PPIscreenML is a method for structure-based screening of protein-protein interactions using AlphaFold

Cellular homeostasis is maintained by a complex protein interaction network that contributes to intricate signaling pathways within the cell (Choi et al., 2019). As the human interactome is estimated to contain between 74,000 and 200,000 protein interactions (Venkatesan et al., 2009), accurate and high-throughput methods to identify these interactions are crucial (Elhabashy et al., 2022). Classically, discovery and cataloging of protein-protein interactions (PPIs) have most commonly employed yeast two-hybrid assays (Fields and Song, 1989) and affinity purification assays (Dunham et al., 2012). More recently, proximity labeling techniques such as BioID (Roux et al., 2018) and TurboID (Cho et al., 2020) have gained popularity. However, each method has its own specific shortcomings, which are further compounded due to sensitive dependence on experimental techniques and cellular conditions (Choi et al., 2019; Qin et al., 2021). Collectively, the variabilities in these approaches ultimately lead to high rates of both false positives and false negatives (Qin et al., 2021; Deane et al., 2002). For example, statistical analysis of representative interactome mapping data showed that recovering 65% of true protein interactions would require combining 10 distinct assays (Choi et al., 2019).

Computational predictions offer a natural means for complementing these experimental techniques. Early methods focused on inferring PPIs based on sequence information, by searching for gene pairs with correlated expression patterns, or distinct genes that can be found fused together in a different organism, or distinct genes that co-evolve with one another (Marcotte et al., 1999). Drawing upon the exciting successes of large language models in many contexts, deep learning methods have also been applied to infer binary PPI pairings using protein sequences (Zhao et al., 2022; Jha et al., 2023; Du et al., 2017). Results from such methods are extremely promising; however, their lack of explainability makes it difficult to determine how such models might perform on PPIs that are dissimilar from the training examples (i.e. their domain of applicability). Moreover, these models are inherently not designed to provide structural information about the bound pose: this information can be extremely valuable, both to provide biological understanding that spurs further experimentation and as a starting point for developing new therapeutic agents that modulate the PPI (Lu et al., 2020).

Broad availability of AlphaFold2 (AF2) (Jumper et al., 2021), especially through the ColabFold framework (Mirdita et al., 2022), has provided the research community with unprecedented access to structural information for monomeric proteins, in many cases at resolution comparable to experimentally determined structures. The startling accuracy of certain AF2 predictions, in turn, enabled structure-based modeling for numerous new applications in biology and medicine (Akdel et al., 2022; Yang et al., 2023). Whereas the original AF2 could be adapted for highly successful predictions of oligomeric assemblies, a subsequent retrained release dubbed AF2-Multimer provided further enhanced performance in this regime (Evans et al., 2021).

While AF2 provides often-accurate predictions for interacting protein pairs, there is not a direct measure to evaluate the likelihood that a given query pair indeed interacts with one another – a set of predicted output structures will be returned regardless. AF2 also provides residue-level confidence measures of the predicted structures, specifically the predicted local-distance difference test (pLDDT) (Mariani et al., 2013), the predicted TM-score (pTM) (Zhang and Skolnick, 2005), and the predicted aligned error (PAE). The accuracy of structures for predicted protein complexes is often defined using DockQ (Basu and Wallner, 2016), a metric that combines three complementary and frequently used quality measures. By combining the number of interface residues with the pLDDT values of interface residues, a confidence measure pDockQ was developed to define the likelihood of an accurate binding mode in protein complexes predicted by AF2 (Bryant et al., 2022). Interestingly, pDockQ also provided some ability to discriminate between ‘true’ (interacting) and ‘decoy’ (noninteracting) protein pairs (Bryant et al., 2022) – performance for this task was only modest, however, presumably because development of pDockQ was fundamentally focused elsewhere (recapitulating DockQ values).

In light of interest in using AF2 to distinguish between interacting and noninteracting protein pairs, the AlphaPulldown (Yu et al., 2023) Python package calculates several measures that have been proposed for this task: iPTM (Evans et al., 2021), PIscore (Malhotra et al., 2021), and pDockQ (Bryant et al., 2022). While each of these measures is intuitively reasonable, none were expressly developed for this classification task.

Here, we describe PPIscreenML, a machine learning classifier that uses AF2 models to distinguish interacting versus noninteracting protein pairs. To do so, PPIscreenML uses descriptors drawn from AF2 confidence measures (Jumper et al., 2021) and complements these with energetic terms from the Rosetta scoring function (Leman et al., 2020). In contrast to methods such as pDockQ, which was trained as a means of evaluating model quality, PPIscreenML is deliberately and intentionally trained/tested to distinguish between models of interacting proteins and compelling decoy pairs. PPIscreenML exhibits superior performance to other methods for identifying interacting pairs in a retrospective screen. Finally, using the tumor necrosis factor superfamily (TNFSF) as an example of a structurally conserved family of ligand/receptor pairings, we demonstrate that PPIscreenML can accurately recapitulate the selectivity profile in this challenging regime.

Computational approach

Dataset for training/testing PPIscreenML

Critical to development of a useful classification model is compiling a set of ‘active’ and ‘decoy’ examples that accurately reflect those that will be encountered in future prospective applications. For PPIscreenML, we employed an overarching strategy that mirrors our previous studies leading to a machine learning classifier that identifies active protein/small molecule complexes likely to engage one another (Adeshina et al., 2020). Specifically, our approach seeks to build highly compelling decoy complexes, so that the resulting model is challenged to distinguish actives from decoys that cannot be resolved using solely trivial criteria.

