Edit · 2026 01 11 181912 Nature S42256 025 01136 1

Saving needs a token. Log in or register to get one. Edits are public and licensed CC-BY-SA-4.0.

# An interaction-derived graph learning framework for scoring protein–peptide complexes | Nature Machine Intelligence

**URL:** https://www.nature.com/articles/s42256-025-01136-1

**Saved:** 2026-01-11T02:19:18.159Z

---

Article
Published: 23 October 2025
An interaction-derived graph learning framework for scoring protein–peptide complexes
Huanyu Tao, Xiaoyu Wang & Sheng-You Huang

Nature Machine Intelligence 
volume
 7, 
pages
1858–1869 (2025)Cite this article

3779 Accesses

1 Citations

19 Altmetric

Metrics
details

Abstract

Accurate prediction of protein–peptide interactions is critical for peptide drug discovery. However, due to the limited number of protein–peptide structures in the Protein Data Bank, it is challenging to train an accurate scoring function for protein–peptide interactions. Here, addressing this challenge, we propose an interaction-derived graph neural network model for scoring protein–peptide complexes, named GraphPep. GraphPep models protein–peptide interactions instead of traditional atoms or residues as graph nodes, and focuses on residue–residue contacts instead of a single peptide root mean square deviation in the loss function. Therefore, GraphPep can not only efficiently capture the most important protein–peptide interactions, but also mitigate the problem of limited training data. Moreover, the power of GraphPep is further enhanced by the ESM-2 protein language model. GraphPep is extensively evaluated on diverse decoy sets generated by various protein–peptide docking programs and AlphaFold, and is compared against state-of-the-art methods. The results demonstrate the accuracy and robustness of GraphPep.

You have full access to this article via your institution.

Download PDF
Similar content being viewed by others
PepNN: a deep attention model for the identification of peptide binding sites
Article Open access
26 May 2022
PepCNN deep learning tool for predicting peptide binding residues in proteins using sequence, structural, and language model features
Article Open access
28 November 2023
Accurate prediction of protein function using statistics-informed graph networks
Article Open access
04 August 2024
Main

Interactions between peptide and protein make up to 40% of all interprotein interactions, which are involved in many cellular functions, including immunization, gene expression and signal transmission1. Therefore, protein–peptide interactions have increasingly attracted researchers’ attention due to their therapeutic potential, as peptides can modulate protein functions by intervening in key interactions2,3,4,5,6. Protein–peptide structures provide fundamental insights into their physicochemical interaction details. However, experimental methods are costly in terms of time and money, making computational modelling methods invaluable. The inherent flexibility of peptide complicates the use of experimental methods, making it more challenging to accurately characterize and predict their interactions and conformations1,7. Therefore, various protein–peptide docking methods have been developed for predicting protein–peptide complex structures8,9,10,11,12,13,14,15,16,17. Docking involves two primary stages: sampling, which generate numerous candidate complex models, known as decoys, and scoring, which evaluates and ranks these models to identify the most accurate structure. The accuracy of a scoring function is crucial, as it determines the overall performance of docking methods. Given the limited number of experimentally solved protein–peptide structures, scoring functions developed specifically for protein–peptide interactions are still rare.

The main categories of current scoring functions for protein–protein or protein–peptide interactions include physics-based potentials16,17,18,19, knowledge-based statistical potentials20,21,22, machine learning-based23,24,25 and deep learning-based methods26,27,28,29,30,31. Physics-based scoring functions incorporate physical energy terms such as van der Waals, electrostatic and hydrophobic energy. Alternatively, knowledge-based scoring functions derive potentials through statistical analysis of the information embedded in experimental structures. Recent advancements in machine learning and deep learning have led to a boom in new scoring functions. Machine learning-based scoring functions utilize random forest25 or support vector machine24 to distinguish near-native from non-native conformations on the basis of a diverse set of features. However, machine learning-based scoring functions depend heavily on feature engineering and training sets. As for deep learning-based methods, there are two deep learning frameworks prevalent in scoring functions: convolutional neural networks (CNNs)26,29 and graph neural networks (GNNs)27,28,30,31,32,33,34,35. The CNN-based approaches grid the interface of complex models to form a three-dimensional (3D) feature for learning, as seen in models such as DOVE29 and DeepRank26. However, CNN-based scoring functions, particularly those using 3D grid representations, are sensitive to the orientation of the input interface and often fail to learn local interactions properly31. In addition, CNNs are also known to be limited in learning long-range interactions because of their localized receptive fields36. Although coordinate-based equivariant graph networks37,38,39 and transformers40,41,42 have recently revolutionized protein structure prediction, traditional GNNs provide a more robust framework in the representation of interactions27,30. The GNN-based scoring functions represent complex structures as node-and-edge graphs and learn interactions by message passing, offering a better performance and rotation invariance27,30. With the emergence of protein language models43,44,45,46, leveraging their embeddings—which capture evolutionary and functional information—as node features can further enhance the performance of GNN-based scoring models47.

Current GNNs for scoring usually construct graphs with residues or atoms as nodes and their contacts as edges27,28,30,31. Graph-level prediction, such as quality assessments of decoys like the fraction of native contacts (Fnat)27, the Local Distance Difference Test (LDDT)48,49 and DockQ50,51, or classification of decoys as correct or incorrect based on interface root mean square deviation (iRMSD) or Critical Assessment of Predicted Interactions (CAPRI) criteria28,30, are performed after processing through the network. However, such models present several challenges. First, in cases where the interacting molecules differ considerably in size—such as in protein–peptide complexes—graph-level predictions tend to be dominated by the larger molecule, as the disproportionate number of nodes can overshadow the contributions of the smaller one. Second, graph-level prediction requires a large number of diverse decoys for training, which often cause a great imbalance between negative and positive decoys. Third, residue-level graphs omit crucial atomic interaction details, and atom-level graphs result in poor performance for unbound receptor docking. In addition, intramolecular and intermolecular interactions are learned together in the network, raising concerns about the ability to capture the interaction pattern at the interface52. Last, the small number of protein–peptide structures in the Protein Data Bank (PDB)53 also makes it challenging to train an accurate scoring function.

To overcome these challenges, we propose a GNN model to predict protein–peptide interactions, named GraphPep, where complexes are represented as an interaction-derived graph at both atom and residue levels. Specifically, GraphPep treats interface interactions between protein and peptide as nodes, and defines an edge if two interactions share one residue or atom. GraphPep performs node-level classification after message passing through a residual gated graph convolutional networks (GCNs)54. Such an interaction-derived graph facilitates an explicit learning of protein–peptide interactions at the interface. In addition, to address the limited training data problem, GraphPep focuses on protein–peptide residue–residue contacts rather than relying solely on a single peptide root mean square deviation (RMSD) in the loss function. GraphPep was extensively evaluated on diverse decoy sets generated with various docking algorithms and AlphaFold-Multimer (AFM)41, demonstrating its accuracy and robustness. Furthermore, GraphPep showed strong performance on an antibody–peptide test set with decoys generated by AlphaFold3 using forced sampling42.

Results
The workflow of GraphPep

Figure 1 shows the workflow of GraphPep. Given a protein–peptide complex structure, two traditional graphs are first generated at atom level and residue level, respectively, where features from ESM (evolutionary scale modeling) protein language models43 are also embedded. Next, the two graphs are converted to interaction-derived graphs. Specifically, the edges that represent protein–peptide interactions in the original graph are treated as nodes in the derived graph, and two edges that share one common vertex in the original graph form an edge in the derived graph (Fig. 1a). Residual gated graph convolution54 is conducted on the interaction-derived graphs at both atom level and residue level during deep learning (Fig. 1b), where the updated atom-level graph node embeddings are pooled to residue-level graph nodes as additional node features (Fig. 1c). The confidence score of each node, which represents a residue–residue interaction between protein and peptide, is obtained from the node embedding in the residue-level graph. The final binding score sums up the numbers of the nodes with different confidence thresholds.

