Systematic discovery of protein interaction interfaces using AlphaFold and experimental validation

“We fi rst inspected the monomeric structural models from the AlphaFold database (Varadi et al, 2022; Jumper et al, 2021) of both interacting proteins to determine the boundaries of their ordered and coiled-coil regions, which were also treated as “ ordered ” .”

“For every NDD-NDD PPI we used Interactome3D (Mosca et al, 2013) and PDB database searches (https://www.rcsb.org/ (Berman et al, 2000)) to identify whether a structure already existed for this PPI.”

“We used the Python libraries pandas (McKinney, 2010) for data analysis, and Matplotlib (Hunter, 2007) and seaborn (Waskom, 2021) for data visualization.”

“Titration curves were fi tted using the leastsq function from the scipy.optimize python package (Virtanen et al, 2020) using the model BRET = ((A/D) * BRETmax)/(BRET50 + (A/D)) described in (Drinovec et al, 2012), which assumes a 1:1 binding mode, to obtain estimates for the BRETmax and BRET50.”

“In cases where the motif adopts partial secondary structure, such as loop-helix-loop or loop-strand-loop, they are treated as helical or strand, respectively.”

“We used the Python libraries pandas (McKinney, 2010) for data analysis, and Matplotlib (Hunter, 2007) and seaborn (Waskom, 2021) for data visualization.”

“ROC and PR statistics were calculated using the Python package sci-kit learn (Pedregosa et al , 2012).”

“Copyright (c) Schrodinger, LLC., for the visualization and superimposition of AlphaFold models.”

“The minimal binding domains of the ELM instances were detected in the interaction partner using Pfam HMMs (Mistry et al , 2021).”

“We used the Python libraries, Biopython (Cock et al, 2009) for sequence similarity computation, pandas (McKin-ney, 2010) for data analysis, and Matplotlib (Hunter, 2007) and seaborn (Waskom, 2021) for data visualization.”

“We used the Python libraries, Biopython (Cock et al, 2009) for sequence similarity computation, pandas (McKin-ney, 2010) for data analysis, and Matplotlib (Hunter, 2007) and seaborn (Waskom, 2021) for data visualization.”

“…(PR) curves using the positive and random datasets (Fig 1E, Fig S2A,B, Table S2) show that the AF-derived metrics model confidence (preprint:Evans et al , 2021), average interface residue pLDDT, average motif interface residue pLDDT, pDockQ (Bryant et al , 2022), and iPAE (Teufel et al ,…”

“We searched for functional annotations and existing structures for the monomers using the PDB, ProViz (Jehl et al, 2016), SMART (Letunic et al, 2021), and the scienti fi c literature to identify enzymatic pockets or binding interfaces for DNA, RNA, or metal ions.”

“Using minimal interacting regions of curated DMI structures from the ELM DB, we obtained sensitivity estimates for AF of around 80%, which is similar compared to previously published work (Tsaban et al , 2022; Johansson-Åkhe et al , 2021).”

“Using yeast two-hybrid (Y2H) as primary screening method, more than 60,000 binary interactions between human proteins have been identified (hereafter referred to as the HuRI dataset, (Luck et al , 2020)).”

“Sensitivities of around 70% were reported using benchmark datasets of structurally resolved protein interactions originally developed to evaluate docking methods (Akdel et al, 2022; Bryant et al, 2022; Johansson-Åkhe et al, 2021; preprint:Evans et al, 2021).”

“Mass spectrometry-based approaches involving affinity purification or co-fractionation have identified around 150,000 co-complex associations between human proteins (Drew et al , 2017; Huttlin et al , 2021).”

“Motifs are short stretches of amino acids in disordered regions of proteins that usually adopt a more rigid structure upon binding to folded domains in interaction partners (Davey et al , 2012).”

“This is likely particularly true for motif-mediated PPIs, which are anticipated to number in the hundreds of thousands or millions (Tompa et al, 2014).”

“For every NDD-NDD PPI we used Interactome3D (Mosca et al, 2013) and PDB database searches (https://www.rcsb.org/ (Berman et al, 2000)) to identify whether a structure already existed for this PPI.”

