as previously reported co-expression and co-occurrence in cellular pathways, highlighting the importance of incorporating functional associations in predicting PPI structures.

Predicted high-confidence interactions with limited or no support in the STRING physical subnetwork To gain an understanding of how indirect, functional associations – such as co-expression – contribute to the predicted high-confidence PPI network, we examined the subset protein pairs with limited or no support in the STRING physical subnetwork (STRING confidence score < 0.4). The analysis identified 692 predicted high-confidence PPIs, involving 918 proteins (Fig 2a, S4 Appendix). Among them, we observed a large, connected subnetwork consisting of 136 proteins and 105 PPIs as well as three median-sized connected subnetworks with 10–50 PPIs each (Fig 2a and 2b). The remaining PPIs formed 82 small, connected subnetworks with 3–9 PPIs and 176 singleton PPIs. To examine the biological relevance of these predicted PPIs, we manually examined several interactions. We identified potential PPIs that were either characterized after the release of the STRING database version 11.5 [14] or suggested by genetic experiments but lacked evidence of physical interaction. For example, Ran/Nup50 and Ptp61F/Stat92E were both scored as high-confidence functional associations but low-confidence physical interactions in the STRING database. However, recent high-resolution structures of the interactions between the orthologs of Ran-Nup50 in humans suggest that they directly interact [38]. Our AlphaFold multimer analysis supports a high-confidence interaction of Ran/Nup50 and Ptp61F/Stat92E (Fig 2c-2e). Notably, the structure of Ran/Nup50 (PDB code 7MO0) was released after the AlphaFold multimer training cutoff date (September 30, 2021) and thus was not included in model training [24]. In the other example, Stat92E is a transcription factor in the JAK/STAT signaling pathway, and Ptp61F is a protein tyrosine phosphatase that regulates the pathway. There is no direct evidence that Ptp61F directly interacts with Stat92E in Drosophila. However, indirect evidence indicates that knockout of Ptp61F in Drosophila increases the level of phosphorylated Stat92E [39,40], suggesting that Ptp61F potentially interacts with phosphorylated Stat92E to mediate its dephosphorylation. Overall, high-confidence PPI predictions will provide valuable hypotheses for future functional and mechanistic studies.

Most predicted dimers involve binding through disordered regions To further understand the interaction details of these PPIs at the genomic scale, we analyzed the interaction interfaces of each PPI and classified them into different binding modes, including binding through ordered regions and IDRs (Methods). The latter were further classified into three modes, which are disordered binding, coil-to-order binding, and conditionally folded (CF) binding (Methods). CF binding was hypothesized to occur predominantly via conformational selection [23,41], while coil-to-order binding is more likely driven by a combined mechanism of conformational selection and induced fit, where induced fit represents an entropy-favored mechanism in which a disordered region folds into a specific structure upon binding, adapting its shape to fit the interacting partner [22]. Our analysis shows that while almost all the predicted high confidence PPIs involve ordered regions (98.5%, 3566/3621), a comparable proportion of PPIs (98.8%, 3580/3621) involve IDRs, either entirely or in part (Fig 3a). PPIs that involve long IDRs (≥30 residues in length) also represent a notable fraction (75.1%, 2719/3621) (Fig 3a). When classifying each PPI that involves IDR to the above mentioned three categories, we found many interactions involving conditionally folded binding (51.7%, 1873/3621) and coil-to-order binding (21.0%, 762/3621) (Fig 2a). The results highlight the important role of IDRs in mediating protein-protein interactions at the genomic scale. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 3. Many PPIs involve IDRs. a) Number of PPIs with disordered (IDR), ordered, short IDR, long IDR, conditionally folded (CF), and coil-to-order interfaces. b) An example of the predicted PPI structure of Ku80/Irbp with its similar PPI structure of their homologs in humans (PDB code 7Z6O). c). pLDDT (blue line) and intrinsic structural disorder (ISD, black line) of Ku80 (left column) and Irbp (right column) in their monomer states. pLDDT values were rescaled to 0 to 1 by dividing by 100. The secondary structures of Ku80 and Irbp in monomers, SS (monomer), and complex states, SS (complex), were shown in the middle panels, with coil in lightblue, helix in purple, and strand in orange. Different interfaces were shown in the bottom panels, including disordered binding interface (labeled as IDR), conditionally folded binding interface (labeled as CF), coil-to-order binding interface (labeled as coil-to-order), and ordered binding interface (labeled as ordered). The presence of one binding mode at a specific residue index is indicated by darker maroon. d) Examples of disordered (left panel), coil-to-order (middle panel), and CF (right panel) interfaces in the predicted Ku80/Irbp structure. The corresponding regions in Ku80 and Irbp were shaded in (c). https://doi.org/10.1371/journal.pcbi.1013899.g003 Note that the three binding modes are not mutually exclusive when two proteins interact, as there might be multiple IDRs in the complex that belong to different binding modes. For example, Ku80 and Irbp are important components in DNA damage response. The predicted PPI structure was highly confident with a pDockQ of 0.73, which matches the homologous human complex structure (PDB code 7Z6O, Fig 3b). Through detailed structural analysis, we found several disordered regions in both Ku80 and Irbp monomers involved in their interactions (Fig 3c). Among these disordered interfaces, residues 246–272 in Ku80 (Ku80_246–272) undergo coil-to-order transitions, where the coil secondary structure in Ku80 monomer becomes a strand in the complex (Fig 3d, left panel). Interestingly, the Irbp interface (residues 276–302) that interacts with Ku80 (residues 246–272) also undergoes a coil-to-order transition (Fig 3d, bottom panel). Besides these interfaces, we also observed that some interfaces between Ku80 and Irbp are conditionally folded regions. For example, two of the interfaces in Ku80, Ku80_423–429 and Ku80_509–534, were predicted to be disordered (Fig 3c, left panel). However, they were also predicted to have high pLDDTs (≥ 70) in the Ku80 monomer and contain helices in the monomer (Fig 3c, left panel). Such structural versatility may enable proteins such as Ku80 and Irbp to fine-tune their interactions in response to cellular conditions, particularly during processes such as DNA damage response. The complex roles and various binding modes of IDRs suggest that a single PPI can involve multiple binding mechanisms simultaneously, highlighting the structural complexity and dynamic nature of disordered regions in mediating protein-protein interactions.

