Researchers at the Institute of Science and Technology Austria (ISTA), in collaboration with international teams from Israel, the United States, and beyond, have developed an experiment-guided variant of the AlphaFold artificial intelligence framework to address its historical limitation in modeling protein dynamics. Published in Nature Biotechnology, the study introduces a methodology that integrates experimental data, such as nuclear magnetic resonance and cryo-electron microscopy, to predict ensembles of protein conformations rather than collapsing heterogeneous structures into a single static shape. Traditional structural biology has long relied on X-ray crystallography, which produces rigid snapshots of molecular architecture and has shaped the training datasets for foundational AI tools like AlphaFold. Consequently, predictive models frequently overlook critical flexible regions, such as loops and dynamic domains, treating them as biologically inert. The newly developed approach explicitly models this structural dynamism. By aligning computational predictions with experimental measurements, the system captures the natural conformational heterogeneity of proteins, even accurately reconstructing interfaces and fibril structures where standard models fail. Led by ISTA professor Alex Bronstein, with contributions from researchers at Tel-Hai University, MIGAL, Princeton University, and the Broad Institute, the team has also introduced a new graphical language to represent structural variability. This visual framework moves beyond static molecular representations, encoding detailed information about molecular fuzziness and motion. The researchers aim to systematically reannotate the Protein Data Bank, extracting and organizing previously discarded or misinterpreted dynamic signals into structured datasets suitable for machine learning refinement. The impact extends beyond academic structural biology. By capturing the full spectrum of protein motion, the model enhances inverse protein design, a critical process for bioengineering and pharmaceutical development where sequences must be engineered to fold into specific, often time-dependent, three-dimensional configurations. The researchers note that many biologically active proteins transition rapidly through functionally essential states, a capability the updated framework is designed to track across millisecond timescales. Following the initial publication, the team has advanced the methodology through additional work accepted for presentation at the 2025 and 2026 International Conference on Machine Learning, focusing on optimizing inference speeds and refining generative diffusion processes. A separate preprint posted to bioRxiv demonstrates the model's ability to extract previously unmodeled conformations of the protein beta2-microglobulin directly from crystallographic data. Researchers emphasize that this evolution represents a foundational shift toward experimentally aware predictive modeling. By treating structural blur and conformational diversity as actionable data rather than noise, the approach establishes a pathway for next-generation AI tools that mirror the physical and biological reality of molecular systems. The team plans to integrate the framework into standard structural prediction pipelines, aiming to transition from proof-of-concept to a routine instrument for computational biology and drug discovery within the coming years.