Deep Neural Architectures for Highly Imbalanced Data in Bioinformatics.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• The post-genome era generates vast amounts of imbalanced biological data, posing significant challenges for computational tasks.
• Classifying precursor microRNA (pre-miRNA) sequences is particularly affected by class imbalance, as known pre-miRNAs are scarce compared to candidate sequences.
• Standard classifiers often fail when dealing with highly imbalanced datasets, impacting real-world applications.

Purpose of the Study:

• To conduct a comparative assessment of recent deep neural network architectures for classifying pre-miRNA sequences.
• To evaluate the performance of these architectures on imbalanced datasets across animal and plant genomes.
• To introduce a novel graphical method for comparing classifier performance under high class imbalance.

Main Methods:

• Utilized a benchmark framework to analyze various deep neural architectures.
• Tested classifiers on animal and plant genomes with imbalance ratios up to 1:2000.
• Developed a new graphical approach for performance comparison in imbalanced scenarios.

Main Results:

• Deep belief neural networks demonstrated the best performance among the evaluated architectures.
• Performance was assessed across increasing levels of data imbalance.
• The proposed graphical method effectively visualizes classifier performance in high-imbalance contexts.

Conclusions:

• Deep belief neural networks are highly effective for pre-miRNA classification with large imbalanced datasets.
• Addressing data imbalance is crucial for developing accurate and reliable bioinformatics classifiers.
• The study provides valuable insights into selecting appropriate deep learning models for genomic sequence classification.