“These short regions were predicted to bind to a pocket on the FHA domain that is known to bind phosphorylated threonines (Durocher et al , 2000), which led us to conclude that these predictions were likely wrong.”

“The calculation of DockQ scores of AlphaFold models was done in reference to their solved structures using the code available on the github repository of DockQ (https://github.com/bjornwallner/ DockQ, (Basu and Wallner, 2016).”

“This points to an important limitation in AF ’ s ability to predict binding speci fi cities and is in line with previous reports on AF ’ s inability to predict the effect of mutations (Buel and Walters, 2022).”

“To mutate the key residues to the ones with opposite physico-chemical properties, we substituted one or two key residues with the ones that are of the largest Miyata distance (Miyata et al , 1979) (Table S2).”

“Such interface type collections can be used to predict occurrences of known interface types in protein interactions (Weatheritt et al, 2012; Mosca et al, 2013).”

“PNKP has dual phosphatase and kinase activity that it uses to modify broken DNA strands preparing them for repair (Aceytuno et al , 2017; Jilani et al , 1999).”

“Peroxisomes mediate these functions via peroxisomal membrane proteins (PMPs) and proteins in the peroxisomal matrix (Islinger et al , 2018).”

“HCFC1 is a transcription factor that associates with other transcription factors (Lu et al , 1997), splice factors (Ajuh et al , 2002), and cell cycle regulators (Freiman & Herr, 1997; Machida et al , 2009).”

“While in-cell XL-MS is a very elegant approach to obtain experimental information on PPI interfaces in unperturbed settings, it is still a method that is only accessible to few experts in the fi eld.”

“To verify that they indeed do not have structural homologues in the PDB, we both used the SIFTS mapping (Dana et al , 2019) between the Pfam domain in ELM and the PDB and also looked at the ELM classes that were listed as homologs on the ELM website.”

“As some domain types can bind to motifs from distinct ELM classes, we manually checked that the randomized pairings did not coincide with actual domain-motif interface types (Dataset EV2).”

“The last decade has seen tremendous progress in the systematic mapping of human protein interactions enabling gene function prediction and the study of genotype-to-phenotype relationships (Luck et al, 2020; Drew et al, 2017; Huttlin et al, 2021).”

“HCFC1 is a transcription factor that associates with other transcription factors (Lu et al , 1997), splice factors (Ajuh et al , 2002), and cell cycle regulators (Freiman & Herr, 1997; Machida et al , 2009).”

“A list of NDD genes was assembled using whole exome and whole genome sequencing studies of cohorts of NDD patients from Gene4Denovo (Zhao et al, 2020) and Deciphering Developmental Disorders (DDD) study (Firth et al, 2011), respectively.”

“TRIM37 is a RING domain-containing E3 ligase known to regulate centriole reduplication, transcriptional repression, and peroxisome import (Kallijärvi et al , 2005; Bhatnagar et al , 2014; Wang et al , 2017).”

“Various studies have highlighted that fl anking sequences of motifs can in fl uence binding af fi nities and speci fi cities (Luck et al, 2012; Bugge et al, 2020).”

“The following proteins were also tested as C-terminal fusions: CSNK2B-NL, ESRRG-NL, CUL3-NL, PEX3-NL, PEX19-NL, PSMC5-NL, PEX3-mCit, PEX19-mCit, PEX16-mCit, RORB-mCit, ESRRG-mCit, PAX6-mCit, CSNK2B-mCit, PSMC5-mCit, KCTD7-mCit (Dataset EV12).”

“Motif-mediated interactions are of moderate binding af fi nity and thus, are particularly suited to mediate dynamic cell regulatory and signaling events (Van Roey et al, 2012).”

“The following proteins were also tested as C-terminal fusions: CSNK2B-NL, ESRRG-NL, CUL3-NL, PEX3-NL, PEX19-NL, PSMC5-NL, PEX3-mCit, PEX19-mCit, PEX16-mCit, RORB-mCit, ESRRG-mCit, PAX6-mCit, CSNK2B-mCit, PSMC5-mCit, KCTD7-mCit (Dataset EV12).”