Fig. 1: The framework of GraphPep.

Atom-level and residue-level interfaces and features of protein–peptide complexes are extracted and converted into interaction-derived graphs, allowing features to be triangle-aware, where only the nodes that represent protein–peptide interactions are retained in the derived graph. a, A GNN of residual gated graph convolution is applied to both atom-level and residue-level interaction-derived graphs. b, Here, atom-level node features are pooled to the corresponding residue-level nodes. c, Finally, GraphPep performs node classification and combines the confidence scores of graph nodes to get the final binding score of a protein–peptide binding pose.

Full size image
Performance on decoys generated by docking programs

We first evaluated GraphPep using the decoys generated by FlexPepDock16,17, HADDOCK18,19 and HPEPDOCK9,10 on two benchmark datasets: LEADS-PEP (bound/unbound)55 and Local_62 (unbound/unbound)56. Its performance is compared with the built-in scoring functions of corresponding docking algorithms and InterPepScore on each decoy set.

Bound benchmark (LEADS-PEP)

GraphPep was first evaluated on the LEADS-PEP test set of 53 protein–peptide complexes using the decoy structures generated by FlexPepDock, HADDOCK and HPEPDOCK (Supplementary Tables 1–3). As shown in Fig. 2, GraphPep consistently outperforms the scoring functions of all three docking algorithms, as well as the graph-based deep learning method InterPepScore35, for different numbers of top-ranked predictions.

Fig. 2: Comparison of GraphPep with other methods on the LEADS-PEP bound test set.

a–c, The DockQ box plots of different methods for top 1, 3 and 5 predictions (n = 53) on the docking decoys generated by FlexPepDock (a), HADDOCK (b) and HPEPDOCK (c). The boxes show the averages as circles, interquartiles as hinges, and 1.5 times the interquartile ranges as whiskers. For reference, the results of the best-sampled decoys are also shown. d–f, Scatter plots for the top 1 DockQ scores of GraphPep versus other methods on the docking decoys generated by FlexPepDock (d), HADDOCK (e) and HPEPDOCK (f).

Source data

Full size image

For FlexPepDock decoys, GraphPep achieves improved average DockQ scores of 0.479, 0.515 and 0.528 for top 1, 3 and 5 predictions, respectively, compared with 0.373, 0.469 and 0.502 for FlexPepDock and 0.405, 0.464 and 0.499 for InterPepScore, respectively (Fig. 2a). The top 1 performance of GraphPep even outperforms the top 3 performances of FlexPepDock and InterPepScore on this decoy set. For HADDOCK decoys, GraphPep also performs the best and obtains the average DockQ scores of 0.549, 0.613 and 0.637 for top 1, 3 and 5 predictions, respectively, compared with 0.438, 0.470 and 0.568 for HADDOCK and 0.535, 0.599 and 0.630 for InterPepScore, respectively (Fig. 2b). For HPEPDOCK decoys, which include the most extensive and accurate binding poses, GraphPep again yields highest average DockQ scores of 0.814, 0.854 and 0.860 for top 1, 3 and 5 predictions, compared with 0.706, 0.794 and 0.816 for HPEPDOCK and 0.747, 0.805 and 0.825 for InterPepScore, respectively (Fig. 2c). In addition, it is shown that InterPepScore is decoy set dependent and improves more on HADDOCK decoys than FlexPepDock and HPEPDOCK decoys, while our GraphPep achieves substantial improvement across three decoy sets (Figs. 2a–c), suggesting the robustness of GraphPep to different docking programs.

Further examination of individual cases reveals that GraphPep achieves comparable or improved performances for most of the complexes. When the top-ranked predictions are considered, GraphPep obtains better DockQ scores for 33 cases than FlexPepDock and 29 cases than InterPepScore (Fig. 2d and Supplementary Table 1). For HADDOCK decoys, GraphPep yields better top 1 DockQ scores for 32 cases than HADDOCK and 27 cases than InterPepScore (Fig. 2e and Supplementary Table 2). For HPEPDOCK decoys, GraphPep achieves better top 1 DockQ scores for 29 cases than HPEPDOCK and 29 cases than InterPepScore (Fig. 2f and Supplementary Table 3).

Unbound benchmark (Local_62)

To further evaluate GraphPep on unbound receptor structures, we tested it on the Local_62 benchmark of 62 protein–peptide complexes (Supplementary Tables 4 and 5), where the docking decoys were generated by HADDOCK and HPEPDOCK, respectively. This benchmark is to evaluate the impact of protein flexibility on GraphPep versus other methods because it presents a major challenge due to the absence of induced-fit conformations in the receptor structures.

For HADDOCK decoys, GraphPep consistently outperforms HADDOCK and InterPepScore (Fig. 3a). When top-ranked predictions are considered, GraphPep achieves a success rate of 32.2% for medium-accuracy criterion, compared with 21.0% for HADDOCK and 16.1% for InterPepScore. In terms of high-accuracy models, GraphPep correctly predicts 4.8% of the cases. By comparison, InterPepScore achieves a success rate of only 1.6%, while HADDOCK fails to predict any high-accuracy models. The performance of GraphPep remains superior when considering the top 3 and top 5 predictions (Fig. 3a). Figure 3b shows the top 1 DockQ scores versus receptor flexibility (apo–holo Cα RMSD) for different methods. It can be seen from the figure that GraphPep can improve the performance across all ranges of receptor RMSDs and is able to predict models with medium or even high quality for complexes with apo–holo RMSD ≥3 Å (Supplementary Table 4). These results demonstrate the robustness of GraphPep to receptor flexibility.

Fig. 3: Comparison of GraphPep with other methods on the Local_62 unbound test set.

a,c, Success rates of different methods for top 1, 3 and 5 predictions on the docking decoys generated by HADDOCK (a) and HPEPDOCK (c). For reference, the results of the best-sampled decoys are also shown. b,d, Scatter plots of top 1 DockQ scores versus Cα RMSD between apo and holo receptor structures for different methods on the docking decoys generated by HADDOCK (b) and HPEPDOCK (d). Dotted lines indicate the thresholds for DockQ classification zones.

Source data

Full size image

For HPEPDOCK decoys, GraphPep also outperforms both HPEPDOCK and InterPepScore for different numbers of top predictions (Fig. 3c). It is known that HPEPDOCK is a rigid docking program, where the protein is treated as a rigid body. Therefore, it is challenging to predict the binding modes for those cases that involve large receptor conformational changes. Nevertheless, GraphPep achieves improved success rates of 61.3% for the medium-accuracy threshold and 19.4% for the high-accuracy criterion when considering the top-ranked predictions, substantially outperforming HPEPDOCK (41.9% and 11.3%) and InterPepScore (38.7% and 9.7%). Similar improvements can also be observed in the success rates of top 2 and 3 predictions. In addition, GraphPep also improves top 1 DockQ scores over other methods across different extents of receptor flexibility and can even improve the DockQ score for a complex with an apo–holo Cα RMSD >6 Å (Fig. 3d and Supplementary Table 5), again suggesting the robustness of GraphPep.

