Exploring the conformational landscape of adenylate kinase and beyond with protein folding models.

Generative Protein Models Outperform Traditional Sampling in Mapping Conformational Landscapes

Proteins are not static sculptures. They are dynamic machines that must bend, twist, and shift to perform biological work. To understand how a protein functions—how it signals a cell or transports a molecule—one must understand its "conformational landscape." This is the collection of different shapes a protein adopts during its lifecycle. Scientists have long struggled to predict these alternative shapes. They often rely on expensive, slow computer simulations to see how a protein moves from an "off" state to an "on" state.

Recent breakthroughs in AI-driven protein folding have changed the game. Most models are still optimized to predict a single, most-likely structure. This leaves a massive gap in our understanding of the transitions between states. A new study presents a systematic benchmark of nine different methods to address this. It reveals that new "generative" models, specifically diffusion-based ones like Chai-1, are significantly better at uncovering the hidden middle steps of protein motion than older methods.

The Problem

The current gold standard for studying protein motion is Molecular Dynamics (MD). This is a physics-based simulation that calculates the movement of every atom over time. While accurate, MD is computationally punishing. It often requires weeks or months of supercomputer time to observe a single functional transition. To bypass this, researchers have attempted to "trick" existing protein folding models, like AlphaFold2, into revealing alternative shapes.

These traditional approaches generally fall into two categories. First, they perturb the Multiple Sequence Alignment (MSA)—the database of evolutionary relatives used by the models. This involves masking or subsampling the data to disrupt co-evolutionary signals. Second, they use "dropout" (randomly disabling parts of the neural network during inference) to induce structural variety.

However, as the authors note, these methods often fail to capture the full breadth of motion. Some methods, like MSA clustering, might find many different shapes but produce "noisy" or structurally unrealistic ones. Others, like alanine mutation (changing specific amino acids to simulate mutations), often produce structures that bear no resemblance to the actual functional states. Consequently, we lack a reliable way to bridge the gap between a protein's static "start" and "end" positions.

How It Works

The researchers conducted a head-to-head comparison of nine sampling methods across 20 different proteins. These targets ranged from enzymes to membrane transporters. To move beyond simple "correct or incorrect" metrics, they introduced an extension of the Aligned Error (AE) matrix. Traditionally, an AE matrix measures the structural difference between two specific protein states. The authors extended this to compute errors across entire ensembles (collections of predicted structures). This allows them to visualize how whole domains move relative to one another.

The breakthrough lies in the shift toward diffusion-based generative modeling. Unlike previous models that attempt to refine a single structure, diffusion models like Chai-1 and Boltz-1 are designed to sample from a distribution of possibilities. Instead of trying to find the "one true shape," they learn the underlying mathematical manifold (the multidimensional space of all physically plausible shapes) of the protein.

The study focuses heavily on Adenylate Kinase (AdK), a classic model of protein motion. AdK undergoes a dramatic transformation where two domains—the LID domain and the NMP domain—close around a substrate. The researchers used Principal Component Analysis (PCA)—a statistical technique that reduces complex, high-dimensional data into a few essential axes of motion—to organize the thousands of shapes predicted by Chai-1. This allowed them to see if the AI was actually "seeing" the same landscape that physics-based MD simulations reveal.

Numbers

The results favor the generative approach with significant margins. In the benchmark of 20 proteins, Chai-1 emerged as the top performer. It achieved a near 90% average recall and precision across all targets . When focusing specifically on the AdK transition, the authors report that Chai-1 achieved a 95% recall rate.

Perhaps more importantly, the study measured the "fill-ratio." This is the ability of a model to cover the geometric space between the active and inactive states. The authors find that Chai-1 achieved a fill-ratio of 0.60 .

This effectively doubles the coverage provided by traditional MSA-perturbation strategies. This means Chai-1 successfully identifies more of the "steps" in a protein's movement than older methods.

The researchers also found that the protein itself matters more than the algorithm. A two-way ANOVA (a statistical test used to determine how much of a result is due to different factors) revealed that the intrinsic properties of the protein explained 47.8% of the performance variance. In contrast, the choice of sampling method only explained 13.2%. For researchers, this implies that choosing a better model is not a universal fix. If a protein is inherently complex or lacks sufficient evolutionary data, even a superior model like Chai-1 may struggle to achieve high accuracy.

What's Missing

Despite the success of Chai-1, the study has clear boundaries. First, the deep validation of intermediate states was limited strictly to Adenylate Kinase. Because the researchers relied on existing MD trajectories (the "ground truth") to prove the AI was correct, they could only perform this high-fidelity check on proteins where such data already existed. We do not yet know if Chai-1's ability to capture "middle steps" translates equally well to the rest of the proteome.

Second, the benchmark was restricted to monomeric proteins (single polypeptide chains). In a real cell, proteins rarely act alone. They interact with ions, ligands, and other proteins. These interactions often drive the very conformational changes the researchers are trying to model. A model that ignores these external "triggers" may provide an incomplete picture.

Finally, it is vital to remember that these models are geometric, not physical. The PCA-derived pathways represent a map of possible shapes. However, they do not provide the temporal ordering or the kinetic rates (how fast the movement happens) that a true MD simulation provides. The AI tells you where the protein can go, but not necessarily how long it stays there.

Should You Prototype This

Yes, but with caution. If your goal is to rapidly explore structural possibilities, Chai-1 is a formidable tool. It can help identify "druggable" intermediate states that traditional crystallography might miss. It offers a massive increase in efficiency compared to the heavy lifting required by MD simulations.

However, do not treat these outputs as definitive physical trajectories. Use them to generate high-quality hypotheses. These hypotheses can then be validated with targeted MD or experimental NMR. For practitioners looking to implement this, the code for the benchmarking framework is available at https://github.com/instadeepai/FoldConfBench.

Figures from the paper