Multi-Branch-CNN: Classification of ion channel interacting peptides using multi-branch convolutional neural network

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Ligand peptides with high affinity for ion channels are crucial for regulating cellular ion flux.
• These peptides are emerging as potential therapeutic agents for diseases like cardiovascular disease and cancers.
• Accurate identification of ion channel peptide binders is essential for drug discovery.

Purpose of the Study:

• To develop and evaluate Multi-Branch-CNN, a novel Convolutional Neural Network (CNN) method, for identifying peptide binders of sodium, potassium, and calcium ion channels.
• To assess the model's performance on both general and novel sequence datasets, simulating real-world applications.
• To compare the efficacy of Multi-Branch-CNN against traditional machine learning algorithms and a standard CNN approach.

Main Methods:

• Developed Multi-Branch-CNN, a CNN architecture with multiple input branches, to analyze intra- and inter-feature types of ion channel peptide sequences.
• Trained and tested the model on distinct datasets, including a general test set and a novel-test set with low sequence similarity to training data.
• Compared Multi-Branch-CNN's performance against thirteen traditional machine learning algorithms (TML13), a Single-Branch-CNN, and an ensemble method (TML13-Stack).

Main Results:

• Multi-Branch-CNN demonstrated superior performance over TML13, achieving accuracy improvements of 3.2%, 1.2%, and 2.3% on general test sets.
• Significant performance gains were observed on the novel-test sets: 8.8%, 14.3%, and 14.6% for sodium, potassium, and calcium ion channels, respectively.
• The study confirmed Multi-Branch-CNN's effectiveness and its advantage over Single-Branch-CNN and TML13-Stack.

Conclusions:

• Multi-Branch-CNN is a highly effective deep learning model for identifying ion channel peptide binders.
• The method shows robust performance, particularly in recognizing novel sequences, which is critical for drug candidate prediction.
• The developed models and resources are publicly available, facilitating further research in ion channel modulation and drug discovery.