Rapid generation of antibody conformational ensembles using machine learning
The paper introduces Cadense, a machine learning framework that rapidly generates diverse antibody conformational ensembles from sequence to overcome the limitations of static structure prediction, thereby enabling more accurate modeling of dynamic antibody-antigen interactions and accelerating drug design.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
Proteins are the workhorses of life, tiny machines that fold into specific shapes to perform tasks like fighting infections or carrying oxygen. For decades, scientists have relied on a method called X-ray crystallography to see these shapes, which involves freezing the protein in a crystal and taking a picture. The problem is that this picture is static; it captures the protein in a single, frozen moment, like a photograph of a runner mid-stride. In reality, proteins are dynamic, constantly wiggling and shifting between different forms. This movement is especially critical for antibodies, the immune system's defenders. Antibodies have flexible loops at their tips that must bend and twist to grab onto specific invaders. If a scientist only sees the antibody in one rigid pose, they might miss the very shape that allows it to bind to a disease-causing virus, or they might incorrectly assume a harmless protein is dangerous because it looks risky in just one frozen snapshot.
For a long time, the only way to see these movements was to run complex computer simulations that mimic the physics of water and atoms, a process that is so computationally heavy it can take days to simulate just a fraction of a second of motion. This makes it impossible to test thousands of potential drug candidates quickly. Researchers at Genentech have now introduced a new tool called Cadense that changes the game. Instead of simulating every single movement from scratch, Cadense uses a type of artificial intelligence trained on millions of previous simulations to instantly generate a whole family of possible shapes for an antibody. It does not just guess one structure; it produces a diverse collection, or ensemble, of structures that show how the antibody might wiggle and bend in real life. This allows scientists to see the full range of motion in seconds rather than days, offering a much clearer picture of how these molecules actually behave.
The researchers tested this new system by asking it to predict the shapes of antibodies where they already knew the answers from experiments. In one set of tests, they looked at antibodies that had been crystallized in two different states: one without a target (the "apo" state) and one bound to a target (the "holo" state). Most existing computer programs, which are trained mostly on the bound versions, tend to predict only the target-bound shape, missing the flexible, unbound form entirely. Cadense, however, successfully generated samples that covered the entire range of motion, producing structures that looked just like the unbound form, the bound form, and everything in between. In a particularly rigorous test, the team used the tool to predict the shape of an antibody they had just crystallized in their own lab, a structure the computer had never seen before. The tool's best guess was accurate to within the width of a single atom, proving it could find the correct shape even when it was completely new to the system.
Beyond just getting the shape right, the team wanted to know if these generated ensembles could help predict how safe and effective an antibody might be as a drug. Antibodies need to have the right balance of electrical charge and stickiness on their surface to work well without clumping together or causing side effects. When scientists look at a single, frozen structure, they often flag a protein as risky because it looks too sticky or charged in that one pose. By using the Cadense ensembles, which show the protein moving through many different shapes, the researchers found that the tool could correct these false alarms. In their tests, the ensemble approach reduced the number of false "high risk" warnings by 3.5 times compared to looking at a single static image. This means that drug developers could stop wasting time and resources on promising candidates that were wrongly rejected because a single snapshot looked bad, while still catching the truly dangerous ones.
The final test involved seeing if these flexible ensembles could help scientists figure out how an antibody grabs onto a virus or bacteria. This process, known as docking, is notoriously difficult because the antibody's flexible loops must move to fit perfectly around the target. The researchers used the Cadense ensembles as input for a physics-based docking program and compared the results to using single, static structures. The ensembles performed significantly better, finding the correct binding pose much more often. This was especially true for antibodies with unusual, novel loops that the computer had never seen before; while static models struggled with these new shapes, the Cadense ensembles maintained high accuracy. This suggests that by providing a cloud of possible shapes rather than a single rigid guess, the tool gives the docking software the flexibility it needs to solve the puzzle of how antibodies bind to their targets.
The work establishes Cadense as a practical bridge between the slow, detailed world of physics simulations and the fast, high-speed needs of drug discovery. While the tool is currently limited to the variable parts of the antibody and cannot yet predict the long-term kinetic behavior of the molecule, it offers a powerful new way to integrate the reality of molecular motion into the design process. By replacing days of calculation with seconds of generation, it allows researchers to consider the dynamic nature of antibodies at a scale that was previously impossible. This shift from static snapshots to dynamic ensembles promises to make the search for new therapies faster and more accurate, ensuring that the flexible, living nature of these molecules is finally accounted for in the design of the next generation of medicines.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.