GENNUS: generative approaches for nucleotide sequences enhance mirtron classification.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Classifying non-coding RNA (ncRNA) sequences, especially mirtrons, is crucial for understanding gene regulation.
• Class imbalance in ncRNA datasets leads to machine learning model overfitting and poor generalization.

Purpose of the Study:

• To introduce GENNUS, a novel approach using generative adversarial networks (GANs) and synthetic minority over-sampling technique (SMOTE) for enhanced mirtron and microRNA (miRNA) classification.
• To address data limitations and improve the performance of machine learning models in ncRNA sequence classification.

Main Methods:

• Utilized GANs to generate high-quality synthetic mirtron sequences, capturing complex patterns and diversity.
• Employed SMOTE in conjunction with GAN-generated data for data augmentation.
• Conducted four experiments to compare classification performance using real data, SMOTE, and GAN-augmented data.

Main Results:

• Models trained with a combination of real and GAN-generated data showed improved classification accuracy compared to traditional SMOTE or real data alone.
• GAN-based data augmentation effectively captured minority class patterns, enhancing model generalization.
• Eliminated the need for extensive feature engineering through effective synthetic data generation.

Conclusions:

• GENNUS, leveraging GANs, offers a powerful solution for addressing class imbalance in ncRNA classification.
• Synthetic data generation significantly enhances mirtron and miRNA classification performance and model generalization.
• This approach provides a scalable pathway for more effective genomic data analysis and understanding gene regulation.