Lessons from Deep Learning Structural Prediction of Multistate Multidomain Proteins-The Case Study of Coiled-Coil...

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Artificial Intelligence in Protein Science

Background:

• Multidomain and multistate proteins present significant challenges for accurate 3D structure prediction.
• Deep learning (AI) models are emerging as powerful tools for protein structure prediction.

Purpose of the Study:

• To evaluate the predictive capabilities of new-generation deep learning models for complex multidomain, multistate proteins.
• To identify limitations and biases in AI-driven protein structure modeling.

Main Methods:

• Case study using coiled-coil Nucleotide-binding Oligomerization Domain-like (NOD-like) receptors from *A. thaliana*.
• Testing deep learning predictors on proteins with well-established and morphing regions.
• Analysis of predictions for multivalued 1D to 3D mappings and secondary structure elements.

Main Results:

• AI predictors accurately model 3D structures of modules with stable topologies.
• Lower performance observed in predicting flexible regions like coiled coils.
• Tendency for AI models to predict compact protein configurations, requiring retraining for sparser models.

Conclusions:

• AI predictors are valuable for modeling multidomain multistate proteins when global templates are available, with caveats.
• Challenges in modeling flexible regions and achieving desired compactness need to be addressed.
• Piecewise modeling with experimental constraints may yield more realistic results in the absence of global templates.