pocketz.doi.bio/2026-01-11_181734_nature_s42256-025-01077-9

Protein–peptide docking with a rational and accurate diffusion generative model | Nature Machine Intelligence

URL: https://www.nature.com/articles/s42256-025-01077-9

Saved: 2026-01-11T02:17:39.930Z

Article Published: 04 August 2025 Protein–peptide docking with a rational and accurate diffusion generative model Huifeng Zhao, Odin Zhang, Dejun Jiang, Zhenxing Wu, Hongyan Du, Xiaorui Wang, Yihao Zhao, Yuansheng Huang, Jingxuan Ge, Tingjun Hou & Yu Kang 

Nature Machine Intelligence volume  7,  pages 1308–1321 (2025)Cite this article

5225 Accesses

5 Citations

2 Altmetric

Metrics details

Abstract

Therapeutic peptides represent the forefront of drug discovery, offering potent and safe alternatives to traditional small molecules. However, their weak and context-dependent nature complicates the efficient virtual screening and structural characterization of protein–peptide patterns. Here we introduce RAPiDock, a diffusion generative model designed for rational, accurate and rapid protein–peptide docking at an all-atomic level. RAPiDock efficiently reduces the sampling space by incorporating physical constraints and uses a bi-scale graph to effectively capture multidimensional structural information while balancing efficiency. In addition, the model uses a Clebsch–Gordan tensor product-based architecture to ensure physical symmetry. RAPiDock outperforms existing tools in prediction of protein–peptide-binding patterns, achieving a 93.7% success rate at top-25 predictions (13.4% higher than AlphaFold2-Multimer), with an execution speed of 0.35 seconds per complex (~270 times faster than AlphaFold2-Multimer). Extensive experiments demonstrate RAPiDock’s remarkable ability to handle 92 types of residue including posttranslational modifications, accurately predict subtle docking patterns, successfully identify multiple potential peptide-binding sites in global docking and serve as a powerful tool for high-throughput virtual screening with structural precision. All these push the boundaries of efficient protein–peptide docking in multiple real-application scenarios.

You have full access to this article via your institution.

Download PDF Similar content being viewed by others Topological deep learning for enhancing peptide-protein complex prediction Article Open access 12 November 2025 Boosting AlphaFold protein tertiary structure prediction through MSA engineering and extensive model sampling and ranking in CASP16 Article Open access 17 November 2025 The power and pitfalls of AlphaFold2 for structure prediction beyond rigid globular proteins Article 21 June 2024 Main

Protein–peptide interactions (PPIs) are ubiquitous in cellular environments and play critical roles in numerous biological processes1. It is believed that up to 40% of all protein–protein interactions in cells, particularly those involving intrinsically disordered regions2,3, are facilitated through PPIs or peptide-like interactions4. Owing to their reduced immunogenicity5 and lower production costs6, peptides are frequently used as potent modulators in protein-related interactions in various biotechnological applications including drug delivery7,8, medical imaging9,10, therapeutic agents11,12, protein-targeted warheads13 and other applications14.

Understanding the molecular mechanisms of protein–peptide recognition is fundamental for developing peptide-based therapeutics15. However, characterizing the intricate structures of protein–peptide complexes poses substantial challenges due to their weak, transient and context-dependent interactions, especially with oligopeptides (2 to approximately 20 amino acids)16. This dynamic nature complicates traditional structural characterization methods such as X-ray crystallography, nuclear magnetic resonance and electron microscopy. As per the statistics, out of more than 223,000 structural data entries in the Research Collaboratory for Structural Bioinformatics (RCSB) database, only about 5.1% are protein–peptide complexes, notably fewer than protein–protein (~49.4%) and protein–small-molecule (~19.6%) complexes.

As important alternatives to complement experimental technologies, various computational approaches have been developed to predict binding patterns of protein–peptide complexes17. Traditional docking programs generally use search algorithms to explore potential binding conformations and use scoring functions to evaluate and rank them. These programs can be subdivided into global and local docking on the basis of whether the binding site is unknown or known. The former, including programs such as HPEPDOCK18, pepATTRACT19 and PIPER-FlexPepDock20, exhaustively search the entire protein surface for potential binding sites, while local docking programs, such as AutoDock CrankPep (ADCP)21, HADDOCK22 and MDockPeP23, focus on predefined regions. Although classic docking can yield helpful results for protein–small-molecule complexes24, it is, not surprisingly, less effective in predicting accurate protein–peptide-binding patterns25. This limitation arises from the weak-affinity-driven nature of PPIs and the intrinsic structural flexibility of peptides, which leads to an exploding conformational space that challenges traditional search algorithms and scoring functions.

Building on the revolutionary achievements of AlphaFold2 (ref. 26) in protein structure prediction, researchers have turned to deep learning algorithms to forecast the binding patterns of PPIs. A previous study innovatively connected proteins and peptides with polyglycine in AlphaFold2, framing PPI binding prediction as a protein folding problem27. AlphaFold2-Multimer (AF2Multi) has further provided fresh avenues for predicting protein–peptide complexes28,29,30, but leans on multiple sequence alignments and requires simultaneous structure predictions for both proteins and peptides, making it time-consuming for large-scale applications such as high-throughput virtual screening. In addition, AF2Multi follows a regression paradigm; it predicts only a single ‘ground state’ structure for each protein–peptide pair, which may overlook structural flexibility and hinder accurate PPI predictions to some extent. In 2022, OmegaFold31 was proposed to predict protein structures without multiple sequence alignments, offering faster predictions, although its PPI prediction performance was demonstrated to be limited in a previous benchmark study29. In 2024, a study demonstrated that AF2Multi’s predictive accuracy for protein–intrinsically disordered region complexes could be improved using refined strategies, such as delineating the interaction region into fragments of decreasing size32. The recent AlphaFold3 (AF3) has sparked considerable excitement33. It is reported to exhibit enhanced proficiency in predicting a broader range of complex structures, and its reduced reliance on multiple sequence alignments results in quicker predictions.

In this study, we propose RAPiDock, a diffusion generative model for rational, accurate and rapid protein–peptide docking at the all-atomic level (Fig. 1). The workflow of RAPiDock is illustrated in Fig. 1a, where peptide sequences and protein structures are input to generate a distribution of multiple peptide-binding conformations as output. Specifically, RAPiDock defines a diffusion process of peptide over the degrees of freedom involved in docking: global translations, global rotations and torsions of both the backbones and side chains of peptides (Fig. 1d). These treatments effectively reduce the sampling space of peptides and make the conformations more realistic by introducing additional physical constraints. A bi-scale graph is introduced for peptide representation at both residual and atomic levels. By executing the learned diffusion process, RAPiDock iteratively transforms the conformations of peptides from an uninformed, noisy prior distribution to the learned, ordered distribution (Fig. 1d,f). In addition, we introduce a peptide initial-structure builder that transforms one-dimensional peptide sequence into a standardized initial three-dimensional structure. This builder ensures that identical amino acid types across different peptides adhere to optimal bond length and angle parameters, which represent the most probable distributions in crystallographic data (Fig. 1e, top). These strategies ensure the capture of intricate interactions at various levels while maintaining excellent structural rationality, compared with distribution of realistic dihedrals (Fig. 1e, bottom).

Fig. 1: The architecture and workflow of RAPiDock.

a, The docking workflow of RAPiDock, where the peptide sequence and protein structures are required as input and a distribution over multiple peptide-binding conformations as output. RAPiDock introduces a bi-scale graph for peptide representation at both residual and atomic level. Residue representations facilitate efficient message passing and enable the prediction of global translational and rotational scores; atomic representations, which combined with updated residue-scale features, are used for predicting the torsional scores of backbones and side chains. b, The construction workflow of the high-quality and large-scale RefPepDB used for model training and testing. c, The process of constructing standardized peptide structures for model training. d, Four types of denoising process. e, Top: identical amino acids across different peptides are unified to consistent bond length and angle parameters. Bottom: RAPiDock generates realistic peptide structures. The asterisk denotes that RAPiDock was used here without the customized FastRelax procedure. f, RAPiDock iteratively denoises the peptide from a chaotic conformation into an ordered peptide conformation that matches the dihedral distribution of the crystal structure and docks the peptide from a random spatial space into binding pockets of the protein. t-SNE, t-distributed stochastic neighbour embedding; PPII, polyproline II helix.

Full size image

RAPiDock surpasses all other state-of-the-art models in terms of both precision and efficiency, achieving a success rate of 93.7% in predicting peptide-binding conformations at the top-25 predictions for RefPepDB-RecentSet (13.4% higher than AF2MultiRS), while maintaining a rapid execution time of 0.35 seconds per complex (~270 times faster than AF2MultiRS). Extensive case studies provide compelling evidence that RAPiDock can handle 92 types of residue including posttranslational modifications (PTMs), and capture the subtle structure information of protein pockets. Furthermore, RAPiDock is able to successfully identify multiple potential peptide-binding pockets within proteins, and serve as a powerful tool for high-throughput virtual screening of peptides with structural precision.

Results RAPiDock yields accurate and robust PPI prediction

RAPiDock’s performance in local docking was evaluated on RefPepDB-RecentSet (containing 523 PPI complexes) and compared with two outstanding baseline models: AF2Multi and ADCP. ADCP was selected for its top local docking performance among traditional docking programs, as noted in a previous assessment25. AF2Multi was chosen for its notable accuracy in predicting protein–peptide complexes29. Our previous study indicated that AF2Multi’s accuracy decreased when predicting more than two chains34. Therefore, we evaluated both AF2MultiCS (using sequences of all protein chains within 20 Å from the peptide) and AF2MultiRS (using sequences of specifically selected interacting protein chains), with detailed comparisons in Supplementary Note 1 and Supplementary Fig. 1.

