Assessing the reliability of point mutation as data augmentation for deep learning with genomic data.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Deep neural networks (DNNs) show promise for genetic disease research but require extensive training data.
• Existing data augmentation methods from other fields often fail with unique genomic data properties.
• Genomic data augmentation is crucial for advancing DNN applications in genetics.

Purpose of the Study:

• To develop a novel data augmentation technique for genomic data.
• To address the data scarcity limitation in training deep neural networks for genetic analysis.
• To enhance the performance of DNNs on genomic tasks.

Main Methods:

• Proposed a new data augmentation technique for genomic data inspired by biological point mutations.
• Utilized point mutations as substitutes for codons within genomic sequences.
• Evaluated the technique's impact on DNN performance for genetic tasks.

Main Results:

• The proposed point mutation-based data augmentation enhances DNN performance on genomic tasks.
• Improvements were observed in tasks involving coding regions, such as translation initiation and splice site detection.
• Silent and missense mutations positively impacted model effectiveness.

Conclusions:

• Point mutation-based data augmentation offers a valuable strategy for improving predictive models of DNA sequences.
• Careful selection of mutation types (e.g., silent, missense) is important for positive outcomes.
• This method presents opportunities to increase the accuracy and reliability of genomic predictive models.