Structural insights into predicted protein–protein interactions and protein complexes in the human proteome

Abstract

Large-scale computational predictions of human protein-protein interactions (PPIs) have recently enabled systematic mapping of the human interactome at near-atomic resolution. In our previous work, we reported nearly 18,000 high-confidence predicted PPIs, along with selected examples that demonstrated the ability of these predictions to reveal both pairwise associations across diverse pathways and multi-subunit complex organization. Here, we extend that study by presenting additional examples of biological and biomedical interest. We highlight novel PPIs across DNA repair, mitochondrial function, and biogenesis of cilia and other organelles, where predicted interfaces provide mechanistic hypotheses for protein function and pathway crosstalk. We also illustrate how binary predictions can be assembled into models of higher-order assemblies, uncovering new candidate subunits and organizational principles for multi-protein complexes. Importantly, we integrate structural mapping of disease-associated mutations, many of which lie directly within predicted interaction interfaces, thereby offering explanations for how genetic variation may disrupt protein networks and contribute to pathology. Together, these case studies underscore how large-scale computational predictions can be distilled into detailed mechanistic insights, generating hypotheses for functional studies and expanding the structural and functional annotation of the human interactome.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-54567-0.

Introduction

Protein-protein interactions (PPIs) lie at the heart of nearly every biological process, from intracellular signaling and organelle organization to DNA repair, immunity, and metabolism¹. Despite decades of biochemical and structural research, the human interactome remains far from complete². Large portions of the proteome, including many poorly characterized or disease-associated proteins, lack structural or functional context in terms of their interaction partners. Understanding PPIs at atomic resolution is therefore essential not only for mapping the architecture of cellular pathways but also for identifying molecular mechanisms that drive disease and therapeutic responses^3,4.

The advent of deep learning–based structure prediction has transformed the study of PPIs^5–7. Tools such as AlphaFold⁸ and AlphaFold-Multimer⁵ have enabled large-scale, high-confidence predictions of binary complexes across the human proteome. These approaches go beyond interaction detection, revealing structural details of interfaces and conformations that were previously inaccessible. By integrating sequence co-evolution, structural templates, and deep neural network inference, these models provide a powerful means to expand interactome maps with mechanistic resolution.

In our recent large-scale study, we systematically predicted PPIs across the human proteome by integrating deep multiple sequence alignments (MSAs) derived from comparative genomics with a deep learning–based interaction prediction framework trained on experimentally validated interacting and non-interacting protein pairs⁹. High-scoring candidate interactions were further evaluated by structural modeling using AlphaFold2 and AlphaFold-Multimer, allowing both interaction likelihood and interface geometry to be assessed. This approach yielded thousands of candidate PPIs spanning diverse biological processes, including DNA repair, mitochondrial function, immunity, and intracellular transport, many of which were not captured by existing experimental interaction databases⁹. Together, these predictions provided a proteome-wide, structure-informed interactome map that forms the foundation for the detailed case studies presented here.

While deep learning–based structure prediction methods such as AlphaFold and AlphaFold-Multimer have substantially advanced the ability to model protein structures and complexes, their application to large-scale PPI prediction has important limitations. Structural confidence metrics produced by these tools—such as predicted local distance difference test (pLDDT) scores or interface predicted TM-scores (ipTM)—may reflect the estimated accuracy of a modeled structure, rather than the biological existence, specificity, or functional relevance of an interaction in vivo. Interpretation of proteome-wide predictions may be confounded by transient or condition-specific interactions, alternative binding partners, conformational plasticity, and uncertainties in stoichiometry or cellular context. Consequently, even predictions with high model confidence should be interpreted as structural hypotheses rather than definitive evidence of biological interaction. Experimental approaches for PPI detection remain essential for validation and contextualization. They also serve as critical benchmarks and positive control sets for computational methods. A comprehensive understanding of the human interactome thus emerges from the integration of predictive structural modeling with experimental data, leveraging the strengths of each while recognizing their respective limitations. The experimental datasets supported our previous study⁹ in two ways. First, they were used to construct the positive control set. Second, they served as priors for calibrating interaction confidence thresholds: protein pairs with prior experimental evidence were considered confident at a lower in silico threshold.

Here, we present selected case studies that demonstrate how structure-based prediction advances our understanding of human protein networks. Focusing on representative examples, we show how predicted PPIs can rationalize disease-associated mutations, reveal structural parallels between distant homologs, and suggest mechanistic links between pathways not previously thought to intersect. By assembling binary predictions into higher-order complexes, we further illustrate how these models capture subunit organization and functional architecture. Together, these analyses provide a framework for exploring the structural interactome at a systems level, bridging individual predictions to network-level biological insight. The case studies presented here were selected from the larger set of predicted interactions based on a combination of interaction confidence, structural model quality, and biological relevance. We prioritized examples that span diverse cellular processes, involve poorly characterized or disease-associated proteins, map pathogenic variants to predicted interfaces, reveal unexpected connections between pathways, or enable assembly into higher-order complexes. These examples are intended to be illustrative of the types of mechanistic insights that can be obtained from large-scale structure-based PPI prediction, rather than a comprehensive survey of all predicted interactions.

Results and discussion

In our previous work, we showed how comprehensive computational pipelines integrating deep MSAs, RF2-PPI, and AlphaFold-based structural modeling enabled the prediction of nearly 18,000 human PPIs at an estimated precision of ~ 90%, with confidence calibrated using large positive and negative control sets and strategy-specific cutoffs⁹. The present study focuses on further mining and analyzing selected examples of these high-confidence PPI predictions to illustrate their biological and mechanistic significance. For the examples discussed here, we assessed structural confidence using the AlphaFold2-derived inter-molecular contact probability score (AF2 score), which reflects the likelihood of specific residue–residue contacts at predicted PPI interfaces based on distance probability distributions (see Materials and methods).

We highlight representative case studies that illustrate the types of insights enabled by these predictions. All case studies presented below were selected from interactions with high AF2 scores (> 0.9), supporting the reliability of the predicted interface residues and increasing confidence that these models capture biologically meaningful interactions. The contact probability score provides a direct measure of residue-level interaction patterns. In contrast, pTM and ipTM metrics reported by AlphaFold-Multimer provide more global confidence estimates and are therefore used in a complementary manner. Specifically, pTM reflects confidence in the overall fold and topology of the modeled proteins, whereas ipTM measures confidence in the relative arrangement and packing of interacting subunits within the complex. Consistent with this, these pairwise examples show generally high pTM scores (19 out of 22 exceeding 0.5, Table S1), supporting the accuracy of the predicted global folds. Likewise, 17 out of 22 interactions exhibit ipTM scores exceeding 0.7 (Table S1), indicating confident prediction of subunit orientation and interface organization. In addition, the best intermolecular predicted aligned error (PAE) values are uniformly low, with all cases displaying such PAE values below 3 (Table S1). These examples include pairwise interactions that connect proteins across diverse cellular processes such as DNA repair, ciliogenesis, and mitochondrial function, as well as assemblies of binary predictions into models of large higher-order multi-subunit complexes. We also integrate structural analysis of disease-associated variants, many of which map to predicted PPI interfaces, thereby providing mechanistic hypotheses for how single amino acid changes may disrupt PPIs and contribute to human disease. Together, these examples highlight not only the ability of large-scale structure-based predictions to recover known interactions and rationalize genetic variation, but also their potential to uncover unexpected functional connections and propose roles for poorly characterized proteins, ultimately bridging large-scale computational screening with biologically and medically meaningful discovery.

