A self-supervised deep learning method for data-efficient training in genomics.

Area of Science:

• Bioinformatics
• Machine Learning
• Genomics

Background:

• Supervised learning in bioinformatics requires extensive labeled data, limiting its application.
• Self-supervised learning (SSL) can utilize unlabeled data to improve model performance, especially with limited labels.
• Existing SSL methods do not effectively leverage the unique properties of genomic data.

Purpose of the Study:

• Introduce Self-GenomeNet, a custom SSL technique for genomic data.
• Improve machine learning model performance on genomic tasks using unlabeled data.
• Develop a method that learns representations generalizable to new datasets and tasks.

Main Methods:

• Developed Self-GenomeNet, an SSL technique tailored for genomic data.
• Utilized reverse-complement sequences to capture genomic data characteristics.
• Implemented prediction of targets with varying lengths to learn dependencies.

Main Results:

• Self-GenomeNet outperforms other SSL methods on data-scarce genomic tasks.
• Achieved superior performance compared to supervised training using ~10 times less labeled data.
• Demonstrated strong generalization of learned representations to new datasets and tasks.

Conclusions:

• Self-GenomeNet is well-suited for large-scale, unlabeled genomic datasets.
• The method can significantly enhance the performance of genomic models.
• SSL tailored for genomic data offers a promising direction for bioinformatics research.