RAPiDock’s docking accuracy

As depicted in Fig. 2a–f, RAPiDock outperformed other competitive methods, achieving the best performance across all three critical assessment of predicted interactions for peptide–protein complexes (CAPRI-peptide) standards. Under the acceptable standard (Fig. 2a,d), RAPiDock achieved an 81.3% success rate considering the top-ranked conformation, and reached 93.7% at the top-25 level, exceeding AF2MultiRS by 13.4% (Fisher’s exact test, P = 9.68 × 10−11). Under the medium standard (Fig. 2b,e), RAPiDock achieved success rates of 61.0%, 68.6% and 71.7% at the top-1, top-5 and top-25 levels, respectively, surpassing AF2MultiRS (56.4%, 58.7%, 61.8%) and ADCP (33.1%, 49.5%, 57.6%). Under the high-quality standard (Fig. 2c,f), RAPiDock improved more notably, achieving a top-100 success rate of 53.3%, which was 17.5% higher than AF2Multi_RS (Fisher’s exact test, P = 1.38 × 10−8) and 29.8% higher than ADCP (Fisher’s exact test, P = 2.53 × 10−23).

Fig. 2: Superior accuracy of RAPiDock for prediction of PPI patterns.

a–c, The top-N success rates of various methods based on the CAPRI-peptide standards in the RefPepDB-RecentSet. d–f, The top-1, top-5 and top-25 success rates of various methods in the RefPepDB-RecentSet. According to CAPRI-peptide standards, three accuracy levels are considered, namely, acceptable quality (a,d), medium quality (b,e) and high quality (c,f). g, Comparison of the DockQ scores of top-1 predictions of RAPiDock and AF2MultiRS using the RefPepDB-RecentSet, with the y axis representing DockQRAPiDock and the x axis for DockQAF2MultiRS. h, Comparison of the DockQ scores of top-1 predictions of RAPiDock and AF3 using the RefPepDB-RecentSet4AF3, with the y axis representing DockQRAPiDock and the x axis for DockQAF3. For g and h, the light purple area in the top left corner indicates complexes where DockQRAPiDock is ≥0.8 and DockQAF2multiRS (or DockQAF3) is <0.8, signifying superior RAPiDock performance. Conversely, the light blue area in the bottom right shows complexes where DockQAF2multiRS (or DockQAF3) is ≥0.8 and DockQRAPiDock is <0.8. Each shaded area also represents the proportion of complexes within these performance ranges. Upper% indicates the ratio of complexes where DockQRAPiDock surpasses DockQAF2multiRS (or DockQAF3). i, Cumulative frequency curves of DockQ scores of RAPiDock, AF3 and AF2MultiRS, using the RefPepDB-RecentSet4AF3.

Source data

Full size image

Moreover, RAPiDock not only exhibited excellent top-1 performance but also showed rather obvious improvement at the top-25 and top-100 levels, which indicates that this generative model has the capability to sample a diverse set of high-quality peptide conformations. In contrast, the conformations predicted by AF2Multi were relatively similar29, making its conformation sampling yield rather limited improvements.

DockQ, a continuous value metric, was further used on each PPI complex to assess the docking performance of RAPiDock and other baseline methods. A DockQ score above 0.8 is typically considered high quality. As shown in Fig. 2g and Supplementary Fig. 2, for far more complexes, RAPiDock yielded DockQ scores higher than AF2MultiRS (63.29%, 70.94% and 72.85% complexes at the top-1, top-5 and top-25 levels, respectively). The cumulative frequency curve of these DockQ scores further illustrates RAPiDock’s superiority over AF2MultiRS and ADCP (Supplementary Fig. 3).

RAPiDock’s robustness from geometric feature perspectives

We further conducted a comprehensive exploration from three structural aspects: peptide length, binding pocket depth and peptide secondary structure compositions, to evaluate the RAPiDock’s robustness. The performance of the traditional docking method ADCP decreased generally with increasing peptide length because of the exploding search space of longer peptides35, while the artificial intelligence-based methods can be less sensitive to it. Across almost all evaluated peptide lengths, RAPiDock demonstrated higher DockQ scores than both AF2MultiRS (Extended Data Fig. 1a) and ADCP (Supplementary Fig. 4a). Regarding the depth of the binding pocket, for all models the performance dropped with shallower binding pockets. However, this influence was minimal on RAPiDock, as its average DockQ scores were above 0.8 across all percentage of interface accessible surface area (ASA%) subsets, whereas those for AF2MultiRS notably dropped with an ASA% less than 52% (Extended Data Fig. 1b). As shown in Extended Data Fig. 1c and Supplementary Fig. 4c, different secondary structure categories seem to have little effect on the docking performance of RAPiDock, which performed stably best, with average DockQ scores consistently above 0.9.

Comparison with AlphaFold3

As the latest version of the AlphaFold series, AF3 is highly anticipated for its prediction performance on PPI patterns. To avoid overlaps in the training datasets of AF3 and our RefPepDB-RecentSet, we finally retained 228 protein–peptide complexes for this comparison, which called RefPepDB-RecentSet4AF3. We compare the server-based version and the local version of AF3, and no significant difference in performance between the two AF3 versions was found (details are provided in Supplementary Note 3 and Supplementary Fig. 8).

As shown in Fig. 2h and Supplementary Fig. 9, RAPiDock outperformed AF3 notably, with 57.02% and 61.84% of the top-1 and top-5 results achieving higher DockQ scores, respectively. Notably, 20.61% of complexes at the top-5 level had DockQ scores below 0.8 in AF3 but above 0.8 in RAPiDock, while only 8.77% showed the opposite. The cumulative frequency curve of these DockQ scores further illustrates RAPiDock’s superiority over AF3 (Fig. 2i). More importantly, we analysed the impact of the similarity between the test dataset and the training dataset on model performance. As shown in Supplementary Fig. 11, when peptide sequence and protein structure similarities fell below 80%, RAPiDock maintained a remarkable success rate of 94.74% in sampling peptide conformations that met acceptable standards, while AF3 had a notably lower success rate of 57.89%. This underscores RAPiDock’s outstanding robustness when applied to data with low similarity. Further analysis stratified by structural similarity (Supplementary Fig. 13) revealed that, although RAPiDock’s performance exhibited an observable decline with decreasing similarity, it remained statistically comparable to ADCP under low-similarity conditions (for example <50%, details are provided in Supplementary Note 8). These results indicate that, under specific low-similarity conditions, conventional approaches such as ADCP still offer competitive accuracy.

RAPiDock shows high accuracy and robustness on challenging tasks

In the cases when only the unbound or other-ligand-bound protein structure is available, this necessitates conducting local docking based on expert knowledge-derived target information or even global docking, greatly increasing the difficulty36. To address this, we used PepSet (containing 185 PPI complexes) as the test set and performed peptide local and global docking on unbound protein structures. For local docking, four traditional docking tools (ADCP, GOLD, HADDOCK, MDockPeP) were used for comparison. For global docking, HPEPDOCK was selected, as it showed the best global docking performance in ref. 25. RAPiDocklocal and RAPiDockglobal represent two different RAPiDock models trained using protein pockets and intact proteins, respectively.

As shown in Extended Data Fig. 2, both RAPiDock models outperformed all tested traditional docking tools under all three CAPRI-peptide standards for both local and global docking. For local docking, under the acceptable standard (Extended Data Fig. 2a,d), RAPiDocklocal achieved success rates of 58.9%, 75.7% and 82.2% at the top-1, top-5 and top-25 levels, respectively, which were 41.1% (Fisher’s exact test, P = 3.01 × 10−16), 33.5% (Fisher’s exact test, P = 7.21 × 10−11) and 20.0% (Fisher’s exact test, P = 2.58 × 10−5) higher than ADCP, the best performing traditional method. Under medium and high-quality standards (Extended Data Fig. 2b,c), the top-100 success rates of RAPiDocklocal are 48.6% and 17.3%, respectively, with a minimum improvement of 28.1% (Fisher’s exact test, P = 1.80 × 10−8) and 15.7% (Fisher’s exact test, P = 1.23 × 10−7) compared with traditional methods. For global docking, as shown in Extended Data Fig. 2a–c, the top-100 success rates of RAPiDockglobal slightly dropped compared with RAPiDocklocal, but notably outperformed HPEPDOCK under all standards.

We also tested the three geometric features, namely, peptide length, binding pocket depth and the secondary structure composition. Still, RAPiDock demonstrated outstanding docking accuracy from all three aspects, notably surpassing ADCP or HPEPDOCK for local or global docking (Supplementary Figs. 16 and 17). Regarding the depth of the binding pockets, as shown in Supplementary Figs. 16b and 17b, all methods demonstrated better docking performance for deeper pockets compared with shallower ones. It may suggest that deeper binding pockets provide a more favourable environment for the prediction of binding peptides, together with the increased interaction surface area and structural constraints that facilitate more precise orientation and positioning of the peptide within the pocket.

RAPiDock generates realistic peptide structures

Peptides, composed of amino acids connected via amide bonds, adhere to specific structural rules as indicated by, for example, classical Ramachandran plots. Therefore, it is critical to investigate whether generated peptide structures align with realistic distributions to ensure the validity of the performance. In this study, we conducted a detailed analysis of the peptide structures generated by RAPiDock and other methods.

As shown in Fig. 1f, RAPiDock begins by initializing the peptide structure using the peptide initial-structure builder module, positioning it at the coordinate origin with one of three default secondary structures: extended (Extend), polyproline II helix or α-helix (Helix). These initial structures exhibit single-peaked dihedral angle distributions due to default dihedrals. Subsequently, RAPiDock introduces noise through four distinct forward diffusion processes: translation, rotation, backbone torsion and side-chain torsion, until it reaches time T, resulting in random conformations with chaotic dihedral distributions. Then, RAPiDock initiates the denoising operation, marking the commencement of the peptide docking process. The denoising phase guides the peptide into the protein’s binding pocket, progressively restoring orderly dihedral angles. Finally, a customized FastRelax procedure could be applied to fine-tune the docked peptide structure into an empirical force field precision.