PPIs in cell cycle, DNA repair and cancer biology

Cell cycle regulation and DNA repair are fundamental processes that maintain genomic integrity, and their disruption is a hallmark of cancer¹⁰. Increasing evidence shows that proteins canonically involved in meiosis, DNA damage repair, and transcriptional control can form unexpected interactions when misexpressed in somatic cells, thereby creating new functional connections that may drive oncogenesis. By predicting PPIs across diverse cellular pathways, we can uncover potential molecular links between complexes such as the mismatch repair (MMR) system, the Fanconi anemia (FA) pathway, nucleotide excision repair (NER), and regulators of cell cycle progression. In particular, the exemplar predicted interactions described below between proteins with established tumor suppressor or cancer-testis antigen roles, as well as metabolic and trafficking regulators, highlight points of crosstalk that may couple DNA repair, replication fidelity, and growth control with processes such as meiotic chromosome organization and protein turnover. These findings suggest that novel or ectopic protein associations could contribute to genomic instability, therapeutic resistance, and proliferative signaling in cancer, warranting further functional investigation.

PMS1 and FAAP24

We predicted an interaction between PMS1 and the Fanconi anemia (FA) complex subunit FAAP24 (AF2 score: 0.999, ipTM score: 0.63, pTM score: 0.42) (Fig. 1A, PMS1 residues 6-340,486–528,542–655,659–709,721–925; FAAP24 residues 11–215). The FA complex functions in the repair of DNA interstrand crosslinks (ICLs), highly toxic lesions that covalently link two DNA strands¹¹. PMS1 participates in the pathway of DNA mismatch repair (MMR), which primarily corrects base mismatches and small insertions or deletions arising during DNA replication¹². The predicted interaction between FAAP24 and PMS1 suggests a potential functional link between the FA and MMR DNA repair pathways. Consistent with this, previous studies have shown that MMR plays a role in sensing and initiation of ICL repair^13,14. Such crosstalk could enhance the coordination of genome surveillance, and in cancer cells may influence sensitivity to DNA-damaging therapies that generate ICLs.

Fig. 1.

Examples of PPIs involved in cell cycle, DNA repair and cancer biology. Colors of the protein names match their corresponding structures. Sidechains of residues with disease-causing mutations are shown in spheres.

ERCC8 and STK19

ERCC8 functions in the transcription-coupled nucleotide excision repair (TC-NER) pathway¹⁵. Its primary roles include recognizing RNA polymerase II stalled at DNA lesions during transcription and recruiting additional repair factors to the damage site. Mutations in ERCC8 cause Cockayne syndrome, a rare genetic disorder characterized by growth failure, neurodevelopmental impairment, and premature aging¹⁶. An interaction (AF2 score: 0.9995, ipTM score: 0.88, pTM score: 0.89) between ERCC8 and STK19 (Inactive Serine/Threonine Kinase 19) was predicted. STK19 contains three tightly packed winged helix-turn-helix (WH) domains but lacks a canonical protein kinase domain, suggesting it may not function as a kinase. Recent studies identified STK19 as a DNA/RNA-binding protein essential for DNA damage repair and cell proliferation, acting within both the nucleotide excision repair (NER) and mismatch repair (MMR) pathways¹⁷. Its predicted interaction with ERCC8 is consistent with experimental findings. Structurally, ERCC8 contains a seven-bladed β-propeller domain, and the predicted STK19–ERCC8 interaction involves all three WH domains of STK19 and the fourth, fifth, and sixth blades of the ERCC8 β-propeller (Fig. 1B, ERCC8 residues 6-365, STK19 residues 146–365).

C1orf50 and AHCY

C1orf50 is a protein of unknown function conserved in Metazoa. We predict an interaction (AF2 score: 0.9976, ipTM score: 0.56, pTM score: 0.69) between C1orf50 and adenosylhomocysteinase (AHCY), which was supported by large-scale experimental PPI studies in human^18,19 and fruit fly²⁰. AHCY is a hydrolase that catalyzes the conversion of S-adenosylhomocysteine into adenosine and homocysteine. Pathogenic variants in AHCY cause hypermethioninemia with S-adenosylhomocysteine hydrolase deficiency (HMAHCHD) [MIM:613752], a rare metabolic disorder. Our structural model (Fig. 1C, AHCY residues 88–186, C1orf50 residues 6-404) places the disease-causing mutation D86G at the predicted C1orf50–AHCY interface. C1orf50 has been implicated in multiple malignancies, where its elevated expression is associated with poor prognosis in melanoma and breast cancer, driving proliferation, stemness, and immune evasion^21,22. AHCY, a key enzyme in methylation metabolism through regulation of S-adenosylhomocysteine, has likewise been tied to tumorigenesis across diverse contexts, including bladder, liver, and breast cancers, where its dysregulation alters proliferation, invasion, and DNA damage responses²³. The predicted C1orf50–AHCY interaction may provide a mechanistic link between oncogenic signaling and cellular methylation potential, coupling cancer cell growth and stemness with epigenetic and metabolic reprogramming.

TMEM127 and LRP10

TMEM127 is an endo-lysosomal membrane protein characterized as a negative regulator of mTOR signaling and as a tumor suppressor²⁴. An interaction (AF2 score: 0.999, ipTM score: 0.7, pTM score: 0.41) was predicted between TMEM127 and LRP10 (low-density lipoprotein receptor-related protein 10), a type I transmembrane receptor involved in the trafficking and processing of amyloid precursor protein²⁵. Mutations in LRP10 have been implicated in neurodegenerative disorders, including alpha-synucleinopathies such as familial Parkinson’s disease and dementia with Lewy bodies²⁶. TMEM127 contains four transmembrane segments, with its third and fourth transmembrane segments predicted to engage the single transmembrane segment of LRP10 (Fig. 1D, TMEM127 residues 31–236, LRP10 residues 21–499). In addition, the extracellular loop between these helices in TMEM127 interacts with the cysteine-rich domain of LRP10 near the membrane. Several pathogenic TMEM127 variants identified in hereditary pheochromocytoma-paraganglioma map to this predicted interface. While the precise consequences of the TMEM127–LRP10 interaction remain unknown, such crosstalk may couple lysosomal signaling and receptor trafficking. In cancer, this could influence mTOR-dependent growth pathways as well as protein processing and turnover, thereby contributing to tumorigenesis and therapeutic resistance.

LIN9 and TESMIN

LIN9 and TESMIN (MTL5) lie at the intersection between cell cycle regulation and meiotic control, and their predicted interaction (AF2 score: 0.9976, ipTM score: 0.38, pTM score: 0.38) may hold significance for cancer biology. LIN9 is a core component of the LINC/DREAM complex, a five-protein assembly that governs the transcription of cell cycle genes; depletion of LIN9 leads to cell cycle arrest²⁷. TESMIN (metallothionein-like 5, MTL5) plays a critical role in early meiosis, where it regulates germ cell development²⁸. We predict an interaction between LIN9 and TESMIN mediated by their C-terminal regions (Fig. 1E, LIN9 residues 141–540, TESMIN residues 301–508), consistent with prior truncation experiments that demonstrated this interface is essential for TESMIN’s nuclear translocation²⁹. The functional consequences of this interaction may extend beyond meiosis, as both LINC/DREAM dysregulation and ectopic expression of meiotic proteins have been linked to oncogenesis. A LIN9–TESMIN axis could therefore integrate transcriptional control of cell cycle genes with aberrant meiotic pathways, potentially driving unchecked proliferation, altered chromatin states, and tumor progression.

