pocketz.doi.bio/2026-01-10_085441_nature_s41592-025-02820-1.md
Integrating diverse experimental information to assist protein complex structure prediction by GRASP | Nature Methods
URL: https://www.nature.com/articles/s41592-025-02820-1
Saved: 2026-01-10T16:54:46.268Z
Article Published: 15 September 2025 Integrating diverse experimental information to assist protein complex structure prediction by GRASP Yuhao Xie, Chengwei Zhang, Shimian Li, Xinyu Du, Yanjiao Lu, Min Wang, Yingtong Hu, Zhenyu Chen, Sirui Liu & Yi Qin Gao
Nature Methods volume 22, pages 2362–2374 (2025)Cite this article
4781 Accesses
2 Citations
13 Altmetric
Metrics details
Abstract
Protein complex structure prediction is crucial for understanding of biological activities and advancing drug development. While various experimental methods can provide structural insights into protein complexes, the knowledge obtained is often sparse or approximate. A general tool is needed to integrate limited experimental information for high-throughput and accurate prediction. Here we introduce GRASP to efficiently and flexibly incorporate diverse forms of experimental information. GRASP outperforms existing tools in handling both simulated and real-world experimental restraints including those from crosslinking, covalent labeling, chemical shift perturbation and deep mutational scanning. For example, GRASP excels at predicting antigen–antibody complex structures, even surpassing AlphaFold3 when using experimental deep mutational scanning or covalent-labeling restraints. Beyond its accuracy and flexibility in restrained structure prediction, GRASP’s ability to integrate multiple forms of restraints enables integrative modeling. We also showcase its potential in modeling protein structural interactome under near-cellular conditions using previously reported large-scale in situ crosslinking data for mitochondria.
You have full access to this article via your institution.
Download PDF Similar content being viewed by others Integrated structure prediction of protein–protein docking with experimental restraints using ColabDock Article 05 August 2024 High-accuracy protein complex structure modeling based on sequence-derived structure complementarity Article Open access 19 November 2025 Generating mutants of monotone affinity towards stronger protein complexes through adversarial learning Article Open access 28 February 2024 Main
Protein–protein interactions (PPIs) are essential in multiple biological processes1. The disparity between known amino acid sequences and solved protein structures highlights the critical need for computational prediction2,3. Current approaches such as AlphaFold-Multimer (AFM)4, AlphaFold3 (AF3)5, ESMFold6 and OmegaFold7 still show limited accuracy in protein complex structure prediction8.
Various experimental approaches can provide sparse but valuable information to assist prediction of protein complex structures. Residue pair restraints (RPRs) come from methods such as crosslinking mass spectrometry (XL-MS)9,10 and nuclear magnetic resonance nuclear Overhauser effect spectroscopy11. Interface restraints (IRs) represent potential interface residues of the complex. Differential covalent labeling (CL) detects reduced covalent modifications when protein complexes transit from unbound to bound states12. Chemical shift perturbation (CSP) identifies ligand interacting regions through shifts in chemical signals on binding13,14. Deep mutational scanning (DMS) highlights mutations affecting binding, often corresponding to interface residues15,16. These methods all provide IR information. In addition, in situ biophysical experiments including cryogenic (cryo) electron tomography17, solution NMR18 and in vivo XL19 can detect protein interactions at their real functional state. Integrating these experimental data to structure modeling will tremendously benefit the understanding of complicated biological systems.
Several tools were developed to integrate experimental data into protein structure prediction. AlphaLink20 embeds XL information directly into the pair representation and provides a protein complex version21. ColabDock22 leverages restraints for protein–protein docking by inverting AlphaFold models but is time-consuming and limited to proteins <1,200 residues. Traditional docking methods such as HADDOCK23,24 and ClusPro25,26,27 also support experiment restraints. HADDOCK transforms restraints into energy term to guide conformation sampling and scoring while ClusPro uses them to filter out FFT-generated structures. One study group28 integrated a CL scoring term into RosettaDock score function to select models aligned with experiments, but is limited to CL data.
In addition, different experimental methods provide cross-validated or complementary information, making them valuable for modeling complicated complexes. Integrative modeling approaches have been developed inspired by this. For example, the integrative modeling platform (IMP)29 samples protein complex conformations through Monte Carlo simulation and estimates the conditional likelihood of structures given restraints in the Bayesian framework, a framework also advantageous for multi-state modeling30. Current integrated modeling approaches are mostly case specific and need much manual effort31,32.
In this paper, we introduce a protein complex structure prediction model, the generalized restraints assisted structure predictor (GRASP). GRASP integrates both RPR and IR restraints directly to accommodate diverse experimental data in model inference. It is fine-tuned with restraint-related loss terms and applies iterative noise filtering in inference to reduce the impact of erroneous restraints. We perform comprehensive evaluations on both simulated and real-world experimental data on different functional protein complexes to assess GRASP’s ability to handle various types of experimental information, both simultaneously and independently, to improve prediction accuracy. Specifically, we applied GRASP to antibody–antigen complex structure determination, integrative modeling that simultaneously uses multiple experimental data and prediction of in situ complex structures in mitochondria. GRASP was shown to provide reliable structures and interactions for complex biological machines, and potentially also molecular details on dynamical protein interactions in cellular environment using experimental information such as XL.
Results GRASP integrates general restraints into protein modeling
GRASP integrates both RPRs and IRs into AFM for protein complex structure prediction. RPRs, derived from experiments such as XL-MS and NMR, provide distance restraints between residue pairs and can be viewed as edge features in a protein complex graph with residues as nodes. IRs, obtained from experimental methods such as CL, CSP and DMS, serve as node features, indicating whether a residue is at the interface. These features are incorporated into AFM via three blocks (Fig. 1a). RPRs are integrated into GRASP as MSA and invariant point attention bias (Fig. 1b) following our previous work33, while IRs are introduced using relative position encoding and interface fusion blocks within the Evoformer and invariant point attention modules (Fig. 1c).
Fig. 1: Scheme of GRASP to integrate generalized restraints.
a, The GRASP architecture shows how two types of restraint from various experiments are integrated using three AFM blocks. b,c, Details of RPR (b) and IR (c) integration are illustrated, with arrows indicating information flow and array shapes noted in parentheses. IPA, invariant point attention; MSA, multiple sequence alignment; NOE, nuclear Overhauser effect; ReLU, rectilinear units; repr., representation.
Full size image
We added four restraint-related loss terms (Supplementary Methods) to the original loss function and fine-tuned GRASP for 22,000 steps using 64 Ascend-910A processors starting from the AFM v.2.3 Model-1 checkpoint. Four inference checkpoints are selected in the fine-tuning process. A multimer prediction model was trained from scratch (Methods) and similarly fine-tuned for 22,000 steps to obtain the fifth model checkpoint to provide diversified checkpoints. In standard GRASP inference, similar to AFM, each of the five models was run with five random seeds. The best predictions were selected based on predicted local distance difference test (pLDDT) scores and restraint recall. We also applied an iterative noise filtering strategy to reduce the impact of erroneous restraints (Methods).
GRASP effectively uses RPRs and IRs on a self-curated dataset
To evaluate GRASP’s performance with the two independent types of restraint, we curated a benchmark dataset by selecting samples released after training set samples. The benchmark set differs from the training set in chain composition, has no internal redundancies, and shows suboptimal AFM predictions (Methods), to prevent data leakage and ensure dataset diversity and quality. This curated dataset contains 713 interfaces from 313 challenging complexes. For RPRs, simulated restraints were sampled from inter-chain residue pairs within 8 Å (contact RPR). The IR restraints were sampled from interface residues whose Cβ atom is within 8 Å of the nearest Cβ atom from the other chain, providing potential contact residues from only one protein. We incorporated varying numbers of simulated restraints and compared the performance of AlphaLink, ClusPro, HADDOCK, ColabDock and GRASP on applicable restraint types, as well as AFM or AF3 without restraints (Supplementary Data).
Without restraints, AF3 outperformed AFM, improving the mean DockQ from 0.17 to 0.23. As the number of restraints increased, the prediction accuracy of the restraint-based methods improved consistently for both restraint types. GRASP consistently ranked first among all methods and delivered the most substantial gains across different quantities of both IRs and contact RPRs (Fig. 2a,b and Supplementary Fig. 1). With just two inter-chain contact RPRs, GRASP reached a mean DockQ of 0.35, with 52.7% of samples exceeding the acceptable threshold of 0.23. For the less informative IRs, GRASP achieved mean DockQ scores and success rates of 0.24 and 35.3%, 0.34 and 51.9%, and 0.41 and 63.2% for 4, 10 and 20 restraints, respectively (Fig. 2a,b and Supplementary Table 1). For the protein complex of Protein Data Bank (PDB) ID 7PBZ (AFM DockQ 0.02), GRASP improved the DockQ to 0.87 using two contact RPRs. Similarly, in the case of PDB 7UBZ, GRASP raised DockQ score from the AFM DockQ of 0.01 to 0.9 with four IRs. In other examples including PDB 7NXX, 7F15 and 7NDF, notable improvements are also observed (Fig. 2c). We note that for both RPR and IR, ColabDock, HADDOCK and ClusPro showed overall better performance than AFM. AlphaLink, trained on 25 Å Sulfo-SDA restraints and not specifically designed to accommodate 8 Å restraints, exhibited limitations in processing restraints of short distances.
