Experimental Data Driven AI Framework for Flexible Protein Conformational Reconstruction.

Area of Science:

• Structural biology
• Biophysics
• Artificial intelligence

Background:

• Deep learning models accurately predict static protein folds from amino acid sequences.
• Proteins function as dynamic ensembles, and current models struggle with conformational states influenced by cellular environments or ligand binding.
• Generative models can explore conformational landscapes but may produce experimentally invalid states.

Purpose of the Study:

• To develop an AI framework that incorporates experimental data to predict accurate protein conformational states.
• To address the limitations of sequence-only models in capturing dynamic protein behavior and ligand-induced changes.
• To bridge the gap between AI-driven structural prediction and experimental biophysical measurements.

Main Methods:

• Developed AlphaSAXS, an end-to-end framework integrating Small Angle X-ray Scattering (SAXS) data into the AlphaFold architecture.
• Utilized real-space pair distance distributions (P(r)) from SAXS data to constrain AI inference.
• Implemented a hybrid inference protocol combining deep learning with biophysical hydration modeling.

Main Results:

• AlphaSAXS successfully steers AI predictions toward experimentally observed protein structures.
• The framework resolves failure modes of sequence-only models, distinguishing between protein states with identical sequences but different scattering profiles (Apo-Holo transitions).
• Reconstructed solution state protein ensembles compatible with experimental SAXS data.

Conclusions:

• AlphaSAXS establishes a paradigm for experimentally guided AI in structural biology.
• The integration of SAXS data enhances the accuracy and reliability of AI-based protein structure prediction for dynamic systems.
• This approach enables the reconstruction of biologically relevant protein ensembles in solution, aligning computational predictions with experimental reality.