SYCP2 and HORMAD1

SYCP2 and HORMAD1 are key regulators of meiotic chromosome organization, and their predicted interaction may reveal functional crosstalk relevant to cancer biology. During meiotic prophase, SYCP2, together with SYCP3 and meiotic cohesins, forms part of the axial element (AE), which organizes sister chromatids and provides the structural framework for homologous recombination and synapsis³⁰. HORMAD proteins, including HORMAD1 and HORMAD2, localize to the AE and regulate meiotic progression³¹. While HORMAD2 has been shown to bind the N-terminus of SYCP2³², and HORMAD1 is known to interact with cohesins such as RAD21L and REC8³³, direct binding between HORMAD1 and SYCP2 has not previously been established. Our prediction of a SYCP2–HORMAD1 interaction (AF2 score: 0.9966, ipTM score: 0.73, pTM score: 0.54) (Fig. 1F, SYCP2 residues 1-427,1381–1430,1466–1520, HORMAD1 residues 11–135,141–240,311–325) therefore suggests a previously unrecognized layer of coordination within the AE. Notably, HORMAD1 is aberrantly expressed in multiple cancers as a cancer-testis antigen, where it contributes to genome instability and therapeutic resistance³⁴, while SYCP2 has also been identified as a misexpressed germline factor in certain tumors³⁵. Their interaction may thus create a functional axis through which meiotic proteins, when ectopically expressed in somatic cells, promote recombination-like processes, replication stress, or altered chromosomal dynamics that fuel tumorigenesis.

PPIs of proteins mediating ciliary function

Cilia are multifunctional organelles essential for sensing, signaling, and protein trafficking, and defects in their assembly or regulation give rise to a diverse group of human disorders known as ciliopathies³⁶. At the molecular level, ciliary function depends on the coordinated activity of multiprotein assemblies such as the BBSome, the intraflagellar transport (IFT) complexes, and the transition zone, together with regulatory factors that ensure proper cargo delivery and signaling output. However, many aspects of how these complexes interface remain poorly understood, particularly how less characterized or noncanonical proteins contribute to ciliary biology. Using our predictive framework, we identified and modeled a series of interactions that highlight both known and novel links within the ciliary network. Some interactions led to the construction of multi-subunit complexes described previously⁹, and several representative examples are illustrated below. These predictions expand our understanding of the structural and functional architecture of the cilium, provide mechanistic explanations for disease-causing mutations, and suggest new pathways by which defects in ciliary protein interactions can lead to human ciliopathies.

OSCP1 and WDR54

OSCP1 was recently identified as a novel ciliary protein localized to both the base and axoneme of the cilium in human cells, and is required for proper ciliogenesis in zebrafish³⁷. WDR54 was proposed as a candidate novel ciliary protein and later confirmed to localize to cilia³⁸. We predicted an interaction (AF2 score: 1.0, ipTM score: 0.93, pTM score: 0.88) between OSCP1 and WDR54, suggesting a functional association within the ciliary compartment. WDR54 adopts the fold of a seven-bladed β-propeller, while OSCP1 is predominantly α-helical. Our model indicates that they form a substantial interaction interface (Fig. 2A, OSCP1 residues 1-346, WDR54 residues 1-334) in which a β-hairpin motif of OSCP1 inserts into and extends the β-sheet of the fifth blade of the WDR54 propeller. This configuration not only provides a binding interface but also suggests a potential mechanism by which OSCP1 may regulate or stabilize WDR54 within the BBSome-related or ciliogenesis machinery. Given that defects in either ciliogenesis or β-propeller scaffold proteins are frequently associated with ciliopathies, the OSCP1–WDR54 interaction may represent an important structural and functional node in ciliary signaling pathways.

Fig. 2.

Examples of PPIs involved in ciliary function. Colors of the protein names match their corresponding structures. Sidechains of residues with disease-causing mutations are shown in spheres.

BBS1 and FAIM

BBS1 (Bardet-Biedl Syndrome 1) is a core subunit of the BBSome complex, which is essential for ciliary function³⁹. The BBSome is involved in protein trafficking to and within cilia, maintenance of ciliary structure and function, and regulation of various signaling pathways associated with cilia⁴⁰. Our method identified several known BBS1 interactions, including other subunits of the BBSome complex (BBS2, BBS4, BBS5 and BBS9), TTC8, and ARL6, consistent with established complex assembly. Interestingly, BBS1 was also predicted to interact with FAIM (Fas apoptotic inhibitory molecule) (AF2 score: 0.997, ipTM score: 0.78, pTM score: 0.69), a protein not previously linked to ciliary functions⁴¹. FAIM consists of two duplicated β-sandwich domains⁴². The interface of FAIM and BBS1 is formed by the C-terminal β-sandwich domain of FAIM and the second and third blades of the β-propeller domain of BBS1 (Fig. 2B, BBS1 residues 21–417, FAIM residues 1-179). FAIM was originally identified as a mediator of resistance to FAS-induced apoptosis⁴³. The predicted interaction between BBS1 and FAIM raises the intriguing possibility that FAIM may influence BBSome activity, potentially linking apoptotic regulation with ciliary signaling and function.

BBS7 and WDR19

We predicted an interaction (AF2 score: 0.993, ipTM score: 0.89, pTM score: 0.87) between BBS7, a core subunit of the BBSome complex⁴⁴, and WDR19, a subunit of the intraflagellar transport A (IFT-A) complex⁴⁵. The BBSome functions as a central organizer of ciliary protein trafficking by coordinating the IFT-A and IFT-B complexes together with kinesin motors to form a supercomplex that mediates the bidirectional transport of cargo along the axoneme. Our structural model (Fig. 2C, BBS7 residues 1-715, WDR19 residues 1-350) provides novel insights into the interface between BBS7 and WDR19, thereby expanding our understanding of how the BBSome engages the IFT machinery. Notably, a pathogenic mutation in WDR19 (V80D) maps directly to this predicted interface. This variant has been linked to Senior–Løken syndrome 8 (SLSN8) [MIM:616307], a rare autosomal recessive oculo-renal ciliopathy characterized by the combination of nephronophthisis, a progressive cystic kidney disease, and retinal dystrophy leading to vision loss⁴⁶. The presence of a disease-causing mutation at this interface suggests that disruption of BBSome–IFT-A contacts may underlie key pathogenic mechanisms in ciliopathies. More broadly, this predicted interaction reinforces the concept that coordinated assembly of trafficking complexes at the ciliary base is essential for ciliary homeostasis, and that its perturbation can drive multisystem disorders.

TMEM218 and TMEM216

TMEM218 and TMEM216 are transmembrane proteins associated with the primary cilium, where they contribute to the structure and function of the ciliary transition zone—a critical gatekeeper for ciliary protein trafficking and signaling⁴⁷. Mutations in TMEM218 have been linked to ciliopathies, including Joubert syndrome and related disorders, which manifest as brain malformations, kidney dysfunction, and retinal degeneration. Similarly, mutations in TMEM216 cause Joubert syndrome⁴⁸, characterized by a distinctive midbrain–hindbrain malformation that results in motor and cognitive impairments, as well as Meckel syndrome, a severe ciliopathy presenting with encephalocele, polydactyly, polycystic kidneys, and liver fibrosis. Our predicted structural model (Fig. 2D, TMEM218 residues 16–145, TMEM216 residues 1-179) reveals an interaction (AF2 score: 0.9995, ipTM score: 0.77, pTM score: 0.64) between TMEM218 and TMEM216, with two pathogenic variants of TMEM218 (R80C and R80H) mapping precisely to the interface of this interaction⁴⁹. These findings suggest that disease-causing mutations may disrupt TMEM218–TMEM216 binding, thereby compromising transition zone integrity and leading to the broad spectrum of ciliopathy phenotypes observed in patients.