Fig. 2: Performance of different methods on benchmark dataset with two types of basic restraint.
a,b, DockQ (a) and restraint recall (b) comparisons across methods for varying numbers of contact RPRs (upper) and IRs (lower). Red dashed lines indicate a DockQ of 0.23. In the boxplots, the center line denotes the median, box edges represent the interquartile range (IQR) and whiskers extend to 1.5× IQR, with points beyond classified as outliers (same for d and f). In violin plots, the white dot marks the median, the thick bar indicates the IQR, the thin line extends to 1.5× IQR and the width reflects data density. Sample sizes are noted above the median lines. ClusPro failed on some samples, and ColabDock could not handle proteins over 1,200 residues, leading to smaller sample sizes. Each restraint type includes 713 cases across three levels. c, Representative structures of GRASP and AFM on selected samples, with DockQ, r.m.s.d., satisfied (satis.) and total restraints shown. Short predicted chains are cryo, long chains yellow, superimposed on the native structure (gray). All IRs are displayed as sticks, while restrained residue pairs are connected by solid lines. Blue indicates satisfied restraints, while red shows violated ones. AFM was run without restraints, GRASP restraints are also shown on the AFM structures only for comparison. d, Boxplot comparing the binding energy of aligned structures, GRASP predictions and the ground truth (713 samples per box). e, Proportion of the selected structures and the best structures (by DockQ) from the five models in GRASP. f, Ablation study for GRASP performance with and without restraints (restr.). GRASP without restraints (GRASP-Null) showed no improvement and is comparable to AFM. Each box contains 713 samples, except the final GRASP box (GRASP-IR and GRASP-RPR combined) with 1,426 samples. Pred., prediction.
Full size image
The model confidence, pLDDT, exhibited positive Pearson correlations with key performance metrics: DockQ (r = 0.39), TM score (r = 0.65) and LDDT (r = 0.87) (Supplementary Fig. 2a,b). In addition, improvements in pLDDT relative to AFM correlated with corresponding gains in DockQ, TM score and LDDT, yielding Pearson correlations of 0.38, 0.58 and 0.77, respectively (Supplementary Fig. 2c). These findings establish pLDDT as a robust confidence measure, both within GRASP predictions and in comparison to other models. To refine prediction selection, we integrated model and restraint confidence to define a ranking score as pLDDT + 1,000 × (recall ≥0.3). Most samples demonstrated positive intra-sample Spearman correlations between ranking scores and performance for both contact RPR and IR restraints (Supplementary Fig. 2d), underscoring the effectiveness of this combined metric in identifying high-quality predictions. GRASP also yielded reliable pLDDT scores for interface residues when using restraints, with median ipLDDT values of 84.3 and 80.1 for RPRs and IRs, respectively.
To examine whether GRASP reconstructs physically meaningful interfaces, three key interface properties were assessed: binding energy, cross-interface hydrogen bonds and buried solvent-accessible surface area (Supplementary Methods)34,35. GRASP-predicted structures exhibited similar binding energy (median −23.2 REU versus −21.4 REU, where 1 REU (Rosetta energy unit) ≈ 0.4 kcal mol−1; ref. 36), hydrogen bond numbers (median 5 versus 4) and interface areas (median 1,587.7 versus 1,515.7) to the experimental structures (Fig. 2d and Supplementary Fig. 3). In contrast, aligning AF2-predicted monomers to GRASP models without interface remodeling resulted in lowered binding affinity (median 5.9 REU; Fig. 2d). The heavy atom root mean-square deviations (r.m.s.d.s) of interface residues were also lower for GRASP-predicted structures than for AF2-aligned monomers (one-sided pairwise t-test, P = 0.003). These observations indicate that GRASP not only positions protein monomers following restraint information, but it also remodels their interfaces, enhancing physical fitness.
We further performed a series of ablation analysis of model number, restraint effect and noise filtering strategy on model performance. To evaluate the contribution of the five models, we examined their contribution to the final selections and best predictions. Each model contributed to the final selections, with the JAX-22k checkpoint contributing the most. The best structures were more evenly distributed among the models than the selected ones (Fig. 2e). Similar to AFM, using multiple models and selecting the top-scoring prediction enhanced both accuracy and robustness, and reduced the risk of failure (Supplementary Fig. 4).
We also evaluated the inference gain of GRASP from restraint by running an ablation experiment of the default restraint fine-tuned GRASP model without inputting restraints, mimicking the AFM protocol. GRASP without restraint shows comparable accuracy to AFM, with mean DockQs of 0.15 and 0.17, respectively (Fig. 2f). Thus, the improvements by GRASP stem from its ability to incorporate restraints rather than fine-tuning. Furthermore, the chain length slightly negatively correlate with prediction accuracy, showing larger proteins require more restraints for accurate predictions (Supplementary Fig. 5). With no erroneous restraints incorporated, iterative noise filtering had minimal impact on performance (Supplementary Fig. 6).
GRASP improves prediction using XL, CL and CSP restraints
GRASP and other methods are then benchmarked on XL, CL or CSP data. XL is one of the most common experiments to obtain RPR information. To evaluate the ability of integrating XL data, we simulated three sets of SDA XL restraints on the 294 complexes from the self-curated dataset and collected experimental XL data for additional seven protein complexes. The number of restraints in the three sets was set to 1%, 2% and 5% of the total sequence length, mimicking scenarios of sparse restraints such as in protein interactome, standard level restraints37, and a mixture of multiple XL reagents, respectively. In all sets, the maximum restraint distance was set to 25 Å, with 20% of restraints violating the threshold to mimic real-world experimental XL data (Methods). Since ColabDock can only handle restraints below 22 Å, and ClusPro only allows restraints between two rigid chains or subcomplexes, for these three datasets we evaluated GRASP, AlphaLink and HADDOCK.
Overall, GRASP demonstrated superior performance for all restraint coverages, consistently outperforming HADDOCK and AlphaLink in both accuracy and restraint satisfaction (Fig. 3a, Supplementary Table 2 and Supplementary Data). For comparison, AFM achieved a mean DockQ of 0.07. With only 1% restraint coverage, GRASP improved the mean DockQ to 0.18, outperforming both HADDOCK (0.06) and AlphaLink (0.12). GRASP obtained mean DockQ scores of 0.21 and 0.27 at 2% and 5% coverage, respectively, higher than HADDOCK’s 0.08 and 0.10, and AlphaLink’s 0.13 and 0.20. GRASP also excelled in restraint satisfaction metrics across all coverage levels. For instance, at 2% restraint coverage, GRASP achieved a median satisfaction rate of 69% for all and 81% for correct restraints, outperforming both AlphaLink (37% and 42%) and HADDOCK (61% and 71%). Even with 20% erroneous restraints, GRASP achieved a median satisfaction rate of over 80% for correct restraints, indicating its robustness against experimental noise. Iterative noise filtering consistently improved DockQ and pLDDT across all levels of restraint coverage (Supplementary Fig. 7a). Examples such as PDB 7SQP show notable improvement (DockQ increasing from 0.06 to 0.83; Supplementary Fig. 7b).
Fig. 3: The performance of GRASP and other methods using single-type experimental restraint data.
a, DockQ (left), restraint satisfaction rate (middle) and correct restraint satisfaction rate (right) for various methods on simulated SDA XL-MS data across different restraint coverages. In boxplots, the center line denotes the median, box edges represent the IQR and whiskers extend to 1.5× IQR, with points beyond classified as outliers. In violin plots, the white dot marks the median, the thick bar indicates the IQR, the thin line extends to 1.5× IQR and the width reflects data density. Each box and violin includes 294 samples. The red dashed horizontal line in the left subfigure indicates a DockQ value of 0.23 (same for subfigures in c). b, 3D visualization of predicted structures by various methods compared to the true structure (gray) on 7R94 in the simulated XL dataset. c, DockQ scores for various methods using single-type real-world experimental restraint data, including XL (left), CL (middle) and CSP (right). Since the number of cases for each experimental approach is lower than ten, all of the evaluated cases are drawn separately as scatters for different methods. d, 3D cartoons of representative predicted and true structures for each method under different experimental restraints. The predicted structures are colored cyan (smaller part) and yellow (larger part), with the latter superimposed on the real structure in gray. Satisfied restraints are shown in blue and violated ones in red. Correct IRs are displayed as sticks, incorrect ones as cartoons. Correct RPRs are connected by solid lines, while incorrect RPRs are shown with dashed lines. AFM was used without restraints, while GRASP restraints are displayed on the AFM structure for comparison.
Full size image
GRASP also performs well on large complexes such as PDB 7R94 (>3,500 residues, Fig. 3b). In this case, AFM mispositioned many chains, and even after adding simulated XLs, HADDOCK cannot rescue the prediction. Both AlphaLink and, especially, GRASP showed marked improvements, yielding DockQs of 0.51 and 0.67, respectively.
We gathered experimental data from various studies (Methods and Supplementary Data) for further tests. We collected nine samples with XL data21,38,39,40,41 (Fig. 3c left). GRASP achieved a mean DockQ of 0.48, outperforming other tools (AFM, 0.31; ClusPro, 0.05; HADDOCK, 0.05; AlphaLink, 0.38) (Supplementary Table 3 and Supplementary Data). For the four non-antigen–antibody complexes, GRASP, AlphaLink and AFM made predictions with DockQ higher than 0.23. Among the five antibody–antigen XL examples, AFM had only one prediction with DockQ>0.23 but with a large r.m.s.d. of 26.37 (Supplementary Fig. 8). After incorporating experimental XL data, GRASP showed improvements in 80% of the cases. For PDB 4G3Y, using 12 DSBU (30 Å crosslink distance) restraints including three erroneous ones, GRASP increased the DockQ from 0.03 (AFM) to 0.77, satisfying all nine correct restraints while ignoring the three incorrect ones. AlphaLink also predicted the correct structure with DockQ of 0.67. In contrast, HADDOCK was biased toward incorrect restraints, resulting in a DockQ below 0.23 (Fig. 3d, upper). Therefore, GRASP not only ensures structure alignment with the provided restraints, but also excels in restraint discrimination, minimizing the impact of erroneous restraints.
