An explainable one-dimensional convolutional neural network with modified Gabor wavelet transform for the...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Accurate identification of exons is crucial for understanding gene function and regulation.
• Existing methods for exon identification face challenges in preserving critical feature information from DNA sequences.
• Digital Signal Processing (DSP) approaches offer potential for extracting informative features from genomic data.

Purpose of the Study:

• To develop and evaluate an effective one-dimensional convolutional neural network (1D-CNN) model for exon identification.
• To investigate the utility of DSP-based feature extraction methods in conjunction with CNNs for genomic sequence analysis.
• To propose a novel CNN architecture that minimizes information loss during feature processing.

Main Methods:

• Utilized one-dimensional convolutional neural network (1D-CNN) architecture.
• Employed Digital Signal Processing (DSP) techniques, including short-time discrete Fourier transform (ST-DFT) and modified Gabor wavelet transform (MGWT), for feature extraction from DNA sequences.
• Implemented various numerical mapping methods for DNA sequence representation.
• Developed a novel CNN model by excluding the pooling layer to preserve feature information.

Main Results:

• The proposed 1D-CNN model demonstrated effective exon identification capabilities.
• The Voss-MGWT feature extraction method, combined with the proposed 1D-CNN, achieved superior performance compared to other methods.
• The exclusion of the pooling layer in the CNN architecture contributed to preserving feature information.
• The HMR195 dataset was used for experimental validation, showing improved identification accuracy.

Conclusions:

• The proposed 1D-CNN model offers an effective approach for exon identification in DNA sequences.
• DSP-based feature extraction, particularly Voss-MGWT, significantly enhances the accuracy of CNN-based exon identification.
• The novel CNN architecture without pooling layers is beneficial for maintaining feature integrity in genomic analysis.