A community effort to optimize sequence-based deep learning models of gene regulation.

Area of Science:

• Computational genomics
• Machine learning in biology
• Bioinformatics

Background:

• Genomics model performance is significantly influenced by architectural choices and training methodologies.
• A systematic evaluation framework is crucial for understanding these impacts in genomics.

Purpose of the Study:

• To systematically evaluate the impact of model architectures and training strategies on genomics model performance.
• To identify optimal neural network designs and training approaches for DNA sequence analysis.

Main Methods:

• Conducted a DREAM Challenge using a large dataset of yeast promoter DNA sequences and expression levels.
• Developed the Prix Fixe framework to modularize and test various neural network architectures and training strategies.
• Evaluated models using a comprehensive suite of benchmarks across different sequence types and species.

Main Results:

• Top-performing models utilized neural networks, with variations in architecture and training strategies.
• The Prix Fixe framework enabled further performance improvements by testing modular building block combinations.
• Models achieved state-of-the-art results on yeast data and surpassed existing benchmarks on Drosophila and human genomic datasets.

Conclusions:

• Neural network architectures and training strategies critically influence genomics model performance.
• The Prix Fixe framework provides a method for dissecting and optimizing these models.
• Gold-standard datasets and systematic evaluations drive significant progress in computational genomics.