RAB28 and WDR31/BBS4

WDR31 is a recently identified ciliary protein implicated in the regulation of the intraflagellar transport (IFT) machinery and in recruitment of the BBSome to the cilium⁵⁰. Our predictions revealed an interaction between WDR31 and RAB28 (AF2 score: 0.998, ipTM score: 0.83, pTM score: 0.91) (Fig. 2E, RAB28 residues 11–190, WDR31 residues 56–367), a small GTPase associated with both the BBSome and IFT processes⁵¹. Structurally, we predict that the WDR31–RAB28 interface is positioned near residue G19 of RAB28, a site where a missense variant (G19R) has been classified as deleterious by AlphaMissense. This suggests that disruption of RAB28 function at the interface may impair its ability to coordinate BBSome–IFT assembly. Consistent with this, RAB28 is also predicted to interact with the BBSome subunit BBS4 (AF2 score: 0.9897, ipTM score: 0.87, pTM score: 0.8), forming a potential tripartite connection between WDR31, RAB28, and the BBSome. Although the full complex structure of WDR31 and RAB28 remains to be experimentally resolved, our model highlights disease-relevant residues: two missense variants in BBS4 (P199S and D251G) map directly to the predicted RAB28–BBS4 interface (Fig. 2F, RAB28 residues 11–190, BBS4 residues 36–447,454–478) and are predicted to be deleterious. Together, these findings suggest that mutations affecting WDR31–RAB28–BBS4 interactions may destabilize BBSome recruitment and IFT regulation, thereby contributing to ciliopathy pathogenesis.

PPIs of proteins involved in intracellular transport and organelle generation

Intracellular transport and organelle biogenesis rely on precisely coordinated PPIs that govern cargo recognition, vesicular trafficking, membrane remodeling, and organelle assembly^52,53. Disruption of these interactions can impair essential cellular processes such as endocytosis, peroxisome formation, and Golgi and endosomal dynamics, often resulting in severe developmental or neurological disorders^54–56. Advances in structure prediction now allow systematic identification of previously uncharacterized PPIs and provide mechanistic insights into how disease-associated mutations may perturb intracellular trafficking pathways. Below, we present representative PPIs involving proteins implicated in intracellular transport and organelle generation, highlighting structural models that suggest functional roles for poorly characterized proteins and reveal potential molecular mechanisms underlying human disease.

C6orf226 and PEX7

C6orf226, a protein with unknown function, is predicted to have an interaction with PEX7 (AF2 score: 0.9995, ipTM score: 0.86, pTM score: 0.93). PEX7 (peroxisomal biogenesis factor 7) plays a vital role in the formation and function of peroxisomes. PEX7 acts as a cytosolic receptor for peroxisomal matrix enzymes targeted to the organelle by the peroxisome targeting signal 2 (PTS2)⁵⁷. It is responsible for transporting several essential enzymes into peroxisomes, particularly in lipid synthesis and breakdown processes. PEX7 is also involved in the import of specific enzymes necessary for the normal assembly and function of peroxisomes. The critical nature of PEX7’s function is evident from the disorders associated with its mutations. Severe mutations in PEX7 can lead to Rhizomelic chondrodysplasia punctata type 1 (RCDP1), characterized by skeletal abnormalities and intellectual disability⁵⁸. Some mutations in PEX7 can cause a small percentage of Refsum disease cases, affecting vision, smell, and other functions⁵⁹. C6orf226 is a small poorly characterized protein predicted to be mostly disordered. In the AF2 model (Fig. 3A, PEX7 residues 11–320, C6orf226 residues 22-64), C6orf226 mainly adopts a loop conformation that wraps around the β-propeller domain of PEX7, while the N- and C-termini of C6orf226 adopts small α-helical conformations. One likely pathogenic mutation of PEX7 according to ClinVar, H39P, lies in the interface of the predicted PEX7-C6orf226 model (Fig. 3A, magenta residue). This mutation is linked to Peroxisome biogenesis disorder 9B and Rhizomelic chondrodysplasia punctata type 1 (RCDP1).

Fig. 3.

Examples of PPIs involved in intracellular transport and organelle generation (A–C), and mitochondrial function (D,E). Colors of the protein names match their corresponding structures. Sidechains of residues with disease-causing mutations are shown in spheres.

C12orf57 and ANKRD50

C12orf57 is a small protein with unknown function that has been linked to disorders affecting brain function such as Temtamy syndrome⁶⁰, colobomatous microphthalmia⁶¹, intellectual disability⁶², and corpus callosum hypoplasia^62,63. We predict an interaction (AF2 score: 0.9995, ipTM score: 0.9, pTM score: 0.89) between C12orf57 and ANKRD50 (Ankyrin Repeat Domain 50). C12orf57 forms a helical bundle with 5 α-helices and interacts with the middle region of ANKRD50 consisting of ankyrin repeats in our model (Fig. 3B, C12orf57 residues 11–110, ANKRD50 residues 1–50,66–505). ANKRD50 is a key component of the SNX27–retromer–WASH super complex responsible for endocytic recycling of various membrane proteins⁶⁴. Retromer dysfunction has been linked to various neurodegenerative diseases⁶⁵. It would be interesting to test if C12orf57 also has retromer-related functions. We identify a likely disease-causing mutation at the interface of C12orf57 and ANKRD50 (Fig. 3B). This mutation (L51Q) on C12orf57 is buried inside the interface and makes several interactions with large hydrophobic residues from ANKRD50. This mutation is linked to the autosomal recessive Temtamy syndrome characterized by intellectual disability, variable craniofacial dysmorphism, ocular coloboma, seizures, and brain abnormalities, including abnormalities of the corpus callosum and thalamus⁶¹.

GPR89A and LMBR1L/RABL3

GPR89A (Golgi PH regulator A) was annotated as a putative G protein coupled receptor. It enables voltage-gated monoatomic anion channel activity, which is crucial for the movement of ions across cell membranes⁶⁶. We predict that GPR89A interacts with LMBR1L (AF2 score: 0.99, ipTM score: 0.6, pTM score: 0.67), a membrane protein involved in receptor-mediated endocytosis and signal transduction. Interestingly, GPR89A and LMBR1L adopt the same fold with 9 transmembrane helices, suggesting that GPR89A is not a classic GPCR. Both GPR89A and LMBR1L have a soluble insertion with two long α-helices in between the fifth and sixth transmembrane helices. Their distant homologous relationship is supported by HHpred similarity search (probability score > 0.98). GPR89A was also predicted to interact with RABL3 (AF2 score: 0.994, ipTM score: 0.83, pTM score: 0.63), a RAB family small GTPase, consistent with a recent experimental study⁶⁷. GPR89A, RABL3 and LMBR1L are important for lymphoid development and function^67,68. GPR89A interacts with LMBR1L by using their transmembrane helices and interacts with RABL3 by using the two long helices in the soluble region (Fig. 3C, GPR89A residues 1-453, LMBR1L residues 1-475, RABL3 residues 6-210).

PPIs of mitochondrial proteins

Mitochondria are central hubs of cellular metabolism, energy production, and signaling, and their function relies on the coordinated activity of hundreds of proteins across multiple compartments. Defects in mitochondrial proteins or their assembly frequently underlie human disease, including neuromuscular disorders, metabolic syndromes, and cardiomyopathies. PPIs are critical for nearly every aspect of mitochondrial biology, from respiratory chain complex assembly to RNA processing and metabolite transport. Mapping these interactions provides mechanistic insight into how mutations disrupt mitochondrial function and cause disease. We identified and modeled a range of mitochondrial PPIs, many of which involve factors essential for the assembly and activity of respiratory chain complexes. The predicted interaction models not only recover known functional associations but also highlight novel contacts that directly link disease-causing mutations to protein interaction interfaces, thereby offering possible structural explanations for mitochondrial pathologies.

