RINAMI: Residue-attributed interpretable neural network for predicting absolute folding free energy by merging...

Area of Science:

• Computational Biology
• Protein Engineering
• Machine Learning

Background:

• De novo protein design generates novel proteins, but assessing folding stability remains a challenge.
• Reliable computational methods are needed to quantify folding free energy (ΔG) and distinguish successful designs.
• Current methods struggle to accurately predict if a designed protein's target structure is thermodynamically favored.

Purpose of the Study:

• To develop an accurate and interpretable machine learning model for predicting absolute protein folding free energy (ΔG).
• To provide a computational tool for evaluating and prioritizing de novo protein designs.
• To address the fundamental challenge of assessing folding stability in novel protein sequences.

Main Methods:

• Developed RINAMI (Residue-attributed Interpretable Neural network for predicting Absolute folding free energy by Merging structure and sequence Information).
• Integrated structure-based and sequence-based representations using ProteinMPNN and Evolutionary Scale Modeling 2 (ESM2).
• Employed a multi-head cross-attention mechanism to contextualize sequence signals within the structural environment.

Main Results:

• RINAMI accurately predicts absolute folding free energy (ΔG) for both natural and designed proteins.
• The model outperforms existing approaches in correlation with experimental measurements and prediction errors.
• An ablation study confirmed the importance of integrating sequence and structure information for accuracy.
• RINAMI demonstrated interpretability by identifying key physicochemical effects influencing protein stability.

Conclusions:

• RINAMI provides an accurate and interpretable framework for predicting protein folding free energy (ΔG).
• The model serves as a practical computational tool for assessing and prioritizing protein designs.
• This advancement facilitates the reliable generation of novel proteins with desired stability.