To build relevant and diverse structures of active protein pairings (Figure 1A), we used the DockGround database (Collins et al., 2022) to assemble the complete set of heterodimeric protein complexes in the PDB, excluding homodimers, antibody/antigen complexes, and complexes that were either very large or very small (see Methods). Upon filtering for a maximum of 30% sequence identity, there remained 1481 nonredundant protein-protein complexes with experimentally defined structures. The planned application of PPIscreenML is to distinguish AF2 models of active pairs from AF2 models of inactive pairs – thus, the model should be trained on AF2 actives, rather than experimentally defined structures. Accordingly, we used the sequences of the component proteins to build five AF2 models for each of these 1481 nonredundant complexes.

Figure 1 with 4 supplements see all

Download asset Open asset

Active protein-protein pairs that are mis-predicted by AF2 should not lead to predictions of ‘active’ from PPIscreenML, since the specific interactions in the AF2 are not those that lead to binding. In the context of training, then, we sought to use only those models that were correctly predicted by AF2: we therefore filtered the models and retained only those with a DockQ score (Basu and Wallner, 2016) of at least 0.23 relative to the corresponding experimentally defined structure (0.23 is the quality cutoff defined to be an ‘acceptable’ prediction for the MOAL and CAPRI sets). Filtering on model quality in this manner excluded 932 models from training, leading to a total of 6473 active models with median DockQ value of 0.88 (Figure 1—figure supplement 1A).

In a prospective application of PPIscreenML, however, the DockQ score will not be known a priori (since this requires an experimentally defined structure, which itself is evidence of an interaction). To ensure a realistic estimate of the performance that can be expected in future prospective experiments, complexes with structures that were mis-predicted by AF2 were not excluded from the test set. Accordingly, active pairs that are not correctly classified by PPIscreenML may occur either because AF2 builds a poor model, or because PPIscreenML fails to recognize a well-built model as an interacting pair. While the test set performance thus reflects more than the success of PPIscreenML, this strategy enhances confidence that the observed test set performance will carry forward to prospective applications in the future.

With respect to creating a set of decoy complexes (for both training and testing), it is essential that the decoys suitably represent the types of inactive complexes expected to be encountered in prospective applications. In particular, decoys that can be dismissed as inactive based on trivial criteria offer little benefit in developing PPIscreenML: rather, we prioritized building a set of extremely compelling decoys (Adeshina et al., 2020), so that training would be impelled to find a more sophisticated and nuanced model for classification.

To build a set of compelling decoys (Figure 1A), we began by using each active complex as a template. For each component protein in the complex, we identified its closest structural analog (based on TM-score; Zhang and Skolnick, 2004) among all other proteins in our collection. Despite the fact that homologs had been excluded from our dataset on the basis of sequence identity, close structural matches could nonetheless be identified in most cases (Figure 1—figure supplement 1B): 87% of the proteins in our set were paired with a structural analog having TM-score greater than 0.5, a cutoff far beyond that which is observed between random protein pairs (Xu and Zhang, 2010). The structural analog for each component protein was then superposed onto the template in the active complex, providing a crude model of these two (presumably) noninteracting proteins that structurally resembles an active complex (Figure 1B). As noted earlier, the planned application of PPIscreenML is to distinguish AF2 models of active pairs from AF2 models of inactive pairs: as for the active pairings, we therefore used the sequences of the component proteins in each of these ‘decoy’ complexes to build five AF2 models for each.

To train PPIscreenML, data was partitioned using a 60/20/20 split. The data reserved for testing was not used in any phase of training or validation. The dataset includes five AF2 models for each (active or decoy) complex: training on one AF2 model of a given complex then using a different model of the same complex in the test set would introduce obvious information leakage. To avoid this, all five models for a given complex were together placed in either the training set or the validation set or the test set (Ong et al., 2023). As noted earlier, all five AF2 models for a given complex were included in the validation and testing sets; however, only those AF2 models of active complexes close to the experimentally derived structures were included in the training set (Figure 1—figure supplement 2).

Having partitioned the AF2 models into training, validation, and test sets, a series of quantitative features were extracted from each model. Specifically, three groups of features were calculated from each model (Figure 1C): 7 features representing AF2 confidence measures, 33 structural ‘counting’ features, and 17 features from the Rosetta energy function. This set of 57 features (Figure 1—figure supplement 3) was initially evaluated in the development of PPIscreenML.

Developing the PPIscreenML model

Using the compiled collection of features for each complex, we evaluated seven standard machine learning frameworks to build classification models. In each case, models were trained using fivefold cross-validation within the test set and applied to a validation set that was not included in each fold of training. Models were then compared using the combined cross-fold AUC score (area under the curve of a receiver operating characteristic [ROC] plot). All seven models had similar performance, with XGBoost (an implementation of gradient-boosted decision trees) yielding the highest AUC value of 0.924 (Figure 2—figure supplement 1).

The model was then applied to the test set, which had been completely held out from model training and evaluation. To better recapitulate a real-world use case, this experiment incorporates two key differences from the earlier evaluation using the validation set (Figure 1—figure supplement 2). First, AF2 models of active complexes with poor DockQ scores relative to the corresponding PDB structure are included in the test set, whereas these had been excluded from the validation set. This reflects the reality that active complexes will occasionally be mis-docked by AF2, and these models cannot be discarded in a true prospective application. Second, rather than evaluating performance using the scores of individual AF2 models, the top-ranked model for each complex is drawn from the scores of all five AF2 models. Here too, in a prospective screening application, one would mitigate the possibility that individual AF2 models may be mis-docked by simply using the model that PPIscreenML finds most compelling. Applying the XGBoost model to the held-out test set with these conditions yielded an AUC value of 0.892 (Figure 2A).

