TL;DR
This paper systematically compares two methods for refining protein structural ensembles using experimental data: restraining during simulations and reweighting after simulations. They found reweighting is computationally efficient and performs comparably to restraining when the initial simulation data is good, and they propose metrics to assess the reliability of reweighted ensembles, especially when experimental data is limited.
What Problem Does This Paper Address?
Proteins aren't static; they exist as dynamic ensembles of different shapes (conformations) to perform their functions. Accurately characterizing these ensembles is crucial for understanding protein behavior. Molecular dynamics (MD) simulations can generate these ensembles, but they are limited by the accuracy of the underlying physical models (force fields). Integrating experimental data with MD is key to improving accuracy. However, there are two main ways to do this: either by applying experimental constraints during the simulation (restraining) or by adjusting the probabilities of conformations after the simulation is complete (reweighting). It's not always clear which approach is better or under what conditions each works best, especially when the initial simulation data (the "prior ensemble") might not be very good.
Methods
The core of this work is a systematic comparison of "restraining" and "reweighting" approaches.
- Restraining: This involves adding extra forces to the MD simulation to guide it towards conformations that better match experimental data.
- Reweighting: This starts with an existing MD simulation (the "prior ensemble") and mathematically adjusts the statistical weights of the sampled conformations to make the ensemble averages better match experimental data. The authors implement a maximum entropy-based reweighting method that allows for the incorporation of experimental and forward model errors. They derive the biasing potential for reweighting from the maximum entropy principle.
- Optimization: The process of finding the correct weights for reweighting is framed as an optimization problem. They use a gradient descent approach and, for better convergence, a conjugate gradient approach to minimize an objective function.
- Error Models: They explore two error models for experimental data: a Gaussian error model and an uninformative Jeffreys prior for the variance.
Datasets
The authors use two systems for their analysis:
- Dialanine: A simple amino acid used to test the reweighting method's ability to shift free energy landscapes and assess computational efficiency. They generated synthetic data by shifting the free energy landscape of dialanine.
- c-Myc402-412: A peptide from the oncogenic protein c-Myc, studied in the presence of a small molecule inhibitor (10058-F4). This system uses experimentally available NMR chemical shifts as the target data. They used multiple force field settings (Amber94, Amber03w, Amber03ws) and simulation techniques (PBMETAD, REMD) to generate prior ensembles.
Key Insights and Ideas
- Reweighting is Efficient: The reweighting approach is computationally efficient, often requiring only minutes on a personal computer to process simulation data.
- Accuracy Depends on Prior Quality: Reweighting performs comparably to restraining when the initial simulation ensemble (the "prior") is already relatively accurate. However, if the prior ensemble poorly overlaps with the true ensemble, reweighting may not significantly improve the results.
- Metrics for Reliability: When validation data is scarce, it's crucial to have ways to assess the reliability of a reweighted ensemble. The authors propose three key indicators:
- Domain Failures: Checking if the range of predicted experimental observables from the ensemble overlaps with the actual experimental data. A lack of overlap suggests a problem.
- Block Error Analysis: Computing the error of an observable across different blocks (segments) of the reweighted simulation. This helps identify inconsistencies.
- Kish Effective Sample Size: A metric that quantifies how much the reweighting process has concentrated the ensemble onto a few frames versus distributing it broadly. Lower scores indicate less reliable reweighting.
Key Results
- Dialanine: Reweighting successfully shifted the free energy landscape of dialanine closer to the true ensemble, reducing the average free energy error from 6.70 kJ/mol to 3.87 kJ/mol. The weighted free energy error was significantly reduced to 0.41 kJ/mol, indicating good agreement in highly populated regions.
- c-Myc402-412: Both reweighting and restraining approaches, when applied to priors generated with different force fields, showed comparable performance in matching experimental Ha chemical shifts. Reweighting reduced the average deviation from experimental Ha chemical shifts to between 0.06 and 0.08 ppm across all force fields.
- Poor Sampling Impact: In tests with deliberately poor sampling (shifted ensembles), the proposed metrics (domain failures, block error, Kish score) were shown to effectively signal unreliable reweighting. For instance, a low Kish score (below -8) strongly indicated that reweighting was less reliable.
Limitations and Future Work
- Limitations of Reweighting: The paper acknowledges that reweighting can be inaccurate when the prior ensemble has poor overlap with the true ensemble. In such cases, it might not recover the true ensemble, especially in less populated regions of conformational space.
- Future Work: The authors anticipate that inventive combinations of restraining and reweighting approaches will lead to more accurate methods for integrating experiments and simulations. This will ultimately improve our mechanistic understanding of protein function by enabling more precise determination of structural ensembles. They also suggest that other experimental data sets might be able to better correct discrepancies between different force fields for systems like c-Myc.
Summary updated