Guiding AlphaFold predictions with experimental knowledge to inform dynamics and interactions with VAIRO

Area of Science:

• Structural Biology
• Computational Biology
• Biophysics

Background:

• Protein structure prediction methods like AlphaFold achieve high accuracy by integrating diverse data sources.
• However, protein function relies on dynamic conformational changes, which current models may not fully capture due to inherent data conflicts.

Purpose of the Study:

• To develop a new computational method that incorporates prior knowledge to predict specific functional states of proteins.
• To overcome limitations in current structural prediction models by resolving conflicting evolutionary and structural information.

Main Methods:

• Developed VAIRO, a program that integrates three information channels: learned parameters, templates, and aligned sequences.
• Implemented a strategy to select state-specific information, thereby delimiting the functional context of predictions.
• Applied VAIRO to study molecular assemblies in bacterial surface layers and membrane protein complexes.

Main Results:

• VAIRO successfully rescues asymmetric and weaker interactions, providing a more complete view of molecular assemblies.
• The method reveals dynamic states in complex protein structures that were previously inaccessible to prediction.
• Demonstrated application in bacterial surface layer architecture and pneumococcal multimeric membrane protein complexes.

Conclusions:

• VAIRO offers a powerful approach to enhance protein structure prediction by incorporating prior knowledge for specific functional states.
• The method advances our understanding of molecular assemblies and dynamic protein behaviors.
• VAIRO is available as a Python package (PyPI) and open-source code (GitHub).