Deepprune: Learning Efficient and Interpretable Convolutional Networks Through Weight Pruning for Predicting...

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) excel at predicting DNA-protein binding.
• Increasing CNN kernels improves performance but hinders interpretation and can cause overfitting.
• Limited motifs in true biological models necessitate kernel number reduction.

Purpose of the Study:

• To develop a high-performance CNN model for motif inference with a limited number of kernels.
• To enhance the interpretability of CNN-based models for DNA-protein binding analysis.
• To present DeepPrune, a novel deep learning framework for efficient motif inference.

Main Methods:

• Developed DeepPrune, a deep learning framework utilizing weight pruning in the dense layer.
• Implemented iterative fine-tuning to train CNNs with reduced kernel numbers.
• Evaluated performance on simulated datasets and ChIP-seq data for DNA-binding site inference.

Main Results:

• DeepPrune significantly improves motif inference performance on simulated datasets.
• The framework enables training CNNs with limited kernels, facilitating model interpretation.
• DeepPrune outperforms baseline methods with limited kernels in inferring DNA-binding sites from ChIP-seq data.

Conclusions:

• DeepPrune offers a viable solution for interpretable and high-performance motif inference using CNNs.
• The method effectively balances predictive accuracy with model simplicity.
• DeepPrune advances the application of deep learning in understanding DNA-protein interactions.