To verify the realism of RAPiDock-generated peptides, we compared the backbone dihedral angles φ, ψ and ω distributions of 708 peptide structures of top docking results to corresponding crystal structures. Figure 3a–c shows that RAPiDock’s dihedral distributions almost exactly recapitulate those of crystal structures, with Kullback–Leibler (KL) divergence values all below 0.1. Furthermore, Ramachandran plots37 (Fig. 3d,e) demonstrate that RAPiDock accurately captures the correlated dihedral angle distributions observed in crystal structures, concentrating in the characteristic regions of the upper-left quadrant between (−160°, +180°) and (−50°, +90°) and bottom-left quadrant between (−120°, 0) and (−50°, −70°). RAPiDock-generated structures with post-FastRelax optimization (Supplementary Fig. 20) exhibit distributions that are highly consistent with crystal data, which are comparable to or better than AF2Multi_RS (with structural optimization by default) and ADCP results (Supplementary Figs. 21 and 22), outperforming most protein–peptide docking methods (Supplementary Fig. 23). Figure 3f further shows that varying initial secondary structures has negligible impact on the RAPiDock’s performance across different peptide secondary structure categories. This suggests that the RAPiDock’s superior performance is not contingent on the accuracy of the initial peptide structure, underscoring its robustness and broad applicability.

Fig. 3: Structural validity analysis of RAPiDock docking results (without using customized FastRelax procedure).

a–c, The distribution of the dihedral angles of φ (a), ψ (b) and ω (c) from the top-1 predicted structures of RAPiDock (shown in blue), compared with that of crystal peptide structures (shown in pink). The asterisk denotes that RAPiDock was used here without the customized FastRelax procedure. The samples include all non-glycine and non-proline residues from 523 + 185 complexes across the RefPepDB-Recent and PepSet datasets. KL divergence values are shown above the plots. d, A two-dimensional Ramachandran plot of the φ and ψ from crystal peptide structures. Two-dimensional dihedral angle distribution of amino acids in the crystal peptide structure is predominantly concentrated in two regions, which are located in the upper-left (φ, ψ) quadrant between (−160°, +180°) and (−50°, +90°) and in the bottom-left quadrant centred between (−120°, 0°) and (−50°, −70°). e, A two-dimensional Ramachandran plot of the φ and ψ from top-1 predicted structures of RAPiDock, where contour lines represent the frequency distribution of these angles in the corresponding crystal structures. Blue points indicate preferred regions, green points denote allowed regions and red points show areas deviating from the distribution observed in the crystal structures. f, The effect of initial structures on RAPiDock docking performance. Mix indicates the combined use of extended (Extend), α-helix (Helix) and polyproline II helix (PPII) as initial structures. The box plots show the 25–75% confidence intervals (box limits), the median (centre line) and the 5–95% confidence intervals (whiskers). Green triangles indicate the mean values. N values report the number of clusters in each band.

Source data

Full size image RAPiDock accurately captures pocket structural information

Given that RAPiDock is an all-atom, structure-based protein–peptide docking method, we sought to evaluate both its overall performance from a statistical perspective and its capability to accurately capture relevant structural features in real-world applications. To this end, we opted to further evaluate the performance of RAPiDock on specific targets and compared it with that of state-of-the-art approaches such as AF2Multi and ADCP. The protein–peptide complexes for chosen targets were strictly excluded from RAPiDock’s training set to ensure an unbiased assessment.

RAPiDock accurately captures PPI patterns with holo PLK1-PBD

Polo-like kinase 1 (PLK1) plays a crucial role in mitosis and is associated with various cancers38. Its Polo-box domain (PLK1-PBD)39,40,41 is a shallow binding site more suitable for peptides than small molecules (more details in Supplementary Note 6.1). Multiple crystal structures of PLK1-PBD bound with different peptides indicate the subtle structural difference on binding (Supplementary Table 2). We examined RAPiDock’s ability to distinguish among different side-chain orientations in PLK1-PBD and to generate correct peptide poses for two representative crystal structures (that is, Protein Data Bank (PDB) IDs 8JOY and 4E67) (Fig. 4a–c) (more details in Supplementary Note 6.2).

Fig. 4: The docking performance of RAPiDock on PLK1-PBD.

a, The two crystal structures of PLK1-PBD, PDB IDs 8JOY and 4E67, with different side-chain orientations of Arg516 and Tyr417, resulting in varied peptide-binding modes. The 8JOY and 4E67 proteins are coloured green and yellow, respectively; with their binding peptides coloured white and grey, respectively. A black circle indicates where the elongated peptide in 4E67 is poised to collide with the Arg516 residue in 8JOY. The black arrow shows that, in 4E67, the Arg516 and Tyr417 residues form a relatively ‘open’ state, with their side chains projecting outwards to accommodate a peptide with more elongated conformations. The red arrow shows that, in 8JOY, these side chains exhibit an inwards orientation, thereby reducing the volume of the binding pocket and causing the peptide to adopt a more curled conformation. b, The interaction visualization in 8JOY, where the peptide in a curled conformation leads to increased interactions between the N-terminal proline and the inward-pointing Arg516. c, The interaction visualization in 4E67, where the conserved phosphorylated threonine (coloured coral red) remains immobilized by salt bridges and serves as an anchor for binding, and the N-terminal residues present more diversity to fit in the binding pocket with multiple conformations. d, RAPiDock-generated peptide conformation (second ranked) reproduces the binding patterns suggested by 8JOY. e, RAPiDock-generated peptide conformation (top ranked) reproduces the binding patterns suggested by 4E67. The peptide conformations generated by RAPiDock are coloured coral red. HP, hydrophobic interactions; HB, hydrogen bonds; SB, salt bridges; π–π, π–π stacking bonds.

Full size image

RAPiDock successfully reproduced the anchor interactions involving the phosphorylated threonine and key side-chain rearrangements in PLK1-PBD. Notably, for PDB 8JOY, the second-ranked conformation already achieved high-quality peptide binding by CAPRI-peptide standards, accurately reflecting the curled conformation of the peptide’s N-terminal residues (Fig. 4d). For 4E67, the top-ranked pose captured a more extended peptide arrangement along with additional stacking interactions (Fig. 4e). These results highlight RAPiDock’s capacity to account for subtle conformational shifts (more details in Supplementary Note 6.3). A comparison with AF2MultiRS and ADCP further underscores RAPiDock’s superior performance and efficiency (Supplementary Table 3). Notably, RAPiDock can handle PTMs such as phosphothreonine directly, whereas AF2MultiRS and ADCP require separate mutation steps for PTM modelling. Further comparisons with AF2Multi_RS and ADCP are provided in Supplementary Note 6.4.

RAPiDock accurately captures PPI patterns across SH2 domains

Src homology 2 (SH2) domains are critical for phosphotyrosine-dependent signal transduction42 and frequently bind phosphotyrosine-containing peptides with lengths of about ten amino acids43,44. We tested RAPiDock on two SH2 domains of Src homology region 2 domain-containing phosphatase-2 (ref. 45) (SHP2), known as N-SH2 (PDB 5X7B) and C-SH2 (PDB 5X94), which exhibit a high degree of structural consistency but a notable difference in sequence (Extended Data Fig. 3a,b; more details in Supplementary Note 7.1).

As shown in Extended Data Fig. 3c,d RAPiDock’s top-ranked pose for 5X7B and 5X94 matched the crystal structures with high- and medium-quality CAPRI-peptide standards, respectively. In both cases, as shown in Supplementary Fig. 30c,d, RAPiDock effectively replicates the key salt bridge interactions involving phosphotyrosine (detailed analysis is provided in Supplementary Note 7.2).

In comparisons with AF2MultiRS and ADCP, RAPiDock again showed higher accuracy and a faster docking speed (Supplementary Table 4). AF2MultiRS mispredicted the structure of specific pockets, particularly the BG loop, leading to notable shifts in the peptide structure when compared with the crystal structure (Supplementary Figs. 31a,b and 32a,b). ADCP failed to perform reasonable docking for both PDB 5X7B and 5X94 (Supplementary Figs. 31c,d and 32c,d), underlining the limitations of traditional protein–peptide docking tools in SH2 domain tasks. More detailed comparisons with AF2Multi_RS and ADCP are provided in Supplementary Note 7.3.

RAPiDock finds multiple peptide-binding pockets in protein

Many proteins possess more than one distinct binding site to accommodate peptides. RAPiDock exhibits the capability to capture multiple pockets and to provide correct binding mode with efficiency, demonstrating a strong advantage for challenging multi-site protein targets. Importin-α46, which recognizes nuclear localization signals, features both a major and a minor binding pocket47 (Fig. 5a). RAPiDock was applied in a global docking mode on importin-α, successfully identifying high-quality poses in both pockets within the top-5 predictions (Fig. 5b). Specifically, the top-ranked pose hit the minor pocket and met the high-quality CAPRI-peptide standard, and the fifth-ranked pose hit the major pocket and also met the high-quality CAPRI-peptide standard. By contrast, AF2Multi_RS located only the major pocket and produced a medium-quality pose (Fig. 5c), while HPEPDOCK did identify both pockets but struggled to sample high-quality conformations and to provide correct ranking (Fig. 5d).