Example of IDR interfaces with different binding mechanisms and evolutionary origins IDR interfaces can either be conserved or lineage specific. Here, we present examples in which the IDR interfaces have different evolutionary origins and undergo different binding mechanisms. One example is the predicted PPI structure between Sec61alpha and CG13426 (Fig 4a, left panel), which resembles the human homolog structure (PDB code 1LDJ). Sec61alpha and CG13426 are both subunits of the translocon; their interactions are important for the transportation of peptides onto the ER membrane. The disordered interface in Sec61alpha is residues 47–55, which does not involve coil-to-order transitions upon binding. The interface in CG13426, residues 46–104, is predicted to be conditionally folded, i.e., it is predicted to be disordered but with high pLDDT (>70) in the monomer state (Fig 4A, right panel). Despite these differences in binding mechanism and intrinsic disorder propensity, both interface regions are highly conserved across diverse organisms (Fig 4b). PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 4. PPI examples illustrate different IDR interfaces with different evolutionary origins. a) interact through a coil region in Sec61alpha and a conditionally folded region in CG13426. b) Both interfaces of Sec61alpha and CG13426 are highly conserved and have very early evolutionary origins in eukaryotes. c) Roc1b undergoes coil-to-order transitions when binding to Cul5. d) The coil-to-order interface for Roc1b is highly conserved and has early evolutionary origins. e) Jra undergoes coil-to-order transitions when binding to Stat92E. f) The coil-to-order interface for Jra has a relatively recent origin and is highly conserved after its origin. https://doi.org/10.1371/journal.pcbi.1013899.g004 In another example, Roc1b/Cul5 is a key interaction in the multisubunit Cullin-RING E3 ubiquitin ligase. The predicted PPI structure showed that the interface between Roc1b and Cul5 involves a long β strand (Fig 4c, left panel). However, this strand was not observed in the monomer structure of Roc1b (Fig 4c, right panel). Thus, we categorized this interface as a coil-to-order interface. We showed that this interface has an early evolutionary origin and is highly conserved across different taxa (Fig 4d, left panel). A third example is a recently originated coil-to-order interface in Jra (Fig 4e and 4f). In this example, Jra residues 163–186 were disordered, forming a coil structure in the monomer state (Fig 4e). It forms an antiparallel β strand structure upon binding to its partner Stat92E. The interface is relatively new in evolution, which can date back to arthropod species (Fig 4f).

[END]
---
[1] Url: https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1013899

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/