Performance on decoys constructed by ADCP and AFM

We further tested GraphPep on the AutoDock CrankPep (ADCP) test set of 99 protein–peptide complexes (Supplementary Table 6). As the original decoys for ADCP99 are not publicly available, we generated the decoys and analysed them locally using the same procedure described in its original study57. Figure 4 shows the performances of GraphPep and different methods on the decoys of the ADCP99 benchmark generated by ADCP11 and AFM41. For the combined decoys, we used the same consensus scoring strategy as described in the literature57, where the AFM predictions are selected if the multiple sequence alignment (MSA) of the peptide yields a non-zero effective sequence count (that is, Neff-pep >0); otherwise, the top-ranked structure is chosen based on the tested scoring function.

Fig. 4: Comparison of GraphPep with other methods on the ADCP99 test set.

a–c, Success rates of different methods on the decoys generated by ADCP (a), AFM (b) and their combined decoys (c). For ranking combined decoys in c, the AFM prediction is used if the Neff of the peptide is >0; otherwise, the top-ranked model from the combined decoy set is selected based on each scoring function. For reference, the results of the best-sampled decoys are also shown. d–f, Scatter plots of top 1 DockQ scores for GraphPep versus other methods on the decoys generated by ADCP (d), AFM (e) and their combined decoys (f). For better illustration, only the complexes with Neff-pep = 0 are shown in d–f.

Source data

Full size image

It can be seen from Fig. 4 that GraphPep consistently outperforms other methods on all three decoy sets. Specifically, on the ADCP decoy set (Fig. 4a), GraphPep achieves a success rate of 56.6% for the medium-accuracy threshold and 31.3% for the high-accuracy criterion, which surpass 47.5% and 23.2% for ADCP and 51.5% and 25.3% for InterPepScore. On the AFM decoy set, the performance differences between scoring functions are less pronounced than those for ADCP decoys, which is probably due to limited structural diversity (Fig. 4b). Nevertheless, GraphPep still ranks first on this decoy set in both medium- and high-accuracy success rates. For the combined decoy set, all scoring methods exhibits improved success rates, highlighting the effectiveness of the consensus strategy (Fig. 4c). Among all scoring functions, GraphPep achieves the highest overall success rate under the consensus strategy and yields medium or high predictions for 74.7% of the cases, outperforming InterPepScore (68.7%) and the consensus energy used in the original paper (71.7%). The advantage of GraphPep in high-accuracy predictions is even more pronounced with a success rate of 58.6% for GraphPep versus 48.5% for InterPepScore and 51.5% for consensus energy.

Figure 4d–f shows the comparison of top 1 DockQ scores of GraphPep versus other methods on different decoy sets. Here, we focused on the 37 complexes with no valid peptide MSA (that is, Neff-pep = 0) where scoring functions will influence the ranking under the consensus strategy. It is shown that GraphPep achieves comparable or higher DockQ scores for most of the cases, compared with the other methods (Supplementary Table 6). For ADCP decoys, GraphPep achieves a better DockQ score in 14 cases compared to ADCP and in 18 cases compared with InterPepScore (Fig. 4d). For AFM decoys, GraphPep outperforms AFM in 15 cases and InterPepScore in 20 cases (Fig. 4e). For the combined decoys, GraphPep yields better DockQ scores for 25 cases than the consensus energy and 22 cases than InterPepScore (Fig. 4e).

Performance on decoys generated by AlphaFold3

GraphPep was also evaluated on the nr_epitope_minus benchmark of 121 antibody–peptide complexes58, for which we generated their decoys using AlphaFold342 with enforced sampling (100 models per complex using 20 seeds) (Supplementary Table 7). The generated models are originally ranked by the interface predicted template modelling (ipTM) score of the peptide chain, which reflects the confidence in interchain positions. For comparison, the performance of InterPepScore was also evaluated on these decoys.

As shown in Fig. 5a, GraphPep achieves the best performance among the evaluated methods. When the top-ranked predictions are considered, GraphPep achieves the success rates of 43.7% and 15.6% under medium-accuracy and high-accuracy criteria, respectively. By comparison, InterPepScore achieves 37.5% and 14.1%, while AlphaFold3 gives 42.2% and 12.5% at the corresponding accuracy levels. Another notable feature in the figure is that there is a large gap between the success rates of scoring methods and the best-sampled models, indicating that scoring remains a challenge for antibody–peptide cases. Figure 5b shows the box plots of DockQ scores for different methods. It can be seen from the figure that GraphPep achieves a higher average DockQ score of 0.469, compared with 0.428 for InterPepScore and 0.441 for AlphaFold3.

Fig. 5: Comparison of GraphPep with other methods on the nr_epitope_minus test set.

The decoys are generated by AlphaFold3 with enhanced sampling. a, Overall success rates of different methods. b, Box plots of the top 1 DockQ scores (n = 64). The boxes show the averages as circles, interquartiles as hinges, and 1.5 times the interquartile ranges as whiskers. c, Success rates for different peptide length ranges. The proportions of short (6–10 a.a.), medium (11–15 a.a.) and long (16–19 a.a.) peptides are shown on the right. Only the top-ranked predictions for each complex are considered in the success rate calculation. For reference, the results of the best-sampled decoys generated by AlphaFold3 are also shown.

Source data

Full size image

Figure 5c shows the top-ranked prediction performances in different peptide length ranges. It can be seen from the figure that all three methods yield decreasing success rates as peptide length increases, reflecting the greater difficulty in accurately predicting the binding poses of long peptides. In addition, GraphPep consistently outperforms InterPepScore for both short- and medium-length peptides and surpasses AlphaFold3 for short peptides. Predicting antibody–peptide complexes with long peptides (>15 amino acids (a.a.)) is challenging due to increased conformational flexibility and binding complexity. All three methods give an identical medium-accuracy success rate of 33.3% for long peptides. However, AlphaFold3 fails to predict any high-accuracy structures for these cases, while both GraphPep and InterPepScore achieve high-accuracy predictions for 11.1% of the complexes.

Examples of predicted binding modes

Figure 6 shows three examples of the top binding modes predicted by HPEPDOCK and GraphPep, compared with the native peptide structures. It can be seen from the figure that the binding modes predicted by HPEPDOCK are different from the experimental structures, for which GraphPep gives low confidence scores. By contrast, the binding poses predicted by GraphPep are closer to the native, and the residues that overlap well with the native peptide have higher predicted confidence scores. GraphPep achieves a successful high-accuracy prediction on the case 4DS1 from LEADS-PEP with a DockQ score of 0.875, while HPEPDOCK fails with a low DockQ score of 0.469. For 2C3I from Local_62, the peptide orientation predicted by HPEPDOCK is opposite to the native structure. However, the binding pose predicted by GraphPep has a right orientation with a DockQ score of 0.771. The peptide in 1T7R from Local_62 is an alpha helix, but HPEPDOCK predicted the peptide in a coil conformation. The incorrect secondary structure for 1T7R results in a poor prediction by HPEPDOCK, with a DockQ score of 0.247. By contrast, GraphPep correctly predicts the peptide as an alpha helix, resulting in a medium-accuracy prediction with a DockQ score of 0.745.

Fig. 6: Examples of the top-ranked binding modes by HPEPDOCK and GraphPep.

a–f, The first row shows the top binding modes predicted by HPEPDOCK for 4DS1 (a), 2C3I (c) and 1T7R (e), and the second row shows those predicted by GraphPep for 4DS1 (b), 2C3I (d) and 1T7R (f).

