The CAHRA Challenge: A Community-Wide Assessment of Cryo-EM Heterogeneous Reconstruction Algorithms

To see the inner workings of life, scientists often turn to a technique called cryo-electron microscopy. Imagine taking a photograph of a single, tiny machine made of proteins, but the machine is frozen in a thin layer of ice so fast that it cannot move or break. Because these machines are so small, a single photograph is just a blurry, two-dimensional shadow. To understand the machine, researchers must take thousands of these blurry pictures from different angles and use a computer to stitch them together into a clear, three-dimensional model. This process has revolutionized biology, allowing us to see the atomic structure of viruses, enzymes, and other essential molecules. However, real biological machines are rarely static; they wiggle, bend, and change shape as they work. When a dataset contains a mix of these different shapes, or even different types of molecules entirely, the computer struggles to decide which picture belongs to which version of the machine. This makes it incredibly difficult to reconstruct the true, moving story of the molecule.

In 2026, a team of researchers from Princeton University, the Flatiron Institute, and other institutions launched a global test to see how well current computer programs can solve this specific problem. They called it the CAHRA Challenge, a community-wide assessment designed to measure the ability of different algorithms to untangle these complex mixtures. The organizers did not simply ask scientists to guess; they built three distinct, highly controlled test cases where the correct answers were known only to the creators. These tests covered the three main ways molecules get complicated: when a sample contains a mix of different species, when a single molecule shifts through a continuous range of shapes, and when the angle from which a molecule is viewed hides its true form. By comparing how well various teams performed against these known truths, the challenge provided a clear map of where the field stands and where it needs to improve.

The first test focused on a crowded room of different molecules. The researchers took four distinct protein complexes—alcohol dehydrogenase, aldolase, glutamate dehydrogenase, and pyruvate kinase—that are similar in size and shape. They also added two other proteins, gamma-globulin and beta-amylase, which acted as distractions because they were damaged or formed flat sheets that could not be reconstructed into 3D models. All of these were mixed together into a single dataset containing over 1.6 million particle images. The challenge for the participants was to look at this jumbled pile of images and figure out which pictures belonged to which of the four main targets, and then build a clear 3D model for each one without being told what the targets were. To make the task even harder, the researchers simulated radiation damage by splitting the images into high-quality and low-quality groups, forcing the algorithms to distinguish between a molecule that is slightly broken and one that is just a different type of molecule.

The second test moved away from mixing different species and instead focused on a single molecule that changes its shape. The researchers used a computer simulation of a human enzyme called presequence protease, which opens and closes like a gate. They captured the molecule at 100 different stages of this motion, creating a smooth transition from an open state to a closed state. From these 100 snapshots, they generated over 1.7 million simulated images. The goal for the participants was not just to see that the molecule moved, but to produce a set of atomic models—detailed blueprints of the atoms—that accurately described this entire range of motion. Unlike the first test, where the goal was to separate distinct groups, this test required the algorithms to understand a continuous flow of change and translate that into precise structural models.

The third test addressed a tricky problem known as pose entanglement. In real experiments, molecules often prefer to land on the microscope slide in certain orientations, leaving other angles empty. If a molecule changes shape along the same axis that is missing from the data, the computer might mistake the missing angle for a change in shape, or vice versa. To test this, the researchers simulated images of a ring-shaped channel called CALHM2, which has an open and a closed state. They ensured that the images were heavily biased toward a side view, which makes it very hard to see the difference between the open and closed states. The dataset contained 30,000 particles with a known ratio of open to closed states, but the participants were not told the ratio or the correct angles. They had to determine if their methods could accurately count how many molecules were open versus closed, even when the viewing angles made the difference nearly invisible.

The results of this challenge, which ran for seven months with submissions from researchers around the world, are expected to be reviewed in a workshop in Oxfordshire later in 2026. While the paper does not yet list the winners, the design of the challenge itself offers a clear path forward. It moves the field beyond simple comparisons of resolution and forces developers to confront the messy reality of experimental data. The organizers hope that by providing these standardized, ground-truth datasets, the community can build workflows that are more efficient and require less manual tuning. This is crucial for the future of the field, as scientists move toward studying more complex samples, such as molecules inside living cells or those changing shape in real time. The CAHRA Challenge serves as a rigorous benchmark, ensuring that the tools used to visualize the machinery of life are as reliable as the science they aim to reveal.