Figure 2 with 4 supplements see all

Download asset Open asset

To mitigate any potential overtraining, sequential backward selection was applied to reduce the number of features, using the validation set (not the test set, which is not used in any model optimization). This approach provided a model with only seven features that maintained the AUC of the model with all features when evaluated using the validation set (Figure 2—figure supplement 2). Specifically, this model uses one feature from the AF2 confidence measures (the average of the top ¼ of the interfacial PAE values), one structural ‘counting’ feature (number of charged amino acids in the interface), and five features from the Rosetta energy function (the average Lennard-Jones attractive score for interfacial residues, the average Lennard-Jones repulsive score for interfacial residues, the average solvation score for interfacial residues, the average electrostatic score for interfacial residues, and the proportion of interfacial residues involved in beta-sheet structures) (Figure 2B).

When applied to the test set that was completely held out from model training and optimization, this 7-feature PPIscreenML model yields an AUC of 0.884, comparable to the full 57-feature model (Figure 2C).

The availability of AF2 and related methods (Jumper et al., 2021; Mirdita et al., 2022; Evans et al., 2021; Baek et al., 2021) has enabled structure-based analysis of proteins and protein complexes to an extent that has rapidly transformed many parts of biology and medicine (Akdel et al., 2022; Yang et al., 2023). Numerous studies have incorporated AF2 models to justify or rationalize the structural basis for a particular PPI, and in many cases, use either iPTM or pDockQ as further support for the relevance of the proposed interaction. Such studies – which range from predicting protein quaternary structures to defining biological functions of peptides and intrinsically disordered regions (Lim et al., 2023; Bret et al., 2024; Danneskiold-Samsøe et al., 2023; Marciano et al., 2022; Martin, 2024; Teufel et al., 2023; Pogozheva et al., 2023) – highlight the value of using AF2 to nominate new candidate PPIs (i.e. in a screening setting), extending it beyond simply providing structures for fully validated protein pairings.

While the early adoption of iPTM or pDockQ has spurred their use in this screening, neither one was intended for this use case. Rather, both were developed with the goal of assessing model quality for AF2 structures, and the ability to distinguish between active and inactive structures on the basis of model quality was a convenient secondary benefit (Evans et al., 2021; Bryant et al., 2022). Due to this history, however, neither method was rigorously benchmarked in a manner that would set expectations for their performance in real-world applications.

To lay the groundwork for this study, we began by building a set of nonredundant active complexes and, from these, a set of highly compelling decoy complexes. The overarching philosophy behind the construction of this set was to realistically mimic conditions that would be encountered in future (prospective) screening applications. Notably, distinguishing actives from decoys in this set is markedly more challenging than in other (less realistic) benchmarks: e.g., pDockQ provided an AUC value of 0.87 in its first-reported classification experiment (Bryant et al., 2022), but only 0.71 when applied to this newer set (Figure 3A).

Developing this challenging and realistic set allowed us to train PPIscreenML, a tool expressly designed and optimized to be used in screening for PPIs. As expected, performance of PPIscreenML is superior to that of iPTM or pDockQ for this screening regime (Figure 3A). Moreover, PPIscreenML also makes use of subtle structural cues to provide its discriminative power, as evidenced from its impressive performance on the TNFSF selectivity benchmark (Figure 4).

Contemporaneously with our study, other groups have developed alternate methods for predicting pairs of interacting proteins. For example, the AF2Complex method also shows superiority over iPTM (Gao et al., 2022), albeit through testing on arbitrary collections of protein pairings (rather than on compellingly built decoys).

The natural extension of using AF2 and related methods to carry out structure-based screening for PPIs, naturally, is using these methods to define whole interactomes. Already, certain studies have applied these tools either within targeted protein families (Weeratunga et al., 2024; Baryshev et al., 2023) or to groups of proteins thought to belong together in large complexes (Gao et al., 2022; Burke et al., 2023; Humphreys et al., 2021). Importantly, careful consideration of the statistical measures needed for productively scaling assessments to whole proteomes must complement these studies; e.g., this will allow for defining thresholds that ideally balance precision and recall (Table 1), and also ensure that the expected results justify these larger-scale analyses.

There is also ample reason for optimism that the performance of PPIscreenML will further improve as the underlying AF2 models improve. Specifically, older versions of the AF2 software yielded slightly worse performance from PPIscreenML (Figure 2—figure supplement 3, Figure 2—figure supplement 4), but the difference in performance could be attributed to the fact that a (slightly) smaller proportion of the active complexes were correctly built. Thus, the true bottleneck in extending PPIscreenML may not lie with distinguishing correct versus incorrect interfaces. Rather, screening for interactions at whole-proteome scale may currently be instead limited by the fact that not all protein pairs are correctly modeled by AF2, and active pairings cannot be reliably identified if their structures are incorrectly built.

All protein structures, code, and machine learning models described in this paper are available on GitHub (https://github.com/victoria-mischley/PPIScreenML copy archived at Mischley, 2025).

