AlphaFold3 Learns from Lab Experiments to Map Proteins’ Shape-Shifting Behavior

Proteins are not the rigid sculptures often depicted in textbook diagrams—they flex, twist, and sample a range of shapes to carry out their functions. Yet the most celebrated protein-structure prediction tools, including AlphaFold3, typically collapse this dynamism into a single static snapshot. A study published today in Nature Biotechnology shows how to nudge AlphaFold3 out of that habit by feeding it real experimental data, allowing it to generate small ensembles of structures that reflect the true conformational variability seen in the lab.

The work, led by Alex M. Bronstein at the Institute of Science and Technology Austria and Technion, with first author Advaith Maddipatla, incorporates experimental measurements directly into AlphaFold3’s diffusion-based generative process for protein structures. The result is a method that can, in typical cases, produce a compact set of protein conformations whose average properties agree with nuclear magnetic resonance (NMR) distance restraints, X-ray electron density maps, cryo-electron microscopy (cryo-EM) electrostatic potential (ESP) maps, or combinations of these data.

When the team guided AlphaFold3 with NMR-derived interatomic distances for the model protein ubiquitin, the resulting ensemble reduced the number of distance-constraint violations compared with the conventionally determined NMR ensemble. Adding spin-relaxation order parameters—reporters of fast internal motion—improved agreement with experimental dynamics in the ubiquitin benchmark (r = 0.93), comparable to ensembles generated by much costlier molecular dynamics simulations. Across a benchmark of 91 protein structures, the guided approach improved constraint satisfaction in 70 out of 91 cases (~77%) relative to the deposited Protein Data Bank (PDB) ensembles, whereas unguided AlphaFold3 outperformed the PDB sets in 17% of cases.

In X-ray crystallography, the method recovered alternate conformations that had been overlooked, filled in regions of missing electron density, and rebuilt poorly predicted loops—including one in HSP90α that adopts different shapes in the two chains of a crystal dimer. For cryo-EM, it improved predictions for large complexes, such as the asymmetric insulin receptor and an amyloid-β fibril. The framework also merges modalities: guiding simultaneously with cryo-EM maps and solid-state NMR dihedral angles improved both density fit and local backbone accuracy of a human RIPK3 amyloid fibril.

These ensembles are not thermodynamic populations—they are best understood as plausible structural explanations that satisfy the provided experimental evidence and the AlphaFold3 prior. The authors stress that different mixtures of conformations could account for the same measurements, and that force-field–based energy reweighting nudges the output closer to physically meaningful distributions, a direction flagged for future work. The present implementation also has limited handling of ligands, metals, post-translational modifications, structural waters, and symmetry-related effects.

Nevertheless, because the computation runs on GPU timescales (minutes per target in typical cases), the approach could slot into everyday crystallographic model building, NMR spectral interpretation, and cryo-EM refinement as a rapid, ensemble-aware assistant. It paves the way for large-scale reanalysis of legacy datasets and may help reveal fleeting protein states relevant to drug binding.

The study is published in Nature Biotechnology (2026). DOI: 10.1038/s41587-026-03166-5.