JanusDDG: a physics-informed neural network for sequence-based protein stability via two-fronts attention.

Area of Science:

• Computational Biology
• Protein Engineering
• Biophysics

Background:

• Predicting protein stability changes from mutations is vital for protein design and understanding disease mechanisms.
• Protein language models (PLMs) have improved computational predictions, but struggle to adhere to thermodynamic laws.
• Sequence-based models face challenges in balancing accuracy with fundamental thermodynamic principles.

Purpose of the Study:

• To develop a novel physics-informed neural network for predicting protein stability changes.
• To accurately predict stability changes for both single and multiple residue mutations using sequence data.
• To ensure predictions satisfy thermodynamic laws while maintaining high accuracy.

Main Methods:

• Developed JanusDDG, a physics-informed neural network integrating PLM embeddings and a cross-attention transformer.
• Employed a physics-informed paradigm to constrain the model to thermodynamic principles (antisymmetry, transitivity).
• Utilized a cross-interleaved attention mechanism to analyze wild-type and mutant sequence embeddings.

Main Results:

• JanusDDG achieves state-of-the-art performance in predicting protein stability changes from sequence.
• The model demonstrates high accuracy for both single and multiple residue mutations.
• JanusDDG's performance matches or surpasses structure-based methods using only sequence information.

Conclusions:

• JanusDDG offers a powerful, sequence-based approach for predicting mutational effects on protein stability.
• The physics-informed design ensures thermodynamic consistency in predictions.
• This method advances rational protein design and the assessment of disease-related mutations.