A scalable approach to investigating sequence-to-function predictions from personal genomes

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Sequence-to-function (S2F) models are crucial for understanding DNA's role in gene regulation.
• Existing S2F models face challenges in capturing individual gene expression variability.
• Personalized genomics offers a potential avenue to enhance S2F model accuracy.

Purpose of the Study:

• To introduce SAGE-net, a scalable framework for training and evaluating S2F models on personal genomes.
• To assess the impact of personal genome training on S2F model performance.
• To investigate the sources of performance improvements in S2F models trained with personal genomic data.

Main Methods:

• Developed SAGE-net, a scalable computational framework.
• Trained and evaluated S2F models using individual personal genomes.
• Analyzed performance gains by focusing on variant identification versus cis-regulatory grammar learning.

Main Results:

• Training S2F models with personal genomes significantly improves gene expression prediction accuracy for unseen individuals.
• Performance enhancements are mainly attributed to the identification of specific predictive genetic variants.
• The models did not generalize a broad cis-regulatory grammar across different genomic loci.

Conclusions:

• SAGE-net provides a scalable approach for advancing S2F models in personal genomics.
• Personalized genomic data can boost S2F model accuracy by pinpointing key regulatory variants.
• Further development of scalable software is essential for the widespread application of S2F models in personal genomics.