1. 1. Adeshina YO
   2. Deeds EJ
   3. Karanicolas J

   (2020) Machine learning classification can reduce false positives in structure-based virtual screening

   PNAS 117:18477–18488.

   https://doi.org/10.1073/pnas.2000585117

   • PubMed
   • Google Scholar
2. 1. Akdel M
   2. Pires DEV
   3. Pardo EP
   4. Jänes J
   5. Zalevsky AO
   6. Mészáros B
   7. Bryant P
   8. Good LL
   9. Laskowski RA
   10. Pozzati G
   11. Shenoy A
   12. Zhu W
   13. Kundrotas P
   14. Serra VR
   15. Rodrigues CHM
   16. Dunham AS
   17. Burke D
   18. Borkakoti N
   19. Velankar S
   20. Frost A
   21. Basquin J
   22. Lindorff-Larsen K
   23. Bateman A
   24. Kajava AV
   25. Valencia A
   26. Ovchinnikov S
   27. Durairaj J
   28. Ascher DB
   29. Thornton JM
   30. Davey NE
   31. Stein A
   32. Elofsson A
   33. Croll TI
   34. Beltrao P

   (2022) A structural biology community assessment of AlphaFold2 applications

   Nature Structural & Molecular Biology 29:1056–1067.

   https://doi.org/10.1038/s41594-022-00849-w

   • PubMed
   • Google Scholar
3. 1. Baek M
   2. DiMaio F
   3. Anishchenko I
   4. Dauparas J
   5. Ovchinnikov S
   6. Lee GR
   7. Wang J
   8. Cong Q
   9. Kinch LN
   10. Schaeffer RD
   11. Millán C
   12. Park H
   13. Adams C
   14. Glassman CR
   15. DeGiovanni A
   16. Pereira JH
   17. Rodrigues AV
   18. van Dijk AA
   19. Ebrecht AC
   20. Opperman DJ
   21. Sagmeister T
   22. Buhlheller C
   23. Pavkov-Keller T
   24. Rathinaswamy MK
   25. Dalwadi U
   26. Yip CK
   27. Burke JE
   28. Garcia KC
   29. Grishin NV
   30. Adams PD
   31. Read RJ
   32. Baker D

   (2021) Accurate prediction of protein structures and interactions using a three-track neural network

   Science 373:871–876.

   https://doi.org/10.1126/science.abj8754

   • PubMed
   • Google Scholar
4. Preprint

   1. Baryshev A
   2. La Fleur A
   3. Groves B
   4. Michel C
   5. Baker D
   6. Ljubetič A
   7. Seelig G

   (2023) Massively parallel protein-protein interaction measurement by sequencing (MP3-Seq) enables rapid screening of protein heterodimers

   bioRxiv.

   https://doi.org/10.1101/2023.02.08.527770

   • Google Scholar
5. 1. Basu S
   2. Wallner B

   (2016) DockQ: A quality measure for protein-protein docking models

   PLOS ONE 11:e0161879.

   https://doi.org/10.1371/journal.pone.0161879

   • PubMed
   • Google Scholar
6. 1. Bossen C
   2. Ingold K
   3. Tardivel A
   4. Bodmer JL
   5. Gaide O
   6. Hertig S
   7. Ambrose C
   8. Tschopp J
   9. Schneider P

   (2006) Interactions of tumor necrosis factor (TNF) and TNF receptor family members in the mouse and human

   The Journal of Biological Chemistry 281:13964–13971.

   https://doi.org/10.1074/jbc.M601553200

   • PubMed
   • Google Scholar
7. 1. Bret H
   2. Gao J
   3. Zea DJ
   4. Andreani J
   5. Guerois R

   (2024) From interaction networks to interfaces, scanning intrinsically disordered regions using AlphaFold2

   Nature Communications 15:597.

   https://doi.org/10.1038/s41467-023-44288-7

   • PubMed
   • Google Scholar
8. 1. Bryant P
   2. Pozzati G
   3. Elofsson A

   (2022) Improved prediction of protein-protein interactions using AlphaFold2

   Nature Communications 13:1265.

   https://doi.org/10.1038/s41467-022-28865-w

   • PubMed
   • Google Scholar
9. 1. Burke DF
   2. Bryant P
   3. Barrio-Hernandez I
   4. Memon D
   5. Pozzati G
   6. Shenoy A
   7. Zhu W
   8. Dunham AS
   9. Albanese P
   10. Keller A
   11. Scheltema RA
   12. Bruce JE
   13. Leitner A
   14. Kundrotas P
   15. Beltrao P
   16. Elofsson A

   (2023) Towards a structurally resolved human protein interaction network

   Nature Structural & Molecular Biology 30:216–225.

   https://doi.org/10.1038/s41594-022-00910-8

   • PubMed
   • Google Scholar
10. 1. Burley SK
    2. Bhikadiya C
    3. Bi C
    4. Bittrich S
    5. Chao H
    6. Chen L
    7. Craig PA
    8. Crichlow GV
    9. Dalenberg K
    10. Duarte JM
    11. Dutta S
    12. Fayazi M
    13. Feng Z
    14. Flatt JW
    15. Ganesan S
    16. Ghosh S
    17. Goodsell DS
    18. Green RK
    19. Guranovic V
    20. Henry J
    21. Hudson BP
    22. Khokhriakov I
    23. Lawson CL
    24. Liang Y
    25. Lowe R
    26. Peisach E
    27. Persikova I
    28. Piehl DW
    29. Rose Y
    30. Sali A
    31. Segura J
    32. Sekharan M
    33. Shao C
    34. Vallat B
    35. Voigt M
    36. Webb B
    37. Westbrook JD
    38. Whetstone S
    39. Young JY
    40. Zalevsky A
    41. Zardecki C

