cryoWEight: Conformational ensembles from cryo-EM particle images with weighted ensemble simulations

Biology is not a collection of static statues, but a world of constant, fluid motion. Proteins, the molecular machines that power life, are never truly still; they twist, bend, and shift between different shapes to perform their tasks. Understanding how they move is just as important as knowing what they look like, because their function depends entirely on these changes. For decades, scientists have used powerful microscopes to take high-resolution snapshots of these molecules, freezing them in time to see their structures. However, a single photograph cannot tell the whole story of a journey. It shows where a protein is, but not how it got there, or how often it visits different spots. To fill in the gaps between these frozen moments, researchers use computer simulations to watch proteins move. Yet, these simulations often struggle to capture the rare, large-scale movements that happen over long periods, leaving scientists with incomplete pictures of how these molecules actually work.

A new approach called cryoWEight offers a way to bridge this gap by combining the best of both worlds: the detailed snapshots from electron microscopy and the dynamic motion of computer simulations. The method was developed by researchers at the Flatiron Institute and the University of California, Irvine, to solve a specific problem. Traditional simulations can get stuck in one shape for too long, missing the rare transitions that are crucial for function. Conversely, simply trying to force a simulation to match experimental images often distorts the natural physics of the molecule. The team created a system that lets the experimental data gently guide the simulation without breaking the rules of physics. They tested this system on three different proteins, ranging from a tiny ten-residue chain to a larger enzyme, and found that it could successfully steer the computer models toward the shapes observed in the microscope, even when the simulation started from a completely different place.

The core of this method relies on a clever feedback loop. Imagine a simulation running thousands of parallel paths, like a swarm of explorers spreading out across a landscape. Some paths wander into areas that are rarely visited, while others crowd into common spots. The cryoWEight system periodically checks these explorers against the real experimental images. If a path leads to a shape that looks very similar to what the microscope sees, that path is given more importance. If a path leads to a shape that does not match, it is given less importance. This process does not force the explorers to move in a specific direction; instead, it simply adjusts the weight of their stories. The paths that match the data are then used to start the next round of exploration, effectively seeding the simulation with the most promising directions. Over many cycles, the entire group of explorers gradually shifts its focus, moving from the starting point toward the target shapes seen in the experiments.

The researchers validated this approach using three distinct proteins to ensure it worked across different levels of complexity. The first was chignolin, a tiny synthetic protein that folds into a simple hairpin shape. They began with a simulation of the folded state and asked the system to find the unfolded state, which was defined by a set of experimental images. Within just a few cycles, the simulation successfully moved the protein from its compact folded form into the extended, unfolded shape, matching the experimental distribution perfectly. The second test involved a slightly larger protein called NTL9, which folds from an unfolded chain into a stable structure. Here, the process worked in reverse: starting with the unfolded chain, the system guided the simulation toward the folded state, again matching the experimental data with high precision.

The most complex test was performed on adenylate kinase, a large enzyme that acts as a cellular energy manager. This protein has a unique ability to swing between a closed shape, where it holds onto its fuel, and an open shape, where it releases it. This movement involves large parts of the molecule bending and rotating. The researchers started the simulation in the closed state and used images of the open state to guide it. The system successfully drove the protein through the transition, capturing the specific intermediate shapes it takes as it swings open. Crucially, the path the simulation took matched the natural pathway observed in long, unbiased computer runs, proving that the method did not just force the protein into a new shape, but found the correct route to get there.

One of the most significant findings was that this method works even when the starting point and the target are completely different. In many previous attempts, if a simulation started in a region of shape-space that had no overlap with the target, it would fail to find the target. CryoWEight, however, was able to bridge this gap. By using the experimental images to reweight the paths, the system could generate new shapes that were never present in the original starting data. It did this by recognizing that certain shapes were more likely to produce the observed images, and then using those shapes to guide further exploration. This allowed the simulation to cross high energy barriers that would normally be impossible to cross in a standard run.

The study also explored how much information is needed from the experimental images for the method to work. They tested the system with images that had varying levels of clarity, simulating conditions where the signal is very strong and conditions where it is buried in noise. When the images were clear, the system converged quickly to the correct shapes. When the images were very noisy, the system failed to find the target, confirming that the guidance comes from the data itself and not from an internal bias of the simulation. This sensitivity to the quality of the data is a strength, as it ensures that the results are driven by the actual experimental evidence.

Furthermore, the researchers showed that the method is robust to how the simulation space is divided. The computer needs to break the complex landscape of possible shapes into smaller regions to manage the calculations. They tested different ways of drawing these boundaries and found that the final result was the same regardless of how the map was drawn. This suggests that the method is reliable and does not depend on arbitrary choices made by the researchers. It also demonstrated that the system could recover mixtures of different shapes. When the target was a combination of folded and unfolded states, the simulation correctly reproduced the ratio of these states, showing that it can handle complex, multi-state scenarios.

The implications of this work extend beyond just these three proteins. The framework is designed to be flexible, meaning it could be applied to other types of experimental data, such as those from nuclear magnetic resonance or X-ray scattering, not just electron microscopy. The key is that the system uses the likelihood of an image to guide the simulation, a principle that can apply to any observable that defines a target distribution. While the current tests used synthetic images generated from known structures, the authors note that the method is ready to be applied to real experimental data. This would allow scientists to take actual microscope images of a protein in action and use them to build a dynamic model of how that protein moves and functions.

By integrating experimental snapshots with the power of enhanced sampling simulations, cryoWEight offers a new way to see the invisible dynamics of life's machinery. It does not replace the need for high-quality experimental data, nor does it replace the need for accurate physical models. Instead, it acts as a bridge, using the data to correct the models and the models to fill in the gaps in the data. The result is a more complete picture of how proteins work, revealing the pathways they take and the states they visit. As the field moves toward understanding larger and more complex molecular assemblies, tools like this will be essential for turning static images into dynamic stories of biological function.