Fig. 5: RAPiDock finds multiple peptide-binding pockets of importin-α.

a, Importin-α contains two distinct peptide-binding pockets. The major pocket and its crystal binding peptide are coloured coral red, while the minor pocket and its crystal binding peptide are coloured blue. b, RAPiDock successfully predicts peptide-binding conformations in both two pockets with high accuracy. The peptide conformations generated by RAPiDock are coloured orange. c, AF2MultiRS only predicts peptide-binding conformations in major pockets. The protein and peptide structures generated by AF2MultiRS are coloured white and purple. d, HPEPDOCK predicts peptide-binding conformations in both two pockets with limited accuracy. Specifically, for the major and minor binding pockets, HPEPDOCK’s best peptide predictions ranked 83th and 90th, respectively, achieving only medium and acceptable quality CAPRI-peptide standards. The peptide conformations generated by HPEPDOCK are coloured grey.

Full size image RAPiDock enables virtual screening of pHLA systems

We further tested RAPiDock’s use in peptide–human leukocyte antigen (pHLA) systems that are central to immunological processes such as antigen presentation48, which is one of the most representative cases of peptide docking. As shown in Fig. 6a and Supplementary Table 6, RAPiDock has demonstrated exceptional accuracy in predicting pHLA binding modes with top-ranked high-quality success rate of 68.9%, surpassing both AF2Multi_RS and ADCP. RAPiDock also achieved the best in all key quality metrics including fraction of native contacts (Fnat), interface root-mean-square deviation (iRMS), ligand root-mean-square deviation (lRMS) and DockQ, demonstrating its superior performance in predicting pHLA binding modes. To illustrate the effectiveness of RAPiDock in modelling peptide binding for HLA targets, we analysed the docking case of PDB 7BBG, as presented in Fig. 6b. In the crystal structure of 7BBG, the native peptide’s terminal ends effectively occupy two key binding pockets of the protein (pocket A and pocket F)49. The top-1 pose generated by RAPiDock closely replicates the binding profiles observed in 7BBG, with a DockQ value of 0.934.

Fig. 6: RAPiDock predicts the binding mode and binding affinity of pHLA at high accuracy.

a, Top-1, top-5 and top-25 success rates of different methods for predicting the binding modes of pHLA. According to CAPRI-peptide standards, three accuracy levels are considered, namely, acceptable quality, medium quality and high quality. b, RAPiDock successfully predicts the binding mode of 7BBG. Two key binding pockets of the HLA are indicated by black arrows. c, The top-ranked peptide conformation predicted by RAPiDock is highly consistent with the peptide structure in the 7BBG crystal structure. d–o, The binding affinity prediction capacities of RAPiDock for 12 different HLA alleles: HLA-A02:01 (d), HLA-A02:06 (e), HLA-A03:01 (f), HLA-A11:01 (g), HLA-A24:02 (h), HLA-A29:02 (i), HLA-B07:02 (j), HLA-B08:01 (k), HLA-B15:01 (l), HLA-B35:01 (m), HLA-B53:01 (n) and HLA-B57:01 (o). The total number of peptides screened for each pHLA allele and the PBD ID of representing protein for each pHLA allele are shown in the left-top corner.

Source data

Full size image

To further assess the validity and effectiveness of the peptide-binding modes generated by RAPiDock, we evaluated the binding affinity for 12 different HLA alleles (Supplementary Fig. 33) by using two typical scoring functions (that is, PyRosetta Ref2015 and Vina score50,51). Both scoring functions suggest that all indicators for RAPiDock outperform those of the template-based methods. As detailed in Fig. 6d–o and Supplementary Tables 7, RAPiDock’s conformations generally led to higher predictive accuracy, with average area under the curve-receiver operating characteristic (AUC-ROC) and AUC-precision recall curve (PRC) values (0.748 and 0.528, respectively) surpassing the template-based method (0.673 and 0.419). Notably, for certain systems such as HLA-B*57:01, RAPiDock achieves an AUC-ROC value of approximately 0.9, highlighting RAPiDock’s promising potential for large-scale virtual screening. Given its rapid and accurate peptide-binding mode sampling capabilities, RAPiDock’s performance heralds new possibilities for future advancements in structure-based pHLA interaction modelling and even for large-scale structure-based peptide screening for specific targets.

Discussion

In this work, we introduced RAPiDock, a diffusion generative model for rational, accurate and rapid protein–peptide docking at all-atomic level. RAPiDock effectively reduces the sampling space of peptide conformations by incorporating a diffusion process over key degrees of freedom, including global translations, rotations and torsions of both peptide backbones and side chains. This process, combined with additional physical constraints, ensures that the generated peptide conformations are both realistic and physically plausible. Moreover, RAPiDock uses a bi-scale graph to capture multidimensional structural information and uses a Clebsch–Gordan (CG) tensor product-based architecture to ensure physical symmetry.

We show that RAPiDock can predict protein–peptide docking patterns with excellent accuracy and high speed, consistently achieving robust docking results across multiple geometric features, and notably outperforms other state-of-the-art approaches including AF2Multi_RS, AF3 and ADCP. Further experiments suggest that RAPiDock can not only successfully capture the subtle structure information of binding pockets and reproduce accurate peptide-binding patterns on various holo structures, but also identify multiple potential peptide-binding pockets within proteins. Finally, we demonstrate that RAPiDock can be a powerful tool for high-throughput virtual screening of peptides with structural precision.

While we demonstrate promising results with RAPiDock, there are certain limitations that motivate opportunities for future work. Although RAPiDock theoretically supports the prediction of peptides composed of up to 92 α-amino acids, its accuracy diminishes when predicting structures of some modified amino acids, such as D-methionine. This limitation primarily stems from the scarcity of crystallographic data available for these amino acids, resulting in insufficient training data for the model. To overcome this challenge, it may be necessary to augment the dataset with more modified amino acids. Another inherent issue with the model is that its speed could be further optimized. As RAPiDock relies on a diffusion model, the iterative nature of the diffusion process can be computationally intensive and time-consuming. It is expected to explore the possibility of reducing the number of required iterations by directly learning a smoother trajectory through the conformational space.

In summary, this work presents RAPiDock as a powerful tool that sets a new standard for protein–peptide docking, offering remarkable accuracy and speed. RAPiDock holds important promise for pushing the boundaries of protein–peptide docking across a wide range of real-world applications.

Methods Peptide initial-structure builder

Almost all proteins or peptides are composed of specific types of amino acid, including 20 canonical amino acids and non-canonical amino acids formed through PTMs or other alterations52. These amino acids not only have different side chains53 but also exhibit subtle variations in their backbone bond lengths and bond angles54. For example, based on the statistical data of amino acids in the crystal structure, the N–Cɑ–C bond angle distributions of glycine, proline, tyrosine and cysteine are each unique (Supplementary Fig. 34). In addition, even within the same type of amino acid, the differences in bond lengths and bond angles among individuals are as unique as fingerprints. As RAPiDock maintains structural integrity by not altering bond lengths and bond angles during the denoising process, variations in bond lengths and bond angles within the same amino acid type could potentially affect the performance of RAPiDock.

Therefore, we constructed a peptide initial-structure builder based on the PeptideBuilder55. This tool aims to eliminate the interference caused by variations in bond lengths and bond angles among different individuals of the same amino acid type on model training and inference, while preserving the differences between different amino acid types by strictly defining the bond lengths and bond angles for each amino acid type (Fig. 1e, top). Specifically, for each amino acid type (t\in T), where T represents the 92 supported amino acid types, we performed kernel density estimation for the ith amino acid individual’s jth bond length or bond angle value ti,j:

$$P({t}{i,\;j})=\frac{1}{{n}{j}{h}{j}\sqrt{2{\mathrm{\uppi}} }}\mathop{\sum }\limits{k=1}^{{n}{j}}{\mathrm{e}}^{(-\frac{{({t}{i,\;j}-{t}{k})}^{2}}{2{h}{j}^{2}})}$$ (1)

where nj represents the number of amino acids that have jth value, hj represents the bandwidth parameter and the P(ti,j) represents the score of ti,j based on the probability distribution function. The average score S(ti) of all bond lengths and angles of ith amino is computed:

$$S({t}{i})=\frac{1}{m}\mathop{\sum }\limits{j=1}^{m}P\left({t}_{i,\;j}\right)$$ (2)

where m represents the total number of bond lengths and angles of the ith amino. The individual t* with the highest score was selected as the representative of this amino acid type:

$${t}^{* }={{\rm{argmax}}}{i\in t}S({t}{i})$$ (3)

Then, the representative amino acids are connected by amide bonds to generate the initial structure of peptides according to the given peptide sequences. The generation of amide bonds is achieved by using predefined backbone dihedral angles, namely, φ, ψ and ω (Supplementary Fig. 35a).

This approach allows our model to support 92 types of α-amino acid, whose geometric properties are collected from peptide structures in RCSB database, as inputs. The specific types of supported amino acid are shown in Supplementary Fig. 35b,c and Supplementary Tables 9 and 10.

Bi-scale graph for multiscale modelling

In many graph learning networks, small molecules are typically represented with atoms as nodes, while macromolecules are represented with residues as nodes56. Therefore, the protein is represented as a residue graph (G({P}{i},{Q}{i})), where P represents the set of residues and Q represents the set of residue–residue edges. Given that the size of peptides lies between that of small molecules and macromolecules, RAPiDock uses both atom-level and residue-level representations for peptides to effectively capture multidimensional information and balance efficiency. In docking tasks, the peptide needs to be docked into the target pocket, followed by searching and optimizing its conformation within the pocket to achieve the optimal result. This process can be decomposed into global structure positioning and local structural optimization. While, in principle, these physical quantities could theoretically be represented by atomic-level information, statistical physics suggests that physical systems follow different laws at varying scale, known as emergent phenomena. From a graph learning perspective, the residue graph is a coarse-grained version of the atom graph. Techniques in recommendation systems have demonstrated that graph coarsening can capture interactions emerging at the group level57. Therefore, we introduced a bi-scale graph representation for peptides, simultaneously modelling both residue and atom graphs to fully exploit the physical laws at different scales. Specifically, for global structure positioning, including translation and rotation predictions of the peptide, we use the residue graph (G({N}{i},{E}{i})) to capture information at the same scale as protein residues. For local structural optimization, we use the atom graph (G({n}{i},{e}{i})) to capture the inductive effects of protein pocket residues on the local conformation of the peptide. To integrate information between the bi-scale graphs, RAPiDock propagates information of (G({N}{i},{E}{i})) to the corresponding atoms within that residue:

$${b}{j}^{i}={B}{i},\forall {n}{j}^{i}\in \left{{n}{j}|{\rm{atoms}}\; {\rm{within}}\; {\rm{residue}}\,{N}_{i}\right}$$ (4)

where Bi represents features of residue Ni, and ({b}_{j}^{i}) represents the features of atom j within residue Ni. The bi-scale graph uses residue feature inheritance, allowing the two-scale represents to achieve information gain from the other scale without compromising the modelling accuracy at each individual scale.

