SSMFN: a fused spatial and sequential deep learning model for methylation site prediction.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in proteomics

Background:

• Traditional methods for predicting protein post-translational modification sites are costly and time-consuming.
• Neural networks (NN) offer an effective computational alternative for site prediction.
• Methylation is a critical post-translational modification impacting protein function.

Purpose of the Study:

• To develop and evaluate a novel neural network model, the Sequential and Spatial Methylation Fusion Network (SSMFN), for predicting protein methylation sites.
• To compare the performance of SSMFN against existing state-of-the-art methylation site prediction models.

Main Methods:

• The SSMFN model integrates spatial information using Convolutional Neural Networks (CNN) and sequential information using Long Short-Term Memory (LSTM).
• Latent representations from CNN and LSTM branches are fused within the model architecture.
• Model performance was evaluated on both balanced and imbalanced datasets, comparing against established prediction tools.

Main Results:

• The SSMFN model demonstrated superior performance across most metrics when trained on a balanced dataset.
• On imbalanced datasets, all models showed improved performance due to larger training data.
• SSMFN outperformed the PRMePred model in several key metrics, notably reducing the need for extensive feature engineering.

Conclusions:

• The developed neural network model achieves high performance in predicting methylation sites under various conditions.
• Training NN models on imbalanced datasets is recommended for methylation site prediction to enhance specificity, reflecting real-world sample distributions.
• The findings highlight the potential of deep learning approaches for efficient and accurate post-translational modification site prediction.