EvoAug

EvoAug enhances the performance and interpretability of deep neural networks (DNNs) for functional genomics prediction by applying evolution-inspired data augmentations.

Key Features:

• Evolution-inspired data augmentation: Applies augmentation strategies motivated by natural evolutionary mechanisms to expand training data diversity.
• Random transformations of DNA sequences: Generates augmented examples by applying random transformations directly to DNA sequences in the training set.
• Increased genetic variation: Expands genetic variation within training datasets to help models learn more robust sequence patterns.
• Fine-tuning with original data: Uses a fine-tuning procedure on the original, non-transformed data to preserve biological relevance and functional integrity.
• Enhances DNN generalization: Improves the generalization performance of deep neural networks on genomic prediction tasks.
• Improves interpretability: Increases the interpretability of established DNN models in regulatory genomics contexts.
• Addresses limited data regimes: Targets scenarios with small or limited genomic datasets to reduce overfitting and improve robustness.

Scientific Applications:

• Functional genomics prediction: Augments training data for DNNs used to predict functional genomic signals from DNA sequence.
• Regulatory genomics prediction: Applied to regulatory genomics tasks to improve model accuracy and interpretability on regulatory element prediction.
• Training robust genomic models with limited data: Enables more accurate genomic deep learning when available labeled genomic data are scarce.

Methodology:

EvoAug generates augmented DNA sequence examples by applying random transformations to existing sequences and then applies a fine-tuning step using the original, non-transformed data to retain biological relevance.