For five cases with experimental CL data28, AFM achieved an average DockQ of 0.45, while HADDOCK and ClusPro yielded average DockQs of 0.3 and 0.04, respectively. ColabDock improved to an average DockQ of 0.56, and GRASP showed the greatest improvement, reaching an average DockQ of 0.58 (Fig. 3c). Notably, for 4INS8, GRASP (DockQ of 0.56) is the only method that outperforms AFM (DockQ of 0.43) (Fig. 3d).
For the four CSP cases from two studies24,42 (Fig. 3c, right), AFM performed well in two cases but struggled with the two antibody–antigen complexes, yielding an average DockQ of 0.39. HADDOCK and ClusPro showed average DockQs of 0.28 and 0.21, respectively. ColabDock improved this value to 0.5, while GRASP further raised it to 0.81. For instance, GRASP accurately predicted PDB 4G6J with a DockQ of 0.79, whereas all other methods failed. In another challenging case of PDB 4G6M, GRASP improved the DockQ to 0.9 using CSP data. ClusPro, HADDOCK and ColabDock also improved but only reached acceptable DockQs (Fig. 3d). Notably, all methods except GRASP struggled with these antibody–antigen complexes, highlighting the challenge of this task (Fig. 3c,d).
GRASP performance on antibody–antigen prediction
Accurate structure prediction of antigen–antibody complexes is critical for antibody design and immunotherapy. However, current methods such as AFM have limited accuracy. AF3 was reported to perform better on such complexes5, and is included in our evaluation in this section.
We first evaluated GRASP on the BM5.5 (ref. 43) antibody–antigen dataset, which comprises 67 samples, each with three sets of simulated DMS IRs (Methods). GRASP surpassed all other methods with a DockQ median of 0.64 and a success rate of 71.6% (Supplementary Fig. 9, Supplementary Table 4 and Supplementary Data). To avoid data leakage, we examined GRASP performance with an additional set of DMS restraints generated for cases curated in ref. 44, comprising 65 antibodies and 107 nanobodies released posttraining (Methods). GRASP yielded median DockQ scores of 0.477 for antibodies and 0.541 for nanobodies with success rates of 60.0% and 88.8%, respectively, substantially surpassing the second-best method, AF3 (median DockQ 0.069 and 0.237, success rate 36.9% and 50.5%, respectively) (Fig. 4a, Supplementary Fig. 10 and Supplementary Table 5). Other docking methods all exhibited lower success rates. These findings underscore GRASP’s superiority over both deep learning-based (AF3) and restraint-guided docking approaches.
Fig. 4: GRASP improves antigen–antibody structure prediction accuracy.
a, Performance on the simulated DMS Hitawala–Gray dataset for antibodies (left) and nanobodies (right). Top, boxplots of DockQ distributions. The center line denotes the median, box edges represent the IQR and whiskers extend to 1.5× IQR, with points beyond classified as outliers. Sample sizes are noted above the medians. Bottom, stacked bar plots of success rates at three accuracy levels for antibody (left) and nanobody (right). Numbers above bars indicate cumulative percentages. b, Scatter plot comparing GRASP and AF3 DockQs for antigen–antibody complexes with XL or CSP restraints. The yellow dot highlights the case in c. c, 3D visualization of GRASP (pTM-based) and AF3 predictions for 6OGE_ABC. Top, two types of crosslinker on the ground truth (GT) structure. Bottom, predicted structures by AF3 and GRASP (antibody, cyan; antigen, yellow), with antigen superimposed on GT (gray). Satisfied restraints are colored blue, and violations red (same in f). Correct RPRs appear as solid lines; incorrect as dashed lines. d, Schematic of GRASP’s restraint strategy on the experimental COVID-19 DMS dataset. e, Violin plot of DockQs on the experimental DMS dataset. The white dot marks the median, the thick bar shows the IQR, the thin line extends to 1.5× IQR; numbers above and below the median indicate case counts and success rates, respectively (same in i). f, 3D predictions on PDB 7K8S (antibody C002). DMS sites are shown, with correct sites shown as sticks. g, Scatter plot comparing AF3 and GRASP DockQs on samples with high (orange) and low (green) AF3 DMS recall. Numbers beside y = x line indicate sample counts on each side. Parentheses in the legend show the total samples per category. h, Schematic of the strategy combining AF3 and GRASP. i, Violin plot of DockQ for the combined strategy versus GRASP alone. The slanted, vertical and horizontal red dashed lines in relevant subfigures indicate y = x, x = 0.23 and y = 0.23, respectively. DSSO, disuccinimidyl sulfoxide; CDI, carbonyldiimidazole.
Full size image
We further evaluated GRASP assisted by experimental data, including 7 CSP or XL samples from the previous section and 84 additional DMS samples. GRASP achieved a success rate of 5 out of 7, outperforming AF3’s 4 out of 7 (Fig. 4b and Supplementary Data). In one of GRASP’s two failure cases, PDB ID 3C09 (Supplementary Fig. 11), 10 of 12 experimental restraints were not satisfied by the X-ray structure, suggesting a possible alternative binding conformation. In the second case PDB 6OGE_ABC, with 5 of 11 XLs unsatisfied by experimentally resolved reference structure, GRASP achieved a DockQ of 0.13, outperforming AF3’s 0.01 and AlphaLink’s 0.004 (Fig. 4c and Supplementary Fig. 12). Notably, GRASP predicted the correct structure with a DockQ score of 0.48, which failed to be identified by the pLDDT-based ranking score but could be selected by a pTM-based score (0.8 × ipTM + 0.2 × pTM), suggesting that adjusting the ranking strategy can further enhance performance in certain cases (Fig. 4b,c). Under both ranking schemes, GRASP consistently met all correct restraints and ignored the incorrect ones, again demonstrating its resistance to experimental noise.
We also compiled a new DMS dataset from experiment45, containing 84 different monoclonal antibodies bound to the Receptor Binding Domain (RBD) of the SARS-CoV-2 spike protein (Methods). We selected residues with escape scores above 0.2 as the antigen-side interface to reduce noise (Supplementary Fig. 13). We used complementarity determining region (CDR) residues as the antibody-side interface and also applied RPRs across the interface residues on both sides (Fig. 4d). GRASP outperformed all other methods with a DockQ median of 0.25 and a success rate of 53.6%. In comparison, AF3 has a DockQ median of 0.07 and a success rate of 39.3% (Fig. 4e, Supplementary Table 6 and Supplementary Data). For PDB 7K8S (Fig. 4f), AFM and AF3 performed poorly, and AF3 misidentified the binding site entirely. ClusPro, HADDOCK and ColabDock met DMS restraints but failed to predict the correct binding orientation. GRASP corrected the interface in assistance of DMS data and accurately predicted the structure with a DockQ of 0.72.
Notably, AF3 excelled in some cases (Fig. 4g), outperforming GRASP in 38 out of 84 cases, 16 of which showed over 70% agreement with DMS restraints. DockQ gains of GRASP over AF3 correlate positively with information gain from restraints (Pearson r = 0.45), defined as the difference between extra information and erroneous information provided by restraints (Supplementary Fig. 14). To take advantage of both GRASP and AF3, we proposed a joint protocol (Fig. 4h): to use AF3 for initial predictions and accept structures that align well with experimental restraints, otherwise use GRASP to improve predictions by restraints. This approach improved the overall prediction to a median DockQ of 0.28 and a success rate of 56.0%, surpassing both GRASP (0.25, 53.6%) and AF3 (0.06, 39.3%) (Fig. 4e,i and Supplementary Table 6).
Antigen–antibody complex prediction is challenging due to CDR variability and absent coevolutionary signals46, limiting MSA-based models such as AFM and AF3. GRASP improves the performance by effectively integrating diverse external restraints (for example, XL, DMS, CSP and CDR), which constrain the conformational search and guide predictions toward native-like binding. Its success underscores the value of experimental and prior knowledge in antibody modeling.
GRASP serves as a toolkit for integrative modeling
Single experimental technique often yields limited structural information, sometimes insufficient to fully determine protein interaction. Integrating multi-source experimental restraints to assist modeling is a promising way to study the structure and function of protein complexes47. To extend GRASP’s applicability, we further expand its usable restraint types to include cryo-electron microscopy (cryo-EM) density map. We iteratively dock chain combinations of the GRASP-predicted structure into the density map using a ‘random-split’ strategy with our graphical processing unit (GPU)-accelerated program Combfit, and then extract the spatial arrangement of subunits as RPRs to guide GRASP prediction (Supplementary Fig. 15 and Methods). Since a high-resolution density map can resolve atomic structures with specialized software packages such as ModelAngelo48, we focus here on low-to-medium resolution maps. We evaluate GRASP’s integrative modeling performance across diverse combinations of experimental approaches (Supplementary Methods).
The first case contains experimental information from XL, mutagenesis and the density map. It characterizes human APOBEC3G (A3G) proteins bound to the substrate receptor module of CRL5 containing HIV-1 Vif, CBFβ, Elongin B and Elongin C (VCBC). We first applied the mutagenesis data between A3G and Vif and XL-MS data containing 13 pairs of crosslinked peptides49 to model this system. Nine out of 13 crosslinked residue pairs satisfy the maximum Cɑ–Cɑ distance of 24 Å in the ground truth structure (PDB 8CX0) (Fig. 5a). AFM mispredicted the A3G-Vif binding interface, violating six XLs. When integrating XL and mutagenesis data, GRASP satisfied all IRs and 10 of 13 XLs with a DockQ of 0.398, but A3G still took a wrong orientation. By incorporating a medium-resolution cryo-EM density map of 7.82 Å, downsampled from the original map (EMD-27032), GRASP performance further improved, yielding a correct binding pose of Vif and a higher quality interface (DockQ 0.552 versus 0.307 for IMP).