Full size image

The three examples indicates the limitations of HPEPDOCK in predicting binding modes, such as far deviations, incorrect orientations and wrong secondary structures. By contrast, GraphPep accurately predicts the binding modes close to the native ones with right directions and secondary structures. Notably, GraphPep can give the confidence score for each interaction residue pair or contact, which can be used to guide the optimization of predicted binding modes. For instance, the tails of the binding modes predicted by GraphPep are all red for 4DS1, 2C3I and 1T7R (Fig. 6), indicating that local refinement is needed on these residues.

GraphPep score versus pose accuracy

An ideal scoring function should not only recognize near-native binding pose but also guide the refinement of binding modes towards the native state. In other words, a ‘funnel-like’ feature59 should be seen in the binding energy landscape, where the binding modes that are closer to the native structure have better scores. The HPEPDOCK decoy sets from the LEADS-PEP and local_62 datasets were selected for evaluation because their larger number of decoys per complex provides more reliable statistics.

We calculated the Pearson correlations of scores versus ligand RMSD (LRMSD) and DockQ for different methods on two decoy sets (Supplementary Tables 8 and 9). It is shown that GraphPep achieves an overall better correlation than the other methods on both the LEADS-PEP and Local_62 decoy sets (Supplementary Fig. 1). For the score–LRMSD relationship, GraphPep obtains a higher average correlation of 0.541, compared with 0.134 for HPEPDOCK and 0.305 for InterPepScore on the LEADS-PEP decoy set (Supplementary Table 8). In addition, GraphPep also gives a better correlation of 0.446, compared with 0.047 for HPEPDOCK and 0.098 for InterPepScore on the local_62 decoy set (Supplementary Table 9). Similar trends can be observed in terms of score-DockQ relationship. Specifically, GraphPep achieves a better average correlation of −0.599, compared with −0.172 for HPEPDOCK and −0.348 for InterPepScore on the LEADS-PEP decoy set (Supplementary Table 8). GraphPep yields a better average correlation of −0.512, compared with −0.063 for HPEPDOCK and −0.118 for InterPepScore on the Local_62 decoy set (Supplementary Table 9). These results suggest that GraphPep may also be used to guide the optimization of initially sampled binding modes.

Ablation experiments

To investigate the impact of different components on the performance of GraphPep, we performed an ablation study on graph levels, protein language model embeddings and intramolecular interactions (Supplementary Tables 10 and 11). Supplementary Fig. 2 shows the DockQ interval plot of different ablation models when considering the top 1 predictions on the decoy set of LEADS-PEP and Local_62 generated by HPEPDOCK. It can be seen from the figure that the multiscale graph learning of the base model GraphPep, which combines atom- and residue-level representations, yields a better performance than single-level graph models, ResGraph or AtomGraph. In addition, incorporation of ESM-2 embeddings improves the performance of GraphPep, especially on Local_62. Interestingly, explicit inclusion of intramolecular interactions, which include intramolecular contacts (w/Intra_contact) or intramolecular covalent bonds (w/Intra_covalent), does not seem to improve the performance (Supplementary Fig. 2), probably due to overwhelming intramolecular signals obscuring key interfacial interactions.

In addition, we also evaluated the impact of embedded protein language models by replacing the ESM-2 model with the latest ESM-3 model in GraphPep. However, such an operation causes a reduced performance (Supplementary Fig. 3). This can be attributed to the fact that ESM-3 incorporates structural data during training, making it more structure-sensitive than ESM-2. However, conformational changes at the interface often occur upon protein–peptide binding, making ESM-3 less robust for realistic interaction prediction. A similar decline in performance is observed when replacing the current GCN with a SE(3)-equivariant network (E3NN) (Supplementary Fig. 4).

Discussion

We have introduced GraphPep, an interaction-derived GNN scoring model for predicting protein–peptide interactions. Due to the limited number of protein–peptide complex structures available in the PDB, training an accurate scoring function for protein–peptide interactions has been challenging. In this study, we integrate three strategies to overcome the small training data issue. First, GraphPep treats protein–peptide interactions instead of traditional residues or atoms as nodes and defines an edge between two nodes if they share one common atom or residue in the graph representations. Second, GraphPep is trained to reproduce true protein–peptide residue–residue contacts instead of a single peptide RMSD for a given protein–peptide complex, reducing the demand of decoy number in the training set. Third, GraphPep uses a multiscale strategy to balance accuracy and robustness by adding the pooled atom graph features to residue graph nodes. Extensive evaluations on diverse test sets show that GraphPep achieves notably improved performances in protein–peptide binding mode prediction, compared with state-of-the-art methods including FlexPepDock, HADDOCK, HPEPDOCK, ADCP, InterPepScore, AFM and AlphaFold3. Moreover, GraphPep score also has a strong correlation with LRMSD and DockQ scores, exhibiting a good funnel-like feature in the binding energy landscape. These results demonstrate the effectiveness of interaction-derived graphs and the accuracy of our GraphPep scoring model.

Docking algorithms often generate models with steric clashes. Therefore, GraphPep provide users an option to perform structural plausibility checks on complex models by using PoseBusters60. It is revealed that, although GraphPep shows a slightly decreased performance for structurally plausible docking decoys, it still outperforms the other scoring methods (Supplementary Fig. 5). In addition, GraphPep also has a certain ability to identify and downrank decoys with improper steric conformations. For example, GraphPep ranks a HPEPDOCK decoy with atomic clashes at #44 despite its high DockQ score of 0.843 and moves it to #7 after removing the clashes with a short refinement (Supplementary Fig. 6). As such, the top-ranked decoys by GraphPep normally contain no severe steric clashes. However, it should be noted that, because GraphPep combines all contact confidence scores in the graph, decoys containing a few clashing contacts may still receive high scores if most interactions are reasonable. This design provides GraphPep with robustness against certain receptor flexibilities by prioritizing models with overall favourable binding interfaces, even if minor local clashes are present.

In addition, the secondary structure types of peptides can also influence the accuracy of algorithms in predicting protein–peptide complex structures. In particular, it is known that peptides, which adopt a helical conformation, are more challenging for docking methods61. To assess this effect on GraphPep, we grouped peptides into helix and non-helix categories and analysed the performances on three benchmark datasets (Supplementary Fig. 7). GraphPep consistently demonstrates superior or comparable performance for both helical and non-helical peptides compared with other methods on docking- and deep learning-based decoy sets, further highlighting its robustness.

Methods
Graph generation

Given a protein–peptide complex structure, we represent it as interaction-derived graphs at both the residue and atom levels, as shown in Fig. 1a. In these graphs, protein–peptide interactions are treated as nodes, with edges connecting the nodes that share one common atom or residue.

Specifically, at the residue level, residue–residue interactions between protein and peptide are considered as nodes in the derived graph. A pair of residues are defined as an interaction when their minimum distance of all heavy atom pairs is less than 6 Å. The node of residue–residue interaction i consists of sequence features, geometric feature and pairwise energy feature as ${X}_{i}^{\;r}=({X}_{S,i}^{\;r},{X}_{G,i}^{\;r},{X}_{E,i}^{\;r})$, and the edge (j, k) contains the geometric features${X}_{G,\;jk}^{\;r}$ between nodes j and k. Of these, the sequence features ${X}_{S,i}^{\;r}$ contain amino acid types of two interacting residues and residue embeddings extracted from ESM-2. The geometric features ${X}_{G,i}^{\;r}$ are derived from the structure by taking the minimum distance and CA (carbon alpha) distance between two residues. The energy calculated by ITScore-PP potential20 is used as the pairwise energy feature ${X}_{E,i}^{r}$. As for edges, the geometric features ${X}_{G,jk}^{r}$ include the angles and distances between the CA atoms of two nodes that represent two pairs of interacting residues. Amino acid types are one-hot encoded, while the distances and potential are encoded as 16 Gaussian radial basis functions, respectively.

