End-to-End Protein Normal Mode Frequency Predictions Using Language and Graph Models and Application to Sonification

Area of Science:

• Computational biology
• Biophysics
• Structural biology

Background:

• Predicting protein mechanical and dynamical properties is crucial for understanding protein function.
• Advances in protein structure availability necessitate efficient prediction methods.

Purpose of the Study:

• To develop and evaluate end-to-end models for predicting protein dynamical properties from amino acid sequences.
• To compare the performance of different neural network architectures for this task.

Main Methods:

• Utilized Natural Language Processing (NLP) and graph-based neural network models.
• Developed Long Short-Term Memory (LSTM), transformer, graph-based transformer, and equivariant graph neural networks.
• Focused on high-throughput normal mode predictions directly from amino acid sequences.

Main Results:

• All four models demonstrated exceptional performance in predicting protein dynamical properties.
• The graph-based transformer achieved the best results but requires graph structure input.
• End-to-end transformer and LSTM models provide efficient sequence-to-property predictions.
• The transformer model outperformed a Principal Neighborhood Aggregation graph neural network and predicted multiple normal mode frequencies simultaneously.

Conclusions:

• End-to-end transformer and LSTM models offer efficient, structure-independent prediction of protein dynamics from sequences.
• These models facilitate protein engineering, analysis, and design.
• Potential applications include scientific sonification for analyzing subtle sequence changes.