MT-CO1 and SMIM20/CMC1/CMC2/COA3

Four complex IV (cytochrome c oxidase complex) assembly factors (SMIM20, CMC1, CMC2 and COA3) classified in the MitoCarta database⁶⁹ exhibited interactions (AF2 scores > 0.92, ipTM scores > 0.51, pTM scores > 0.78) with the major complex IV subunit MT-CO1 (cytochrome­ c oxidase subunit I). These assembly factors form distinct interfaces with MT-CO1, enabling the joint modeling of them (full-length proteins) by AlphaFold3 (ipTM: 0.66, pTM: 0.71) (Fig. 3D). SMIM20 and COA3 are single transmembrane proteins that mainly interact with MT-CO1 by using their transmembrane regions. CMC1 and CMC2 belong to the CX9C family proteins characterized by two disulfide bonds formed between two α-helices⁷⁰. They are soluble proteins predicted to interact with MT-CO1 from the intermembrane space.

SCO1 and COA6

SCO1 (Synthesis of Cytochrome c Oxidase 1) is a mitochondrial protein that plays crucial roles in the assembly of cytochrome c oxidase (COX) and copper homeostasis⁷¹. It acts as a copper metallochaperone, helping to incorporate copper ions into the COX complex. SCO1 is essential for the biogenesis of the CuA site in the cytochrome c oxidase (COX) subunit Cox2. Mutations in the SCO1 gene can lead to mitochondrial disorders characterized by complex IV (cytochrome c oxidase) deficiency, resulting in conditions such as hepatic failure, encephalopathy, hypertrophic cardiomyopathy, and fatal infantile encephalopathy. We predict an interaction (AF2 score: 0.9976, ipTM score: 0.78, pTM score: 0.79) between SCO1 and COA6, another complex IV assembly factor involved in the biogenesis of CuA. Two residues involving disease-causing mutations (P174L of SCO1 and W59C of COA6, both causing complex IV deficiency) reside in the interface and interact with each other (Fig. 3E, SCO1 residues 96–295, COA6 residues 51–120).

YBEY and MRPS11

We predict an interaction (AF2 score: 0.995, ipTM score: 0.86, pTM score: 0.88) between YbeY, an endoribonuclease that functions in rRNA maturation and ribosomal control, and 28S ribosomal protein S11 in Mitochondria (MRPS11) (Fig. 3F, YBEY residues 1-167, MRPS11 residues 61–190). In our prior bacterial interaction work⁷², we predicted an interaction between YbeY and the 30 S ribosomal protein S11, which demonstrates how our methodology is able to detect highly conserved interactions across domains of life despite our MSAs only containing eukaryotic sequences.

Multi-subunit complex modeling based on pairwise interactions

While binary interaction predictions are invaluable for identifying potential partners and binding interfaces, most cellular machines operate as multi-subunit assemblies. To bridge this gap, we developed a systematic strategy to assemble binary predictions into higher-order complexes using AlphaFold3 and iterative superposition methods. These multi-subunit complexes are constructed from proteins that exhibit high-confidence pairwise interactions. This approach ensures that the assemblies are grounded in reliable binary interaction evidence prior to higher-order modeling. The resulting complexes are further evaluated using AlphaFold3-derived pTM and ipTM scores, which provide a measure of overall structural confidence. The relatively low ipTM scores observed for some complexes may arise from several factors, including structural flexibility between subunits, partial or transient interfaces, uneven confidence across different regions of the complex, or challenges in modeling larger assemblies. These factors can reduce global confidence scores without necessarily undermining the reliability of specific local interactions, and may also reflect alternative conformational states, partial occupancy of interfaces, or the accumulation of uncertainty when combining multiple subunits into larger assemblies.

This approach allowed us to model a diverse set of complexes, ranging from membrane-tethering assemblies like the NRZ complex to cytosolic scaffolds such as the SNAPc transcription factor and large vesicle trafficking complexes like TRAPPIII. The resulting models not only recapitulate known structural features of these assemblies but also extend them by identifying uncharacterized subunits, alternative binding arrangements, and interfaces enriched for pathogenic mutations. For example, we uncovered candidate new components of the NRZ tethering complex, revealed structural details missing from experimental SNAPc structures, and proposed architectural models for poorly characterized complexes such as BLOC-1 and BORC. Importantly, these predictions provide mechanistic hypotheses about how mutations disrupt multi-protein interactions and how subunit organization dictates function. Taken together, our work shows that multi-subunit complex modeling is a powerful extension of binary PPI prediction, enabling a deeper understanding of molecular assemblies that underlie human biology and disease.

The NRZ-syntaxin-LMBR1L complex

The NRZ tethering complex, comprising NBAS, RINT1, and ZW10, is essential for vesicle tethering to the ER and Golgi membranes, maintaining Golgi integrity, and facilitating both anterograde and retrograde transport between ER and Golgi⁷³. Its function is crucial for proper protein sorting, cellular homeostasis, and overall vesicle trafficking within the cell. Our method predicted interactions between C19orf25 and NRZ subunits RINT1 and ZW10, suggesting that C19orf25 is a candidate subunit of the NRZ complex. C19orf25 is an uncharacterized small protein (118 residues) that has been found to bind ZW10 in tandem affinity purification screens⁷⁴. Multi-subunit modeling of the NRZ complex and its interacting proteins by AlphaFold3 (ipTM: 0.37, pTM: 0.41) (Fig. 4A and Figure S1) suggests that C19orf25 forms coiled coil interactions with RINT1 and ZW10. Although the global confidence metrics (ipTM and pTM) are modest—likely reflecting the involvement of flexible helices and coiled-coil regions—our biological interpretation focuses on the consistency of predicted interface residues rather than precise geometric details. Notably, the interface residues identified in the AlphaFold3 models agree with those inferred from independent modeling using five AlphaFold2 and five AlphaFold-Multimer models, based on residue-level contact probability. In addition, multiple interactions between subunits of the NRZ complex and subunits of a syntaxin complex (STX18, BNIP1, and USE1) suggest they assemble into a higher-order complex, consistent with previous studies⁷³. Interestingly, LMBR1L, a multi-pass transmembrane protein, interacts with both NRZ subunit ZW10 and syntaxins STX18 and BNIP1, suggesting its roles in membrane tethering and transport between ER and Golgi.

Fig. 4.

Examples of multi-subunit complex modeling. (A) The NRZ-syntaxin-LMBR1L complex modeled by AlphaFold3. (B) The sarcoglycan complex modeled by AlphaFold3. Colors of the subunit names match corresponding structures.

The sarcoglycan complex

The sarcoglycan complex is made up of four sarcoglycan subunits (α, β, γ, and δ) that integrate into the cell membrane of skeletal muscle and cardiac muscle⁷⁵. The proper assembly and function of the sarcoglycan complex are essential for maintaining the structural integrity of muscle fibers and their resistance to mechanical stress. Mutations in the genes encoding these subunits can lead to various forms of limb-girdle muscular dystrophy (LGMD)⁷⁶, highlighting their critical role in muscle health and function. Several pairwise interactions were predicted between sarcoglycan complex subunits. The sarcoglycan complex with the α, β, γ, and δ subunits (SGCA, SGCB, SGCG, and SGCD) is modeled by AlphaFold3 (ipTM: 0.64, pTM: 0.64) (Fig. 4B and Figure S2). The β, γ, and δ subunits are homologous type II single transmembrane proteins with their extracellular C-termini consisting of repeating β-strands. They form a long and intertwined triple-stranded β-helical structure resembling phage tail fiber proteins⁷⁷ and the spike region of the type VI secretion system spike protein VgrG⁷⁸. Such a structural resemblance suggests that the sarcoglycan complex could have originated from bacteria or bacteriophages. The α-sarcoglycan is a type I transmembrane protein with a long N-terminal region in the extracellular space and a short C-terminal tail. It has two globular extracellular domains: an Ig-like domain and a SEA domain with the ferredoxin-like fold, consistent with a homology-based prediction⁷⁹. The interactions between α-sarcoglycan and the triple β-helix are mainly mediated by the Ig-like domain.

