Protein Structure and Function: Applications of Bioinformatics Methods - John Rigden 2014

Protein Dynamics: From Structure to Function
Conclusions and Perspectives

Computational Methods are gaining widespread recognition in structural biology and protein research. Protein function is typically a dynamic process involving structural rearrangements and conformational transitions between stable states. Because such dynamic processes are difficult to study experimentally, in silico methods can make a significant contribution to understanding protein function with atomic resolution.

The most prominent method for studying protein dynamics is Structure/8.html">Molecular Dynamics (MD), in which atoms are treated as classical particles and their interactions are approximated by empirical force fields. At each discrete time step, Newton's equations of motion are solved to generate a trajectory describing the dynamic behavior of the system. Despite the growing popularity of MD, its scope is constrained by heavy computational demands. Over the next decade, accessible simulation times for medium-sized Proteins are unlikely to reach the microsecond range for most biomolecular systems. However, because functionally relevant protein dynamics typically manifest as low-frequency motions occurring on a micro- to millisecond timescale, standard MD simulations are ill-suited for broad application in studying the conformational dynamics of large macromolecules.

To alleviate this sampling bottleneck plaguing standard MD, various approaches have been proposed. One strategy involves reducing the number of particles, either by grouping atoms into pseudo-atoms (coarse-grained representation) or by replacing explicit solvent molecules with an implicit continuum model. In both cases, the particle count is substantially reduced, enabling much longer simulation times than achievable in full-atom explicit-solvent calculations. Nevertheless, the loss of resolution inherent to both methods may limit their accuracy and, consequently, their applicability. Other approaches retain the all-atom description while implementing alternative sampling strategies.

Generalized ensemble algorithms, such as Replica Exchange (REX), exploit the fact that conformational transitions occur more frequently at elevated temperatures. In the REX method, simulations are run at a standard Temperature for multiple copies (replicas) of the system using MD at various temperatures, with frequent exchanges between replicas. This allows low-energy replicas to borrow the enhanced barrier-crossing capabilities of high-temperature replicas. Although dynamical information is lost in this setup, each replica still represents a Boltzmann ensemble at its respective temperature, providing valuable insights into the Thermodynamics and stability of various conformational substates. While all-atom replica exchange simulations are frequently used in protein folding studies, they quickly become computationally prohibitive for systems containing more than a few thousand atoms.

While the REX method is a non-distorting sampling technique, several methods deliberately perturb the system to enhance sampling along specific collective degrees of freedom. Functionally important protein motions often correspond to the eigenvectors with the largest eigenvalues in the atomic fluctuation covariance matrix. If these vectors are known from Principal Component Analysis (PCA), experimental data, or previous calculations, they can be utilized in simulation protocols such as Conformational Flooding or Essential Dynamics (ED). However, in both methods, improved sampling comes at the cost of losing the canonical Properties of the resulting trajectory.

The recently developed TEE-REX protocol combines the advantages of the REX algorithm with those arising from the targeted excitation of functionally relevant modes (as in ED), while avoiding the aforementioned drawbacks of both methods. Specifically, it approximately preserves the canonical integrity of the reference ensemble while significantly enhancing sampling along the principal collective motion modes. Consequently, the resulting reference ensemble can be used to calculate equilibrium system properties, enabling direct comparison with experimental data.

Although significant progress has been made in developing enhanced sampling methods, the computational demands of MD-based techniques remain high, often requiring weeks or months of CPU time on modern multi-processor clusters. However, for many structural biology questions, simply having An Overview of possible protein Conformations and functional modes is sufficient, without The Need for detailed timescales and energies. In such cases, Elastic Network Models offer a straightforward way to estimate potential functional protein motions. Although this approach relies on major simplifications and does not yield atomic-level movement details, the predicted collective motions often show good qualitative agreement with experimental results. Another computationally efficient approach that preserves the atomic representation is the CONCOORD method, which describes proteins through a set of geometric constraints. Based on a connectivity map derived from the initial structure, an ensemble of structures is generated, representing an exhaustive sampling of the conformational space compatible with the given constraints, though no information on timescales or energies is provided.

Currently, there is no single universal method ready for routine prediction of functionally relevant protein motions based on Spatial Structure. Nevertheless, a wide array of methods addresses different aspects of this challenge, collectively advancing our understanding of protein function. Thus, combining existing techniques is likely the most direct path to enhancing the predictive power of in silico methods.