TL;DR
This paper introduces ABSURDer, a novel computational method that integrates molecular dynamics (MD) simulations with NMR relaxation data to more accurately model protein dynamics. By reweighting simulation trajectories based on experimental NMR relaxation measurements, ABSURDer improves the agreement between simulations and experiments, leading to better insights into protein structure and function.
What Problem Does This Paper Address?
Understanding the dynamic behavior of proteins is crucial for comprehending their function. While molecular dynamics (MD) simulations offer a powerful tool to study these dynamics, and Nuclear Magnetic Resonance (NMR) experiments can probe them at atomic resolution, bridging the gap between these two approaches has been challenging. Specifically, accurately capturing the fast (ps-ns) timescale motions of protein side chains using MD simulations has been difficult, often leading to discrepancies with experimental NMR relaxation data. These discrepancies can arise from limitations in sampling, inaccuracies in force fields, or issues with the "forward model" used to interpret simulation data. This paper aims to address this by developing a method to directly fit MD simulations to experimental NMR relaxation data, thereby improving the accuracy of simulated protein dynamics.
Methods
The core of this work is the ABSURDer (Average Block Selection Using Relaxation Data with Entropy Restraints) method. It builds upon the ABSURD approach by:
- Spectral Density Mapping: Calculating deuterium NMR relaxation rates from MD simulations. This involves:
- Dividing long MD trajectories into shorter, non-overlapping blocks (10 ns in this study).
- Calculating internal correlation functions for methyl C-H bonds within each block.
- Incorporating global protein tumbling using an axially symmetric model derived from backbone N-H dynamics.
- Transforming these correlation functions into spectral density functions, J(ω).
- Extracting three specific deuterium NMR relaxation rates (R(D₂), R(D₁), and R(3D₂-2)) from J(ω).
- Reweighting with Entropy Restraint: This is the novel aspect of ABSURDer.
- The 1500 blocks from the MD simulations are assigned weights.
- These weights are optimized to minimize a cost function that balances the agreement between simulated and experimental NMR relaxation rates (χ²) with an entropy restraint term.
- The entropy restraint helps prevent overfitting by penalizing large deviations from the initial, uniform weights, thus maintaining a connection to the original force field and sampling. The functional minimized is T(w) = χ²(w) – θS(w), where θ controls the balance between fitting the data and preserving the prior information.
Datasets
The study utilizes the following datasets:
- T4 Lysozyme (T4L) protein: This protein was used as the model system.
- Synthetic NMR relaxation data: Generated from MD simulations using the same force field (Amber ff99SB*-ILDN) and a different force field (Amber ff15ipq) to mimic experimental data and test the method's robustness to force field differences.
- Experimental NMR relaxation data: Deuterium relaxation measurements for T4 lysozyme, specifically for 73 methyl groups, as previously reported (Hoffmann et al., 2018b).
Key Insights and Ideas
- Bridging Dynamics and Structure: Fitting dynamic quantities (NMR relaxation rates) can lead to improved agreement with static properties (like dihedral angle distributions), suggesting a deeper connection between how simulations are refined and their overall accuracy.
- The Power of Reweighting: ABSURDer demonstrates that reweighting MD trajectories based on experimental NMR relaxation data can significantly improve the agreement between simulations and experiments, even when starting from good quality simulations.
- Entropy Restraint is Crucial: The inclusion of an entropy restraint in the reweighting process is vital to avoid overfitting, ensuring that the reweighted ensemble remains representative of the underlying simulations and force field.
- Force Field Matters, but Can Be Mitigated: While differences in force fields introduce larger discrepancies, ABSURDer can still improve agreement, albeit to a lesser extent than when using the same force field for simulation and data generation.
- Global vs. Local Fitting: The paper explores fitting all NMR relaxation data globally versus fitting subsets. Global fitting is more challenging due to the "curse of dimensionality" but offers a more comprehensive view of protein dynamics.
Key Results
- Synthetic Data (Same Force Field): ABSURDer achieved a substantial reduction in reduced χ² (from ~15-22 to ~1-3) when fitting synthetic data generated with the same force field as the simulations. This reweighting also improved dihedral angle distributions.
- Synthetic Data (Different Force Field): When fitting synthetic data generated with a different force field, ABSURDer still improved agreement, reducing χ² by about two-fold (compared to a ten-fold reduction with the same force field).
- Experimental Data: For experimental NMR data, ABSURDer provided a more modest improvement, decreasing χ² by an average of 30%. This suggests that while the method is effective, limitations in sampling, force field accuracy, or the forward model still play a role.
- Block Size Optimization: A block size of 10 ns was found to be a good compromise for balancing the accuracy of correlation functions and the number of blocks for reweighting.
- Convergence: MD simulations of several microseconds provided relatively precise estimates of relaxation rates.
Limitations and Future Work
- "Curse of Dimensionality": Fitting a large number of NMR relaxation rates simultaneously can be challenging, especially when there's limited overlap between simulated and experimental rates.
- Overlap is Key: The effectiveness of reweighting is dependent on the overlap between the distribution of simulated rates and the experimental values. If experimental rates fall outside the sampled range, reweighting has limited impact.
- Force Field and Sampling: While ABSURDer can mitigate some issues, fundamental inaccuracies in force fields or insufficient sampling can limit the achievable agreement.
- Future Directions:
- Exploring blocks of different sizes to capture motions across various timescales.
- Combining relaxation data from different magnetic field strengths.
- Developing more localized models of protein dynamics to overcome the high-dimensionality of global fitting.
- Increasing sampling and improving force fields remain crucial for further advancements.
Summary updated