The BLOC-1 complex

The BLOC-1 (biogenesis of lysosome-related organelles complex 1) complex contributes to the sorting, trafficking, and targeting of proteins and lipids to lysosome-related organelles (LROs) such as melanosomes (involved in pigment production and storage in melanocytes), platelet-dense granules (important for blood clotting), and lytic granules^80,81. Disruptions in BLOC-1 components can lead to defects in LROs, resulting in disorders such as Hermansky-Pudlak syndrome (HPS), which is characterized by albinism, bleeding disorders, and other symptoms due to impaired LRO function. The BLOC-1 complex consists of eight subunits: BLOC1S1, BLOC1S2, BLOC1S3, BLOC1S4, BLOC1S5, BLOC1S6, SNAPIN and DTNBP1. The atomic structure of the BLOC-1 complex is not available. Our method detected numerous interactions between BLOC-1 subunits, all of which adopt mainly coiled coil structures. Some pairwise interactions involve overlapping long α-helices, while other interactions are restricted between the ends of α-helices. The AlphaFold3 structural model of the BLOC-1 complex (ipTM: 0.53, pTM: 0.52) revealed an elongated and bent overall shape (Fig. 5A and Figure S3), consistent with electron microscopy studies⁸². The intertwined coiled coil complex structure consists of two four-helical bundles connected by their ends. One four helical coiled coil bundle consists of BLOC1S1, BLOC1S3, BLOC1S4 and BLOC1S6, and the other helical bundle involves BLOC1S2, BLOC1S5, SNAPIN and DTNBP1. The α-helices of the complex all have the same N- to C-direction.

Fig. 5.

Models of the BLOC-1 and BORC complexes involved in lysosomes and LROs. (A) The BLOC-1 complex functioning in LROs. (B) The BORC complex functioning in lysosomes. Both complexes share three subunits—SNAPIN, BLOC1S1, and BLOC1S2 (underlined) and differ in other subunits. Protein names are colored to match their corresponding subunit structures. Structurally flexible regions were not displayed in these models.

The BORC complex

The BORC (BLOC-One Related Complex) is a protein complex that plays a crucial role in the positioning and movement of lysosomes within cells⁸³. BORC achieves this by recruiting kinesin-1 motor proteins to lysosomes, which move lysosomes along microtubules to the cell periphery. This positioning is crucial for cellular processes like nutrient sensing and signal transduction. By positioning lysosomes at the cell periphery, BORC plays a role in cell migration and invasion⁸⁴, as lysosomes can release enzymes that degrade extracellular matrix components, facilitating these processes. BORC is also involved in autophagy, the process by which cells degrade and recycle their components⁸⁵. The BORC complex is related to the BLOC-1 complex as they share three subunits: BLOC1S1, BLOC1S2 and SNAPIN (underscored in Fig. 5). The BORC complex also has five additional subunits: BORCS5, BORCS6, BORCS7, BORCS8, and KXD1. Numerous pairwise interactions were found for BORC complex. The AlphaFold3 model of the eight subunits together (ipTM: 0.61, pTM: 0.61) revealed an elongated structure similar to the BLOC-1 complex. Two coiled-coil four helical bundles were formed and joined together: one consisting of BLOC1S1, BORCS6, BORCS8 and KXD1 and the other consisting of BLOC1S2, BORCS5, BORCS7 and SNAPIN (Fig. 5B and Figure S4).

The BLOC-2 complex

The BLOC-2 (Biogenesis of Lysosome-related Organelles Complex-2) complex is another key protein complex involved in the biogenesis and function of LROs⁸⁰. The BLOC-2 complex is composed of three subunits: HPS3 (Hermansky-Pudlak Syndrome 3 protein), HPS5, and HPS6⁸⁶. The primary function of the BLOC-2 complex is to facilitate the later stages of LRO formation and maturation. Our method detected two pairwise interactions between BLOC-2 complex subunits: HPS3-HPS5 and HPS3-HPS6, allowing us to build a model of the BLOC-2 complex by AlphaFold3 (ipTM: 0.41, pTM: 0.44) (Fig. 6A and Figure S5). The three subunits adopt the same domain architecture consisting of an N-terminal β-propeller domain followed by C-terminal α-helical repeats.

Fig. 6.

Examples of multi-subunit complex modeling. (A) The BLOC-2 complex modeled by AlphaFold3. (B) The SNAPc complex modeled by AlphaFold3. (C) The URI1 complex modeled by AlphaFold3. Colors of the subunit names match corresponding structures.

The SNAPc complex

The snRNA-activating protein complex (SNAPc) is essential for the transcription of small nuclear RNA (snRNA) genes, which are crucial components of the spliceosome machinery involved in pre-mRNA splicing⁸⁷. The complex specifically facilitates the transcription of snRNA genes by RNA polymerase II (Pol II) and RNA polymerase III (Pol III). The SNAPc complex is composed of five subunits (SNAPC1-SNAPC5), which together bind to the proximal sequence element (PSE) in the promoter regions of snRNA genes. Structures of parts of the SNAPc complex involving SNAPC1, SNAPC3, SNAPC5 and the N-terminal part of SNAPC4 have been characterized experimentally⁸⁸. However, experimental structures are not available for SNAPC2. Our method detected an interaction between SNAPC2 and the C-terminal part of SNAPC4. The AlphaFold3 model (ipTM: 0.52, pTM: 0.45) revealed detailed structures of the α-helical domain formed by the N-terminal part of SNAPC1 (Fig. 6B and Figure S6). In addition, the N-terminal part of SNAPC4 and SNAPC5 are modeled, for which a significant portion is missing in the cryo-EM structure.

The URI1 complex

The URI1 (Unconventional prefoldin RPB5 Interactor 1) complex is a specialized molecular chaperone complex that plays a crucial role in the proper folding and assembly of nascent polypeptides and protein complexes^89,90. The key subunits of the URI1 complex include URI1, UXT, PDRG1, ASDURF, PFDN2 and PFDN6. These subunits are evolutionarily related and share the same prefoldin fold. Of these components, PFDN2 and PFDN6 are shared with the classical prefoldin complex. The URI1 subunit (green in Fig. 6C) is different from other URI1 complex subunits as it contains a long C-terminal disordered region. Multiple interactions between the URI1 complex subunits were detected by our method. In addition, multiple interactions were found between this C-terminal disordered region of URI1 and other proteins, including DNAAF10, POLR2E, GPN1 and GPN3. We built a multisubunit model of the URI1 complex and its interaction partners (ipTM: 0.36, pTM: 0.39) (Fig. 6C and Figure S7). All six subunits of the URI1 complex adopt the prefoldin fold with two α-helices forming a helical hairpin and one or two β-hairpins inserted in between the helices. The β-hairpins form two intertwined β-barrels in the complex and the α-helices interact with each other to form a ring-like core structure. GPN1 and GPN3 are two small GTPases that form a heterodimer and interact with PDRG1, PFDN6 and the disordered region of URI1.

The TRAPPIII complex

The TRAPPIII complex is an essential multi-subunit protein complex involved in intracellular vesicle trafficking⁹¹. It functions primarily as a guanine nucleotide exchange factor (GEF) for Rab1 GTPase, which is crucial for autophagy and trafficking between ER and the Golgi apparatus. The TRAPPIII complex is composed of both core and TRAPPIII-specific subunits. The core subunits (TRAPPC1, TRAPPC2, TRAPPC2L, TRAPPC3, TRAPPC4, TRAPPC5, TRAPPC6A and TRAPPC6B) are shared across different TRAPP complexes. The subunits unique to the TRAPPIII complex are TRAPPC8, TRAPPC11, TRAPPC12, and TRAPPC13. While the CryoEM structure of TRAPPIII complex is available for the core subunits and parts of TRAPPC8 and TRAPPC11, the detailed structure of the subcomplex formed by TRAPPC11, TRAPPC12 and TRAPPC13 is not known⁹². AlphaFold detected interactions between these subunits and allowed us to build the subcomplex for them (ipTM: 0.49, pTM: 0.51) (Fig. 7A and Figure S8).

