Neural network-derived Potts models for structure-based protein design using backbone atomic coordinates and tertiary...

Area of Science:

• Computational Biology
• Protein Engineering
• Synthetic Biology

Background:

• Designing functional proteins is a key challenge in synthetic biology.
• Existing computational methods include energy-based frameworks using tertiary motifs (TERMs) and neural networks using backbone coordinates.

Purpose of the Study:

• To improve neural network-based protein design models by integrating TERM-derived features.
• To develop novel architectures for generating protein sequence designs.

Main Methods:

• Developed two neural network architectures, TERMinator and COORDinator, to generate Potts models.
• TERMinator integrates both TERM-based and coordinate-based features, while COORDinator uses only coordinate-based features.
• Evaluated models on native sequence recovery and predicted protein folding using AlphaFold.

Main Results:

• TERMs significantly improve the native sequence recovery performance of neural models.
• Sequences designed by TERMinator are predicted to fold correctly by AlphaFold.
• Both models learn underlying energetics and can be fine-tuned with experimental data.

Conclusions:

• Combining TERM-based and coordinate-based features enhances protein design.
• Structure-based neural models generating Potts energy tables offer versatile applications in protein science.