At the atomic level, it is similar to the residue-level representation. Two atoms with distance less than 6 Å at the interface form an interaction, representing a node in the derived graph. The node feature ${X}_{i}^{a}=({X}_{S,i}^{a},{X}_{G,i}^{a},{X}_{E,i}^{a})$ include the types of two interacting atoms ${X}_{S,i}^{a}$, the distance between them ${X}_{G,i}^{a}$ and the ITScore-PP pairwise potential ${X}_{E,i}^{a}$. The atom types follow the classification used in ITScore-PP. The edge features ${X}_{G,\;jk}^{a}$ between node j and node k consist of the angles and distances between the two interacting atom pairs corresponding the two nodes.

As illustrated in the Fig. 1a, the three atoms or residues of two connected nodes form a triangle. The interaction pair and contact-derived graphs capture the characteristics of this triangle, including the sides and angles. Specifically, the sides between protein and peptide (shown in red) are incorporated into the node features of the interaction-derived graph, while the intraprotein or intrapeptide side and the angle (shown in green) are included in the edge features. The conversion from atom- or residue-level graphs to interaction-derived graphs preserves both interfacial and intramolecular interactions within the protein–peptide complex, ensuring a lossless transfer of structural information.

Model architecture

GraphPep accepts a protein–peptide complex structure as input and outputs a binding energy score. After data preprocessing, the protein–peptide complex structure is converted to a interaction-derived graph. Residual gated graph convolution54 is used in GraphPep, which is an advanced variant of GCNs. Residual gated GCN can effectively capture and process complex relational data within graph structures by combination of residual connections and gated mechanism. Residual connections add the input of a layer to its output, which helps to retain important information from earlier layers. In addition, gated mechanism selectively filters and propagates the most relevant information while ignoring the less useful data, which can be beneficial in dealing with the noisy interface interactions. Similar to a standard GCN, the residual gated GCN consists of three steps: message aggregation, update and readout. The hidden state hi of each node will aggregate and update iteratively as follows:

$${{h}_{i}}^{{\prime} }={W}_{1}{h}_{i}+\sum _{j\in {\mathcal{N}}(i)}{\eta }_{i,\;j}\odot {W}_{2}({h}_{j}| | {e}_{i,\;j}),$$
(1)

where ⊙ means multiplying the elements at corresponding positions of the matrixes, ∣∣ represents the concatenation of different features, ${\mathcal{N}}(i)$ denotes the neighbours of the ith node, ei,j is the edge feature between node i and j, and the gate ηi,j is defined as

$${\eta }_{i,\;j}=\sigma ({W}_{3}({h}_{i}| | {e}_{i,\;j})+{W}_{4}({h}_{j}| | {e}_{i,\;j})),$$
(2)

where σ denotes the activation function. W1, W2, W3 and W4 are all learnable parameters of the network.

We conduct multiscale graph learning at the atomic and residue levels (Fig. 1). First, we performed atomic-level graph convolution, and then pooled the updated atom-level graph node embeddings to residue-level graph nodes as additional node features (Fig. 1c). Then, we perform residue-level graph convolution to learn the residal interaction patterns. During the convolution, the features of neighbour node j and the edge features between nodes i and j are used to update the features of node i. Node and edge features contain the distances and angles between and within two pairs of interacting atoms/residues. Therefore, the updating process is triangle-aware and considers both intermolecular and intramolecular interactions, enabling a more comprehensive study of protein–peptide interaction. We perform dropout and layer normalization after graph convolution. Three layers of residual gated GCN are applied in GraphPep. The convolution results of each layer are concatenated and linearly transformed to obtain the final node representations. At the final step, we perform node classification rather than graph-level prediction, to reduce the need for large amounts of training data and to learn the interaction patterns between residues. The node embeddings are directly used in the readout process to obtain the contact confidence score of each node. The final binding score is a weighted summation for the numbers of the contacts at different confidence thresholds as follows:

$$\begin{array}{l}{\mathrm{GraphPep}}=\\-\log \left(1+{\omega }_{1}\sum _{i\in {\mathcal{V}}}{\mathbb{I}}(\;{p}_{i} > 0.5)+{\omega }_{2}\sum _{i\in {\mathcal{V}}}{\mathbb{I}}(\;{p}_{i} > 0.7)+{\omega }_{3}\sum _{i\in {\mathcal{V}}}{\mathbb{I}}(\;{p}_{i} > 0.9)\right),\end{array}$$
(3)

where ${\mathcal{V}}$ represents all nodes in the interaction contact-derived graph. ${\mathbb{I}}({\mathrm{condition}})$ equals 1 if the condition is satisfied or 0 otherwise. The weights are empirically selected with the ratio of 1:2:4.

Model training

The training set is randomly split into a validation set and a training set with a 1:9 ratio of complexes, rather than decoys. There are 2,007 protein–peptide complexes with 32,666 decoys in the training set and 224 complexes with 3,635 decoys in the validation set. The model was optimized using the AdamW62 optimizer, with a batch size of 32, a dropout rate of 0.1, a learning rate of 10−3 and learning rate step decay with ReduceLROnPlateau. Model training continues until the validation loss does not decrease for ten consecutive epochs.

The label of a node is defined based on whether its corresponding contact exists in the native structure, where two residues are considered to be in contact if the distance between them is within 5 Å (ref. 63). It should be noted that the residue–residue interaction cut-offs for the graph at the atom level and residue level are both 6 Å, which is larger than the labelling distance. One reason is that a longer distance cut-off allows the construction of a larger interaction graph that captures more long-range interactions. Another is to balance the positive and negative samples in a graph to a certain extent and force the model to distinguish the difference between contacts of small distance differences. For example, if the interaction-derived graphs of native complex structures were generated with a cut-off of 5 Å, all nodes would be classified as positive samples and no negative samples, resulting in highly imbalanced data that hinders effective model learning.

GraphPep score is derived from the node classification of residue-level graph, incorporating detailed atomic interactions through atom-level graph convolution. It has the capability to tolerate some inaccuracies in unbound receptor structures. Multitask neural networks have proven to be helpful in improving the accuracy of graph models64. Here, we introduce an auxiliary task based only on atom-level graphs. The updated atom node embeddings are pooled to residue-level node representation, which is used to perform the same prediction as the main task. As positive samples (that is, native contacts in the decoy) are fewer than negative samples in the training set, used Focal Loss as the loss function. The final loss65 is computed as the sum of the node classification losses from both residue-level and atom-level predictions, and each component has equal weight.

Training sets

For training, we screened the pepBDB database66 according to the following criteria: (1) X-ray structures with resolutions better than 2.5 Å; (2) the length of peptide ranges from 3 to 15 a.a.; (3) peptide without non-standard residues; (4) peptide without breaks; (5) proteins with no more than two chains and at least one chain longer than 50 a.a.; (6) more than 20 interresidue pairs contacting within 6 Å at the interface; (7) peptide not covalently bound to the protein. The remaining 2,645 complexes were clustered by MMSeq267 with a 40% sequence identity cut-off for proteins and 70% for peptides. For each cluster, the complex with the best resolution was selected as the representative. The same criteria were also applied to remove those cases that are redundant with the test set LEADS-PEP and Local_62, resulting in a total of 2,231 complexes in the final training set. The decoys in the training set were generated directly by docking the bound peptide to the binding site of receptor using HPEPDOCK. The native binding modes were added to the decoy set to ensure that each complex has at least one correct binding mode for the model to learn. To create a small yet diverse decoy set, we divided the decoys into 20 bins based on Fnat, with each bin having a width of 0.05. Fnat is the fraction of native interaction pairs or contacts reproduced by a docked pose and is commonly used to evaluate the performance of a docking algorithm. For each Fnat bin, the model with the best backbone root mean square deviation (bRMSD) was selected as the representative. Finally, we selected up to 20 decoys per protein–peptide complex and ensured the contact diversity in the training set, resulting in a training set of 36,301 decoys. On average, each of the 2,231 complexes in training set contributes only 16 decoys.