Fig. 7.

Examples of multi-subunit complex modeling. (A) The TRAPPIII complex modeled by AlphaFold3. The core subunits (TRAPPC1, TRAPPC2, TRAPPC2L, TRAPPC3, TRAPPC4, TRAPPC5, TRAPPC6A and TRAPPC6B) are shown in gray. Other subunits are labeled with names. (B) The DPM complex modeled by AlphaFold3. (C) The SAMM50 complex modeled by AlphaFold3. Colors of the subunit names match corresponding structures.

The DPM complex

Proper glycosylation is vital for cell-cell interactions, signal transduction, and immune response. The DPM complex, or Dolichol-Phosphate Mannose Synthase complex, is a critical enzyme complex involved in the process of glycosylation, which is the attachment of sugar molecules to proteins and lipids⁹³. This process is essential for the proper folding, stability, and function of many proteins. The DPM complex catalyzes the synthesis of Dol-P-Man from GDP-mannose and dolichol phosphate. Dol-P-Man serves as a mannose donor in various glycosylation reactions. The DPM complex is composed of three subunits: DPM1, DPM2, and DPM3. DPM1 is the catalytic subunit responsible for the synthesis of Dol-P-Man. DPM2 and DPM3 are regulatory subunits that stabilize DPM1 and enhance its activity. DPM2 and DPM3 are transmembrane proteins with two transmembrane segments that anchor the complex to the ER membrane and are necessary for the stability of DPM1. DPM2 is also a subunit of the GPI-GnT complex that interacts with PIGP and PIGY. These proteins are homologous, suggesting a common evolutionary origin of the DPM complex and the GPI-GnT complex. The experimental structure of the DPM complex is not available. We detected interactions between the DPM subunits, which allowed modeling of this complex by AlphaFold3 (Fig. 7B and Figure S9). DPM1 interacts with the soluble α-helices of DPM2 and DPM3, while DPM2 and DPM3 form interactions by using their transmembrane helices. The model (ipTM: 0.47, pTM: 0.53) also captures the dimerization of the DPM1-DPM2-DPM3 complex, mainly through interactions of the transmembrane helices.

The SAMM50 complex

The SAMM50 complex, also known as the SAM (Sorting and Assembly Machinery) complex, is a crucial component of the mitochondrial outer membrane⁹⁴. It plays a significant role in the assembly and maintenance of the mitochondrial outer membrane proteins, particularly the β-barrel proteins. The SAM complex is essential for mitochondrial biogenesis and proper functioning of the mitochondrial outer membrane. AlphaFold detected several interactions of SAMM50. DNAJC11, another transmembrane β-barrel protein⁹⁵, was predicted to interact with the transmembrane barrel of SAMM50 by using a transmembrane helix (AF2 score: 0.9395). SAMM50 also interacts with MTX2 (AF2 score: 0.995) from the outside of the outer mitochondrial membrane and CHCHD6 (AF2 score: 0.9883) from the intermembrane space. The interaction of SAMM50 and MTX2 is mediated by the loop regions between the membrane β-strands of SAMM50. The N-terminal alpha-beta-domain of SAMM50 is involved in the interaction between SAMM50 and CHCHD6, which is a subunit of the MICOS complex⁹⁵. The interaction between SAMM50 and CHCHD6 provides a bridge between the outer membrane and the inner membrane of mitochondria. Our method also identified the interaction between DNAJC11 and ARMC1 and the interaction between DNAJC11 and MTX1. These pairwise interactions allowed modeling of the SAMM50 complex (ipTM: 0.48, pTM: 0.53) (Fig. 7C and Figure S10).

Conclusions

This study extends our recent large-scale prediction of the human interactome by focusing on representative case studies that demonstrate how PPI predictions can be leveraged to generate biological and biomedical insights. By examining pairwise interactions in DNA repair, mitochondrial biology, and ciliogenesis, and by reconstructing multi-subunit assemblies from binary predictions, we illustrate the capacity of structure-guided methods to illuminate both local interaction interfaces and higher-order complex organization. The integration of disease-associated variants further reveals how disruptions at predicted interfaces may underlie human genetic disorders, providing mechanistic explanations that link molecular structure to phenotype. While our main study established the overall framework and scope of proteome-wide PPI prediction, the present work emphasizes the depth of insight that can be extracted from individual examples, ranging from functional hypotheses for poorly characterized proteins to architectural models of protein complexes. Together, these findings highlight the utility of systematic PPI prediction as both a discovery tool and a hypothesis generator, bridging large-scale computational screening with experimental and clinical investigation.

Materials and methods

This study builds on the proteome-wide PPI predictions, trained RF2-PPI models, deep MSAs, and AlphaFold-based pairwise complex models reported previously⁹. The present work focuses on biological analysis of selected high-confidence interactions and on new modeling of multi-subunit complexes assembled from the predicted pairwise interactions. The sections below summarize procedures described in detail in our previous work⁹.

Assembly of the human protein set

The reference human proteome was obtained from UniProt (release UP000005640, version 2021_02)⁹⁶. Structural models for these proteins had been predicted by AlphaFold2 and made available through the AlphaFold Protein Structure Database (AFDB)⁹⁷. To ensure computational feasibility during downstream modeling, each protein was segmented into domains, allowing long proteins to be represented by smaller independently folded elements⁹. A small subset of proteins that were modeled in multiple frames were excluded. The final dataset contained 20,296 proteins and was used for all analyses⁹.

Benchmark interaction datasets and control sets

Experimentally supported PPIs were collected from UniProt⁹⁶, BioGRID⁹⁸, and STRING⁹⁹. From UniProt, 75,739 candidate pairs were retrieved⁹. BioGRID version 4.4.225 provided 771,432 candidate interactions, with 77,147 supported by multiple experiments. STRING version 12.0 contributed 626,513 candidate PPIs, of which 68,761 were considered high confidence (combined score > 700). Only 3,988 interactions were common to all three sources, and these were designated as a stringent positive control set. An extended set of 22,184 interactions supported by two sources was used as an additional positive benchmark.

To construct negative controls, all possible pairs among the human protein set were enumerated, and those with evidence of interaction or functional association were excluded⁹. Homologous protein pairs were also removed, yielding approximately 140 million candidate negative pairs without supporting evidence, which served as the background set for training and evaluation.

Genomic data acquisition

Genomic and transcriptomic data from diverse eukaryotic species were assembled to enable construction of deep MSAs. A total of 9,337 high-quality genome assemblies were downloaded from NCBI, filtered to ensure contig or scaffold N50 values above 50 kb⁹. Among these, 2,451 assemblies contained annotated proteomes. For species lacking complete assemblies, whole-genome and transcriptomic reads were downloaded from the NCBI Sequence Read Archive. Datasets were restricted to Illumina platforms and filtered by size to exclude extremely small or overly large entries. Species from phyla close to humans, particularly Chordata and Arthropoda, were emphasized due to their evolutionary proximity and large dataset representation.

Ortholog detection and sequence assembly

Orthologs were identified using multiple complementary strategies⁹. For annotated proteomes, human proteins were aligned with BLAST and reciprocal best hit criteria applied. For unannotated genomes, coding sequences were assembled from scaffolds using MMseqs¹⁰⁰, exonerate¹⁰¹, and miniprot¹⁰². Transcriptomic and genomic reads were aligned against reference protein or exon sets using DIAMOND¹⁰³, filtered by reciprocal best hit, and clustered to remove redundancy before consensus sequences were generated. Outgroup comparisons were used to set lineage-specific identity cutoffs, ensuring that only true orthologs were retained while minimizing the inclusion of paralogs or contaminants.