    (2023) RCSB Protein Data Bank (RCSB.org): delivery of experimentally-determined PDB structures alongside one million computed structure models of proteins from artificial intelligence/machine learning

    Nucleic Acids Research 51:D488–D508.

    https://doi.org/10.1093/nar/gkac1077

    • Google Scholar
11. Conference

    1. Chen T
    2. Guestrin C

    (2016) XGBoost

    Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining.

    https://doi.org/10.1145/2939672.2939785

    • Google Scholar
12. 1. Cho KF
    2. Branon TC
    3. Udeshi ND
    4. Myers SA
    5. Carr SA
    6. Ting AY

    (2020) Proximity labeling in mammalian cells with TurboID and split-TurboID

    Nature Protocols 15:3971–3999.

    https://doi.org/10.1038/s41596-020-0399-0

    • PubMed
    • Google Scholar
13. 1. Choi SG
    2. Olivet J
    3. Cassonnet P
    4. Vidalain P-O
    5. Luck K
    6. Lambourne L
    7. Spirohn K
    8. Lemmens I
    9. Dos Santos M
    10. Demeret C
    11. Jones L
    12. Rangarajan S
    13. Bian W
    14. Coutant EP
    15. Janin YL
    16. van der Werf S
    17. Trepte P
    18. Wanker EE
    19. De Las Rivas J
    20. Tavernier J
    21. Twizere J-C
    22. Hao T
    23. Hill DE
    24. Vidal M
    25. Calderwood MA
    26. Jacob Y

    (2019) Maximizing binary interactome mapping with a minimal number of assays

    Nature Communications 10:3907.

    https://doi.org/10.1038/s41467-019-11809-2

    • PubMed
    • Google Scholar
14. 1. Collins KW
    2. Copeland MM
    3. Kotthoff I
    4. Singh A
    5. Kundrotas PJ
    6. Vakser IA

    (2022) Dockground resource for protein recognition studies

    Protein Science 31:e4481.

    https://doi.org/10.1002/pro.4481

    • PubMed
    • Google Scholar
15. Preprint

    1. Danneskiold-Samsøe NB
    2. Kavi D
    3. Jude KM
    4. Nissen SB
    5. Wat LW
    6. Coassolo L
    7. Zhao M
    8. Santana-Oikawa GA
    9. Broido BB
    10. Garcia KC
    11. Svensson KJ

    (2023) AlphaFold2 enables accurate deorphanization of ligands to single-pass receptors

    bioRxiv.

    https://doi.org/10.1101/2023.03.16.531341

    • Google Scholar
16. 1. Deane CM
    2. Salwiński Ł
    3. Xenarios I
    4. Eisenberg D

    (2002) Protein interactions: two methods for assessment of the reliability of high throughput observations

    Molecular & Cellular Proteomics 1:349–356.

    https://doi.org/10.1074/mcp.m100037-mcp200

    • PubMed
    • Google Scholar
17. 1. Du X
    2. Sun S
    3. Hu C
    4. Yao Y
    5. Yan Y
    6. Zhang Y

    (2017) DeepPPI: Boosting prediction of protein-protein interactions with deep neural networks

    Journal of Chemical Information and Modeling 57:1499–1510.

    https://doi.org/10.1021/acs.jcim.7b00028

    • PubMed
    • Google Scholar
18. 1. Dunham WH
    2. Mullin M
    3. Gingras AC

    (2012) Affinity-purification coupled to mass spectrometry: basic principles and strategies

    Proteomics 12:1576–1590.

    https://doi.org/10.1002/pmic.201100523

    • PubMed
    • Google Scholar
19. 1. Elhabashy H
    2. Merino F
    3. Alva V
    4. Kohlbacher O
    5. Lupas AN

    (2022) Exploring protein-protein interactions at the proteome level

    Structure 30:462–475.

    https://doi.org/10.1016/j.str.2022.02.004

    • PubMed
    • Google Scholar
20. Preprint

    1. Evans R
    2. O’Neill M
    3. Pritzel A
    4. Antropova N
    5. Senior A
    6. Green T
    7. Žídek A
    8. Bates R
    9. Blackwell S
    10. Yim J
    11. Ronneberger O
    12. Bodenstein S
    13. Zielinski M
    14. Bridgland A
    15. Potapenko A
    16. Cowie A
    17. Tunyasuvunakool K
    18. Jain R
    19. Clancy E
    20. Kohli P
    21. Jumper J
    22. Hassabis D

    (2021) Protein complex prediction with alphafold-multimer

    bioRxiv.

    https://doi.org/10.1101/2021.10.04.463034

    • Google Scholar
21. 1. Fields S
    2. Song O

    (1989) A novel genetic system to detect protein-protein interactions

    Nature 340:245–246.

    https://doi.org/10.1038/340245a0

    • PubMed
    • Google Scholar
22. 1. Gao M
    2. Nakajima An D
    3. Parks JM
    4. Skolnick J

    (2022) AF2Complex predicts direct physical interactions in multimeric proteins with deep learning

    Nature Communications 13:1744.