“HCFC1 is a transcription factor that associates with other transcription factors (Lu et al, 1997), splice factors (Ajuh et al, 2002), and cell cycle regulators (Freiman and Herr, 1997; Machida et al, 2009).”

“A more detailed description of the curation procedure and information on the pairs will be soon published elsewhere (Geist et al, in preparation).”

“Various studies have highlighted that fl anking sequences of motifs can in fl uence binding af fi nities and speci fi cities (Luck et al, 2012; Bugge et al, 2020).”

“For donor saturation experiments various donor DNA amounts (1, 2, 4 and 8 ng) encoding NL-fused proteins were co-transfected with increasing amounts of acceptor DNA (12.5, 25, 50, 100, 200 ng) encoding mCitrine-fused proteins.”

“Key conserved residues of the motifs in the DMI benchmark dataset were identi fi ed computationally using the regular expression of the corresponding ELM class in the ELM DB and SLiMSearch (Krystkowiak and Davey, 2017).”

“Unfortunately, no assertion criteria for the annotation of this mutation to be pathogenic is provided by ClinVar meaning that the mutation is either not pathogenic after all or its pathogenicity is mediated via another perturbed function not tested in this study.”

“However, longer disordered fragments or fragments containing ordered and large disordered regions generally decrease AF prediction accuracies as also reported in a recent preprint (preprint:Bret et al , 2023).”

“TRIM37 is a RING domain-containing E3 ligase known to regulate centriole reduplication, transcriptional repression, and peroxisome import (Kallijärvi et al , 2005; Bhatnagar et al , 2014; Wang et al , 2017).”

“The complex between STX1A, VAMP2, and SNAP25 has been solved structurally and shows a 4-helix bundle with a 1:1:2 stoichiometry, respectively (PDB:1N7S (Ernst & Brunger, 2003)).”

“These contacts are also observed in solved structures of the proteasome (i.e. PDB:5VFT (Zhu et al , 2018)).”

“Interestingly, this hydrophobic pocket does not exist in the previously solved structure of the GYF domain of CDBP2 binding to a proline-rich peptide that is fl anked by positively charged residues establishing important contacts with the domain (PDB:1L2Z, (Freund et al, 2002)).”

“TRIM37 is a RING domain-containing E3 ligase known to regulate centriole reduplication, transcriptional repression, and peroxisome import (Kallijärvi et al , 2005; Bhatnagar et al , 2014; Wang et al , 2017).”

“This retest rate is actually higher compared to retest rates of gold standard PPI datasets used in the past to benchmark various binary PPI assays including this BRET assay, attesting the overall detectability of PPIs from HuRI (Braun et al, 2009; Trepte et al, 2018; Choi et al, 2019).”

“We searched for functional annotations and existing structures for the monomers using the PDB, ProViz (Jehl et al , 2016) , SMART (Letunic et al , 2021), and the scientific literature to identify enzymatic pockets or binding interfaces for DNA, RNA, or metal ions.”

“From Gene4Denovo, we selected genes linked to autism-spectrum disorders (ASD), intellectual disability (ID), epilepsy (EE), undiagnosed developmental disorders (UDD) and NDDs in general.”

“…two distinct interfaces between PEX19 and PEX3, one of which, believed to be the dominant one, was crystallized and is described in the ELM DB as an interaction between an N-terminal motif in PEX19 that binds to the cytosolic alpha-helical domain of PEX3 (PDB:3MK4, (Schmidt et al , 2010)).”

“PNKP has dual phosphatase and kinase activity that it uses to modify broken DNA strands preparing them for repair (Aceytuno et al , 2017; Jilani et al , 1999).”

“Mutations in any of the three PEX proteins are associated with severe developmental phenotypes referred to as peroxisome biogenesis disorders (Fujiki et al, 2022).”

“In particular, the motif ranging from 110-116 attracted our attention because phosphorylation of S114 has been reported to stabilize PNKP protein levels (Parsons et al , 2012).”

