Deep-learning prediction of gene expression from personal genomes

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Predicting gene expression from DNA sequence is challenging with personal genomes due to lack of variation data in training.
• Existing models trained on reference genomes fail to generalize to individual genetic variations.

Purpose of the Study:

• To develop a deep learning model, Variformer, that accurately predicts gene expression from personal genome sequences.
• To assess the model's ability to generalize across individuals and identify functional genetic variants.

Main Methods:

• Developed Variformer, a deep sequence-based neural network.
• Utilized fine-tuning with personal genomes and tissue-specific gene expression data.
• Evaluated model performance on held-out individuals and genes.

Main Results:

• Variformer achieved high accuracy in predicting gene expression across individuals, approaching cis-heritability.
• The model prioritized functional genetic variants, including those disrupting motifs.
• Variformer demonstrated limitations in generalizing to unseen genes.

Conclusions:

• Fine-tuning with personal genomes enhances gene expression prediction across individuals.
• The model does not learn a generalizable regulatory grammar for unseen loci.
• Fine-tuned models show similar performance and limitations compared to linear models, indicating an existing research gap.