    https://doi.org/10.1038/s41467-022-29394-2

    • PubMed
    • Google Scholar
23. 1. Humphreys IR
    2. Pei J
    3. Baek M
    4. Krishnakumar A
    5. Anishchenko I
    6. Ovchinnikov S
    7. Zhang J
    8. Ness TJ
    9. Banjade S
    10. Bagde SR
    11. Stancheva VG
    12. Li X-H
    13. Liu K
    14. Zheng Z
    15. Barrero DJ
    16. Roy U
    17. Kuper J
    18. Fernández IS
    19. Szakal B
    20. Branzei D
    21. Rizo J
    22. Kisker C
    23. Greene EC
    24. Biggins S
    25. Keeney S
    26. Miller EA
    27. Fromme JC
    28. Hendrickson TL
    29. Cong Q
    30. Baker D

    (2021) Computed structures of core eukaryotic protein complexes

    Science 374:eabm4805.

    https://doi.org/10.1126/science.abm4805

    • Google Scholar
24. 1. Jha K
    2. Karmakar S
    3. Saha S

    (2023) Graph-BERT and language model-based framework for protein-protein interaction identification

    Scientific Reports 13:5663.

    https://doi.org/10.1038/s41598-023-31612-w

    • PubMed
    • Google Scholar
25. 1. Jumper J
    2. Evans R
    3. Pritzel A
    4. Green T
    5. Figurnov M
    6. Ronneberger O
    7. Tunyasuvunakool K
    8. Bates R
    9. Žídek A
    10. Potapenko A
    11. Bridgland A
    12. Meyer C
    13. Kohl SAA
    14. Ballard AJ
    15. Cowie A
    16. Romera-Paredes B
    17. Nikolov S
    18. Jain R
    19. Adler J
    20. Back T
    21. Petersen S
    22. Reiman D
    23. Clancy E
    24. Zielinski M
    25. Steinegger M
    26. Pacholska M
    27. Berghammer T
    28. Bodenstein S
    29. Silver D
    30. Vinyals O
    31. Senior AW
    32. Kavukcuoglu K
    33. Kohli P
    34. Hassabis D

    (2021) Highly accurate protein structure prediction with AlphaFold

    Nature 596:583–589.

    https://doi.org/10.1038/s41586-021-03819-2

    • PubMed
    • Google Scholar
26. 1. Leman JK
    2. Weitzner BD
    3. Lewis SM
    4. Adolf-Bryfogle J
    5. Alam N
    6. Alford RF
    7. Aprahamian M
    8. Baker D
    9. Barlow KA
    10. Barth P
    11. Basanta B
    12. Bender BJ
    13. Blacklock K
    14. Bonet J
    15. Boyken SE
    16. Bradley P
    17. Bystroff C
    18. Conway P
    19. Cooper S
    20. Correia BE
    21. Coventry B
    22. Das R
    23. De Jong RM
    24. DiMaio F
    25. Dsilva L
    26. Dunbrack R
    27. Ford AS
    28. Frenz B
    29. Fu DY
    30. Geniesse C
    31. Goldschmidt L
    32. Gowthaman R
    33. Gray JJ
    34. Gront D
    35. Guffy S
    36. Horowitz S
    37. Huang P-S
    38. Huber T
    39. Jacobs TM
    40. Jeliazkov JR
    41. Johnson DK
    42. Kappel K
    43. Karanicolas J
    44. Khakzad H
    45. Khar KR
    46. Khare SD
    47. Khatib F
    48. Khramushin A
    49. King IC
    50. Kleffner R
    51. Koepnick B
    52. Kortemme T
    53. Kuenze G
    54. Kuhlman B
    55. Kuroda D
    56. Labonte JW
    57. Lai JK
    58. Lapidoth G
    59. Leaver-Fay A
    60. Lindert S
    61. Linsky T
    62. London N
    63. Lubin JH
    64. Lyskov S
    65. Maguire J
    66. Malmström L
    67. Marcos E
    68. Marcu O
    69. Marze NA
    70. Meiler J
    71. Moretti R
    72. Mulligan VK
    73. Nerli S
    74. Norn C
    75. Ó’Conchúir S
    76. Ollikainen N
    77. Ovchinnikov S
    78. Pacella MS
    79. Pan X
    80. Park H
    81. Pavlovicz RE
    82. Pethe M
    83. Pierce BG
    84. Pilla KB
    85. Raveh B
    86. Renfrew PD
    87. Burman SSR
    88. Rubenstein A
    89. Sauer MF
    90. Scheck A
    91. Schief W
    92. Schueler-Furman O
    93. Sedan Y
    94. Sevy AM
    95. Sgourakis NG
    96. Shi L
    97. Siegel JB
    98. Silva D-A
    99. Smith S
    100. Song Y
    101. Stein A
    102. Szegedy M
    103. Teets FD
    104. Thyme SB
    105. Wang RY-R
    106. Watkins A
    107. Zimmerman L
    108. Bonneau R

    (2020) Macromolecular modeling and design in Rosetta: recent methods and frameworks

    Nature Methods 17:665–680.

    https://doi.org/10.1038/s41592-020-0848-2

    • PubMed
    • Google Scholar
27. 1. Lim Y
    2. Tamayo-Orrego L
    3. Schmid E
    4. Tarnauskaite Z
    5. Kochenova OV
    6. Gruar R
    7. Muramatsu S
    8. Lynch L
    9. Schlie AV
    10. Carroll PL
    11. Chistol G
    12. Reijns MAM
    13. Kanemaki MT
    14. Jackson AP
    15. Walter JC

    (2023) In silico protein interaction screening uncovers DONSON’s role in replication initiation

    Science 381:eadi3448.

