Genomic benchmarks: a collection of datasets for genomic sequence classification

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Deep neural networks show promise in biological fields, exemplified by AlphaFold's success in protein structure prediction.
• However, the advancement of deep learning in genomics is hindered by the lack of curated benchmark datasets, unlike in protein folding.
• Genomics faces challenges in genome annotation and functional element identification, requiring standardized benchmarks.

Purpose of the Study:

• To create a collection of curated and accessible sequence classification datasets for genomics research.
• To provide a baseline model and framework for machine learning in genomics.
• To enhance the comparability, reproducibility, and accessibility of machine learning applications in genomics.

Main Methods:

• Compiled novel datasets by mining public databases and incorporating existing published datasets.
• Focused on regulatory elements, including promoters, enhancers, and open chromatin regions.
• Developed a Python package 'genomic-benchmarks' containing nine datasets for three model organisms (human, mouse, roundworm) and a baseline convolutional neural network model.

Main Results:

• A collection of nine curated genomic sequence classification datasets is now available.
• These datasets cover regulatory elements across human, mouse, and roundworm.
• A baseline convolutional neural network model and associated code are provided for immediate use.

Conclusions:

• The proposed benchmark datasets and baseline model facilitate machine learning in genomics.
• This effort aims to make machine learning for genomics more comparable and reproducible.
• The repository is intended to reduce researcher overhead and foster competition and discovery in the field.