Integrative protein modeling in RosettaNMR from sparse paramagnetic restraints

RosettaNMR offers a versatile platform for integrative modeling with paramagnetic NMR restraints

The success of protein structure modeling from limited NMR data can be markedly improved if different types of NMR data are combined in one calculation. To realize an integrative modeling approach, we extend the suite of NMR tools in Rosetta, which we collectively term RosettaNMR, by paramagnetic NMR data in combination with CSs and NOEs in one single framework (Figure 1).

We focused our implementation on PCSs, RDCs, and PREs since they are most informative in terms of the protein structure representing long-range distance and/or orientational restraints (Figure 1A). Each data type was implemented as a restraint energy penalty reporting the sum of squared errors between experimental and model-predicted data (details in Data S1). During scoring, the PCS score is obtained by computing the ΔΧ-tensor by singular value decomposition coupled with a grid search over the metal ion coordinates to find the best possible fit. The RDC score is computed analogously by determining the molecular alignment tensor, which is related to the ΔΧ-tensor by a constant pre-factor if alignment originates solely from the paramagnetic metal ion but is different for an external anisotropic medium (e.g. bicelles). To generalize score calculation for any para- or diamagnetic data, we decided to fit PCSs and RDCs separately. Moreover, effective tensor values obtained by fitting PCSs and RDCs tend to be not identical, because the ΔΧ-tensor and the alignment tensor average in different ways. The ΔΧ-tensor is affected by any tag motions that relocate the metal ion relative to the protein whereas RDCs are independent of the distance from the metal ion. Considering structural flexibility becomes even more important when interpreting PREs due to the many degrees of freedom (DOFs) of the spin-label; we have adapted the ensemble averaging approach from Iwahara and Clore for computation of the PRE score (Iwahara et al., 2004). The spin-label is modeled as an ensemble of side-chain conformations which are approximated from a pre-computed rotamer library. An effective distance <r⁻⁶> is calculated between the protein nuclear spin and all spin-label conformations, whose structural variability is represented by an order parameter S².

Rosetta allows manipulating the protein through modeling objects accessible from high-level scripting languages such as PyRosetta (Chaudhury et al., 2010) and Rosetta Scripts (Fleishman et al., 2011). This allows users to build new protocols by mixing and matching different modeling strategies (Figure 1B). The main tools fall into two categories: Movers and Filters. Movers change the conformation of a biomolecule and Filters decide whether the conformation should go into the next stage. One can easily devise new strategies of how paramagnetic restraints can be used to guide conformational sampling or model filtering. We exemplify this flexibility in method development by applying PCSs to ligand docking for which we have designed a PCS ligand scoring grid and a rigid-body sampling Mover, and by providing a supporting PyRosetta protocol (Data S2) for assessing collinearity between ΔΧ-tensors calculated from PCSs and RDCs. Furthermore, RosettaNMR allows adding spin-labels to the chemical database in a straightforward manner to accommodate new spin-label designs.

Paramagnetic NMR data enhance sampling of native-like models in de novo structure prediction

We tested RosettaNMR with RosettaAbinitio (Simons et al., 1997) for de novo structure prediction of 28 monomeric proteins (Table S1) using different types and combinations of NMR data: CSs, RDCs, PCSs, and PREs. We included further RDCs from external alignment media as proxies for paramagnetically-induced RDCs (see footnotes of Table 1 and S1) because only a limited number of RDCs measured with lanthanides could be found. However, the computational treatment of RDCs from both sources is identical. Table 1 summarizes the RMSD₁₀₀ of the best 1% of models ranked by either RMSD or score. Score-vs-RMSD plots of selected targets are displayed in Figure 2, and for all targets in Figure S1 and S2. Noticeably, paramagnetic data greatly improved sampling of near-native models. In 29 of 34 test cases, the best 1% of models ranked by RMSD had a lower RMSD₁₀₀ to the native structure when paramagnetic data were included (Figure 3A and 3F), and the average RMSD₁₀₀ improved from 4.5 Å to 3.5 Å (last row in Table 1; see Figure S3A for statistical analysis of RMSD change).