Fig. 5: Applying GRASP in integrative modeling results in reasonable structures.
a, Visualization of experimental restraints for A3G-VCBC complex on ground truth (PDB ID 8CX0, left), AFM-predicted structure (middle left), GRASP-predicted structure using XL and mutagenesis (middle), IMP-predicted structure and GRASP-predicted structure using XL, mutagenesis and density map (middle right, right). Blue lines refer to satisfied XL restraints and red lines refer to violated ones (same in c and d). XL recall refers to the ratio of crosslinked residue pairs with a Cɑ–Cɑ distance below 24 Å in a, and below 25 Å in c and d (Supplementary Table 7). Mutagenesis recall indicates ratio of restrained residues whose Cβ atoms are within 10 Å of the Cβ atoms of residues from another chain. b, Predicted structures for the KCNE1–KCNQ1 complex from GRASP with contact restraints and custom templates (left two panels). S1–S11 refer to the helices arranged from N terminal to C terminal. The right two panels are enlarged interface labeled in left middle panel. The Cɑ–Cɑ distance computed for each the used RPRs are labeled. Two violated restraints (F340/KCNQ1-T58/KCNE1; A344/KCNQ1-Y65/KCNE1) are colored in red. Compared with the open state, the KCNQ1 S11 tail in the closed state twists toward the ion channel center and more contacts exist between KCNE1 and the KCNQ1 S9 or S10. c, Visualization of modeling structure of XcpHIJK from AFM, HADDOCK, IMP and GRASP using XL and PRE data. PRE recall is the ratio of restrained residues whose Cβ atoms are within 10 Å of the Cβ atoms of residues from another chain. d, Visualization of experimental restraints for AFM-predicted structure (left), GRASP-predicted structure using XL (middle left), IMP-predicted structure (middle right) and GRASP-predicted structure (right) using both XL and the density map. Map cross-correlation measures the degree of agreement between the structure and the density map.
Full size image
In the second case, we used experimental information from disulfide XL, Cd(II)-cysteine bridging and double mutant cycle analysis from ref. 50 to model the open and closed state of voltage-gated potassium (K) channel KCNQ1 and its regulatory protein KCNE1 with a stoichiometry of 4:4. These restraints were organized in a state-dependent manner and unified into RPRs. The structures of KCNQ1 from the KCNQ1–KCNE3 complex51 in an open or closed state were used as single-chain templates to bias GRASP’s prediction toward the corresponding state. Notably, AFM-predicted structures are similar to the closed state with low confidence on the position of KCNE1 (Supplementary Fig. 16a). Integrating multi-source RPRs and using custom single-chain templates, GRASP successfully predicted the binding interface in both states (Fig. 5b). IMP failed to keep the KCNE1 to the binding cavity formed by KCNQ1 (Supplementary Fig. 16b). The GRASP-generated structures satisfied most of the restraints in both states except two in the closed state. Models built by the group in ref. 50 using Rosetta protein–protein docking satisfied the same restraints with GRASP in both states (Supplementary Fig. 16c–e). Compared with them, GRASP predicts smaller motion amplitudes in the peripheral domain between open and closed conformations.
In the third case, we modeled the type II secretion system, where XcpH is assembled into the complex XcpIJK, using XL data and NMR paramagnetic relaxation enhancement (PRE) data from ref. 52. The AFM-predicted structure satisfied 15 of 20 XL and 11 of 15 PRE restraints with one end of the N-terminal helix of XcpH positioned far from XcpIJK (Fig. 5c). One research group52 built a model using HADDOCK with the XcpH helix in parallel to the XcpJ helix, satisfying more PRE restraints. IMP predicted a completely different pose of XcpH but it satisfies more restraints than that predicted by HADDOCK. The GRASP-built structure is similar to the HADDOCK model but the helix of XcpH is slightly removed from XcpIJK, with 18 of 20 XLs and all 15 PRE restraints satisfied.
The fourth modeled system is a 5-subunit CRL4DCAF1-CtD/Vprmus assembly in state-1 integrating XL data and the density map (EMD-10612) from ref. 53. The accessory viral protein Vpr of the simian immunodeficiency virus can redirect the specificity of host Cullin4-RING ubiquitin ligases to the antiviral host factor SAMHD1. The AFM-predicted structure conflicted with the density map, placing the CUL4A head close to other subunits and satisfying only 58 out of 85 XLs (Fig. 5d, left panel). By integrating XLs alone, GRASP-predicted structure satisfied 76 XLs and obtained a cross-correlation of 0.352 with the predicted head of CUL4A fitting into the density map (Fig. 5d, middle left panel). By integrating both XL and density map, GRASP further improved to 77 satisfied XLs and a cross-correlation of 0.436, outperforming the former two models. In contrast, the IMP-predicted structure satisfied 62 XLs with a lower density map cross-correlation of 0.218 (Fig. 5d, right two panels).
Structural insights into in situ mitochondrial PPI
Traditional structure determination methods, such as X-ray crystallography and cryo-EM, require purified protein complexes and stable subunit interactions, making it difficult to capture transient and dynamic PPIs. Recently, improvements in chemical XL techniques have enabled the detection of protein interactomes, even at the whole-cell scale37,54. While in-cell XL-MS reveals valuable in situ interaction information, it lacks atomic details. Here, we used GRASP to build the corresponding structure models.
In the following example, we collected XL-MS data from ref. 55 in which a targeted delivery system was developed to map protein interactions in mitochondria in situ. A total of 158 inter-protein crosslinks were mapped to 144 PPIs using the human MitoCarta v.3.0 database56. We transformed these crosslinked peptides into RPRs with an upper distance limit of 30 Å to predict protein interactions. Among the 144 predicted PPI pairs, 121 showed pLDDT above 75 (Fig. 6a and Supplementary Fig. 17). In addition, 29 out of 69 predicted PPIs with an direct interaction exhibited an interface pLDDT above 75 (Supplementary Data). Compared to AFM, the addition of XLs into GRASP leads to structure adjustment to varying degrees (Supplementary Fig. 18a and Supplementary Data). GRASP satisfied XLs in 140 out of 144 predicted PPIs, whereas AFM satisfied only 31. Further, the distance distribution of GRASP’s predicted crosslinked residues closely matched the experimental statistics57 of reagent disuccinimidyl suberate while AFM predictions resembled a uniform distribution (Fig. 6b).
Fig. 6: Dimeric structure models are predicted by GRASP for PPIs captured by targets XL-MS in mitochondria.
a, Visualization for PPIs colocated in the same subcellular localization for mitochondria outer membrane (MOM, upper), mitochondria inner membrane (MIM, middle) and matrix (lower) with pLDDT >75. Blue lines refer to the well-satisfied restraints and red lines refer to violated restraints. b, Cβ–Cβ distance distribution of crosslinked residues in the GRASP-predicted structure, AFM-predicted structure and experimental statistics. c, Mapping of 17 pair of GRASP-predicted interactions on ground truth. Ground truths are colored in gray d, Structure model of the SDHB–SDHC interaction. Blue lines refer to the well-satisfied restraints and red lines refer to violated restraints. The predicted structure is aligned with a homologous protein (PDB 2WDQ) in the right panel with the ground truth structure colored in gray. e, Visualization of prediction structures for ATP5F1A-ATP5PF by AFM, GRASP (left two panels). GRASP-predicted four-body interaction including extra two sequences are show in the right panel. Distances of crosslinked residues are displayed around each panel. The ground truth structure (chains C and L of PDB 8IK3) is colored in gray. f, Visualization of PPIs related to F0-F1 ATP synthase. PPIs within ATP synthase are shown on the left panel and representative predicted PPIs between ATP synthase and other proteins are shown on the right panel. The ground truth structure of ATP synthase (PDB 8H9V) was colored in gray and crosslinked chains in ATP synthase are in different colors.
Full size image
We then validated the predictions by mapping them to available ground truth structures (Fig. 6c). Experimentally resolved PPI structures in mitochondria are very limited and only 17 pairs were identified from the Protein Data Bank. GRASP accurately predicted most PPIs with a median TM score of 0.881 (0.838 for AFM; Supplementary Fig. 18b), and exhibited binding energies and interface sizes similar to the native ones (Supplementary Fig. 18c,d). The SDHB–SDHC interaction, involved in the tricarboxylic acid cycle, was modeled by GRASP with K141 of SDHC close to K261 of SDHB, satisfying the XL restraint with a Cɑ–Cɑ distance of 8.7 Å (Fig. 6d left). Despite the absence of ground truth structure for this interaction, we identified a pair of remote homologs from Escherichia coli (PDB 2WDQ) and observed a highly conserved interface (TM score 0.916, Fig. 6d, right).
In some cases, we observed that crosslinked protein pairs lacked direct interaction (Fig. 6c), suggesting the presence of additional binding partners mediating these interactions. Incorporating such mediators may improve structural prediction. For example, ATP5PF–ATP5F1A is a part of the 16-subunit F0-F1 ATP synthase. In modeling the ATP5PF–ATP5F1A dimer, AFM predicted an incorrect orientation of ATP5F1A, while GRASP provided a more reasonable orientation but failed to predict the precise pose (Fig. 6e, left two panels). On including two additional interacting subunits, ATP5PB and ATP5PO, in the GRASP modeling, the relative position of ATP5PF and ATP5F1A improved with a TM score of 0.865 (Fig. 6e, right panel).
F0-F1 ATP synthase (complex V) is the most crosslinked complex in the experiment, with 11 intra-complex restraints and 11 additional restraints linking it to other proteins. Although we only modeled two-body interactions, 6 out of 7 PPIs aligned well with the reference structure across the ɑ, β, γ, δ, ε and o subunits in F1 and b1, d, cf6 subunits in the rod (Fig. 6d). In addition, we built three interfaces between other proteins and the mature F0-F1 ATP synthase. TRAP1 binds to a d subunit, while STOM and CB5R3 bind to an ɑ subunit at different sites. All three complexes exhibited minimal clash with the existing ATP synthase, indicating the ability of GRASP to model transient protein interactions. The relative orientation between TRAP1 and d subunit is consistent with the location of TRAP1 in the matrix. For STOM and CB5R3 located in the inner membrane, their predicted orientations relative to ATP synthase also align with their location. TRAP1 is believed to function by balancing the oxidative phosphorylation and aerobic glycolysis58, while STOM may influence the activity of ion channel according to Gene Ontology annotation59, and CYB5R3 serves as NADH-cytochrome b5 reductase60. Associating the predicted interactions with specific function may offer new insights on the regulatory mechanism.
