Personal transcriptome variation is poorly explained by current genomic deep learning models.

Area of Science:

• Genomics
• Computational Biology
• Systems Biology

Background:

• Deep learning models can predict genomic features and gene expression from DNA sequence.
• Current models excel at predicting expression from a reference genome across cell types.

Purpose of the Study:

• To evaluate the performance of state-of-the-art deep learning models in explaining gene expression variation between individuals.
• To assess the models' ability to capture the effects of cis-regulatory genetic variants on gene expression.

Main Methods:

• Utilized paired personal genome and transcriptome data.
• Evaluated four leading deep learning models on this individual-specific data.

Main Results:

• Models demonstrated limited performance in explaining expression variation between individuals.
• Models frequently failed to predict the correct direction of effect for cis-regulatory genetic variants.

Conclusions:

• Current genomic deep learning models have limitations in explaining inter-individual expression variability.
• Further development is needed to accurately model the impact of genetic variation on gene expression.