CG tensor product-based equivariant architecture

In 3D structure prediction tasks, it is a useful inductive bias to assign physical equivariance to the model58. Specifically, we focus on roto-translational equivariance59:

$${Rf}\left(x\right)+{\bf{T}}=f\left({Rx}+{\bf{T}}\right)$$ (5)

where x is the input structure, R is a rotation matrix and T is a translational vector; the f satisfying equation (5) is called a roto-translational equivariant neural network. Roto-translational equivariance, also called SE(3) equivariance, can be decomposed into translational equivariance and rotational equivariance, namely, SO(3) equivariance59. It is simple to incorporate translational equivariance into the model by simply aligning the system centre to the original point. To enforce SO(3) equivariance, we used the Clebsch–Gordan (CG) tensor product to align the model with this physical symmetry60.

Equation (5) only describes the roto-translational equivariance of type-1 tensor. For the equivariance of general type-l tensors, we have:

$$f\left({D}^{l}{a}^{l}\right)={D}^{l}f\left({a}^{l}\right)$$ (6)

where al is the l-order feature and Dl is the l-order feature’s Wigner-D matrix, which describes the response of the type-l feature. The translation and rotation tensors are classified as type-1 tensors, whereas the dihedral rotation (also called torsion) tensors are categorized as type-0 tensors61. Ideally, the model should predict type-l tensors that correspond to various physical quantities. However, within the RAPiDock’s architecture, we aim to map lower-order geometric features to higher-order features to better capture the higher-order symmetries and physical properties of the geometric input. To ensure the equivariance of the operations involving the two type-l tensors, we use the CG tensor product for processing:

$$(u{\otimes }{{cg}}v{)}{m}^{\left(l\right)}=\mathop{\sum }\limits{{m}{1}=-{l}{1}}^{{l}{1}}\mathop{\sum }\limits{{m}{2}=-{l}{2}}^{{l}{2}}{C}{\left({l}{1},{m}{1}\right)\left({l}{2},{m}{2}\right)}^{\left(l,m\right)}{u}{{m}{1}}^{\left({l}{1}\right)}{v}{{m}{2}}^{\left({l}_{2}\right)}$$ (7)

where ((u\otimes v{)}{m}^{\left(l\right)}) represents the mth component of the output type-l tensor, ({C}{\left({l}{1},{m}{1}\right)\left({l}{2},{m}{2}\right)}^{\left(l,m\right)}) represents the corresponding CG parameter, and l1 and l2 represent the feature orders of u and v, respectively. The range of l,m is (\left|{l}{1}-{l}{2}\right|\le l\le {l}{1}+{l}{2}), (-l\le m\le l). To fully leverage the fitting capabilities, we use the neural network to parameterize the CG tensor product:

$$(u{\otimes }{{cg}}^{W}v{)}{m}^{\left(l\right)}={w}{m}\mathop{\sum }\limits{{m}{1}=-{l}{1}}^{{l}{1}}\mathop{\sum }\limits{{m}{2}=-{l}{2}}^{{l}{2}}{C}{({l}{1},{m}{1})({l}{2},{m}{2})}^{(l,m)}{u}{{m}{1}}^{({l}{1})}{v}{{m}{2}}^{({l}{2})}$$ (8)

where ({\otimes }_{{cg}}^{W}) is the CG tensor product parameterized by a set of weights wm. At this point we construct the message-passing function in the graph neural network:

$${m}{{ij}}={{h}{i}^{{l}{1}}\otimes }{{cg}}^{W({d}{{ij}}{||}{h}{i}{||}{h}{j})}{Y}^{{l}{2}}\left(\frac{{\bf{e}}{{ij}}}{{||}{\bf{e}}{{ij}}{||}}\right)$$ (9)

where ({h}{i}^{{l}{1}}) is the l1-order tensor of node i; the W and || in (W({d}{{ij}}{\rm{||}}{h}{i}{\rm{||}}{h}_{j})) represent multilayer perceptron and the splicing operation, respectively; Yl2 is the l2-order spherical harmonics function; and eij is the vector of edge between i and j.

Then we conducted message updating and score predicting on the introduced bi-scale graph:

$$G\left({N}{i}^{\prime} ,{E}{i}^{\prime} \right)={\rm{CG}}-{{\rm{MPNN}}}{1}(G\left({N}{i},{E}{i}\right),G\left({P}{i},{Q}{i}\right))$$ (10) $$G\left({P}{i}^{\prime} ,{Q}{i}^{\prime} \right)={\rm{CG}}-{{\rm{MPNN}}}{2}(G\left({N}{i},{E}{i}\right),G\left({P}{i},{Q}{i}\right))$$ (11) $$t,r={\rm{CG}}-{{\rm{Predictor}}}{1}\left(G\left({N}{i}^{\prime} ,{E}{i}^{\prime} \right)\right)$$ (12) $$G\left({n}{i}^{\prime} ,{e}{i}^{\prime} \right)={\rm{AtomEncoder}}\left(G\left({n}{i},{e}{i}\right),G\left({N}{i}^{\prime} ,{E}{i}^{\prime} \right)\right)$$ (13) $${\tau }{b}={\rm{CG}}-{{\rm{Predictor}}}{2}\left(G\left({n}{i}^{\prime} ,{e}{i}^{\prime} \right)\right)$$ (14) $${\tau }{{\rm{s}}}={\rm{CG}}-{{\rm{Predictor}}}{3}\left(G\left({n}{i}^{\prime} ,{e}_{i}^{\prime} \right)\right)$$ (15)

where t, r and τ represent predicted scores of translation, rotation and torsion, respectively. Each of CG-based message-passing neural networks (CG-MPNNs) includes intra-message passing and inter-message passing. AtomEncoder is the module that extracts and updates atom-level features for the bi-scale graph. These three geometric scores correspond to the model’s final predictions: odd type-1, even type-1 and type-0 that include both odd and even parties. This is because translations show odd parity under coordinate inversion, while rotations show even parity. As torsions are type-0 vectors, they behave the same way with even and odd parity. It is worth noting that in RAPiDock, τ is predicted separately as the backbone torsion score τb and the side-chain score τs. Specifically, we used the shared torsional encoder as the feature extractor for τb and τs, and then two independent heads were used to predict τb and τs. This approach ensures the distribution of these two predictions consistent in the shared torsional encoder, but different after the independent head.

Diffusion on geometries

The diffusion model was used as a generation framework to supervise the predicted geometric variables of t, r, τb and τs in the bi-scale graph. The supervision method of the diffusion model can be conceptualized as denoising score matching62, that is, learning the gradient (-{\nabla }{{x}{k}}\log {p}{\theta }({x}{k}|{x}{k+1})) of the probability distribution of the noise added to the geometric variable (\theta \in {t,r,{\tau }{\mathrm{b}},{\tau }_{\mathrm{s}}}) at the k timestep. The probability distribution of the added noise needs to be determined by the space where the different geometric variables θ are located. For the geometric variable t, the endpoint distribution pT(t) is a Gaussian distribution ({\mathscr{N}}{\mathscr{(}}t)):

$${p}_{T}(t)={\mathscr{N}}(t)\propto \exp\left(-\frac{{{||t||}}^{2}}{2}\right)$$ (16)

At this time, the supervisory signal for t is:

$$-{\nabla }{{t}{k}}{\log p}{t}\left({t}{k},|,{t}{k+1}\right)=-{\nabla }{{t}{k}}\log {\mathscr{N}}({t}{k})$$ (17)

For the three geometric variables r, τb and τs, as the manifold they are in is not a flat Euclidean space, the distribution of the endpoints and the added noise differ accordingly. Previous studies have shown that for geometric variable r, the endpoint distribution is a Gaussian distribution in spherical space63. After uniformly sampling the unit rotation axis (\hat{r}), the distribution ({\mathscr{I}}{\mathscr{G}}{\mathscr{(}}r)) is:

$${\mathscr{I}}{\mathscr{G}}(r)=\frac{1-\cos r}{\pi }\mathop{\sum }\limits_{l=0}^{\infty }\;(2l+1){\mathrm{e}}^{-l(l+1){\epsilon }^{2}}\frac{\sin ((l+\frac{1}{2})r)}{\sin (r/2)}$$ (18)

At this time, the supervisory signal for r is:

$$-{\nabla }{{r}{k}}{\log p}{r}\left({r}{k},|,{r}{k+1}\right)=\left(\frac{{\rm{d}}}{{{\mathrm{d}}r}}\log {\mathscr{I}}{\mathscr{G}}\left({r}{k}\right)\right){\hat{r}}_{k}$$ (19)