Discussion
In this work we propose GRASP, a generalized model that integrates a variety of experimental restraints to enhance protein complex structure prediction. By incorporating RPR and IR information directly into the GRASP model and fine-tuning both the AFM and the self-trained MEGAFold-Multimer checkpoints with additional restraint-related loss terms, GRASP can simultaneously process one or more types of experimental data as restraints to improve the accuracy and reliability of a predicted protein complex structure. This approach allows the integration of multimodal data to complement and/or cross-validate each other. An iterative noise filtering process is also introduced to reduce the impact of erroneous restraints. GRASP infers structure directly and only needs one-time training to obtain an affordable inference time (Supplementary Fig. 19 and Supplementary Data) and AFM comparable GPU usage (Supplementary Fig. 20), supporting diverse restraint forms especially for large complexes.
GRASP supports both RPRs at various distances and IRs. It demonstrates excellent performance in complex structure prediction by effectively integrating various simulated and experimental restraint data. Under the simulated XL or real-world XL, CL and CSP restraints, GRASP achieves higher prediction accuracy than other methods, demonstrating its efficiency in integrating independent data types. For the crucial but highly challenging antigen–antibody prediction task, GRASP leverages sparse restraints from CSP, XL or DMS data to surpass AF3 in accuracy. A combined strategy of GRASP with AF3 further improves prediction. Specifically, the performance advantage of GRASP over AlphaLink may arise from several methodological reasons, including the integration of varied RPRs and IRs across all Evoformer layers and invariant point attention modules, additional restraint-specific loss terms, iterative noise filtering and ensemble prediction from five diverse checkpoints—collectively enabling effective usage of restraints for complex structure prediction.
GRASP naturally serves as an integrative modeling tool. By combining varied information across cryo-EM, XL, NMR PRE, mutagenesis and Cd(II)-cysteine bridging experiments in four cases, we demonstrated the potential for GRASP to perform complicated restrained modeling tasks. With this tool, we are able to model structures with actively developing in situ or dynamic detection techniques, such as in-cell XL, solution NMR and in situ cryo-electron tomography, which helps to reveal potential transient interaction structures and molecular mechanisms near physiological conditions.
For now, GRASP usage is limited to information formats of pairwise distance (RPRs) and IRs. Preliminary attempts to convert cryo-EM density maps into RPRs in combination with Combfit have shown promise. Directly integrating data from methods such as small angle X-ray scattering and cryo-EM will further broaden the applicability of this method.
In the post-AlphaFold stage, structural biologists are focusing on more complex protein interactions with essential biological applications. GRASP can act as a new toolkit for integrating experimental information with different levels of structural details and accuracy. We are optimizing the integrative strategy of GRASP to enable solution of more challenging application scenarios. Moreover, GRASP can also be consolidated into the workflow of existing integrative modeling tools such as CombFold61, thus combining the advantages of sampling and/or searching-based methods and deep learning-based methods.
Methods Datasets Training dataset
We used the PSP dataset62 as the basis for the GRASP and MEGAFold-Multimer training set. The PSP dataset contains a 570,000 true structure sets released before 8 October 2021 with protein sequences, experimental structures, MSAs and templates, along with a 760,000 distillation set with predicted structures, MSAs and templates. Complex samples were constructed by pairing single-chain MSAs with the same PDB ID in the true structure set using the pairing protocol within AFM4 and the rest of the single-chain information combined.
Self-curated test dataset
To assess the ability of GRASP to integrate two types of restraint (contact RPR and IR), we curated a test dataset and benchmarked GRASP against several other methods.
Sample collection
To ensure a comprehensive and unbiased evaluation, we created a new challenging and high-quality dataset to avoid data leakage and redundancy. We retrieved the protein sequences for each PDB ID released before 20 March 2023, split them into individual chains, and excluded nonprotein chains and those shorter than 20 residues, resulting in about 707,901 single-chain protein sequences. Using ‘easy cluster’ in MMSeqs2 (ref. 63) with a 70% sequence identity threshold, we clustered these sequences into 58,671 clusters.
To avoid data leakage, we removed all protein complexes released before 8 October 2021, as well as complexes with the same cluster composition as those released before this date. We note here that we considered only the cluster type. Only biological assembly 1 was used. We then filtered complexes by length, resolution and chain information (Supplementary Methods). This process resulted in 1,336 protein complexes. We further refined this set to 313 complexes where the AFM prediction had a DockQ score below 0.23. We focus solely on interfaces in this part. For complexes with three or more chains, we randomly sampled up to five nonredundant pairs of contacting chains, resulting in a total of 713 interfaces.
Contact RPR restraint sampling
The 313 protein complexes were randomly divided into 3 groups, and each interface in the 3 groups was assigned with 2, 5 or 10 contact RPRs, respectively. Contact RPRs were sampled from inter-chain residue pairs with Cβ distances within 8 Å, and no erroneous restraints were introduced.
IR sampling
Similarly, the 313 protein complexes were randomly divided into 3 groups, with interfaces in each group assigned with 4, 10 or 20 IRs. IRs were randomly sampled from interface residues whose Cβ atoms are within 8 Å of the Cβ atoms of another chain, without introducing erroneous restraints.
For both contact RPR and IR, we treat each chain as a distinct entity and apply restraints directly to the chains from which they were sampled, without distinguishing between homomeric and heteromeric chains.
Simulated XL dataset
To assess the performance of GRASP on simulated XL data, we generated XLs for the 313 complexes curated in the previous section. We performed inter-chain XL simulations using Xwalk64, applying a Cβ–Cβ cutoff of 25 Å and trypsin digestion. Three sets of restraints were simulated, with crosslinks (XLs) amounting to 1%, 2% and 5% of the total sequence length. We treat all chains as distinct entities and apply restraints directly to them, regardless of homomeric or heteromeric type. To mimic real experimental conditions, we included 20% erroneous restraints (false discovery rate of 20%), defined as residue pairs with Cβ–Cβ distances exceeding 25 Å, where at least one residue is Lys, Ser, Thr or Tyr. After excluding 19 complexes having fewer than 40% of the desired correct XLs, 294 complexes remained for analysis.
Experimental datasets with single restraint type
We constructed datasets with single experimental restraint types to compare the performance of different methods. These include nine XL, five CL and four CSP cases.
XL dataset
The dataset consists of five antigen–antibody complexes38,39 (PDB 1YY9, 3C09, 4G3Y, 6OGEABC and 6OGEADE) and four large protein complexes21,40,41 (PDB 4B2T, 5L4K, 6WVK and 6ZX9). XLs were transformed into RPRs with distance range determined by the specific XL reagents, for example, 24 Å for disuccinimidyl sulfoxide (see Supplementary Table 7 for the XL distances for different reagents). For PDB 6OGE, We predicted HER2 extracellular domain binding to Pertuzumab and Trastuzum, respectively, and named the two complexes PDB 6OGEABC and 6OGEADE following ref. 38. PDB 4B2T has a full length of 8,682 residues, which is too large for our model. Since it is of C2 symmetry, we chose to make structure prediction on the nonsymmetric half.
CL dataset
The five cases of CL were collected from ref. 28 (PDB 1YAG, 2F8O, 4INS for 4INS4, 4INS8 and 4INS12), as we did in our ColabDock model22. Covalent-labeling data provide information on residues buried in the binding interface and can be naturally converted into IRs. The prediction results of ColabDock, HADDOCK and ClusPro are available from ref. 22.
CSP dataset
We obtained four CSP examples, two from ref. 24 (PDB 1GGR and 3EZA) and the other two from ref. 42 (PDB 4G6M and 4G6J). The CSP data are transferred into the form of IRs. Since PDB 4G6M and 4G6J are antibody–antigen systems and only have restraints on the antigen side, we selected the residues in the CDR region as IRs on the antibody side. The antibody CDR regions were identified using the ‘Chothia’ scheme (www.bioinf.org.uk/abs/abnum/)65. The prediction results of 1GGR and 3EZA of ColabDock, HADDOCK and ClusPro were acquired from ref. 22. For PDB 4G6M and 4G6J, we ran calculations using these tools with the same settings as used in ref. 22 to complete the benchmark.
Simulated DMS dataset BM5.5
To evaluate GRASP on antibody–antigen systems, we simulated three sets of DMS restraints for each of the 67 antibody–antigen complexes in the BM5.5 (ref. 43) dataset. Of these, 45 complexes had already been assessed by ref. 22, and we reused the simulated restraints for them. The remaining 22 complexes were processed using the same restraint simulation strategy as in ref. 22.
Hitawala–Gray dataset
To prevent data leakage, we simulated DMS restraints using the case list curated in ref. 44, which includes 118 antibodies and 70 nanobodies released after the 2,021 AF3 training cutoff. Two antibodies (PDB 7YIZ and 8CBY) were excluded due to the absence of antigen chains, and 9 antibodies and 5 nanobodies were removed due to duplication, resulting in 107 antibodies and 65 nanobodies. A random selection of 5–10 interface sites was sampled on each side for the simulation. We also evaluated AF3’s performance on this dataset.
Experimental DMS dataset
We obtained DMS profiles for 3,051 SARS-CoV-2 wild-type RBD-targeting monoclonal antibodies from ref. 45. We then searched the PDB3 database using the monoclonal antibody names to identify protein complexes that met our criteria, specifically those containing the spike glycoprotein, light chain and heavy chain. This search resulted in a final set of 84 complexes (Supplementary Data). The spike glycoproteins were aligned to the Wuhan-Hu-1 RBD sequence (residues N331 to T531, GenBank MN908947), and mutation escape sites for each complex were determined. We trimmed the spike protein to include only the RBD regions for further analysis. The antibody CDR regions were identified using the Chothia scheme (www.bioinf.org.uk/abs/abnum/)65.
