Predicting RNA-seq coverage from DNA sequence as a unifying model of gene regulation

Area of Science:

• Genomics and Computational Biology
• Molecular Biology and Genetics

Background:

• Machine-learning models using genomics data aid genetic variant interpretation by predicting effects on the cis-regulatory code.
• Existing tools face challenges in predicting RNA-seq expression profiles due to modeling complexities.

Purpose of the Study:

• To introduce Borzoi, a novel model capable of predicting cell-type- and tissue-specific RNA-seq coverage directly from DNA sequence.
• To leverage Borzoi's predictions for scoring DNA variant impacts across transcription, splicing, and polyadenylation.

Main Methods:

• Developed Borzoi, a sequence-based machine-learning model trained on DNA sequence to predict RNA-seq coverage.
• Utilized statistics derived from Borzoi's predicted coverage to isolate and score variant effects.
• Applied attribution methods to extracted statistics to identify cis-regulatory motifs influencing gene expression and post-transcriptional regulation.

Main Results:

• Borzoi accurately predicts cell-type- and tissue-specific RNA-seq coverage from DNA sequence.
• Variant effect scoring using Borzoi's statistics demonstrated competitive or superior performance compared to state-of-the-art models on quantitative trait loci.
• Identified key cis-regulatory motifs driving RNA expression and post-transcriptional regulation in normal tissues.

Conclusions:

• Borzoi offers a powerful new approach for predicting RNA expression and assessing variant impacts across multiple regulatory levels.
• The model's ability to utilize extensive RNA-seq data across diverse conditions and species highlights its potential for deciphering DNA sequence to regulatory function relationships.