Test sets

Two widely used protein–peptide benchmarks, LEADS-PEP55 and peptiDB56, are used for test our GraphPep model on docking decoys. The LEADS-PEP dataset contains 53 complexes with peptide lengths ranging from 3 to 12 a.a. The complexes in the benchmark are non-redundant with a sequence identity below 30%. In addition, they are all X-ray crystal structures with resolutions better than 2.0 Å. For peptiDB, two benchmarks are available: one for blind global peptide docking and the other for local peptide docking with the binding site. Here, we use only the local peptide docking benchmark, named Local_62, which consists of 62 unboud/unbound protein–peptide complexes. LEADS-PEP is used as the test set with holo receptor structure and Local_62 as the test set with apo receptor structure. The peptide conformations are unknown during docking.

To test the performance of GraphPep, various docking methods were used to generate decoys, including FlexPepDock16,17, HADDOCK18,19 and HPEPDOCK9,10. FlexPepDock and HADDOCK 2.4 web servers were used for docking proteins and peptides in the test sets. The bound/unbound receptor structures on LEADS-PEP/local_62 and the extended peptide conformations were used as input. The default parameters of the two web servers were used for protein–peptide docking. FlexPepDock ranks models based on Rosetta energy and provides the top 10 predicted models for download. The Rosetta energy is an empirical energy score that combines physics-based and statistical-based potentials, which is a linear combination of weighted score terms of van der Waals, hydrogen bonds, electrostatic, disulfide bonds, solvation, backbone torsion angles, sidechain rotamer energies and an average reference energy of the unfolded state. HADDOCK provides the best 4 models from the top 10 clusters for download. The clusters are ranked based on the average HADDOCK score of the best 4 models for each cluster. We reranked the downloaded models according to their HADDOCK scores, regardless of the cluster. The analysis for original HADDOCK docking results are all based on HADDOCK scores, which is a weighted sum of the intermolecular van der Waals energy, electrostatic energy, desolvation energy and distance restraints energy. HPEPDOCK is a hierarchical flexible peptide docking approach developed by our group. We used our in-house program with default parameters, and the top 100 predicted models were used to test GraphPep. HPEPDOCK evaluates the binding poses using an iterative knowledge-based scoring function named ITScore-PP20. All three scoring functions measure the binding free energy between proteins and peptides, where a lower score indicates a better binding mode.

Another test set, named ADCP99, is a benchmark of 99 protein–peptide complexes that has been used to evaluate a consensus docking strategy integrating ADCP and AFM57. We generated the docking decoys using the same protocols and parameters described in the original study57. The consensus energy was calculated using the OpenMM_functions.py script provided in their study, following the definition of Econsensus = Ecomplex − Ereceptor as described in the literature57.

In addition, we also selected a benchmark of antibody–peptide complexes, named nr_epitope58, to test GraphPep on the decoys generated by AlphaFold3 with forced sampling. The dataset consists of 121 high-quality, non-redundant antibody–peptide complexes, with peptide lengths ranging from 5 to 20 a.a. It includes only antibodies with paired heavy and light chains, excluding single-chain nanobodies, and all structures have a resolution of 3.0 Å or better. Decoys were generated locally using AlphaFold342 (version 3.0.0) with default parameters. The MSA configuration followed the original default settings, and MSA and template information were carefully retrieved for all chains. For each antibody–peptide complex, AlphaFold3 generates five predicted models per random seed. To generate diverse decoy sets, we performed repeated sampling using 20 different random seeds (seeds 1–20), resulting in a total of 100 predicted structures per complex. The ipTM score of peptide chain, as predicted by AlphaFold3, was used to rank the models. The ipTM score measures the accuracy of the predicted relative positions between protein subunits within a complex. Values higher than 0.8 represent confident high-quality predictions. For such confidently predicted complexes, further sampling and reranking are typically unnecessary. To evaluate the ranking performance of GraphPep against other scoring functions, we excluded complexes with peptide-chain ipTM scores above 0.8. This filtering resulted in a subset of 64 antibody–peptide complexes with decoys, referred to as nr_epitope_minus.

Evaluation criteria

Predicted models were evaluated using the DockQ50 score, a continuous quality metric that integrates the main quality measures of structural accuracy: interface RMSD (I-RMSD), LRMSD and Fnat. I-RMSD represents as the backbone RMSD calculated over the interface residues after the structural superposition. An interface residue is defined as such, when any of its atoms (hydrogen atoms excluded) are located within 10 Å of any of the atoms of the binding partner. L-RMSD is the backbone RMSD calculated over the ligand (peptide) residues after superimposing the receptor structures. Fnat denotes the fraction of native contacts reproduced by a predicted structure. Here, the distance cut-off for defining a contact between two residues is 5 Å (ref. 63). Based on the DockQ score, predicted structures were classified into four accuracy classes: high (0.8–1.0), medium (0.49–0.8), acceptable (0.23–0.49) and incorrect (0–0.23). A prediction is considered successful for a given protein–peptide complex if the prediction has a medium accuracy or better, that is, DockQ ≥0.49. This is the default success criterion, unless otherwise specified. The success rate was defined as the percentage of the test cases with at least one successful prediction compared with all cases in a protein–peptide benchmark when a certain number of top predictions are considered.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Data availability

The raw data of the evaluation results are provided in the Article and its Supplementary Information. The protein–peptide structures for testing in this Article were taken from the PDB. The training decoy set and test decoy sets are available via Zenodo at https://doi.org/10.5281/zenodo.17097750 (ref. 68). Source data are provided with this paper.

Code availability

The GraphPep package is freely available for academic or non-commercial users at http://huanglab.phys.hust.edu.cn/GraphPep/or via Zenodo at https://doi.org/10.5281/zenodo.17099863 (ref. 69).

References

Petsalaki, E. & Russell, R. B. Peptide-mediated interactions in biological systems: new discoveries and applications. Curr. Opin. Biotechnol. 19, 344–350 (2008).

Article
 
Google Scholar

Muttenthaler, M., King, G. F., Adams, D. J. & Alewood, P. F. Trends in peptide drug discovery. Nat. Rev. Drug Discov. 20, 309–325 (2021).

Article
 
Google Scholar

Lei, Y. et al. A deep-learning framework for multi-level peptide–protein interaction prediction. Nat. Commun. 12, 5465 (2021).

Article
 
Google Scholar

Zhao, Z., Peng, Z. & Yang, J. Improving sequence-based prediction of protein–peptide binding residues by introducing intrinsic disorder and a consensus method. J. Chem. Inf. Model. 58, 1459–1468 (2018).

Article
 
Google Scholar

Taherzadeh, G., Zhou, Y., Liew, A. W. & Yang, Y. Structure-based prediction of protein–peptide binding regions using Random Forest. Bioinformatics 34, 477–484 (2018).

Article
 
Google Scholar

Weng, G. et al. Comprehensive evaluation of fourteen docking programs on protein–peptide complexes. J. Chem. Theory Comput. 16, 3959–3969 (2020).