Mitochondria dataset
We collected targeted XL-MS data from ref. 55 and kept only the inter-protein XLs. We merged XLs with the same gene name to improve the restraint quantity for each pair of PPIs. We renumbered the crosslinked sites according to corresponding sequence in human MitoCarta v.3.0 database56 by sequence alignment and prepared the sequences for structure prediction (Supplementary Data). This led to a dataset including 144 PPIs with 1–4 pairs of RPRs for each. We set the maximum Cβ–Cβ distance to 30 Å for disuccinimidyl suberate to be consistent with ref. 55.
Model training strategy MEGAFold-Multimer training
We trained a MEGAFold-Multimer model from scratch using 128 Ascend-910A (32G) AI chips. MEGAFold-Multimer is a reimplementation of AFM within the MindSpore framework. The training process consisted of two stages (46,200 total steps over 17 days): an initial training phase (33,000 steps over 12 days) and a fine-tuning phase (13,200 steps over 5 days). Hyperparameters and training details are provided in the Supplementary Methods.
GRASP model fine-tuning
We fine-tuned on both AFM v.2.3 Model-1 checkpoint and MEGAFold-Multimer checkpoint over 9 days separately using 64 Ascend-910A (32G) AI chips. To achieve a more stable prediction, we generated five models. Four of these were fine-tuned from AFM v.2.3 Model 1, taken at 8,000 (JAX-8k), 14,000 (JAX-14k), 20,000 (JAX-20k) and 22,000 (JAX-22k) steps during the fine-tuning process. The fifth model (MS-22k) was fine-tuned 22,000 steps from the MEGAFold-Multimer model.
We adjusted the multimer/monomer ratio to 4/1 in GRASP training, set the learning rate to 1 × 10−4, and used a crop size of 384. All other configurations were consistent with those of AFM unless specified otherwise. We also introduced four restraint-related loss terms during fine-tuning, including RPR FAPE loss, RPR distance r.m.s.d. loss, RPR distogram loss and the interface loss (details in Supplementary Methods).
We simulated intra-chain RPRs (intra-RPRs), inter-chain RPRs (inter-RPRs) and IRs during fine-tuning, separately. Intra-RPR and inter-RPR were handled separately considering their differences in distance distribution, quantity and impact on the predicted structure. The details for sampling number, order and form of each type of restraints are provided in Supplementary Methods and Supplementary Fig. 21.
Iterative restraint filtering
To address noise in experimental restraint signals in model inference, which can lead to inaccurate structure predictions, we implement an iterative restraint filtering strategy (Supplementary Fig. 22a). Each round of restraint filtering evaluates the validity of the current restraints based on the predicted structure. Two key metrics are assessed in the order of priority: (1) chain break at the restrained residues and (2) the extent to which the structure violates the restraint.
We regard the Cɑ distance between consecutive residues higher than 5 Å as a chain break. The neighboring Cɑ distance (dNBCA) for a residue is defined as the average Cɑ distance to its two (one for terminal residues) flanking residues. The dNBCA for a RPR restraint is defined as the average dNBCA of the two restrained residues, while for an IR restraint it is the dNBCA of the interface residue.
The violation distance (dviol) measures how much the predicted structure violates the restraint. For example, if an IR specifies residue i as an interface residue but its closest distance to another chain is 20 Å, dviol is 12 Å. For an RPR restraint requiring residues i and j to be within 25 Å, dviol is 0 if their predicted distance is 10 Å, and 5 Å if the distance is 30 Å.
Using these two distances, the irrationality score (IS) of a restraint is defined as:
$${\mathrm{IS}}=\max \left({d}{\mathrm{NBCA}}-\mathrm{5,0}\right)\times 1000+\max \left({d}{\mathrm{viol}}-\mathrm{5,0}\right)$$ (1)
To avoid excessive restraint removal, each iteration filters up to 20% of current restraints, ensuring at least 20% of the original restraints are retained. The filtering stops when the maximum iterations are reached, no restraints are removed or fewer than 20% of the original restraints remain. IR and RPR filtering are executed independently, and the final predicted structure is then output.
Convergence is typically reached within 2–3 iterations, but a larger number may be needed with a large number of restraints. To improve efficiency, we introduced a ‘fast inference mode’ (Supplementary Fig. 22b). This mode reduces the number of recycles between iterations and recycles the structure after each filtering step, allowing up to 20 inference rounds and making it up to 4 times faster than the standard version. We used this mode in the DMS dataset evaluation, where RPRs across both sides of the interface introduced m × n RPRs, with m and n representing the number of interfaces on each side.
Integrative modeling using Combfit
To smoothly integrate cryo-EM density map information into GRASP, we developed Combfit, an advancement over the GPU-based matching software Powerfit66, using clustering and genetic algorithms (Supplementary Methods). Implemented in up-to-date Python v.3.10, Combfit offers excellent readability and is easy to integrate with other models.
In integrative modeling involving cryo-EM, we first converted all restraints, except for the density map, into either RPRs (for XL and NMR nuclear Overhauser effect spectroscopy) or IRs (for CL, mutagenesis and CSP) to predict the initial structure. The initial structure was then split into random or predefined chain combinations, which we refer to as ‘random-split’ strategy. These chain combinations were docked into density map using Combfit. We next extracted residue pairs from the docked structure where the Cβ–Cβ distance was ≤25 Å and had decreased by at least 10 Å compared to the initial structure. Using a min–max algorithm67, up to 30 representative residue pairs were chosen from the extracted residue pairs as RPRs. These RPRs, along with the existing RPRs and IRs from other experiments, were then used to guide the GRASP’s second round of prediction. In doing so, we obtained the final structure for integrative modeling with cryo-EM data.
Setup for GRASP inference
The AFM protocol4 was used to search for MSA and templates and to prepare input features for GRASP. The maximum template date was set to ‘2021-10-08’. We used reduced databases in cases where HHblits raised errors during MSA searches. The standard iterative restraint filtering strategy was applied with a maximum of 5 iterations for cases with fewer than 3,072 residues and 2 iterations for larger ones. Each checkpoint runs with 5 seeds for complexes under 3,072 residues to align with the AFM setup and 1 seed for larger complexes. For all the predicted structures, we select the predicted structure with the highest pLDDT score as the final prediction, prioritizing those with a recall of at least 30%. Due to memory limitations, the Ascend-910A (32G) could not handle inference for sequences longer than 1,280 residues. To address this, we migrated GRASP’s code and checkpoint to JAX (GRASP-JAX). For complexes exceeding 1,280 in length, we performed inference using GRASP-JAX on A100 (80G) GPUs. To reduce computational costs, structure relaxation was disabled by default, except when evaluating interface-specific physicochemical features (for example, hydrogen bonding), in which case the default relaxation protocol was applied to add hydrogen atoms. The basic setup described above was used for all GRASP inference, with some unique settings as follows:
For the experimental DMS dataset, we used both RPR and IR (Fig. 4) and applied quick mode iterative filtering to speed up inference (‘Iterative restraint filtering’ section). We also evaluated performance when IR and RPR were used separately (Supplementary Fig. 23).
For the simulated DMS dataset, we divided all complexes into antigen and antibody for docking. We used GRASP in Dimer Mode, which converts complexes with more than two chains into dimers by setting the same ‘asym_id’ for chains in each group and adding a 200-residue gap between chains61. This approach improves IR precision, narrows the search space, and simplifies the task. GRASP’s performance without Dimer Mode was also evaluated (Supplementary Fig. 9). For fair comparison, all other methods were also evaluated by splitting complexes into antigen and antibody.
For the integrative modeling cases, we applied the integrative strategy with Combfit mentioned above to incorporate additional information from density map for A3G-VCBC. For the rest two cases standard setup of GRASP was used. For the scoring stage of all three cases, we normalized each structure’s pLDDT and recall by subtracting the mean and dividing by the standard deviation, then selected the one with the highest average normalized score.
For the inference of the dimeric interactions in the mitochondria, we disabled the iterative restraint filtering strategy due to the sparsity of the restraints, the number of which is only 1–4 per complex.
Configurations for other structure prediction tools AFM
We used the same input features for AFM as GRASP. Five v.2.3 models were used, with each model predicting 5 structures for sequences up to 3,072 residues and 1 structure for longer sequences due to limitation in resource.
AlphaLink
We used the same input features for AlphaLink as GRASP. Inference was performed using the AlphaLink-MultimerSDAv2.pt checkpoint selected based on a prebenchmark on the 5% coverage simulated Sulfo-SDA restraints (Supplementary Fig. 24). To ensure a fair comparison, AlphaLink was run with an adjusted number of random seeds to match the number of generated structures in GRASP: 25 random seeds for complexes with up to 3,072 residues and 5 seeds for larger complexes.
ColabDock
The settings of ColabDock is the same as in ref. 22. We use AF2 (ref. 68) to generate single-chain templates and AFM to generate templates for the antibody side. If only one side of the interface was available, the IRs for the other side were generated by selecting residues with solvent accessibility ≥40% using the FreeSASA69 software. These single-chain templates and processed restraints also serve as inputs of HADDOCK and ClusPro. Please note that since GRASP does not necessarily require single-chain templates and readily handles one-sided interfaces, these two types of information are not needed. For complexes whose total length exceeds 700 residues, we ran segment-based optimization with recommended settings.
HADDOCK
We installed a local version HADDOCK v.2.4 to perform benchmark experiments. We followed the default setting of HADDOCK and picked the best structure by HADDOCK score from ‘water’ refinement stage as the final structure. For cases that failed in the ‘it1’ or ‘water’ refinement stage due to energy optimization issues, we picked the best structure in the last stage. All RPRs were set as unambiguous restraints, while IRs were set as ambiguous restraints.
