Leveraging Protein Dynamics to Identify Functional Phosphorylation Sites using Deep Learning Models.

Area of Science:

• Computational Biology
• Bioinformatics
• Molecular Dynamics

Background:

• Accurate prediction of post-translational modifications (PTMs) is crucial for understanding protein function and cellular processes.
• Current machine learning models for PTM prediction often lack comprehensive analysis of protein dynamics.
• This limitation hinders a deeper understanding of the functional consequences of PTMs.

Purpose of the Study:

• To develop novel deep learning models that integrate sequence, structure, and dynamics-based features for PTM prediction.
• To elucidate the molecular basis and functional landscape of PTMs by considering dynamic properties.
• To improve the accuracy and functional prediction of key PTMs.

Main Methods:

• Developed two dynamics-centric deep learning models: cDL-PAU for phosphorylation, acetylation, and ubiquitination (PAU) sites, and cDL-FuncPhos for functional phosphorylation (FuncPhos) sites.
• Incorporated sequence, structural, and dynamics-based features into the models.
• Utilized feature selection to identify key contributing factors for PTM prediction.

Main Results:

• cDL-PAU achieved Area Under the Curve (AUC) scores ranging from 0.804 to 0.888 for PAU site prediction.
• cDL-FuncPhos demonstrated a reliable AUC of 0.771 for predicting functional phosphorylation sites.
• Feature analysis highlighted the significant contribution of dynamics-based coupling and commute ability in PTM site discovery.

Conclusions:

• The developed models, cDL-PAU and cDL-FuncPhos, show improved performance in predicting PTM sites by integrating protein dynamics.
• Dynamics-based features are critical for understanding the allosteric propensity and functional significance of PTMs.
• The models provide valuable insights into the physical basis of PTM functions, particularly in oncoproteins.