For the geometric variables τb and τs, the endpoint distribution pT(τ) is a standard wrapped normal distribution ({\mathscr{W}}{\mathscr{N}}{\mathscr{(}}\tau )):

$${p}{T}(\tau )={\mathscr{W}}{\mathscr{N}}(\tau )\propto \mathop{\prod }\limits{i=1}^{N}\sum _{{{d}}\in {\mathbb{Z}}}\exp \left(-\frac{{{||}{\tau }^{(i)}+2{{\uppi}} {d||}}^{2}}{2}\right)$$ (20)

where N is the total number of rotatable bonds of backbone or side chain and d is an integer shift variable. At this time, thesupervisory signal for τ and τ is:

When generating peptide conformations during inference, random geometric variables ({\hat{t}},{\hat{r}},{\hat{\tau}}{b},{\hat{\tau}}{s}) are first sampled from ({\mathscr{N}}), ({\mathscr{I}}{\mathscr{G}}{\mathscr{(}}r)), ({\mathscr{W}}{\mathscr{N}}{\mathscr{(}}\tau )) and applied to the 3D conformation of the peptide:

$${x}{T}={\hat{\tau}}{\mathrm{b}}\circ {\hat{\tau}}_{\mathrm{s}}({\hat{r}}x+{\hat{t}}\;)$$ (22)

where x is the initial structure of peptide and xT is the randomized structure of peptide. Then the bi-graph-based CG tensor product ‘GNN’ is iteratively used to predict the denoising geometric scores tk, rk, τbk and τsk, at timestep k, which decreases from T to 0, and these scores are applied to the 3D conformation of the peptide:

$${{{x}}}{{{0}}}={\tau }{{{\mathrm{b}}T}}\circ {\tau }{{{\mathrm{s}}T}}\cdots {\tau }{\mathrm{b}0}\circ {\tau }{{\mathrm{s}}0}({R}{T}\cdots {R}{0}x+{t}{T}\cdots {t}_{0})$$ (23)

At this point, x0 is the final predicted peptide conformation, which follows the conformational distribution of peptide learned by RAPiDock. The final predicted peptide conformation is relaxed using a customized procedure (leveraging the FastRelax module of PyRosetta64) and scored using the ‘Ref2015’ scoring function in PyRosetta.

Conformer matching during training

During inference, we use the peptide initial-structure builder to construct the initial peptide structure. As a result, different individuals of the same amino acid type in the initial peptides will have identical bond lengths and angles. This leads to a distribution shift between predicted peptide structures in the inference time and crystal peptide structures used in the training time. To eliminate this shift, we are supposed to use the peptide initial-structure builder to construct the peptide structure based on the peptide sequence during training. However, the structures obtained this way would differ from the crystal structures. Therefore, we developed the peptide conformation optimizer to refine the peptide structures generated by the initial-structure builder, making them as close as possible to the crystal structures (Fig. 1c). Specifically, for each peptide, we use the differential evolution algorithm to explore the conformational space of all rotatable bonds, finding structures that closely match the crystal structures65. Through this approach, we constructed near-native structures root mean square deviation (r.m.s.d.) ≤0.5 Å) for each peptide to replace the crystal structures during training.

Dataset preparation

In this study, protein–peptide complex structures for RAPiDock training and performance evaluating were collected from the RCSB PDB database66 with structures released before 10 February 2023. To ensure precision evaluation and optimal model performance, we constructed a high-quality dataset through a series of meticulous steps. First, entries from the RCSB PDB database have to meet the following criteria: (1) the entry must be labelled as ‘prot’ to exclude structures that contain nucleic acids; (2) the entry must contain at least 2 chains, with one being a peptide chain of 20 or fewer residues to ensure PPI; and (3) only NMR structures or the structures with X-ray diffraction or electron microscopy resolution ≤3.0 Å were selected. All qualifying entries were then standardized using the Protein Preparation Wizard module67 in Schrödinger 2021, which included removing water molecules, assigning bond orders, re-adding hydrogen atoms and using Prime to fill in missing side chains and loops. Each standardized protein–peptide complex structure was then double-checked to ensure: (1) the peptide formed a continuous chain through the 4 backbone atoms (N, CA, C, O); (2) the peptide contained at least 3 residues; (3) no backbone atoms (N, CA, C, O) were missing from any protein residue; and (4) the protein had to have at least 50 residues or was 3 times the peptide to ensure sufficient binding surface. After excluding entries from the PepSet dataset25, a total of 8,846 protein–peptide complex structures met the aforementioned criteria, forming RefPepDB. Among these, 7,779 structures that were released before 21 September 2020 were designated for training and named the RefPepDB-TrainSet. The remaining 1,067 structures were further filtered on the basis of the contents of the ‘Structure Title’, specifically by searching for keywords denoted as ‘complex with’ or ‘bound to’. This filtration yielded 523 entries, which were designated for evaluation and referred to as the RefPepDB-RecentSet.

In addition, PepSet25, which contains 185 protein–peptide complex structures paired with their unbound protein structures, was used to further evaluate the performance of RAPiDock on more challenging scenarios when the accurate protein structures in bound state are ‘unknown’. The distribution of peptide characteristics, that is, the length, the ASA% and the secondary structure categories across RefPepDB-TrainSet, RefPepDB-RecentSet and PepSet datasets can be found in Supplementary Fig. 36. As for homology filtering, we apply both sequence identity and structure similarity thresholds (details are provided in Supplementary Note 8).

Evaluation protocol

RAPiDock’s performance in predicting protein–peptide complexes is evaluated from multiple perspectives, including the performance on the latest solved structures; the performance on the more challenging protein–peptide docking tasks; the performance on multiple real-world tasks, such as PLK1-PBD, SHP2 and pHLA systems, and the performance on targeting multiple binding pockets. In addition, for the pHLA task, we further explore RAPiDock’s potential in virtual screening by comparing the scores of RAPiDock predicted structures computed by Ref2015 and Vina with experimental affinity values, and benchmarking its performance against the template-based method. Experimental affinity data of pHLA targets are sourced from the IEDB database (http://tools.iedb.org/mhci/download/). Template-based peptide construction involves initial-structure mutation using PyMOL, followed by energy minimization using Amber2268. Subsequently, the energy-minimized structure is scored using the Ref2015 scoring function in PyRosetta and Vina. Performance metrics, including AUC-ROC and AUC-PRC, are computed using the Sklearn software package (https://scikit-learn.org/), while Pearson and Spearman correlation coefficients are calculated using Scipy software (www.scipy.org/).

Metrics

A ligand r.m.s.d. of 2 Å or less is typically used for small-molecule docking as most atoms in these molecules interact with the macromolecule. However, peptides tend to bind in shallow grooves on the protein surface and have many atoms that do not interact with the macromolecule. To comprehensively evaluate peptide docking results, we adopted a multi-metric strategy, including Fnat, iRMS, lRMS69, DockQ70 and CAPRI-peptide standards71. Fnat, iRMS, lRMS and DockQ were computed using DockQ’s default parameters. The CAPRI-peptide standard, which applies different cutoffs on Fnat, iRMS and lRMS, was used to classify the predicted docking results into the four quality classes: incorrect (Fnat < 0.2 or (lRMS > 5.0 and iRMS > 2.0)), acceptable ((Fnat ≥ 0.2 and Fnat < 0.5) and (lRMS ≤ 5.0 or iRMS ≤ 2.0) or (Fnat ≥ 0.5 and lRMS > 2.0 and iRMS > 1.0)), medium ((Fnat ≥ 0.5 and Fnat < 0.8) and (lRMS ≤ 2.0 or iRMS ≤ 1.0) or (Fnat ≥ 0.8 and lRMS > 1.0 and iRMS > 0.5)) and high (Fnat ≥ 0.8 and (lRMS ≤ 1.0 or iRMS ≤ 0.5)).

Baseline models AF2Multi

When using AF2Multi to predict protein–peptide complex structures, the cutoff date for the template search was set to 21 September 2020, and all other parameters were set to default values. Five docking poses were required for each of the five AF2Multi models, so there were 25 docking poses for each complex in the test dataset. AF2Multi is termed either AF2Multi complete sequences (AF2MultiCS) when using the sequences of the peptide chain and all protein chains within 20 Å for prediction, or AF2Multi refined sequences (AF2MultiRS) when using the sequences of the peptide chain and specifically selected interacting protein chains.

AlphaFold3

The refined sequences of protein–peptide complex structures mentioned above were uploaded to the AlphaFold server to generate docking results of AlphaFold3. As AlphaFold3 was trained on data before 30 September 2021, we only tested the 228 complexes from RefPepDB-RecentSet released after this date. In addition, due to usage limits, only five structures were generated for each complex.

ADCP

The AutoDock affinity maps were generated using AutoGridFR72 with default parameters and the docking calculation was executed within a 4-Å padding on every side of the cocrystallized peptide. Docking was performed with default ADCP parameters, involving 80 Monte Carlo searches with 3 million evaluations per residue in the peptide.

Other docking methods

The docking peptide conformations of ADCP, GOLD73, HADDOCK, MDockPeP and HPEPDOCK for 185 more challenging protein–peptide complexes from the PepSet were obtained from data provided by Weng et al.25. For MDockPeP, ‘SA’ indicates docking with standard accuracy, with a sampling exhaustiveness value set to 100, while ‘HA’ indicates docking with high accuracy, where the exhaustiveness value for peptides ranging from 5 to 10 residues remains at 100, but increases by 10 for each additional residue exceeding 10.

Other protocols Calculations of structure and sequence similarity

Template modelling score (TM-score) software74 was used to analyse structural similarity between protein structures in the training and test sets. NWalign software (http://zhanglab.ccmb.med.umich.edu/NW-align) was used to analyse sequence similarity between peptide sequences in the training and test sets.

Identification of the secondary structure of peptides

DSSP75 is used to compute the secondary structure of amino acids within peptide residues at the PPI interface. In this context, the interface is defined as residues located between the protein and peptide with heavy atoms within a 5-Å radius. The most frequently occurring secondary structure is selected as the peptide’s representative secondary structure.