ClusPro
We installed the program PIPER provided on the ClusPro website and convert the restraints into json format using official scripts. We set the restraints satisfaction ratio to 50% for IRs and decreased this ratio to 30% for cases without available output structure at 50% ratio.
AF3
We used the web server (https://golgi.sandbox.google.com/) for our predictions on experimental XL, CL and DMS datasets. After the code was open-sourced, we also benchmarked AF3 on the self-curated test set and the simulated DMS dataset (Hitawala–Gray dataset) using the released code (https://github.com/google-deepmind/alphafold3). Five random seeds generated 25 predictions, with the highest-ranking result selected. All other settings were default.
IMP
We installed IMP v.2.22.0 in local machine. Single templates are predicted by AlphaFold. We treated each chain in a protein complex as a rigid body and transformed the initial density map to three-dimensional (3D) gaussians to guide the modeling. The hyperparameters follow the tutorial of RNA Pol II on the IMP website unless otherwise stated.
Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
Data availability
All relevant evaluation datasets and results are available via OSF at https://doi.org/10.17605/OSF.IO/6KJUQ (ref. 70). Further training data for MEGAFold-Multimer and GRASP are released at http://ftp.cbi.pku.edu.cn/pub/psp/.
Code availability
The GRASP code is available via GitHub at https://github.com/xiergo/GRASP-JAX and via Zenodo at https://doi.org/10.5281/zenodo.15347070 (ref. 71) under the Apache v.2.0 license. Model weights have been deposited on OSF at https://doi.org/10.17605/OSF.IO/6KJUQ (ref. 70). In addition, a Colab notebook is available via GitHub at https://colab.research.google.com/github/xiergo/GRASP-JAX/blob/main/GRASP-JAX.ipynb for ease of use.
References
Pearce, R. & Zhang, Y. Toward the solution of the protein structure prediction problem. J. Biol. Chem. 297, 100870 (2021).
CAS PubMed PubMed Central Google Scholar
UniProt, Consortium. UniProt: the Universal Protein Knowledgebase in 2023. Nucleic Acids Res. 51, D523–D531 (2023).
Google Scholar
Berman, H. M. et al. The Protein Data Bank. Nucleic Acids Res. 28, 235–242 (2000).
CAS PubMed PubMed Central 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).
Abramson, J. et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493–500 (2024).
Lin, Z. et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science 379, 1123–1130 (2023).
CAS PubMed 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).
Bryant, P. Deep learning for protein complex structure prediction. Curr. Opin. Struct. Biol. 79, 102529 (2023).
CAS PubMed Google Scholar
Graziadei, A. & Rappsilber, J. Leveraging crosslinking mass spectrometry in structural and cell biology. Structure 30, 37–54 (2022).
CAS PubMed Google Scholar
Leitner, A., Faini, M., Stengel, F. & Aebersold, R. Crosslinking and mass spectrometry: an integrated technology to understand the structure and function of molecular machines. Trends Biochem. Sci. 41, 20–32 (2016).
CAS PubMed Google Scholar
Vogeli, B. The nuclear Overhauser effect from a quantitative perspective. Prog. Nucl. Magn. Reson. Spectrosc. 78, 1–46 (2014).
CAS PubMed Google Scholar
Limpikirati, P., Liu, T. & Vachet, R. W. Covalent labeling-mass spectrometry with non-specific reagents for studying protein structure and interactions. Methods 144, 79–93 (2018).
CAS PubMed PubMed Central Google Scholar
Williamson, M. P. Using chemical shift perturbation to characterise ligand binding. Prog. Nucl. Magn. Reson. Spectrosc. 73, 1–16 (2013).
CAS PubMed Google Scholar
Nerli, S., McShan, A. C. & Sgourakis, N. G. Chemical shift-based methods in NMR structure determination. Prog. Nucl. Magn. Reson. Spectrosc. 106-107, 1–25 (2018).
CAS PubMed PubMed Central Google Scholar
Fowler, D. M. & Fields, S. Deep mutational scanning: a new style of protein science. Nat. Methods 11, 801–807 (2014).
CAS PubMed PubMed Central Google Scholar
Starr, T. N. et al. Deep mutational scanning of SARS-CoV-2 receptor binding domain reveals constraints on folding and ACE2 binding. Cell 182, 1295–1310.e1220 (2020).
CAS PubMed PubMed Central Google Scholar
Berger, C. et al. Cryo-electron tomography on focused ion beam lamellae transforms structural cell biology. Nat. Methods 20, 499–511 (2023).
CAS PubMed Google Scholar
Puthenveetil, R. & Vinogradova, O. Solution NMR: a powerful tool for structural and functional studies of membrane proteins in reconstituted environments. J. Biol. Chem. 294, 15914–15931 (2019).
PubMed PubMed Central Google Scholar
Chavez, J. D. et al. Systems structural biology measurements by in vivo cross-linking with mass spectrometry. Nat. Protoc. 14, 2318–2343 (2019).
CAS PubMed PubMed Central Google Scholar
Stahl, K., Graziadei, A., Dau, T., Brock, O. & Rappsilber, J. Protein structure prediction with in-cell photo-crosslinking mass spectrometry and deep learning. Nat. Biotechnol. 41, 1810–1819 (2023).
Stahl, K. et al. Modelling protein complexes with crosslinking mass spectrometry and deep learning. Nat. Commun. 15, 7866 (2024).
CAS PubMed PubMed Central Google Scholar
Feng, S. et al. Integrated structure prediction of protein–protein docking with experimental restraints using ColabDock. Nat. Mach. Intell. 6, 924–935 (2024).
Google Scholar
de Vries, S. J., van Dijk, M. & Bonvin, A. M. The HADDOCK web server for data-driven biomolecular docking. Nat. Protoc. 5, 883–897 (2010).
PubMed Google Scholar
Dominguez, C., Boelens, R. & Bonvin, A. M. J. J. HADDOCK: a protein−protein docking approach based on biochemical or biophysical information. J. Am. Chem. Soc. 125, 1731–1737 (2003).
CAS PubMed Google Scholar
Kozakov, D. et al. The ClusPro web server for protein-protein docking. Nat. Protoc. 12, 255–278 (2017).
CAS PubMed PubMed Central Google Scholar
Comeau, S. R., Gatchell, D. W., Vajda, S. & Camacho, C. J. ClusPro: a fully automated algorithm for protein–protein docking. Nucleic Acids Res. 32, W96–W99 (2004).
CAS PubMed PubMed Central Google Scholar
Comeau, S. R., Gatchell, D. W., Vajda, S. & Camacho, C. J. ClusPro: an automated docking and discrimination method for the prediction of protein complexes. Bioinformatics 20, 45–50 (2004).
CAS PubMed Google Scholar
Drake, Z. C., Seffernick, J. T. & Lindert, S. Protein complex prediction using Rosetta, AlphaFold, and mass spectrometry covalent labeling. Nat. Commun. 13, 7846 (2022).
CAS PubMed PubMed Central Google Scholar
Russel, D. et al. Putting the pieces together: integrative modeling platform software for structure determination of macromolecular assemblies. PLoS Biol. 10, e1001244 (2012).
CAS PubMed PubMed Central Google Scholar
Molnar, K. S. et al. Cys-scanning disulfide crosslinking and Bayesian modeling probe the transmembrane signaling mechanism of the histidine kinase, PhoQ. Structure 22, 1239–1251 (2014).
CAS PubMed PubMed Central Google Scholar
Bayrhuber, M. et al. Structure of the human voltage-dependent anion channel. Proc. Natl Acad. Sci. USA 105, 15370–15375 (2008).
CAS PubMed PubMed Central Google Scholar
Dauden, M. I. et al. Architecture of the yeast Elongator complex. EMBO Rep. 18, 264–279 (2017).
CAS PubMed Google Scholar
Sirui, L. et al. Assisting and accelerating NMR assignment with restrained structure prediction. Commun. Biol. 8, 1067 (2025).
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).
CAS PubMed PubMed Central Google Scholar
Alford, R. F. et al. Correction to ‘The Rosetta all-atom energy function for macromolecular modeling and design’. J. Chem. Theory Comput. 18, 4594 (2022).
CAS PubMed Google Scholar
Lai, J. K., Ambia, J., Wang, Y. & Barth, P. Enhancing structure prediction and design of soluble and membrane proteins with explicit solvent-protein interactions. Structure 25, 1758–1770 e1758 (2017).
CAS PubMed PubMed Central Google Scholar
Bartolec, T. K. et al. Cross-linking mass spectrometry discovers, evaluates, and corroborates structures and protein-protein interactions in the human cell. Proc. Natl Acad. Sci. USA 120, e2219418120 (2023).
CAS PubMed PubMed Central Google Scholar
Wang, L., Xie, L. & Zhang, Z. Determination of HER2 binding domain in antigen-antibody complexes based on chemical crosslinking mass spectrometry. J. Proteom. 286, 104954 (2023).
CAS Google Scholar
Di Ianni, A., Di Ianni, A., Cowan, K., Barbero, L. M. & Sirtori, F. R. Leveraging cross-linking mass spectrometry for modeling antibody-antigen complexes. J. Proteome Res. 23, 1049–1061 (2024).
PubMed Google Scholar
Wang, X. et al. Molecular details underlying dynamic structures and regulation of the human 26S proteasome. Mol. Cell Proteom. 16, 840–854 (2017).
CAS Google Scholar
Kalisman, N., Adams, C. M. & Levitt, M. Subunit order of eukaryotic TRiC/CCT chaperonin by cross-linking, mass spectrometry, and combinatorial homology modeling. Proc. Natl Acad. Sci. USA 109, 2884–2889 (2012).
