New AI Model Goes Beyond AlphaFold to Forecast Protein Shape Shifts

Scientists at Japan's Institute for Molecular Science (IMS) and the Graduate University for Advanced Studies have unveiled an artificial‑intelligence system that can predict how proteins rearrange themselves, addressing a gap left by the widely celebrated AlphaFold3 platform.

Proteins rarely remain in a single static form; they often undergo conformational changes that are essential for processes such as enzyme catalysis, signal transduction, and immune recognition. Capturing these dynamic shifts is crucial for deciphering disease mechanisms and designing effective therapeutics, yet existing computational tools have struggled to model them reliably.

Since its debut, AlphaFold has transformed structural biology by delivering highly accurate predictions of a protein's most probable three‑dimensional structure. However, the model is fundamentally oriented toward a single, energetically favored conformation and does not inherently describe the range of motions a protein may explore during its functional cycle.

The new approach integrates deep‑learning techniques with physics‑based simulation data, training the algorithm on experimentally resolved ensembles that illustrate multiple functional states. By learning patterns that link sequence information to structural flexibility, the system can generate plausible alternative shapes and map potential transition pathways.

Benchmark tests on a diverse set of proteins—including a kinase, a G‑protein‑coupled receptor, and a molecular chaperone—showed that the method reliably reproduced known conformations that AlphaFold3 missed or mischaracterized. In several cases, the predicted intermediate forms matched structures later confirmed by cryo‑electron microscopy, underscoring the model's practical relevance.

Researchers say the capability to anticipate protein dynamics could accelerate drug discovery by highlighting transient binding pockets and informing the design of molecules that stabilize or inhibit specific states. The team plans to expand the training dataset, refine the algorithm's speed, and make the software publicly available, aiming to complement existing structural prediction pipelines and bridge the gap between static models and real‑world biology.