Optimized model architectures for deep learning on genomic data.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Deep learning model performance relies on task-specific architecture design.
• Optimal deep learning architectures for computational biology, especially genome sequences, remain undefined.
• Current practices often adapt computer vision architectures, neglecting genomic data's unique characteristics.

Purpose of the Study:

• To introduce GenomeNet-Architect, an automated framework for optimizing deep learning architectures for genome sequence data.
• To develop a specialized search space within the framework tailored for genomic applications.
• To enhance model performance and efficiency through automated architecture and hyperparameter optimization.

Main Methods:

• GenomeNet-Architect employs an automated neural architecture search (NAS) approach.
• The framework optimizes the overall network layout using a genomics-specific search space.
• It also fine-tunes hyperparameters for individual layers and the training process.

Main Results:

• On a viral classification task, GenomeNet-Architect reduced read-level misclassification by 19%.
• Inference speed increased by 67%, with an 83% reduction in model parameters.
• Similar contig-level accuracy was achieved using approximately 100 times fewer parameters than baseline models.

Conclusions:

• GenomeNet-Architect offers an effective solution for designing high-performance deep learning models for genomic data.
• The automated framework significantly improves efficiency (speed, parameters) while maintaining or enhancing accuracy.
• This approach addresses the need for domain-specific architecture optimization in computational biology.