Comparing models generated with either paramagnetic data or CSs (Figure S4), we find that the former had a slightly better performance: for 72% of the proteins paramagnetic data led to a lower RMSD₁₀₀. Model sampling was most enhanced when RDCs, PCSs, and PREs were combined with CSs (Figure 3B), suggesting that these data have an orthogonal beneficial effect on the conformational search. In 29 of 34 cases, the best 1% of models ranked by RMSD had a lower RMSD₁₀₀ to the native structure when paramagnetic NMR and CS data were used (Figure 3F), and the average RMSD₁₀₀ decreased from 4.5 Å to 3.1 Å (last row in Table 1; see Figure S3A for statistical analysis of RMSD change).

The gradual improvement in model sampling by including CSs, paramagnetic restraints, or a combination of both can be seen in the score-vs-RMSD and density plots in Figure 2, S1, and S2. The increase in density of near-native models is apparent in the low RMSD₁₀₀ region. The fraction of proteins for which a high-accuracy model (RMSD₁₀₀ ≤ 2 Å) could be generated, increased from 9% (no NMR) to 35% (CSs), 35% (paramagnetic data), and 50% (CSs and paramagnetic data). Medium-accuracy models (≤ 3 Å) could be sampled in 62% (no NMR), 68% (CSs), 76% (paramagnetic data), and 76% (CSs and paramagnetic data) of cases.

Paramagnetic NMR data enrich for native-like models in scoring

A similar improvement in model quality was observed when they were ranked by score (Table 1, Figure 3D and 3E). This is more realistic since in a blind prediction models must be identified mostly by their score. Here, we calculated the score as combination of the Rosetta all-atom energy and the RDC, PCS, or PRE score with a weighting factor that was optimized for each NMR dataset and protein as described under Methods. For those cases where PCSs and RDCs were used, both the PCS and RDC score was added to the Rosetta energy with weights proportional to the logarithm of the number of PCS and RDC values.

For 28 of 34 and 30 of 34 targets the ten best-scoring models had a lower average RMSD₁₀₀ when paramagnetic restraints were used individually or combined with CSs (Figure 3D and 3E). The average RMSD₁₀₀ of the ten lowest-scoring models over all proteins improved from 7.6 Å (no NMR) to 5.0 Å (paramagnetic data) and 4.2 Å (CSs and paramagnetic data) (last row in Table 1, Figure S3B). The number of proteins for which the RMSD₁₀₀ of the top ten models was ≤ 3 Å increased from two (no NMR) to 14 (paramagnetic data) and 18 (CSs and paramagnetic data).

To investigate the ability of the RDC, PCS, and PRE score to recognize near-native models we calculated model enrichment (maximum value is 10, see Method Details) displayed in Figure 3C. To this end, models built without paramagnetic data were rescored with RDCs, PCSs, or PREs and the Rosetta energy function. The average enrichment over the benchmark set increased from 2.5 to 3.7 (RDCs: 2.5 → 3.8; PCSs: 2.2 → 4.1; RDCs + PCSs: 1.9 → 4.2). No significant increase in enrichment was seen for the PRE score (2.5 → 2.6), showing that it failed to further improve model selection with Rosetta’s all-atom energy function. Although we only have experimental PREs from three proteins, the observations match those made with simulated PREs for symmetric proteins (see below). We speculate that the lower discriminative power of the PRE score is because PREs are purely distance-dependent whereas PCSs are sensitive to orientation and distance. In addition, PCSs exhibit a more favorable r⁻³ distance dependence (r⁻⁶ for PREs) and are less affected by the presence of minor protein states or encounter complexes (Clore and Iwahara, 2009). Consequently, the PCS score drastically improves model discrimination when included in Rosetta’s energy function.