“The BRET assay used here has been shown in previous studies to be sensitive enough to quantify weakening of binding introduced by point mutations and to detect motif-mediated PPIs (Ebersberger et al , 2023; Trepte et al , 2018; Mo et al , 2022).”

“This retest rate is actually higher compared to retest rates of gold standard PPI datasets used in the past to benchmark various binary PPI assays including this BRET assay, attesting the overall detectability of PPIs from HuRI (Braun et al, 2009; Trepte et al, 2018; Choi et al, 2019).”

“Such interface type collections can be used to predict occurrences of known interface types in protein interactions (Weatheritt et al, 2012; Mosca et al, 2013).”

“This disordered region is particularly rich in prolines and glycines and has been shown previously to bind to the OCRE domain of RBM5 (Mourão et al , 2016).”

“To select a subset of these motif classes, where we can do experimental testing, we also used the InParanoid resource (Persson & Sonnhammer, 2023) to map ELM instances where both proteins are from mouse to their human orthologs.”

“Titration curves were fi tted using the leastsq function from the scipy.optimize python package (Virtanen et al, 2020) using the model BRET = ((A/D) * BRETmax)/(BRET50 + (A/D)) described in (Drinovec et al, 2012), which assumes a 1:1 binding mode, to obtain estimates for the BRETmax and BRET50.”

“Using minimal interacting regions for interface prediction we reached sensitivities of up to 80% similar to previously published work (Tsaban et al, 2022; Johansson-Åkhe et al, 2021).”

“…fi dence (preprint:Evans et al, 2021), average interface residue pLDDT, average motif interface residue pLDDT, pDockQ (Bryant et al, 2022), and iPAE (Teufel et al, 2023) discriminated well between both reference datasets when rando-mizing domain-motif pairs or introducing two motif mutations (max…”

“We also evaluated whether the top 5 reported models by AF tend to be more similar to each other when corresponding to a correct structural model (Pozzati et al, 2022) and found that this feature has moderate predictive power (Appendix Fig.”

“It has been shown previously that PSMC5 is recruited to promoters of nucleic acid receptor target genes and to interact with the nucleic acid receptor RARA thereby regulating transcription (Ferry et al , 2009).”

“HCFC1 is a transcription factor that associates with other transcription factors (Lu et al, 1997), splice factors (Ajuh et al, 2002), and cell cycle regulators (Freiman and Herr, 1997; Machida et al, 2009).”

“Further increases in retest rate can be obtained by testing all possible combinations of N-and C-terminal fusions to the NanoLuc luciferase and mCitrine (preprint:Trepte et al , 2021; preprint:Trepte et al , 2023).”

“AF interface predictions for individual PPIs have occasionally been experimentally corroborated (Mishra et al, 2023; Bronkhorst et al, 2023).”

“Further increases in retest rate can be obtained by testing all possible combinations of N-and C-terminal fusions to the NanoLuc luciferase and mCitrine (preprint:Trepte et al , 2021; preprint:Trepte et al , 2023).”

“Many of these can probably be eliminated by increased sampling 67 or segmenting proteins 31,68 .”

“Some of these can probably be eliminated by folding pairs in all 5 models, increased sampling 74 , or segmenting proteins 75,76 .”

“In a comprehensive study, Lee et al. found shorter fragments of peptides binding to protein domains often scored better than longer fragments or full-length proteins (15).”

“AF2sc values indicate the best AF2 con fi dencescore values obtained foreach complex. correctly delineate and cut the intrinsically disordered regions have also proven useful in increasing the success rate of predictions 46 .”

“This phenomenon has been reported in single-structure prediction, where small protein fragments obtain about 5% higher sensitivity and specificity than full-length proteins (Lee et al. 2024).”

“The rapidly evolving AI tools may help to address this challenge 153,272 .”

“The peptides were predicted with three different lengths: only the motif region, adding flanking regions of same length as the motif, or adding flanking regions twice the motif length (Lee et al. 2024).”

