A Transfer-Learning-Based Deep Convolutional Neural Network for Predicting Leukemia-Related Phosphorylation Sites...

Area of Science:

• Biochemistry
• Bioinformatics
• Computational Biology

Background:

• Phosphorylation, a key post-translational modification (PTM), regulates diverse molecular functions.
• Aberrant phosphorylation is linked to various human diseases, including leukemia.
• Predicting disease-specific phosphorylation sites is crucial for understanding disease mechanisms.

Purpose of the Study:

• To develop reliable prediction models for leukemia-related phosphorylation sites.
• To leverage deep convolutional neural networks (CNNs) combined with transfer learning for improved prediction accuracy.
• To investigate the practical feasibility of these models in analyzing leukemia pathogenesis.

Main Methods:

• Utilized deep convolutional neural networks (CNNs) for automatic feature extraction from protein sequences.
• Employed transfer learning to adapt models to smaller datasets for different leukemia types.
• Validated model performance using Area Under the Curve (AUC) and compared with other machine learning methods.

Main Results:

• Achieved high AUC values for predicting Serine (S), Threonine (T), and Tyrosine (Y) phosphorylation sites in myelogenous leukemia (0.8784, 0.8328, 0.7716, respectively).
• Demonstrated promising performance for T-cell and lymphoid leukemia prediction using transfer learning.
• CNN models outperformed five other machine learning methods in prediction accuracy.

Conclusions:

• The developed CNN models offer a reliable and effective tool for predicting leukemia-related phosphorylation sites.
• The models provide practical feasibility for analyzing leukemia pathogenesis and protein phosphorylation patterns.
• This approach facilitates a comprehensive understanding of disease mechanisms driven by aberrant phosphorylation.