Paramagnetic NMR data facilitate model selection

Experimental data can be compared to predicted data for model selection and validation. The top RosettaNMR models satisfy their experimental data remarkably well (Figure 4A-4C for proteins 2K61, 5T1N, and 1CDL, Figure 4D and Table S2). A direct measure of the agreement provides the NMR quality factor (Q-factor), which we calculated for the ten best-scoring models of each target. We saw a steady improvement of the Q-factor with the inclusion of CSs and paramagnetic data; a combination of both restraint types yielded Q-factors comparable to experimental structures (Figure 4D). While model cross-validation against data deliberately left out from the calculation is beyond the scope of this study, we note that we observed improved Q-factors also in cases where paramagnetic data were only used for final model evaluation. For the set of models created with CSs the Q-factors improved on average from 0.58 to 0.52 (RDCs), 0.20 to 0.18 (PCSs), and 0.63 to 0.61 (PREs) relative to the unrestrained calculation (Figure 4D). Improved Q-factors in those cases may reflect increased structure accuracy of RosettaNMR models even when alternative restraints were used in building them.

We used two measures to assess the quality and estimate success of structure prediction: a) low Q-factors confirm good agreement with the experimental data and b) model convergence indicates that a topology is repeatedly sampled and satisfies the Rosetta and NMR score. We defined convergence as the average RMSD between the best-scoring model and the next nine low-scoring models. We found that paramagnetic restraints and CSs were highly beneficial for model convergence and a steady RMSD improvement was observed when including more NMR data (Figure S5A). In 28 of 34 cases, the model created with CSs and paramagnetic restraints had a convergence ≤ 4 Å (Table S2, Figure S5A). This RMSD cutoff is used in the calculation of other protein structure similarity metrics (e.g. GDT-TS and MaxSub) and below it models usually have the same global topology. The other six proteins (2KYW, 2KYY, 2KD7, 1RW5, 1CDL, 1J54) had limited convergence, identifying their model as unreliable (Figure S6).

Iterative resampling with paramagnetic NMR data further improves model quality

Inspection of these six prediction failures according to our convergence criterion revealed no high-accuracy model (≤ 2 Å) and only a small number of intermediate-accuracy models (≤ 3 Å) after de novo prediction. Insufficient sampling was also noticed for another three proteins (1CMZ, 1RJV, 2K5U) with consequently high RMSD₁₀₀ for their top-scoring models. To test if enhanced conformational sampling could improve model accuracy, we applied the Rosetta Iterative Hybridize refinement protocol (Ovchinnikov et al., 2017), incorporating paramagnetic restraints.

We ran 30 rounds of refinement and selected the offspring generation of models at each step by their combined Rosetta and NMR score. The RMSD₁₀₀ steadily improved for eight of nine proteins; the RMSD₁₀₀ of the ten lowest-scoring models decreased by 3.2 Å on average, improving the number of proteins with RMSD₁₀₀ ≤ 5 Å for the ten lowest-scoring models from one of nine to six of nine (Table S3, Figure S7). For two of nine proteins at least one high-accuracy model was present in the pool after refinement.

RosettaNMR with paramagnetic NMR data improves structure prediction of symmetric proteins

We further applied RosettaNMR to symmetric proteins, focusing on systems with cyclical (Cn) and dihedral (Dn) symmetry as they are most abundant in nature. We hypothesized that RDCs, PCSs, and PREs provide sufficient information to fold the asymmetric monomer and sample symmetric rigid body DOFs toward low-energy and experimentally valid conformations. We incorporated point symmetry into the RDC, PCS, and PRE scoring methods, allowing the use of experimental data collected on the whole protein assembly through symmetric tagging. Hence, laborious and time-consuming asymmetric tagging of single subunits to break the symmetry for the ease of assignment is not required.

We combined RosettaNMR and the Fold-and-Dock protocol (Das et al., 2009), which simultaneously models the asymmetric monomer and symmetric complex. The protocol was tested with published RDCs and simulated PCSs and PREs for six dimeric proteins, one protein trimer, and one tetramer (Table 2 and Table S1). In addition, we employed a two-step protocol for proteins 1RJJ and 2M89 for which Fold-and-Dock failed to build models < 10 Å to the native structure. Here, models of the monomer were first built with RosettaAbinitio and then used for symmetric docking.