Construction of multiple sequence alignments

MSAs were constructed for each human protein by iteratively expanding alignments with HMMER¹⁰⁴ using progressively relaxed thresholds⁹. Fragmented alignments were merged, and for each source species only the highest-scoring hit was retained. Sequences with poor coverage or low similarity were removed. Because intrinsically disordered regions (IDRs) are difficult to align across evolutionary distances, they were systematically filtered using conservation statistics. Regions of poor coverage and low similarity were excluded, while conserved motifs and structured domains were retained. In brief, we scanned multiple sequence alignments to find short regions outside known domains that are more conserved than expected—based on gap fraction and amino-acid similarity (average BLOSUM62 score)—and merged nearby conserved windows to define inter-domain conserved motifs (details described in supplementary materials in⁹⁾. This refinement improved alignment quality and increased the sensitivity of downstream coevolution-based predictions.

Training datasets for interaction prediction

Training datasets were prepared at both the protein and domain level⁹. Positive examples were derived from heterodimeric complexes in the Protein Data Bank with resolution better than 5 Å. Chains containing large fractions of missing residues or non-standard amino acids were removed, and redundant entries were clustered. Pairs homologous to human proteins were excluded to prevent data leakage, resulting in nearly 200,000 high-quality non-human heterodimeric pairs. Negative training examples were generated by random pairing of proteins from the same organism that lacked STRING-supported functional associations, down-sampled and clustered to yield 65,911 non-interacting pairs.

Domain–domain interactions were extracted from AlphaFold-predicted models of multi-domain proteins in the AFDB. Proteins were parsed into structural domains using a parser that combined InterPro annotations with AlphaFold-predicted inter-residue distances and errors. Domain pairs forming confident structural contacts were retained if they mapped to distinct InterPro¹⁰⁵ entries, producing over half a million positive DDIs. Negative DDIs were generated by randomization and filtering against STRING associations, resulting in more than 360,000 high-quality non-interacting examples.

Neural network development

An adapted deep learning model, RF2-PPI, was developed from the RoseTTAFold2 framework¹⁰⁶. The 3D structure prediction track was removed, leaving a simplified two-track architecture optimized for interaction detection. The network was trained using loss functions for masked residue recovery, distance distributions, orientation predictions, and disorder annotation⁹. Positive and negative PPIs and DDIs were provided in balanced ratios. Validation was performed with independent test sets of 3,000 positive and 30,000 negative human protein pairs.

Selection of example protein–protein interactions

We carried out proteome-wide predictions of PPIs using a hierarchical workflow described in detail in the Supplementary Materials of⁹. Briefly, we implemented two complementary pipelines: a de novo PPI prediction pipeline and an experiment-guided pipeline. The de novo pipeline integrates subcellular localization filters with DCA, RF2-PPI, and AlphaFold2 in a staged manner, such that only high-scoring pairs at each step advance to the next. The experiment-guided pipeline starts with protein pairs reported in physical and genetic interaction databases and uses RF2-PPI and AlphaFold2 to enrich confident interactions from noisy experimental datasets. Pipeline performance was monitored using positive and negative control sets, and strategy-specific thresholds were chosen to target an estimated 90% precision⁹.

We assessed model confidence by calculating residue–residue contact probabilities from AlphaFold outputs, defined as the sum of the predicted probabilities of intermolecular residue pairs falling within distance bins below 12 Å. For each candidate pair, we generated five AlphaFold2 and five AlphaFold-Multimer models using the ColabFold pipeline¹⁰⁷ and our MSAs. Inter-protein contacts were defined as residue pairs with contact probability > 0.5 and predicted distance < 6 Å. We aggregated contacts across the 10 models and retained those observed in at least 50% of models as consistent contacts. The representative structure for each PPI was selected as the model containing the largest number of consistent contacts.

Final selections emphasized interactions involving poorly characterized proteins, connecting distinct pathways, or providing structural context for disease-associated variants. Contact probability scores for models generated by AlphaFold2 (named AF2 score in this study), ColabFold-AF2, and ColabFold-AF-Multimer are reported in Supplementary Tables S1 and S2.

Assembly of multi-subunit complexes

To extend binary predictions into higher-order assemblies, pairwise models were systematically combined. Binary complexes were generated with AlphaFold-Multimer and then merged stepwise by superimposing shared subunits. At each stage, geometric consistency, interface integrity, and absence of steric clashes were required. For complexes involving multiple proteins, assembly was guided by iterative superposition. Final assemblies were refined and validated using the AlphaFold3 server¹⁰⁸, which allowed larger multi-subunit structures to be modeled directly. The AlphaFold3 server was used both to confirm assemblies constructed from binary predictions and to build de novo models of complexes that could not be reliably extended by pairwise superposition alone. Assemblies were evaluated for compatibility with available structural information and internal coherence of subunit organization. These multi-subunit complex models were evaluated using confidence metrics provided by AlphaFold3, including the predicted TM-score (pTM), which reflects the overall structural coherence of the assembled complex, and the interface predicted TM-score (ipTM), which reports the confidence of inter-subunit interfaces. These scores were used to assess the reliability of both global assembly and individual interaction interfaces.

Integration of disease-associated mutations

To investigate the impact of genetic variation on PPIs, we mapped disease-associated single amino acid variations (SAVs) onto predicted structural interfaces⁹. Information on human SAVs and their clinical annotations was obtained from UniProt variant files. For quality control, the wild-type amino acid at each SAV position was compared with the amino acid in our sequences to ensure consistency. Proteins for which more than 5% of residues were mismatched were excluded, resulting in the removal of SAV annotations for 246 proteins (approximately 1.3% of the proteome). Predictions from AlphaMissense¹⁰⁹ were also integrated after verifying sequence consistency with our reference set. Three proteins (0.01%) were excluded due to greater than 5% mismatched residues. To define pathogenic SAVs, we considered three lines of evidence: (i) prediction as pathogenic by AlphaMissense, (ii) annotation as “pathogenic” or “likely pathogenic” in UniProt, and (iii) association with disease in UniProt. Only variants supported by at least two of these three criteria were retained, yielding a total of 10.4 million SAVs classified as potentially pathogenic. All SAVs passing these filters were mapped onto protein structures. Residues were considered part of an interaction interface if their side chains were within 6 Å of atoms from a partner protein in either predicted or experimental complexes. Variants at interfacial positions were highlighted as candidates most likely to disrupt or modulate interactions. This integration of large-scale mutational data with structural models enabled mechanistic hypotheses to be generated about how pathogenic variation alters protein networks and contributes to disease.

Structural and functional annotation

All predicted interactions and assemblies were integrated with prior knowledge from UniProt, BioGRID, and STRING. Structural models were cross-referenced with experimentally solved complexes in the Protein Data Bank. Predictions not in experimental databases were annotated as novel. Disease mutations mapped to predicted interfaces were highlighted as candidates for mechanistic follow-up, as their structural positioning suggested potential roles in destabilizing interactions, preventing assembly, or altering pathway connectivity.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

J.P., J.Z., and Q.C. conceived and designed the study. J.P. and J.Z. performed computational analyses and generated structural models of protein–protein interactions and complexes. Q.C. supervised the project and provided guidance on data interpretation and manuscript preparation. J.P. wrote the main manuscript text and prepared the figures. All authors discussed the results, contributed to data analysis, and reviewed and approved the final manuscript.

Funding

This research is supported by National Institute of General Medical Sciences (1R35GM160468-01) to Q.C., the Welch Foundation grant V-I-0004-20230731 to Q.C., and National Institute of Allergy and Infectious Diseases (1K99AI180984-01) to J.Z.

Data availability

Structural models of the predictions of PPIs in the human proteome are available at: http://prodata.swmed.edu/humanPPI.

Declarations

Competing interests

The authors declare no competing interests.