“Although complex prediction has not yet reached experimental accuracy and still exhibits biases 27,28 , these models can generate strong initial hypotheses; however, large-scale dry-lab structural characterisation of protein interactions remains intractable due to experimental cost and time.”

“In this study, researchers employed AlphaFold to generate predicted protein complex structures and integrated these predictions with bio-physical experiments, such as surface plasmon resonance and yeast two-hybrid, to validate the interactions [189].”

“AlphaFold prediction of crosslinking: AlphaFold2 multimer (version 2.3) (Evans et al, 2021; Jumper et al, 2021) to model the interaction between TDP-43 and TAU using a fragmented approach (Lee et al, 2024).”

“We analyzed the predicted structural features of C-terminal FtsZ fragment interactions with divisome proteins.”

“Candidates with model confidence ≥ 0.7 and average motif pLDDT ≥ 70 were classified as high confidence 41 .”

“…spawning hundreds of subsequent papers including high-throughput structure prediction [3, 4, 5], ensemble sampling [6, 7, 8], molecular docking [9, 10, 11, 12], stability prediction [13, 14, 15, 16], further method development [17, 18, 19, 20], and a host of other applied studies [21, 22, 23, 24].”

“Finally, even the most accurate protein structure prediction models are also limited in their ability to model proteins whose conformation is dynamic, e.g., in response to a switch between the cofactor-free apo-and the cofactor-bound holo-states.”

“In addition, us-ing AlphaFold-Multimer for a systematic screen of potential interactors faces challenges of multiple testing as its false positive rate reaches 20% in some cases [38 ].”

“Coming from a different angle, various new approaches have been developed to make use of AlphaFold2 Multimer’s ability in the prediction of interactions between proteins and peptides or protein fragments [34,35].”

“AF structures provide insight into protein–protein interaction networks and their interfaces (Lee et al., 2024).”

“Although AFM can model such transient complexes with high local accuracy, the presence of extensive non-interacting flexible segments can reduce overall confidence scores, pushing them below commonly accepted thresholds.”

“Beyond adopting PROTAC-derived machine learning models for ligand prediction, linker design, and ADMET profiling, emerging tools such as AlphaFold-Multimer show high accuracy in predicting protein — protein and multicomponent complex structures 187 — 189 .”

“To this end, peptides encompassing the DDVF motif were in silico-folded with a fragment from the human RSK2 N-terminal kinase domain [21].”

“…and physicochemical features of proteins in specific areas, such as protein stability [4], conformational heterogeneity [5], protein function [6], structure, [7] or interfaces [8] of protein complexes, domain definitions [9], or disordered regions that do not fold into unique structures [10].”

“Previous studies reported significant decreases in the sensitivity of AlphaFold for complex structure predictions when using full-length sequences [78].”

“Looking ahead, integrating this framework with models of sequence variation [32, 65], protein–protein interactions [66, 67], and functional constraints (e.g., allosteric transitions and ligand binding) [68–70] A. Entropic Measures —”

“…complex prediction by AlphaFold-Multimer may be enhanced, e.g. by testing different domains or fragments of the interacting proteins (Bret et al. 2024, Lee et al. 2024), or by enforcing a dropout at inference time to increase sampling of different conformations (Johansson-Åkhe and Wallner 2022).”

“…simulations enables in silico exploration of vancomycin scaffold modifi - cations and prediction of their impact on binding energetics, guiding the design of next-generation derivatives intended to overcome D-Ala-D-Lac–mediated resistance (Zhu et al. 2024; Talat and Khan 2023 ; Lee et al. 2024 ).”

“Subsequent benchmarking by Lee et al. provided a systematic comparison of different AAF-based metrics for SLiM–domain interactions [5].”

“This information can be used to delimit the protein sequences for AlphaFold-Multimer modeling, which enhances the accuracy of the structure [7,8].”

“Third, AFM ’ s success rate is lower when predicting full-length proteins compared to protein fragments, such as interacting domains 42 .”

“This capability enables proteome-scale computational screening for PPIs [136,137], and its strength lies in synergy with experimental validation.”

“Modeling the shortest region required for an interaction has been shown to be effective and give higher confidence models (Lee et al., 2024).”