Table 2 summarizes the average RMSD₁₀₀ of the best 1% of models by RMSD and of the ten bestscoring models. Score-vs-RMSD plots of selected proteins are displayed in Figure 5 and for all targets in Figure S8. We found a clear improvement in model sampling and scoring with paramagnetic data. The number of cases with a lower RMSD₁₀₀ among the best 1% of models was seven (RDCs), six (PCSs), and seven (PREs) out of eight (considering the symmetric docking results for 1RJJ and 2M89), and the average RMSD₁₀₀ over the best 1% of models for all proteins decreased from 5.2 Å (no NMR) to 3.6 Å (RDCs), 3.8 Å (PCSs), and 4.2 Å (PREs). The sampling density of low-RMSD models improved considerably for many targets; the number of cases with at least one high-accuracy model (≤2 Å) improved from one (no NMR) to five (RDC), six (PCS), and two (PRE). As in our analysis for monomeric proteins, paramagnetic NMR data had a beneficial effect on scoring and convergence. The average enrichment increased from 2.3 to 2.7 (RDCs), 3.1 (PCSs), and 2.4 (PREs) (Figure 5E), and the mean convergence improved from 5.4 Å to 4.2 Å (RDCs), 2.9 Å (PCSs), and 5.1 Å (PREs) (Figure S5B). Using a combination of convergence and NMR Q-factor to assess prediction reliability, five (RDCs), six (PCSs), and three (PREs) out of eight cases could be classified as successful (Figure S6B).

PCSs are highly beneficial to identify native interfaces in protein-protein docking

PCSs are particularly attractive for protein-protein docking because they combine distance and angular information and are easy to interpret in terms of the coordinates of a nucleus within the ΔΧ-tensor frame. This strategy was first realized by implementing PCSs into rigid-body docking approaches (Pintacuda et al., 2006; Ubbink et al., 1998), and later into the docking program HADDOCK (Schmitz and Bonvin, 2011). Here, we combined PCSs in RosettaNMR with RosettaDock (Gray et al., 2003) and show that protein-protein interfaces can be accurately identified using PCS data as the only source of experimental information.

Method performance was assessed on two systems with experimental PCSs (Table 3): a) a ternary complex of FKBP12 (FK506-binding protein), rapamycin and the FKBP12-rapamycin-binding (FRB) domain of FRAP (PDB 1FAP) (Kobashigawa et al., 2012) (Figure 6A and 6B) and b) a homodimer of the PB1 domain of p62 (PDB 2KTR) (Saio et al., 2010) (Figure S9). For FKBP12-rapamycin-FRB , we obtained a best-scoring model with an interface RMSD of 0.5 Å to the X-ray structure (Figure 6A). The final ensemble of ten models was highly converged with an average interface RMSD of 0.6 ± 0.2 Å to the native structure and was in excellent agreement with the PCS data (Q = 0.07). In contrast, docking without PCS data failed to converge; the final ensemble had an average RMSD of 13.2 ± 6.6 Å (Figure 6B).

For the p62 PB1 homodimer, our calculations converged to a best-scoring model with an interface RMSD of 1.3 Å (Figure S9). The final ensemble of ten models had an average RMSD of 1.7 ± 0.3 Å and was in good agreement with the experimental PCSs (Q = 0.17). In contrast, docking without PCSs failed to converge and the interface RMSD of the best ten models was 14.1 ± 2.4 Å (Figure 6B) (Q = 0.21).

PCSs efficiently guide protein-ligand docking to native-like complexes

PCSs are beneficial for ligand docking because they can be measured for ligands with a broad range of affinities and provide atomic detail information about the ligand binding mode (Guan et al., 2013). Here, we tested the performance of RosettaNMR with RosettaLigand (Meiler and Baker, 2006) to leverage PCSs for modeling protein-ligand interactions. Using PCSs for three protein-ligand systems with different affinities and no prior knowledge about where the ligand binds, our method recovered the native ligand binding mode with an accuracy of up to 1.1 Å ligand RMSD.