Article
 
Google Scholar

Neduva, V. et al. Systematic discovery of new recognition peptides mediating protein interaction networks. PLoS Biol. 3, e405 (2005).

Article
 
Google Scholar

Ciemny, M. et al. Protein–peptide docking: opportunities and challenges. Drug Discov. Today 23, 1530–1537 (2018).

Article
 
Google Scholar

Zhou, P., Jin, B., Li, H. & Huang, S. Y. HPEPDOCK: a web server for blind peptide–protein docking based on a hierarchical algorithm. Nucleic Acids Res. 46, W443–W450 (2018).

Article
 
Google Scholar

Zhou, P. et al. Hierarchical flexible peptide docking by conformer generation and ensemble docking of peptides. J. Chem. Inf. Model. 58, 1292–1302 (2018).

Article
 
Google Scholar

Zhang, Y. & Sanner, M. F. AutoDock CrankPep: combining folding and docking to predict protein–peptide complexes. Bioinformatics 35, 5121–5127 (2019).

Article
 
Google Scholar

Schindler, C. E., de Vries, S. J. & Zacharias, M. Fully blind peptide–protein docking with pepATTRACT. Structure 23, 1507–1515 (2015).

Article
 
Google Scholar

Lee, H., Heo, L., Lee, M. S. & Seok, C. GalaxyPepDock: a protein–peptide docking tool based on interaction similarity and energy optimization. Nucleic Acids Res. 43, W431–W435 (2015).

Article
 
Google Scholar

Yan, C., Xu, X. & Zou, X. Fully blind docking at the atomic level for protein–peptide complex structure prediction. Structure 24, 1842–1853 (2016).

Article
 
Google Scholar

Kurcinski, M. et al. CABS-dock standalone: a toolbox for flexible protein–peptide docking. Bioinformatics 35, 4170–4172 (2019).

Article
 
Google Scholar

Raveh, B., London, N. & Schueler-Furman, O. Sub-angstrom modeling of complexes between flexible peptides and globular proteins. Proteins 78, 2029–2040 (2010).

Article
 
Google Scholar

London, N., Raveh, B., Cohen, E., Fathi, G. & Schueler-Furman, O. Rosetta FlexPepDock web server—high resolution modeling of peptide–protein interactions. Nucleic Acids Res. 39, W249–W253 (2011).

Article
 
Google Scholar

Trellet, M., Melquiond, A. S. & Bonvin, A. M. A unified conformational selection and induced fit approach to protein–peptide docking. PLoS ONE 8, e58769 (2013).

Article
 
Google Scholar

Honorato, R. V. et al. The HADDOCK2.4 web server for integrative modeling of biomolecular complexes. Nat. Protoc. 19, 3219–3241 (2024).

Article
 
Google Scholar

Huang, S. Y. & Zou, X. An iterative knowledge-based scoring function for protein–protein recognition. Proteins 72, 557–579 (2008).

Article
 
Google Scholar

Feliu, E., Aloy, P. & Oliva, B. On the analysis of protein–protein interactions via knowledge-based potentials for the prediction of protein–protein docking. Protein Sci. 20, 529–541 (2011).

Article
 
Google Scholar

Liu, S. & Vakser, I. A. DECK: distance and environment-dependent, coarse-grained, knowledge-based potentials for protein–protein docking. BMC Bioinf. 12, 280 (2011).

Article
 
Google Scholar

Fink, F., Hochrein, J., Wolowski, V., Merkl, R. & Gronwald, W. PROCOS: computational analysis of protein–protein complexes. J. Comput. Chem. 32, 2575–2586 (2011).

Article
 
Google Scholar

Geng, C. et al. iScore: a novel graph kernel-based function for scoring protein–protein docking models. Bioinformatics 36, 112–121 (2020).

Article
 
Google Scholar

Jung, Y., Geng, C., Bonvin, A. M., Xue, L. C. & Honavar, V. G. MetaScore: a novel machine-learning-based approach to improve traditional scoring functions for scoring protein–protein docking conformations. Biomolecules 13, 121 (2023).

Article
 
Google Scholar

Renaud, N. et al. DeepRank: a deep learning framework for data mining 3D protein–protein interfaces. Nat. Commun. 12, 7068 (2021).

Article
 
Google Scholar

Rèau, M., Renaud, N., Xue, L. C. & Bonvin, A. M. DeepRank-GNN: a graph neural network framework to learn patterns in protein–protein interfaces. Bioinformatics 39, btac759 (2023).

Article
 
Google Scholar

McFee, M. & Kim, P. M. GDockScore: a graph-based protein–protein docking scoring function. Bioinform. Adv. 3, vbad072 (2023).

Article
 
Google Scholar

Wang, X., Terashi, G., Christoffer, C. W., Zhu, M. & Kihara, D. Protein docking model evaluation by 3D deep convolutional neural networks. Bioinformatics 36, 2113–2118 (2020).

Article
 
Google Scholar

Wang, X., Flannery, S. T. & Kihara, D. Protein docking model evaluation by graph neural networks. Front. Mol. Biosci. 8, 647915 (2021).

Article
 
Google Scholar

Mastropietro, A., Pasculli, G. & Bajorath, J. Learning characteristics of graph neural networks predicting protein–ligand affinities. Nat. Mach. Intell. 5, 1427–1436 (2023).

Article
 
Google Scholar

Hamilton, W., Ying, Z. & Leskovec, J. Inductive representation learning on large graphs. Adv. Neural Inf. Process. Syst. 30, 1024–1034 (2017).

Google Scholar

Xu, K., Hu, W., Leskovec, J. & Jegelka, S. How powerful are graph neural networks? In Proc. 7th International Conference on Learning Representations https://openreview.net/pdf?id=ryGs6iA5Km (ICLR, 2019).

Johansson-åkhe, I., Mirabello, C. & Wallner, B. InterPepRank: assessment of docked peptide conformations by a deep graph network. Front. Bioinform. 1, 763102 (2021).

Article
 
Google Scholar

Johansson-åkhe, I. & Wallner, B. InterPepScore: a deep learning score for improving the FlexPepDock refinement protocol. Bioinformatics 38, 3209–3215 (2022).

Article
 
Google Scholar

Linsley, D. et al. Learning long-range spatial dependencies with horizontal gated recurrent units. Adv. Neural Inf. Process Syst. 31, 152–164 (2018).

Google Scholar

Satorras, V. G., Hoogeboom, E. & Welling, M. E(n) equivariant graph neural networks. In International Conference on Machine Learning 9323–9332 (PMLR, 2021).

Fuchs, F., Worrall, D., Fischer, V. & Welling, M. SE(3)-transformers: 3D roto-translation equivariant attention networks. In 34th Conference on Neural Information Processing Systems (NeurIPS 2020) https://papers.neurips.cc/paper/2020/file/15231a7ce4ba789d13b722cc5c955834-Paper.pdf (NeurIPS, 2020).

Baek, M. et al. Accurate prediction of protein structures and interactions using a three-track neural network. Science 373, 871–876 (2021).

Article
 
Google Scholar

Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021).

Article
 
Google Scholar

Evans, R. et al. Protein complex prediction with AlphaFold-Multimer. Preprint at bioRxiv https://doi.org/10.1101/2021.10.04.463034 (2021).

Abramson, J. et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493–500 (2024).

Article
 
Google Scholar

Lin, Z. et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science 379, 1123–1130 (2023).

Article
 
MathSciNet
 
Google Scholar

Madani, A. et al. Large language models generate functional protein sequences across diverse families. Nat. Biotechnol. 41, 1099–1106 (2023).

