A DNN Biophysics Model with Topological and Electrostatic Features.

Area of Science:

• Computational Biophysics
• Machine Learning in Structural Biology
• Protein Property Prediction

Background:

• Predicting protein properties is crucial for understanding biological function.
• Existing methods often struggle with scalability and capturing complex interactions.
• Deep neural networks offer a promising avenue for advanced biophysical modeling.

Purpose of the Study:

• To develop a novel deep neural network (DNN)-based model for predicting protein properties.
• To introduce multiscale and uniform topological and electrostatic features for enhanced prediction accuracy.
• To demonstrate the model's efficiency and fidelity in representing protein structure and force fields.

Main Methods:

• Utilized element-specific persistent homology (ESPH) for topological feature generation.
• Employed a novel Cartesian treecode for electrostatic feature generation, incorporating underlying electrostatic interactions.
• Trained DNN models on large protein structure databases (over 17,000 for Coulomb energy, over 4,000 for solvation energy).

Main Results:

• Achieved high accuracy in predicting Coulomb energy (MSE ≈ 0.024, MAPE ≈ 0.073, R² ≈ 0.976).
• Demonstrated strong performance in predicting solvation energy (MSE ≈ 0.064, MAPE ≈ 0.081, R² ≈ 0.926).
• Validated the efficiency and fidelity of the generated features in representing protein characteristics.

Conclusions:

• The proposed DNN model with novel features effectively predicts key protein properties.
• The feature generation algorithms show potential as general tools for machine learning-based protein analysis.
• This work advances the integration of computational biophysics and machine learning for biological insights.