BPS2026 – Integrating experimental data with AI to generate physiologically relevant protein ensembles

Deep learning models like AlphaFold have transformed static protein structure prediction, yet a fundamental challenge remains: capturing the dynamic ensembles of conformations that govern biological function. Many proteins contain flexible regions that are crucial for their roles in signaling and disease but are often averaged out in computational predictions. To address this gap, we developed AlphaSAXS, a deep learning framework that integrates experimental small-angle X-ray scattering (SAXS) data to guide an AI model in generating physiologically relevant protein ensembles. AlphaSAXS innovatively incorporates a new cross-attention model into the OpenFold architecture to enable the direct input of SAXS pairwise-distance data, P(r), which inherently captures distance information of protein conformations under different physiological environments. By training the model on NMR data with over 10,000 proteins and 200,000 conformations that exemplify conformational diversity, our model learns to produce structures consistent with the input SAXS P(r) data despite the same sequence. When benchmarked against challenging targets from the test data set, AlphaSAXS consistently generates a single, accurate conformation that is in better agreement with the experimental data than AlphaFold and existing methods. Furthermore, we extended this framework to an ensemble prediction method by using inference-time optimization to enhance conformational sampling. This method successfully generates an ensemble with an internal RMSD larger than 1Å without compromising the overall prediction accuracy. Our work establishes a robust framework for integrating biophysical data into AI predictions, representing a vital step toward solving the ensemble problem and achieving a truly dynamic understanding of protein structure and function.