We employed a two-step protocol for global protein-ligand docking. A search of the ligand’s six rigid body DOFs in a grid around the protein identifies positions with low PCS score, which are then used as input into RosettaLigand. Subsequent high-resolution docking refines the structure to atomic detail. Furthermore, we implemented a PCS scoring grid (DeLuca et al., 2015) into RosettaLigand, which evaluates the ligand position given the experimental PCSs and a ΔΧ-tensor as input. The tensor components can be calculated from the proteins’ PCSs prior to docking.

We tested this method on three protein-ligand complexes (Table 3): the SH2 domain of Grb2 bound to a low-affinity phosphorylated pYTN tripeptide as well as a high-affinity macrocyclic inhibitor (Saio et al., 2011), and a complex of dengue virus NS2B-NS3 protease (DENpro) with a high-affinity ligand (Chen et al., 2016). PCS-assisted docking of the two SH2 ligands was converged. The models of the pYTN ligand (Figure 6C) and the macrocyclic inhibitor (Figure S9) with the lowest combined Rosetta interface and PCS score had a ligand RMSD of 1.1 Å in both cases, and were in good agreement with the NMR data (Q = 0.14 and Q = 0.23). The average RMSD of the ten best models was 1.1 ± 0.1 Å and 1.4 ± 0.6 Å, respectively (Figure 6D). In contrast, ligand docking without PCSs failed to identify the binding pocket and led to ligand RMSDs of 14.7 ± 6.9 Å and 9.8 ± 4.5 Å, respectively. Similar improvements were found for the third system, DENpro with a high-affinity ligand. Our lowest scoring model had a PCS Q-factor of 0.11 and a RMSD of 2.1 Å to the model by Chen et al. (Figure S9) which bears high similarity to the X-ray structure of a related peptide-DENpro complex (PDB 3U1I) (Noble et al., 2012).

Lead Contact and Materials Availability

For information how to download and install Rosetta please visit https://www.rosettacommons.org/software/license-and-download and https://www.rosettacommons.org/docs/latest/Home. Instructions for using RosettaNMR with paramagnetic restraints can be found in the protocol capture (Data S3) accompanying this publication. Any further requests for resources and help should be directed to and will be fulfilled by the lead contact, Georg Kuenze (georg.kuenze@gmail.com).

Method Details

Quantification and Statistical Analysis

Calculation of the mean (± S.D.) Cα-atom RMSD of the best 1% models by RMSD and the ten top-scoring models listed in Table 1, 2, and 3 and displayed in Figure 3, 5, and 6 was carried out with Numpy as available with Anaconda. Statistical significance of the difference in model enrichment and RMSD for calculations with and without paramagnetic restraints was evaluated by a two-tailed Wilcoxon signed-rank test and carried out with Scipy as available with Anaconda. Significance values (* p < 0.05, ** p < 0.01, *** p < 0.001) are listed in Table 1 and displayed in Figure 3, 5 and S3, where n refers to the number of benchmark proteins per type of NMR restraint (i.e. RDC, PCS or PRE).

Data and Code Availability

The Rosetta software is freely available for academic users at https://www.rosettacommons.org/. A license must be obtained from the University of Washington through the Express Licensing Program (https://www.rosettacommons.org/software/license-and-download). The paramagnetic RosettaNMR framework is part of Rosetta as of version v2019.02-dev60579 and later (for information see also: https://www.rosettacommons.org/docs/latest/application_documentation/RosettaNMR-with-Paramagnetic-Restraints). A protocol capture containing detailed instructions how to run RosettaNMR calculations with paramagnetic NMR data as well as accompanying input files and command lines is available with file Date S3. An example script for how to use RosettaNMR through PyRosetta is provided with file Data S2.