Calculation of ASA percentage

For each protein–peptide complex, the solvent ASAs of the protein S1, peptide S2 and the whole complex S3 were calculated using mdtraj software. Then the interface area between the protein and peptide S4 was calculated as ({S}{4}=\frac{({S}{1}+{S}{2}-{S}{3})}{2}), and the ASA% (ref. 76) was defined as ({\rm{ASA}} \% =\frac{{S}{4}}{{S}{2}}\times 100 \%).

Data availability

The processed datasets, as well as pretrained models, are available via Zenodo at https://doi.org/10.5281/zenodo.14193621 (ref. 77). Unbound protein–peptide complexes were taken from the PepSet benchmark package (http://cadd.zju.edu.cn/pepset/statics/binaryDownload/benchmark.tar.gz). Source data are provided with this paper.

Code availability

The source code of this study is freely available via GitHub at https://github.com/huifengzhao/RAPiDock and Zenodo at https://doi.org/10.5281/zenodo.14568726 (ref. 78).

References

Stein, A. & Aloy, P. Contextual specificity in peptide-mediated protein interactions. PLoS ONE 3, e2524 (2008).

Article   Google Scholar

Babu, M. M., van der Lee, R., de Groot, N. S. & Gsponer, J. Intrinsically disordered proteins: regulation and disease. Curr. Opin. Struct. Biol. 21, 432–440 (2011).

Article   Google Scholar

Wright, P. E. & Dyson, H. J. Intrinsically disordered proteins in cellular signalling and regulation. Nat. Rev. Mol. Cell Biol. 16, 18–29 (2015).

Article   Google Scholar

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

Wang, L. et al. Therapeutic peptides: current applications and future directions. Signal Transduct. Target. Ther. 7, 48 (2022).

Article   Google Scholar

Kumar, V. B. et al. Peptide self-assembled nanocarriers for cancer drug delivery. J. Phys. Chem. B 127, 1857–1871 (2023).

Article   Google Scholar

La Manna, S., Di Natale, C., Onesto, V. & Marasco, D. Self-assembling peptides: from design to biomedical applications. Int. J. Mol. Sci. 22, 12662 (2021).

Article   Google Scholar

Li, L., Chen, X., Yu, J. & Yuan, S. Preliminary clinical application of RGD-containing peptides as PET radiotracers for imaging tumors. Front. Oncol. 12, 837952 (2022).

Article   Google Scholar

Zhou, Y. et al. Molecularly stimuli-responsive self-assembled peptide nanoparticles for targeted imaging and therapy. ACS Nano 17, 8004–8025 (2023).

Article   Google Scholar

Sharma, K., Sharma, K. K., Sharma, A. & Jain, R. Peptide-based drug discovery: current status and recent advances. Drug Discov. Today 28, 103464 (2023).

Article   Google Scholar

Vadevoo, S. M. P. et al. Peptides as multifunctional players in cancer therapy. Exp. Mol. Med. 55, 1099–1109 (2023).

Article   Google Scholar

Dai, M. Y. et al. High-potency PD-1/PD-L1 degradation induced by peptide-PROTAC in human cancer cells. Cell Death Dis. 13, 924 (2022).

Article   Google Scholar

Nhàn, N. T. T., Yamada, T. & Yamada, K. H. Peptide-based agents for cancer treatment: current applications and future directions. Int. J. Mol. Sci. 24, 12931 (2023).

Article   Google Scholar

Lee, A. C., Harris, J. L., Khanna, K. K. & Hong, J. H. A comprehensive review on current advances in peptide drug development and design. Int. J. Mol. Sci. 20, 2383 (2019).

Article   Google Scholar

Caporale, A., Adorinni, S., Lamba, D. & Saviano, M. Peptide-protein interactions: from drug design to supramolecular biomaterials. Molecules 26, 1219 (2021).

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

de Vries, S. J., Rey, J., Schindler, C. E. M., Zacharias, M. & Tuffery, P. The pepATTRACT web server for blind, large-scale peptide-protein docking. Nucleic Acids Res. 45, W361–W364 (2017).

Article   Google Scholar

Alam, N. et al. High-resolution global peptide-protein docking using fragments-based PIPER-FlexPepDock. PLoS Comput. Biol. 13, e1005905 (2017).

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

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

Xu, X., Yan, C. & Zou, X. MDockPeP: an ab-initio protein-peptide docking server. J. Comput. Chem. 39, 2409–2413 (2018).

Article   Google Scholar

Zhang, X. et al. Efficient and accurate large library ligand docking with KarmaDock. Nat. Comput. Sci. 3, 789–804 (2023).

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

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

Article   Google Scholar

Tsaban, T. et al. Harnessing protein folding neural networks for peptide-protein docking. Nat. Commun. 13, 176 (2022).

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 (2022).

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

Johansson-Åkhe, I. & Wallner, B. Improving peptide-protein docking with AlphaFold-Multimer using forced sampling. Front. Bioinform. 2, 959160 (2022).

Article   Google Scholar

Wu, R. et al. High-resolution de novo structure prediction from primary sequence. Preprint at bioRxiv https://doi.org/10.1101/2022.07.21.500999 (2022).

Bret, H., Gao, J., Zea, D. J., Andreani, J. & Guerois, R. From interaction networks to interfaces, scanning intrinsically disordered regions using AlphaFold2. Nat. Commun. 15, 597 (2024).

Article   Google Scholar

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

Article   Google Scholar

Ge, J. et al. Deep-learning-based prediction framework for protein-peptide interactions with structure generation pipeline. Cell Rep. Phys. Sci. 5, 101980 (2024).

Article   Google Scholar

Wang, Z. et al. Comprehensive evaluation of ten docking programs on a diverse set of protein–ligand complexes: the prediction accuracy of sampling power and scoring power. Phys. Chem. Chem. Phys. 18, 12964–12975 (2016).

Article   Google Scholar

Mondal, A., Chang, L. & Perez, A. Modelling peptide–protein complexes: docking, simulations and machine learning. QRB Discov. 3, e17 (2022).

Article   Google Scholar

Ramachandran, G. T. & Sasisekharan, V. Conformation of polypeptides and proteins. Adv. Protein Chem. 23, 283–437 (1968).

Article   Google Scholar

Liu, Z., Sun, Q. & Wang, X. PLK1, a potential target for cancer therapy. Transl. Oncol. 10, 22–32 (2017).

Article   Google Scholar

Śledź, P. et al. From crystal packing to molecular recognition: prediction and discovery of a binding site on the surface of Polo-like kinase 1. Angew. Chem. Int. Ed. 50, 4003–4006 (2011).

Article   Google Scholar

Lin, T.-Y. et al. Design, synthesis and biological evaluation of phosphopeptides as Polo-like kinase 1 Polo-box domain inhibitors. Bioorg. Med. Chem. 26, 3429–3437 (2018).

Article   Google Scholar

Elia, A. E. et al. The molecular basis for phosphodependent substrate targeting and regulation of PLKs by the Polo-box domain. Cell 115, 83–95 (2003).

Article   Google Scholar

Diop, A. et al. SH2 domains: folding, binding and therapeutical approaches. Int. J. Mol. Sci. 23, 15944 (2022).

Article   Google Scholar

Neduva, V. & Russell, R. B. Peptides mediating interaction networks: new leads at last. Curr. Opin. Biotechnol. 17, 465–471 (2006).

Article   Google Scholar

Marasco, M. & Carlomagno, T. Specificity and regulation of phosphotyrosine signaling through SH2 domains. J. Struct. Biol. X 4, 100026 (2020).

Google Scholar

Asmamaw, M. D., Shi, X.-J., Zhang, L.-R. & Liu, H.-M. A comprehensive review of SHP2 and its role in cancer. Cell. Oncol. 45, 729–753 (2022).

Article   Google Scholar

Lange, A. et al. Classical nuclear localization signals: definition, function, and interaction with importin α. J. Biol. Chem. 282, 5101–5105 (2007).

Article   Google Scholar

Goswami, R. et al. Nuclear localization signal-tagged systems: Relevant nuclear import principles in the context of current therapeutic design. Chem. Soc. Rev. 53, 204–226 (2024).

Article   Google Scholar

Joglekar, A. V. & Li, G. T cell antigen discovery. Nat. Methods 18, 873–880 (2021).

Article   Google Scholar

Shen, Y., Parks, J. M. & Smith, J. C. HLA class I supertype classification based on structural similarity. J. Immunol. 210, 103–114 (2023).

Article   Google Scholar

Trott, O. & Olson, A. J. AutoDock Vina: improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading. J. Comput. Chem. 31, 455–461 (2010).

Article   Google Scholar

Zhang, O. et al. Learning on topological surface and geometric structure for 3D molecular generation. Nat. Comput. Sci. 3, 849–859 (2023).

Article   Google Scholar

Hampton, J. T. & Liu, W. R. Diversification of phage-displayed peptide libraries with noncanonical amino acid mutagenesis and chemical modification. Chem. Rev. 124, 6051–6077 (2024).

Article   Google Scholar

Zhu, X., Lopes, P. E., Shim, J. & MacKerell, A. D. Jr Intrinsic energy landscapes of amino acid side-chains. J. Chem. Inf. Model. 52, 1559–1572 (2012).

Article   Google Scholar

Malathy Sony, S. M., Saraboji, K., Sukumar, N. & Ponnuswamy, M. N. Role of amino acid properties to determine backbone tau(N-Calpha-C’) stretching angle in peptides and proteins. Biophys. Chem. 120, 24–31 (2006).

Article   Google Scholar

Tien, M. Z., Sydykova, D. K., Meyer, A. G. & Wilke, C. O. PeptideBuilder: a simple Python library to generate model peptides. PeerJ 1, e80 (2013).

