ASU Method Cracks Protein Motion, Speeds Drug Design

Proteins are not the rigid blocks that introductory biology textbooks make them look like. They are flexible molecular machines whose function depends on motion. An enzyme is useful because it can change shape to grip a substrate. A receptor on the surface of a cell is useful because it can shift its conformation when something binds to it. A transport protein is useful because it can rock back and forth across a cell membrane. Stripping away the motion is like trying to understand a violin by looking only at a photograph.

Drug designers know this in theory. In practice, they have struggled for decades to predict the full range of shapes a protein can adopt, because the slow motions that matter most happen on timescales that are punishing for computer simulation. Traditional molecular dynamics tools were built for the fast, tiny vibrations of a guitar string. The slow, sweeping motions of a protein look more like the gentle sway of a tall building in the wind, and they are messy, uneven, and irregular.

That mismatch has consequences. Many of the most important drug targets, from enzymes implicated in cancer to receptors involved in inflammation, work through "allosteric" effects where binding at one site triggers a change far away on the protein. Catching that long-distance communication on camera has been a notoriously hard problem.

Heyden's group found a way to extract the slow rhythms from short simulation snapshots that last only billionths of a second. The method analyses the natural fluctuations that proteins exhibit at room temperature, the tiny background motions caused by molecular collisions, and identifies which low-frequency vibrations are guiding larger conformational changes. The team then uses those vibrations as scaffolding to nudge the protein along its natural pathways in subsequent simulations.

Heyden offered an analogy to News-Medical that captures the trick neatly.

This can be compared to an unlocked door. We can feel quickly if we need to push or pull, while trying to yank the door up and off of its hinges is always hard. The key is that we don't need to execute the full motion to realize these differences. On a molecular scale, it is even enough to observe tiny fluctuations that are always present at room temperature.Matthias Heyden, Associate Professor, Arizona State University

Run the technique on the same protein again and again, the team found, and it tells the same story each time. The reproducibility matters. Drug development cannot tolerate methods that produce a different answer on every run.

The other significant claim in the paper is speed. Using graphics-processing units on ASU's "Sol" supercomputer, the group can now watch proteins undergo meaningful shape changes in less than a day. The same calculation, attempted with conventional methods, can take weeks or months and frequently fails to converge. The first author of the published paper is M. A. Sauer, and the work was funded by the National Science Foundation under grant CHE-2154834 and the National Institutes of Health under grant R01GM148622.

The acceleration matters because it changes what is possible to screen. A drug discovery campaign might want to test how a candidate molecule perturbs the dynamics of a target enzyme. If the simulation takes a month per candidate, only a handful of molecules ever get tested. If it takes a day, an entire library can be run.

The speed gain also opens up the prospect of building large datasets of protein dynamics, which is exactly what next-generation machine learning models will need. Heyden told ASU that the approach will let researchers extend the "sequence-to-structure" relationship captured by tools like AlphaFold into a "sequence-to-structure-to-dynamics" relationship. AlphaFold can already predict the shape a protein folds into. What it cannot do, on its own, is predict how that shape will move.

The most immediate application is allosteric drug discovery, the search for molecules that can switch a protein on or off by binding somewhere other than the active site. Allosteric drugs have long appealed to medicinal chemists because they tend to produce fewer off-target side effects than traditional inhibitors. They have also proven extraordinarily difficult to design, because identifying the right binding site requires understanding the protein's internal communication network.

With faster, more revealing simulations, that communication network becomes visible. A researcher can now ask which surface pockets are dynamically connected to the active site, which residues move when something binds at a distance, and which conformations are accessible only under specific conditions. That kind of granular dynamic picture is what allosteric drug design has been waiting for.

The method also has implications outside therapeutics. Synthetic biologists who design new proteins from scratch have struggled to build molecules that move as elegantly as natural enzymes. Most designed proteins today are stable but inert. A reliable way to predict and tune motion could let designers build sensors that switch on when a target binds, catalysts that mimic natural enzymes, and biological circuits that respond dynamically to their environment.

It is worth being honest about what the method does not do. The simulations still assume a working structural model of the protein in question, which means proteins that resist crystallization or that AlphaFold predicts poorly remain hard to study. The technique also does not eliminate the need for experimental validation. A predicted conformation is not a measured one, and the slow motions that matter most are also the ones hardest to confirm in the laboratory.

And for all the speed gains, the underlying physics has not changed. Conformational sampling remains a hard problem in computational chemistry, and edge cases where the slow vibrations of a protein are tightly coupled to its environment, rather than driven from within, will still trip the method up.

That said, the broader trajectory of computational biology is now clearly moving from static snapshots toward dynamic landscapes. AlphaFold gave the field its first reliable map of where proteins go. Methods like Heyden's are building the rules for how they get there. Tools like the ones used by researchers tracking new biotech drug pipelines and the early-stage discovery work documented in recent microbiome studies stand to benefit immediately.

Two things to track. The first is whether Heyden's group, or others, can scale the method to large membrane proteins, which are some of the most pharmacologically important targets and also some of the hardest to simulate. The second is whether the technique shows up in drug discovery pipelines outside academic labs. Pharmaceutical companies guard their computational workflows closely, but methods that genuinely accelerate sampling tend to migrate from preprint to production within a year or two.

The longer-term question is what happens when dynamic protein data becomes plentiful enough to train machine learning models on. AlphaFold's leap was made possible by the Protein Data Bank, decades of painstaking experimental structures finally aggregated and machine-readable. A comparable dataset for protein dynamics does not exist yet. Methods like this one are how it gets built.