Article
 
Google Scholar

Brandes, N., Ofer, D., Peleg, Y., Rappoport, N. & Linial, M. ProteinBERT: a universal deep-learning model of protein sequence and function. Bioinformatics 38, 2102–2110 (2022).

Article
 
Google Scholar

Yang, K. K., Fusi, N. & Lu, A. X. Convolutions are competitive with transformers for protein sequence pretraining. Cell Syst. 15, 286–294 (2024).

Article
 
Google Scholar

Xu, X. & Bonvin, A. M. DeepRank-GNN-esm: a graph neural network for scoring protein–protein models using protein language model. Bioinform. Adv. 4, vbad191 (2024).

Article
 
Google Scholar

Mariani, V., Biasini, M., Barbato, A. & Schwede, T. lDDT: a local superposition-free score for comparing protein structures and models using distance difference tests. Bioinformatics 29, 2722–2728 (2013).

Article
 
Google Scholar

Zhang, L. et al. ComplexQA: a deep graph learning approach for protein complex structure assessment. Brief. Bioinform. 24, bbad287 (2023).

Article
 
Google Scholar

Basu, S. & Wallner, B. DockQ: a quality measure for protein–protein docking models. PLoS ONE 11, e0161879 (2016).

Article
 
Google Scholar

Chen, X., Morehead, A., Liu, J. & Cheng, J. A gated graph transformer for protein complex structure quality assessment and its performance in CASP15. Bioinformatics 39, i308–i317 (2023).

Article
 
Google Scholar

Yang, Z., Zhong, W., Lv, Q. & Dong, T. Geometric Interaction Graph Neural Network for predicting protein–ligand binding affinities from 3D structures (GIGN). J. Phys. Chem. Lett. 14, 2020–2033 (2023).

Article
 
Google Scholar

Berman, H. M. et al. The Protein Data Bank. Nucleic Acids Res. 28, 235–242 (2000).

Article
 
Google Scholar

Bresson, X. & Laurent, T. Residual gated graph ConvNets. Preprint at https://arxiv.org/abs/1711.07553 (2017).

Hauser, A. S. & Windshügel, B. LEADS-PEP: a benchmark data set for assessment of peptide docking performance. J. Chem. Inf. Model. 56, 188–200 (2015).

Article
 
Google Scholar

London, N., Movshovitz-Attias, D. & Schueler-Furman, O. The structural basis of peptide–protein binding strategies. Structure 18, 188–199 (2010).

Article
 
Google Scholar

Shanker, S. & Sanner, M. F. Predicting protein–peptide interactions: benchmarking deep learning techniques and a comparison with focused docking. J. Chem. Inf. Model. 63, 3158–3170 (2023).

Article
 
Google Scholar

Lee, J. H., Yin, R., Ofek, G. & Pierce, B. G. Structural features of antibody–peptide recognition. Front. Immunol. 13, 910367 (2022).

Article
 
Google Scholar

Su, M. et al. Comparative assessment of scoring functions: the CASF-2016 update. J. Chem. Inf. Model. 59, 895–913 (2019).

Article
 
Google Scholar

Buttenschoen, M., Morris, G. M. & Deane, C. M. PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences. Chem. Sci. 15, 3130–3139 (2023).

Article
 
Google Scholar

Santos, K. B., Guedes, I. A., Karl, A. L. & Dardenne, L. E. Highly flexible ligand docking: benchmarking of the DockThor program on the LEADS-PEP protein–peptide data set. J. Chem. Inf. Model. 60, 667–683 (2020).

Article
 
Google Scholar

Loshchilov, I. & Hutter, F. Decoupled weight decay regularization. Preprint at https://arxiv.org/abs/1711.05101 (2017).

Janin, J. et al. CAPRI: a critical assessment of predicted interactions. Proteins 52, 2–9 (2003).

Article
 
Google Scholar

Shen, C. et al. Boosting protein–ligand binding pose prediction and virtual screening based on residue-atom distance likelihood potential and graph transformer. J. Med. Chem. 65, 10691–10706 (2022).

Article
 
Google Scholar

Lin, T. Y., Goyal, P., Girshick, R., He, K. & Dollar, P. Focal loss for dense object detection. IEEE Trans. Pattern Anal. Mach. Intell. 42, 318–327 (2020).

Article
 
Google Scholar

Wen, Z., He, J., Tao, H. & Huang, S. Y. PepBDB: a comprehensive structural database of biological peptide–protein interactions. Bioinformatics 35, 175–177 (2019).

Article
 
Google Scholar

Steinegger, M. & Söding, J. MMseqs2 enables sensitive protein sequence searching for the analysis of massive data sets. Nat. Biotechnol. 35, 1026–1028 (2017).

Article
 
Google Scholar

Tao, H., Wang, X. & Huang, S. Y. An interaction-derived graph learning framework for scoring protein–peptide complexes. Zenodo https://doi.org/10.5281/zenodo.17097750 (2025).

Tao, H., Wang, X. & Huang, S. Y. GraphPep program. Zenodo https://doi.org/10.5281/zenodo.17099863 (2025).

Download references

Acknowledgements

This work is supported by the National Natural Science Foundation of China (grant nos. 32161133002, 32430020 and 62072199), the Major Project of Guangzhou National Laboratory (GZNL2023A0300) and the startup grant of Huazhong University of Science and Technology.

Author information
Authors and Affiliations

School of Physics and Key Laboratory of Molecular Biophysics of MOE, Huazhong University of Science and Technology, Wuhan, People’s Republic of China

Huanyu Tao, Xiaoyu Wang & Sheng-You Huang

Contributions

S.-Y.H. conceived and supervised the project. H.T. performed the experiments. S.-Y.H. and H.T. analysed the data. H.T. and X.W. tested the program. H.T. and S.-Y.H. wrote the manuscript. All authors read and approved the final version of the manuscript.

Corresponding author

Correspondence to Sheng-You Huang.

Ethics declarations
Competing interests

The authors declare no competing interests.

Peer review
Peer review information

Nature Machine Intelligence thanks Jianyi Yang and the other, anonymous, reviewer(s) for their contribution to the peer review of this work.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary information
Supplementary Information

Supplementary Figs. 1–7.

Reporting Summary
Supplementary Tables

Supplementary Tables 1–11.

Supplementary Data 1

Source data for supplementary figures.

Source data
Source Data Fig. 2

Statistical source data.

Source Data Fig. 3

Statistical source data.

Source Data Fig. 4

Statistical source data.

Source Data Fig. 5

Statistical source data.

Rights and permissions

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.

Reprints and permissions

About this article
Cite this article

Tao, H., Wang, X. & Huang, SY. An interaction-derived graph learning framework for scoring protein–peptide complexes. Nat Mach Intell 7, 1858–1869 (2025). https://doi.org/10.1038/s42256-025-01136-1

Download citation

Received
25 December 2024

Accepted
17 September 2025

Published
23 October 2025

Version of record
23 October 2025

Issue date
November 2025

DOI
https://doi.org/10.1038/s42256-025-01136-1

Share this article

Anyone you share the following link with will be able to read this content:

Get shareable link

Provided by the Springer Nature SharedIt content-sharing initiative

Subjects
Machine learning
Protein structure predictions
Software
This article is cited by
AFP-GFuse: an antifungal peptide identification model with structural information fusion via multi-graph neural networks and cross-attention mechanism
Xiaomeng LinRuiqi LiuZilong Zhang

Molecular Diversity (2025)

Cancel

Preview · live