Article   Google Scholar

Xiong, X., Zhou, B. & Wang, Y. G. Graph representation learning for interactive biomolecule systems. Preprint at https://arxiv.org/abs/2304.02656 (2023).

Isinkaye, F. O., Folajimi, Y. O. & Ojokoh, B. A. Recommendation systems: principles, methods and evaluation. Egypt. Inform. J. 16, 261–273 (2015).

Article   Google Scholar

Deng, C. et al. Vector neurons: a general framework for SO(3)-equivariant networks. Proc. IEEE/CVF Int. Conf. Comput. Vis. 2021, 12200–12209 (2021).

Google Scholar

Thomas, N. et al. Tensor field networks: rotation- and translation-equivariant neural networks for 3D point clouds. Preprint at https://arxiv.org/abs/1802.08219 (2018).

Kondor, R., Lin, Z. & Trivedi, S. Clebsch–Gordan nets: a fully Fourier space spherical convolutional neural network. In Advances in Neural Information Processing Systems (eds Bengio, S. et al.) (Curran Associates, 2018).

Han, J., Rong, Y., Xu, T. & Huang, W. Geometrically equivariant graph neural networks: a survey. Preprint at https://arxiv.org/abs/2202.07230 (2022).

Corso, G., Stärk, H., Jing, B., Barzilay, R. & Jaakkola, T. S. DiffDock: diffusion steps, twists, and turns for molecular docking. In Proc. International Conference on Learning Representations (ICLR, 2023); https://openreview.net/forum?id=kKF8_K-mBbS

Nikolayev, D. I. & Savyolov, T. I. Normal distribution on the rotation group SO(3). Texture Stress Microstruct. 29, 201–233 (1997).

Article   Google Scholar

Chaudhury, S., Lyskov, S. & Gray, J. J. PyRosetta: a script-based interface for implementing molecular modeling algorithms using Rosetta. Bioinformatics 26, 689–691 (2010).

Article   Google Scholar

Jing, B., Corso, G., Chang, J., Barzilay, R. & Jaakkola, T. Torsional diffusion for molecular conformer generation. Adv. Neural Inf. Process. Syst. 35, 24240–24253 (2022).

Google Scholar

Rose, P. W. et al. The RCSB Protein Data Bank: integrative view of protein, gene and 3D structural information. Nucleic Acids Res. 45, D271–D281 (2017).

Madhavi Sastry, G., Adzhigirey, M., Day, T., Annabhimoju, R. & Sherman, W. Protein and ligand preparation: parameters, protocols, and influence on virtual screening enrichments. J. Comput. Aided Mol. Des. 27, 221–234 (2013).

Article   Google Scholar

Salomon‐Ferrer, R., Case, D. A. & Walker, R. C. An overview of the Amber biomolecular simulation package. Wiley Interdiscip. Rev. Comput. Mol. Sci. 3, 198–210 (2013).

Article   Google Scholar

Méndez, R., Leplae, R., De Maria, L. & Wodak, S. J. Assessment of blind predictions of protein–protein interactions: current status of docking methods. Proteins 52, 51–67 (2003).

Article   Google Scholar

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

Article   Google Scholar

Lensink, M. F., Nadzirin, N., Velankar, S. & Wodak, S. J. Modeling protein‐protein, protein‐peptide, and protein‐oligosaccharide complexes: CAPRI 7th edition. Proteins 88, 916–938 (2020).

Article   Google Scholar

Zhang, Y., Forli, S., Omelchenko, A. & Sanner, M. F. AutoGridFR: improvements on AutoDock affinity maps and associated software tools. J. Comput. Chem. 40, 2882–2886 (2019).

Article   Google Scholar

Jones, G., Willett, P., Glen, R. C., Leach, A. R. & Taylor, R. Development and validation of a genetic algorithm for flexible docking. J. Mol. Biol. 267, 727–748 (1997).

Article   Google Scholar

Zhang, Y. & Skolnick, J. Scoring function for automated assessment of protein structure template quality. Proteins 57, 702–710 (2004).

Article   Google Scholar

Kabsch, W. & Sander, C. DSSP: definition of secondary structure of proteins given a set of 3D coordinates. Biopolymers 22, 2577–2637 (1983).

Article   Google Scholar

Shrake, A. & Rupley, J. A. Environment and exposure to solvent of protein atoms. Lysozyme and insulin. J. Mol. Biol. 79, 351–371 (1973).

Article   Google Scholar

Zhao, H. RAPiDock. Zenodo https://doi.org/10.5281/zenodo.14193621 (2024).

Zhao, H. huifengzhao/RAPiDock: RAPiDock_v1.1. Zenodo https://doi.org/10.5281/zenodo.14568726 (2024).

Download references

Acknowledgements

This study was financially supported by the National Key R&D Program of China (grant number 2024YFA1307501 to T.H.), National Natural Science Foundation of China (grant number 82373791 to Y.K.), National Natural Science Foundation of China (grant number 22220102001 to T.H.), National Natural Science Foundation of China (grant number 22307112 to D.J.) and the Young Scientists Fund of Natural Science Foundation of Hunan Province of China (grant number 2025JJ60651 to D.J.). We thank the Information Technology Center and State Key Laboratory of CAD&CG, Zhejiang University.

Author information Authors and Affiliations

College of Pharmaceutical Sciences, Zhejiang University, Hangzhou, China

Huifeng Zhao, Zhenxing Wu, Hongyan Du, Xiaorui Wang, Yihao Zhao, Yuansheng Huang, Jingxuan Ge, Tingjun Hou & Yu Kang

Paul G. Allen School, University of Washington, Seattle, WA, USA

Odin Zhang

Xiangya School of Pharmaceutical Sciences, Central South University, Changsha, China

Dejun Jiang

Contributions

Y.K., T.H., H.Z. and O.Z. designed the research study. H.Z. developed the method and wrote the code. H.Z., O.Z., D.J., Z.W., H.D., X.W., Y.Z., Y.H. and J.G. performed the analysis. H.Z., Y.K., T.H. and O.Z. wrote the paper. All authors read and approved the paper.

Corresponding authors

Correspondence to Odin Zhang, Tingjun Hou or Yu Kang.

Ethics declarations Competing interests

The authors declare no competing interests.

Peer review Peer review information

Nature Machine Intelligence thanks Irina Hashmi 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.

Extended data Extended Data Fig. 1 Effects of various factors on the docking performance of RAPiDock in the RefPepDB-RecentSet (compared to AF2Multi_RS).

(a) shows the effect of peptide lengths. (b) shows the effect of ASA%. (c) shows the effect of different second structures. Secondary structure categorization follows DSSP standards: ‘Loop’ includes Turn, Bend, and Coil; ‘Helix’ encompasses 310 helix, π helix, and α helix; ‘PPII’ is defined as polyproline II helix; ‘Strand’ covers beta bridge and strand structures. The box plots show the 25–75% confidence intervals (box limits), the median (centre line), and the 5–95% confidence intervals (whiskers). Green triangles indicate the mean values. N values report the number of clusters in each band.

Source data

Extended Data Fig. 2 Superior local and global docking performance of RAPiDock on more challenging tasks.

(a-c) illustrate the Top N success rates of various methods based on the CAPRI-peptide standards. Solid lines represent local docking approaches, while dashed lines represent global docking ones, (d-f) show the Top1, Top5 and Top25 success rates of various methods. The bars above the dashed line represent the performance of local docking approaches, while the bars below show that of global docking approaches. According to CAPRI-peptide standards, three accuracy levels are considered, namely Acceptable Quality (a and d), Medium Quality (b and e), and High Quality (c, f).

Source data

Extended Data Fig. 3 The docking performance of RAPiDock on SHP2.

(a) The sequences of 5X7B (N-SH2 domain of SHP2) and 5X94 (C-SH2 domain of SHP2). (b) left: The overlay of protein residues of 5X7B and 5X94. In SH2 domains, two structural regions determine the PpI: (1) the phosphotyrosine binding site, a groove-like structure formed by the αA helix, βB/βC/βD strands, and BC loop, and (2) the specificity pocket, which is a relatively large hydrophobic pocket mainly delimited by residues of αB helix, βD strand, and BG and EF loops. The phosphotyrosine binding site is relatively conserved, whereas the specificity pocket is more flexible; right: the peptides of 5X7B and 5X94 share a common 7-residue sequence PIpYATID. Given the pTyr residue serving as a reference, the residues from −2 to +3 align well in the pocket. The residues at +4 of the peptides exhibit obvious deviation, influenced by structural differences in the protein near the BG loop. Specifically, the proteins of 5X7B and 5X94 are colored green and yellow, respectively, while their peptides are colored white and gray, respectively. (c) Top-ranked peptide conformation generated by RAPiDock for 5X7B. (d) Top- ranked peptide conformation generated by RAPiDock for 5X94. The peptide conformations generated by RAPiDock are colored coral red.

Supplementary information Supplementary Information

Supplementary Notes 1–8, Figs. 1–37 and Tables 1–10.

Source data Source Data Fig. 2

Statistical source data.

Source Data Fig. 3

Statistical source data.

Source Data Fig. 6

Statistical source data.

Source Data Extended Data Fig. 1

Statistical source data.

Source Data Extended Data Fig. 2

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

Zhao, H., Zhang, O., Jiang, D. et al. Protein–peptide docking with a rational and accurate diffusion generative model. Nat Mach Intell 7, 1308–1321 (2025). https://doi.org/10.1038/s42256-025-01077-9

Download citation

Received 18 September 2024

Accepted 13 June 2025

Published 04 August 2025

Version of record 04 August 2025

Issue date August 2025

DOI https://doi.org/10.1038/s42256-025-01077-9

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 Structure-based drug design Virtual screening This article is cited by Seed perception learning for weakly supervised semantic segmentation Wanchun SunShujia LiXinyu Duan

Complex & Intelligent Systems (2026)