“Modeling domains and peptide fragments rather than full-length proteins have been shown to improve prediction accuracy in AlphaFold-based approaches 58,59 .”

“AlphaFold has been shown to be able to predict novel interface types which could have potential in extending definitions for existing interface types (Lee et al. 2024).”

“To identify highly confident structural models, we applied a model confidence cutoff of 0.8 for which it was previously estimated that the false and true positive rate equaled zero and 31%, respectively (Lee et al. 2024).”

“…also been successfully applied to study domain–motif interactions in the context of neurodevelopmental disorders, although predictive accuracy was shown to decrease when full-length proteins containing the motif were used as input, highlighting the need for tailored fragmentation strategies [27].”

“Indeed, not only AlphaFold performs poorly on proteins lacking related sequences and on disordered regions, but AF-Multimer predictions of domain-motif interactions are not specific when using small sequences 72 .”

“They require protein fragmentation strategies, as AF-Multimer largely fails to identify these interfaces when full-length sequences are used [81].”

“While ipTM has been applied in docking studies involving peptide-protein complexes (Bret et al., 2024; Lee et al., 2024; Teufel et al., 2023), it was originally trained on structured PDB entries and is sensitive to sequence length, treating all residues equally regardless of their involvement in…”

“Because accurate prediction of novel peptide-protein interactions remains a challenging regime for structure prediction programs, including AlphaFold,(24, 25) we examined the Cia1-Cia2 structure to identify any overlooked features that could be serving as a structural template for the AlphaFold prediction.”

“…we note that the large region of intrinsic disorder is expected to reduce AlphaFold scores and interaction metrics (Banhos Danneskiold-Samsøe et al. , 2024; Bret et al. , 2024; Homma et al. , 2024; Lee et al. , 2024; Martin, 2024; Pándy-Szekeres et al. , 2024; Dunbrack, 2025; Mitic et al. , 2025).”

“Lee et al. 20 demonstrated that including flanking residues around motifs in AlphaFold predictions could improve the accuracy of predicted models.”

“Such emergent specialization mirrors hierarchical binding principles observed in experimental PPI networks, where structured receptors recognize recurring motifs across multiple partners (Landgraf et al. 2004, Stein et al. 2009, Uversky 2016).”

“Nevertheless, successful strategies have been developed to increase the accuracy of AF for de novo identification of SLiMs (Bret et al. 2024; Lee et al. 2024; Veinstein et al. 2025).”

“Additionally, the interactions appear to be motif-driven, which has been reported to present challenges in AlphaFold-Multimer predictions [28] To overcome this limitation, we incorporated experimentally derived XL-MS restraints and performed integrative modeling using AlphaLink2, a method that…”

“To investigate how SLP2 target the AQ12 silk domain, we computationally mapped its putative cleavage sites within the AQ12 sequence using an AlphaFold2-based fragment docking approach evaluated with the actifpTM confidence metric (18, 19) (see Methods).”

“Similar behavior has been noted in comparative evaluations of deep-learning structure predictors versus disorder-centric binding models (Basu et al., 2023; Lee et al., 2024; Majila et al., 2026).”

“This limitation could be particularly critical for protein-peptide complexes where the peptide contains disordered regions [46, 47].”

“For most tested cases, AF3Complex colocalized the isolated SLiM to its cognate PRD within the full - length receptor; however, consistent with previous work (26), this was not the case when the full - length SLiM - containing protein was used.”

“In addition, we aligned M. musculus and R. norvegicus proteins with their H. sapiens orthologs (sharing the same gene name) and transferred the motif boundaries from orthologous sequences onto the human counterparts (STable 2).”

“Protein fragmentation for AlphaFold predictions has been shown to improve sensitivity but it can also reduce specificity ( 32 ).”

“The experimental data were used as constraints to achieve the prediction of protein structure [52,53].”

“Notably, AlphaFold 3’s multimer mode, when used with peptide fragments, has shown greater success in modeling intrinsically disordered regions of protein ligands compared to full polypeptide ligands [36].”