    https://doi.org/10.1126/science.adi3448

    • PubMed
    • Google Scholar
28. 1. Lu H
    2. Zhou Q
    3. He J
    4. Jiang Z
    5. Peng C
    6. Tong R
    7. Shi J

    (2020) Recent advances in the development of protein-protein interactions modulators: mechanisms and clinical trials

    Signal Transduction and Targeted Therapy 5:213.

    https://doi.org/10.1038/s41392-020-00315-3

    • PubMed
    • Google Scholar
29. 1. Malhotra S
    2. Joseph AP
    3. Thiyagalingam J
    4. Topf M

    (2021) Assessment of protein-protein interfaces in cryo-EM derived assemblies

    Nature Communications 12:3399.

    https://doi.org/10.1038/s41467-021-23692-x

    • PubMed
    • Google Scholar
30. 1. Marciano S
    2. Dey D
    3. Listov D
    4. Fleishman SJ
    5. Sonn-Segev A
    6. Mertens H
    7. Busch F
    8. Kim Y
    9. Harvey SR
    10. Wysocki VH
    11. Schreiber G

    (2022) Protein quaternary structures in solution are a mixture of multiple forms

    Chemical Science 13:11680–11695.

    https://doi.org/10.1039/D2SC02794A

    • Google Scholar
31. 1. Marcotte EM
    2. Pellegrini M
    3. Thompson MJ
    4. Yeates TO
    5. Eisenberg D

    (1999) A combined algorithm for genome-wide prediction of protein function

    Nature 402:83–86.

    https://doi.org/10.1038/47048

    • PubMed
    • Google Scholar
32. 1. Mariani V
    2. Biasini M
    3. Barbato A
    4. Schwede T

    (2013) lDDT: a local superposition-free score for comparing protein structures and models using distance difference tests

    Bioinformatics 29:2722–2728.

    https://doi.org/10.1093/bioinformatics/btt473

    • PubMed
    • Google Scholar
33. 1. Martin J

    (2024) AlphaFold2 predicts whether proteins interact amidst confounding structural compatibility

    Journal of Chemical Information and Modeling 64:1473–1480.

    https://doi.org/10.1021/acs.jcim.3c01805

    • PubMed
    • Google Scholar
34. 1. Mirdita M
    2. Schütze K
    3. Moriwaki Y
    4. Heo L
    5. Ovchinnikov S
    6. Steinegger M

    (2022) ColabFold: making protein folding accessible to all

    Nature Methods 19:679–682.

    https://doi.org/10.1038/s41592-022-01488-1

    • PubMed
    • Google Scholar
35. Software

    1. Mischley V

    (2025) PPIScreenML, version swh:1:rev:1b2def1668351cc5239d4facb63bd80089206a56

    Software Heritage.

    https://archive.softwareheritage.org/swh:1:dir:7cbf9dfc5922d5aa7367639f93aa8e93215704c4;origin=https://github.com/victoria-mischley/PPIScreenML;visit=swh:1:snp:8195dcac766adff4e920376970bbff3cf43c2782;anchor=swh:1:rev:1b2def1668351cc5239d4facb63bd80089206a56
36. Preprint

    1. Ong WJG
    2. Kirubakaran P
    3. Karanicolas J

    (2023) Poor generalization by current deep learning models for predicting binding affinities of kinase inhibitors

    bioRxiv.

    https://doi.org/10.1101/2023.09.04.556234

    • Google Scholar
37. 1. Pogozheva ID
    2. Cherepanov S
    3. Park SJ
    4. Raghavan M
    5. Im W
    6. Lomize AL

    (2023) Structural modeling of cytokine-receptor-JAK2 signaling complexes using AlphaFold multimer

    Journal of Chemical Information and Modeling 63:5874–5895.

    https://doi.org/10.1021/acs.jcim.3c00926

    • PubMed
    • Google Scholar
38. 1. Qin W
    2. Cho KF
    3. Cavanagh PE
    4. Ting AY

    (2021) Deciphering molecular interactions by proximity labeling

    Nature Methods 18:133–143.

    https://doi.org/10.1038/s41592-020-01010-5

    • PubMed
    • Google Scholar
39. Book

    1. Raschka S
    2. Liu Y
    3. Mirjalili V

    (2022)

    Machine Learning with PyTorch and Scikit-Learn

    Packt Publishing Ltd.

    • Google Scholar
40. 1. Roux KJ
    2. Kim DI
    3. Burke B
    4. May DG

    (2018) BioID: A screen for protein-protein interactions

    Current Protocols in Protein Science 91:cpps.51.

    https://doi.org/10.1002/cpps.51

    • PubMed
    • Google Scholar
41. Conference

    1. Strande S
    2. Cai H
    3. Tatineni M
    4. Pfeiffer W
    5. Irving C
    6. Majumdar A
    7. Mishin D
    8. Sinkovits R
    9. Norman M
    10. Wolter N
    11. Cooper T
    12. Altintas I
    13. Kandes M
    14. Perez I
    15. Shantharam M
    16. Thomas M
    17. Sivagnanam S
    18. Hutton T

    (2021) Expanse: Computing without Boundaries

    Practice and Experience in Advanced Research Computing. pp. 1–4.

    https://doi.org/10.1145/3437359.3465588

    • Google Scholar
42. 1. Teufel F
    2. Refsgaard JC
    3. Kasimova MA
    4. Deibler K
    5. Madsen CT
    6. Stahlhut C
    7. Grønborg M
    8. Winther O
    9. Madsen D

