What will computational modelling approaches have to say in the era of atomistic cryo-EM data?

Abstract

The focus of this viewpoint is to identify, in the era of atomistic resolution cryo-electron microscopy data, the areas in which computational modelling and molecular simulations will bring valuable contributions to structural biologists and to give an overview of some of the existing efforts in this direction.

Over the last decade, cryo-electron microscopy (cryo-EM) has become an invaluable technique in structural biology. Thanks to the recent developments in instrumentation, sample preparation, and image-processing software, cryo-EM has now reached atomistic resolution (a resolution high enough to allow structural modelling of unique positions for most atoms in a protein). The progress has been rapid: of the 1818 single-particle EM maps with resolution better than 15 Å deposited during the course of 2019 in the EMDB database, 61% and 86% were resolved at resolution better than 4 Å and 6 Å, respectively. The resolution record is currently held by a 1.54 Å resolution structure of apoferritin (EMD-9865). As of today, 45% of all the single-particle EM maps deposited over the years reported a resolution better than 5 Å.¹

As the number of atomistic cryo-EM datasets rapidly increases, one can wonder what computational modelling approaches will bring to the table. The focus of this viewpoint is to identify the areas in which these techniques can complement the vast amount of information provided by cryo-EM. Simulation approaches have traditionally been leveraged to derive single structures that are good fits to the density maps.^2–4 Here, we highlight some of the new efforts to expand this direction, while acknowledging that more exhaustive overview of traditional modeling and validation methods have recently been published.^{5, 6}

Cryo-EM has uses beyond being a powerful technique for determining PDB structures. Due to the single molecule nature of the experiment, cryo-EM presents new opportunities to study the conformational landscape of dynamic macromolecular systems. Such characterizations can be directly obtained from the raw data, i.e. the set of single-particle, two-dimensional (2D) images of the system deposited on a thin layer of vitreous ice. While there have been initial attempts to define conformational landscapes^7–9, currently the most common practice is to derive multiple three-dimensional (3D) density maps in distinct conformational states. Distinct atomistic models can then be built into these maps and some information about their relative populations can be obtained from the number of particles underlying their reconstruction.

Ensemble modeling¹⁰ may be able to bridge the current gap between the long term ideal of “conformational landscapes” and the current reality of “single structures”. Computational advances will play a key role in this new direction. While image classification techniques are often able to distinguish distinct conformational states at the level of 2D class averages, highly-dynamic parts of the system are sometimes difficult to identify even with focused classification approaches.¹¹ Low resolution regions of a cryo-EM map might therefore hide multiple different, but modellable, conformations whose densities have been averaged out in the processing of the raw data.

Cases where these regions exhibit continuous dynamics are particularly challenging. Highly-flexible parts of otherwise well-ordered systems, such as short loops or other disordered regions, are hard to resolve by cryo-EM image-processing alone, yet they are often crucial for specific biological functions. In these cases, traditional modelling approaches that yield a single structure or multiple independently refined models¹² into a density map may not be helpful, as they may not faithfully represent the underlying dynamics of the system.

Recently, several different computational approaches aimed at determining conformational ensembles consistent with ensemble-averaged experimental data have been developed.¹³ These methods have traditionally been used in combination with solution experiments, such as Nuclear Magnetic Resonance (NMR) spectroscopy or small-angle X-ray scattering, either on-the-fly during a molecular dynamics (MD) simulation to improve the quality of the underlying force field or a posteriori to refine an ensemble previously generated using MD or other modelling techniques. These methods can now be extended to generate structural ensembles from cryo-EM density maps.

Metainference¹⁴ is a method for determining structural ensembles based on a Bayesian probabilistic framework to integrate noisy, ensemble-averaged experimental data into MD simulations. This is a powerful approach for characterizing the conformational heterogeneity in maps that resist further 3D classification, as recently illustrated by the studies of the dynamics of the gating region of the ClpP protease¹⁵ and the effect of acetylation on α-tubulin.¹⁶ Molecular mechanics force fields used in MD simulations are becoming more and more accurate in the description of different environments and their interaction with macromolecules. Therefore, there is an emerging opportunity for integrative methods that combine cryo-EM data with MD to provide a more accurate description of the interactions between macromolecules and other smaller components (lipids, ions, solvent, ligands, etc) as they become visible in atomistic maps. Although these approaches usually come at a higher computational cost compared to standard real-space refinement techniques, they can provide crucial insights into the interaction of proteins with their environment.

