Protein Design Using Physics Informed Neural Networks.

Area of Science:

• Computational Biology
• Protein Engineering
• Biophysics

Background:

• The inverse protein folding problem, or protein sequence design, aims to create amino acid sequences for specific protein structures and functions.
• Machine learning methods have advanced protein design but struggle with robustness and interoperability under non-ambient conditions (e.g., extreme temperatures, pH, or ionic solvents).

Purpose of the Study:

• To develop a robust protein sequence design framework capable of creating proteins that function under non-ambient conditions.
• To improve upon existing machine learning methods by integrating physical principles for enhanced protein stability and design accuracy.

Main Methods:

• Utilized Physics-Informed Neural Networks (PINNs) integrated with all-atom molecular dynamics (MD) simulations.
• Developed a PINNs MD surrogate model for efficient simulation.
• Employed a relaxation of binary programming to optimize protein energy and structural stability.

Main Results:

• Successfully designed proteins predicted to function under non-ambient conditions.
• The PINNs-based approach demonstrated superior performance compared to traditional energy function-based and standard machine learning methods.
• The framework optimizes both energetic stability and structural integrity for designed protein sequences.

Conclusions:

• The proposed PINNs-based framework offers a powerful and robust solution for protein sequence design, particularly for proteins intended for challenging environments.
• This approach advances the field of protein engineering by enabling the design of resilient and functional proteins.
• Future work can explore broader applications of PINNs in molecular design and simulation.