EvoAug: improving generalization and interpretability of genomic deep neural networks with evolution-inspired data...

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Deep neural networks (DNNs) show potential in functional genomics but struggle with limited data.
• Generalization is a key challenge for DNNs in biological sequence analysis.

Purpose of the Study:

• To introduce EvoAug, a novel data augmentation technique for training genomic DNNs.
• To enhance the generalization and interpretability of DNNs in functional genomics prediction tasks.

Main Methods:

• EvoAug employs evolution-inspired augmentations to increase genetic variation in DNA sequences.
• A fine-tuning procedure with original data ensures functional integrity is maintained.

Main Results:

• EvoAug significantly improves the generalization capabilities of established DNNs.
• The method enhances the interpretability of DNNs across various regulatory genomics prediction tasks.
• EvoAug provides a robust solution for improving genomic DNN performance.

Conclusions:

• EvoAug is an effective strategy for overcoming data limitations in genomic DNNs.
• The approach offers a promising avenue for advancing functional genomics predictions.
• EvoAug enhances both performance and understanding of DNNs in genomics.