As the resolution of cryo-EM maps increases, more components of the system are becoming more and more visible at an unprecedented level of detail, such as ordered water molecules, lipids, and ions. Software for single-structure refinement typically provide either none or simplified physico-chemical models of the environment surrounding biological systems. For example, soluble proteins are traditionally refined in 3D density maps using energy functions that describe only basic stereochemical properties and not the surrounding environment. Furthermore, even for the ensemble-modelling approaches that use more accurate molecular mechanics force fields, such as metainference, modelling of ordered water and lipid densities is still challenging and will require further methodological developments.

One of the major challenges for ensemble-modelling approaches is the ability to distinguish conformational heterogeneity from noise in the data. Both these causes can result in the presence of regions at lower resolution in atomistic maps. To overcome this challenge, a modelling approach that accounts for the simultaneous presence of both structural heterogeneity and noise is required along with i) structural priors that can (alone) describe the dynamics of a system sufficiently well and ii) accurate estimates of the experimental errors at play.

Rather than relying on 3D maps, new approaches are emerging that use the raw 2D particle stacks, which are sometimes available in the public EMPIAR database.¹⁷ Notable examples are manifold embedding⁷, BioEM¹⁸, and a variational autoencoder to connect unlabelled 2D cryo-EM images with continuous distributions over 3D densities.¹⁹ The major advantage of these approaches consists in using the raw data prior to any clustering or averaging procedure, therefore fully embodying the single-molecule nature of the cryo-EM experiment. Currently, methods development in this area is limited by the sporadic practice of depositing raw data in the EMPIAR database. These methods are mostly limited by the low signal-to-noise ratio of individual particles, which will be mitigated in time as detectors continue to develop.

Computational modelling can also provide information on several aspects of the cryo-EM experiment that are needed to relate the results to the solution, room temperature ensemble. For example, how particles interact with the air-water interface^{20, 21} could be studied by multiscale methods. Simulations could also be leveraged to determine the vitrification process on the effective “temperature” of the resulting ensemble of molecules. Prior to data collection, the sample is prepared in solution at room temperature and then rapidly cooled down to cryogenic temperatures. The timescale of freezing is not fully known, but may take from hundreds of microseconds to a few milliseconds. In this timescale, scarcely populated “excited” states might collapse in neighboring free-energy minima, while more populated, stable states should be less affected. On a more local scale, rotamers and loops are often highly mobile on the microsecond timescale, and thus they may have time to structurally reorganize during freezing. Therefore, the conformational landscape represented by the cryo-EM single-particle images might differ in subtle, but potentially important ways from the room temperature, biologically relevant ensemble.

One potential approach to study these effects is non-equilibrium MD. By simulating the freezing process starting from a set of conformers extracted from a solution, room-temperature ensemble, which can be modelled using equilibrium MD at 300K, down to the cryogenics ensemble, such simulations can highlight the potential differences between cryo-EM and room temperature ensembles. Recent experiments have begun to address these questions experimentally by incubating the samples at different temperatures prior to the freezing and vitrification process^{22, 23}, and will provide useful points of comparison for molecular simulations.

While the rate of deposition of atomistic cryo-EM maps in the EMDB database is rapidly increasing, a substantial part of the available data is still at medium-low resolution. Integrative modelling approaches aimed at combining different types of experiments provide an excellent way to complement the scarce information content of cryo-EM data in this resolution regime and thus to enable determining more accurate and precise single-structure models. A recent example of combination of cryo-EM with NMR data is the determination of the structure of the 468 kDa aminopeptidase TET2 to a precision below 1 Å starting from a 4.1 Å resolution map.²⁴ At this resolution, it was difficult to trace the backbone and assign the sequence using the cryo-EM data alone, but by combining it with secondary structures modelled using NMR data it was possible to determine precise models using both the original data and data artificially truncated to 8 Å.

Rather intriguing is also the possibility of using integrative ensemble-modelling approaches to combine cryo-EM data with other experiments to obtain more accurate protein conformational ensembles. For example, one could envision incorporating NMR ensemble-averaged data to improve the characterization of highly-flexible parts of biological systems that are often averaged out in the cryo-EM classification and reconstruction processes.

The major challenge in single-structure and ensemble integrative approaches is how to balance the information provided by different types of experimental data. In these regards, Bayesian statistics^{25, 26} is an effective framework that can be used to combine all sources of information available on the system, being experimental data or physico-chemical knowledge, by weighting them based on accuracy and information content.

In conclusion, while we are in the middle of an explosion of the number of atomistic cryo-EM data available, computational modelling and molecular simulations can still play an important role in the future. These methods will certainly provide essential contributions in many areas of structural biology, from improving the description of protein conformational ensembles, to elucidating the effect of freezing on the behavior of biological systems, accurately characterizing complex physico-chemical environments, and integrating cryo-EM with other types of experimental data.