CAS PubMed PubMed Central Google Scholar
Blech, M. et al. One target—two different binding modes: structural insights into gevokizumab and canakinumab interactions to interleukin-1β. J. Mol. Biol. 425, 94–111 (2013).
CAS PubMed Google Scholar
Vreven, T. et al. Updates to the integrated protein-protein interaction benchmarks: Docking Benchmark Version 5 and Affinity Benchmark Version 2. J. Mol. Biol. 427, 3031–3041 (2015).
CAS PubMed PubMed Central Google Scholar
Hitawala, F. N. & Gray, J. J. What has AlphaFold3 learned about antibody and nanobody docking, and what remains unsolved? Preprint at bioRxiv https://doi.org/10.1101/2024.09.21.614257 (2024).
Cao, Y. et al. Imprinted SARS-CoV-2 humoral immunity induces convergent Omicron RBD evolution. Nature 614, 521–529 (2023).
CAS PubMed Google Scholar
Ruffolo, J. A., Chu, L. S., Mahajan, S. P. & Gray, J. J. Fast, accurate antibody structure prediction from deep learning on massive set of natural antibodies. Nat. Commun. 14, 2389 (2023).
CAS PubMed PubMed Central Google Scholar
Kosinski, J. et al. Molecular architecture of the inner ring scaffold of the human nuclear pore complex. Science 352, 363–365 (2016).
CAS PubMed PubMed Central Google Scholar
Jamali, K. et al. Automated model building and protein identification in cryo-EM maps. Nature 628, 450–457 (2024).
CAS PubMed PubMed Central Google Scholar
Kaake, R. M. et al. Characterization of an A3G-VifHIV-1-CRL5-CBFβ; structure using a cross-linking mass spectrometry pipeline for integrative modeling of host-pathogen complexes. Mol. Cell. Proteomics 20, 100132 (2021).
CAS PubMed PubMed Central Google Scholar
Kuenze, G. et al. Allosteric mechanism for KCNE1 modulation of KCNQ1 potassium channel activation. eLife 9, e57680 (2020).
CAS PubMed PubMed Central Google Scholar
Sun, J. & MacKinnon, R. Structural basis of human KCNQ1 modulation and gating. Cell 180, 340–347 e349 (2020).
CAS PubMed Google Scholar
Escobar, C. A. et al. Structural interactions define assembly adapter function of a type II secretion system pseudopilin. Structure 29, 1116–1127 e1118 (2021).
CAS PubMed Google Scholar
Banchenko, S. et al. Structural insights into Cullin4-RING ubiquitin ligase remodelling by Vpr from simian immunodeficiency viruses. PLoS Pathog. 17, e1009775 (2021).
CAS PubMed PubMed Central Google Scholar
Wheat, A. et al. Protein interaction landscapes revealed by advanced in vivo cross-linking-mass spectrometry. Proc. Natl Acad. Sci. USA 118, e2023360118 (2021).
CAS PubMed PubMed Central Google Scholar
Chen, Y. et al. Targeted cross-linker delivery for the in situ mapping of protein conformations and interactions in mitochondria. Nat. Commun. 14, 3882 (2023).
CAS PubMed PubMed Central Google Scholar
Rath, S. et al. MitoCarta3.0: an updated mitochondrial proteome now with sub-organelle localization and pathway annotations. Nucleic Acids Res. 49, D1541–D1547 (2021).
CAS PubMed Google Scholar
Merkley, E. D. et al. Distance restraints from crosslinking mass spectrometry: mining a molecular dynamics simulation database to evaluate lysine-lysine distances. Protein Sci. 23, 747–759 (2014).
CAS PubMed PubMed Central Google Scholar
Yoshida, S. et al. Molecular chaperone TRAP1 regulates a metabolic switch between mitochondrial respiration and aerobic glycolysis. Proc. Natl Acad. Sci. USA 110, E1604–E1612 (2013).
CAS PubMed PubMed Central Google Scholar
Thomas, P. D. The gene ontology and the meaning of biological function. Methods Mol. Biol. 1446, 15–24 (2017).
CAS PubMed PubMed Central Google Scholar
Bulbarelli, A., Valentini, A., DeSilvestris, M., Cappellini, M. D. & Borgese, N. An erythroid-specific transcript generates the soluble form of NADH-cytochrome b5 reductase in humans. Blood 92, 310–319 (1998).
CAS PubMed Google Scholar
Shor, B. & Schneidman-Duhovny, D. CombFold: predicting structures of large protein assemblies using a combinatorial assembly algorithm and AlphaFold2. Nat. Methods 21, 477–487 (2024).
Liu, S. et al. PSP: million-level protein sequence dataset for protein structure prediction. Preprint at https://arxiv.org/abs/2206.12240 (2022).
Steinegger, M. & Söding, J. MMseqs2 enables sensitive protein sequence searching for the analysis of massive data sets. Nat. Biotechnol. 35, 1026–1028 (2017).
CAS PubMed Google Scholar
Kahraman, A., Malmström, L. & Aebersold, R. Xwalk: computing and visualizing distances in cross-linking experiments. Bioinformatics 27, 2163–2164 (2011).
CAS PubMed PubMed Central Google Scholar
Abhinandan, K. R. & Martin, A. C. Analysis and improvements to Kabat and structurally correct numbering of antibody variable domains. Mol. Immunol. 45, 3832–3839 (2008).
CAS PubMed Google Scholar
Gydo, C. Pv. Z. & Alexandre, M. J. J. B. Fast and sensitive rigid-body fitting into cryo-EM density maps with PowerFit. AIMS Biophys. 2, 73–87 (2015).
Google Scholar
Eldar, Y., Lindenbaum, M., Porat, M. & Zeevi, Y. Y. The farthest point strategy for progressive image sampling. IEEE Trans. Image Process. 6, 1305–1315 (1997).
CAS PubMed Google Scholar
Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021).
CAS PubMed PubMed Central Google Scholar
Mitternacht, S. FreeSASA: an open source C library for solvent accessible surface area calculations. F1000Res. 5, 189 (2016).
PubMed PubMed Central Google Scholar
Xie, Y. et al. Datasets and model weights for GRASP. OSF https://doi.org/10.17605/OSF.IO/6KJUQ (2025).
Xie, Y. et al. Source code for GRASP. Zenodo https://doi.org/10.5281/zenodo.15347070 (2025).
Download references
Acknowledgements
This work was supported by the National Science and Technology Major Project (grant no. 2022ZD0115001 to Y.Q.G. and S. Liu), the National Natural Science Foundation of China (grant nos. 92353304 and T2495221 to Y.Q.G.) and New Cornerstone Science Foundation (grant no. NCI202305 to Y.Q.G.). We thank K. Stahl from Technische Universität Berlin for recommending AlphaLink configuration and providing simulated CASP15 restraints for debugging. We thank F. N. Hitawala from Johns Hopkins University for sharing the PDB IDs and sequence data of their curated benchmarking dataset. We thank Y. Kaiguang from Dalian Institute of Chemical Physics to provide guidance for usage of the targeted XL-MS in the mitochondria. We also thank Z. Zhu for helpful discussions on binding energy calculation.
Author information Author notes
These authors contributed equally: Yuhao Xie, Chengwei Zhang.
Authors and Affiliations
Changping Laboratory, Beijing, China
Yuhao Xie, Sirui Liu & Yi Qin Gao
Academy for Advanced Interdisciplinary Studies, Peking University, Beijing, China
Yuhao Xie
Biomedical Pioneering Innovation Center (BIOPIC), Peking University, Beijing, China
Chengwei Zhang, Yanjiao Lu & Yi Qin Gao
School of Life Sciences, Peking University, Beijing, China
Chengwei Zhang & Yanjiao Lu
New Cornerstone Science Laboratory, Beijing National Laboratory for Molecular Sciences, College of Chemistry and Molecular Engineering, Peking University, Beijing, China
Shimian Li, Xinyu Du, Zhenyu Chen & Yi Qin Gao
Huawei Technologies Co. Ltd, Hangzhou, China
Min Wang & Yingtong Hu
Contributions
Y.Q.G. and S. Liu designed and developed overall concepts in the paper and supervised the project. Y.X., C.Z. and X.D. developed the GRASP and Combfit methods. M.W. and Y.H. trained and evaluated MEGAFold-Multimer. Y.X., C.Z., S. Li, X.D., Y.L. and Z.C. performed data collection, evaluation and analysis. X.Y., C.Z., S. Li and X.D. wrote the initial draft of the paper. All authors contributed ideas to the work and assisted in editing and revision.
Corresponding authors
Correspondence to Sirui Liu or Yi Qin Gao.
Ethics declarations Competing interests
Changping Laboratory and Huawei Technologies Co., Ltd are in the process of applying for a patent covering the GRASP method, which lists authors including S. Liu, Y.X., C.Z. and Y.Q.G. as inventors. The other authors declare no competing interests.
Peer review Peer review information
Nature Methods thanks the anonymous reviewers for their contribution to the peer review of this work. Primary Handling Editor: Arunima Singh, in collaboration with the Nature Methods team.
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 1–24, Tables 1–7, Methods 1–7 and References.
Reporting Summary Supplementary Data
Detailed evaluation results and analytical data for each case across the six datasets.
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
Xie, Y., Zhang, C., Li, S. et al. Integrating diverse experimental information to assist protein complex structure prediction by GRASP. Nat Methods 22, 2362–2374 (2025). https://doi.org/10.1038/s41592-025-02820-1
Download citation
Received 11 October 2024
Accepted 08 August 2025
Published 15 September 2025
Version of record 15 September 2025
Issue date November 2025
DOI https://doi.org/10.1038/s41592-025-02820-1
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 Cryoelectron microscopy Machine learning Protein structure predictions Solution-state NMR This article is cited by In-situ cross-linking mass spectrometry reveals compartment-specific proteasomal interactions and structural heterogeneity Lili ZhaoRuntao ZhaoChun Tang
Nature Communications (2025)