Protein folding stability estimation with explicit consideration of unfolded states

Area of Science:

• Computational biology
• Protein structure and stability analysis
• Artificial intelligence in bioinformatics

Background:

• Protein folding stability (ΔG) is essential for protein function, yet current computational prediction methods struggle with quantitative accuracy.
• Existing AI-driven protein structure prediction tools have limitations in reproducing experimental folding stability values.

Purpose of the Study:

• To develop an advanced deep neural network, IFUM, for accurate prediction of protein folding stability (ΔG) and the equilibrium ensemble of folded/unfolded states.
• To improve upon existing computational methods for predicting protein folding stability and mutational effects.

Main Methods:

• IFUM employs a deep neural network architecture for joint estimation of ΔG and residue-pair distance probability distributions.
• The model was trained on a diverse dataset encompassing small proteins, disordered proteins, and natural proteins to ensure robustness.

Main Results:

• Joint learning of ΔG and ensemble states significantly enhances prediction accuracy compared to predicting ΔG alone.
• IFUM demonstrates robustness across various protein types and accurately predicts complex mutational effects, including insertions and deletions.
• The model shows a strong correlation with experimental melting temperatures in protein engineering applications.

Conclusions:

• IFUM represents a significant advancement in computational prediction of protein folding stability.
• The tool effectively guides protein design challenges and outperforms existing AlphaFold-based metrics in de novo design selection.
• IFUM's ability to predict ensemble states alongside stability offers deeper insights into protein behavior.