    (2023) Deorphanizing peptides using structure prediction

    Journal of Chemical Information and Modeling 63:2651–2655.

    https://doi.org/10.1021/acs.jcim.3c00378

    • PubMed
    • Google Scholar
43. 1. Venkatesan K
    2. Rual J-F
    3. Vazquez A
    4. Stelzl U
    5. Lemmens I
    6. Hirozane-Kishikawa T
    7. Hao T
    8. Zenkner M
    9. Xin X
    10. Goh K-I
    11. Yildirim MA
    12. Simonis N
    13. Heinzmann K
    14. Gebreab F
    15. Sahalie JM
    16. Cevik S
    17. Simon C
    18. de Smet A-S
    19. Dann E
    20. Smolyar A
    21. Vinayagam A
    22. Yu H
    23. Szeto D
    24. Borick H
    25. Dricot A
    26. Klitgord N
    27. Murray RR
    28. Lin C
    29. Lalowski M
    30. Timm J
    31. Rau K
    32. Boone C
    33. Braun P
    34. Cusick ME
    35. Roth FP
    36. Hill DE
    37. Tavernier J
    38. Wanker EE
    39. Barabási A-L
    40. Vidal M

    (2009) An empirical framework for binary interactome mapping

    Nature Methods 6:83–90.

    https://doi.org/10.1038/nmeth.1280

    • PubMed
    • Google Scholar
44. 1. Weeratunga S
    2. Gormal RS
    3. Liu M
    4. Eldershaw D
    5. Livingstone EK
    6. Malapaka A
    7. Wallis TP
    8. Bademosi AT
    9. Jiang A
    10. Healy MD
    11. Meunier FA
    12. Collins BM

    (2024) Interrogation and validation of the interactome of neuronal Munc18-interacting Mint proteins with AlphaFold2

    The Journal of Biological Chemistry 300:105541.

    https://doi.org/10.1016/j.jbc.2023.105541

    • PubMed
    • Google Scholar
45. 1. Xu J
    2. Zhang Y

    (2010) How significant is a protein structure similarity with TM-score = 0.5?

    Bioinformatics 26:889–895.

    https://doi.org/10.1093/bioinformatics/btq066

    • PubMed
    • Google Scholar
46. 1. Yang Z
    2. Zeng X
    3. Zhao Y
    4. Chen R

    (2023) AlphaFold2 and its applications in the fields of biology and medicine

    Signal Transduction and Targeted Therapy 8:115.

    https://doi.org/10.1038/s41392-023-01381-z

    • PubMed
    • Google Scholar
47. 1. Yu D
    2. Chojnowski G
    3. Rosenthal M
    4. Kosinski J

    (2023) AlphaPulldown-a python package for protein-protein interaction screens using AlphaFold-Multimer

    Bioinformatics 39:749.

    https://doi.org/10.1093/bioinformatics/btac749

    • PubMed
    • Google Scholar
48. 1. Zhang Y
    2. Skolnick J

    (2004) Scoring function for automated assessment of protein structure template quality

    Proteins 57:702–710.

    https://doi.org/10.1002/prot.20264

    • PubMed
    • Google Scholar
49. 1. Zhang Y
    2. Skolnick J

    (2005) TM-align: a protein structure alignment algorithm based on the TM-score

    Nucleic Acids Research 33:2302–2309.

    https://doi.org/10.1093/nar/gki524

    • PubMed
    • Google Scholar
50. 1. Zhao N
    2. Zhuo M
    3. Tian K
    4. Gong X

    (2022) Protein-protein interaction and non-interaction predictions using gene sequence natural vector

    Communications Biology 5:652.

    https://doi.org/10.1038/s42003-022-03617-0

    • PubMed
    • Google Scholar

Author details

1. Victoria Mischley

   1. Cancer Signaling and Microenvironment Program, Fox Chase Cancer Center, Philadelphia, United States
   2. Molecular Cell Biology and Genetics, Drexel University, Philadelphia, United States

   Contribution

   Conceptualization, Data curation, Formal analysis, Validation, Investigation, Methodology, Writing – original draft, Writing – review and editing

   Competing interests

   No competing interests declared

2. Johannes Maier

   Triana Biomedicines, Lexington, United States

   Contribution

   Conceptualization, Methodology, Writing – review and editing

   Competing interests

   Johannes Maier is affiliated with Triana Biomedicines, Lexington. The authors has no other competing interests to declare

3. Jesse Chen

   Triana Biomedicines, Lexington, United States

   Contribution

   Conceptualization, Methodology, Writing – review and editing

   Competing interests

   Jesse Chen is affiliated with Triana Biomedicines, Lexington. The authors has no other competing interests to declare

4. John Karanicolas

   1. Cancer Signaling and Microenvironment Program, Fox Chase Cancer Center, Philadelphia, United States
   2. Moulder Center for Drug Discovery Research, Temple University School of Pharmacy, Philadelphia, United States

   Contribution

   Conceptualization, Supervision, Funding acquisition, Validation, Methodology, Writing – original draft, Writing – review and editing

   For correspondence

   johnkaranicolas1@gmail.com

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0003-0300-726X

• 2,548

  views

• 148

  downloads

• 7

  citations

Views, downloads and citations are aggregated across all versions of this paper published by eLife.

Citations by DOI

• 3

  citations for umbrella DOI https://doi.org/10.7554/eLife.98179

• 4

  citations for Reviewed Preprint v1 https://doi.org/10.7554/eLife.98179.1