Enhancing personalized gene expression prediction from DNA sequences using genomic foundation models.

Area of Science:

• Genomics
• Bioinformatics
• Artificial Intelligence

Background:

• Deep learning (DL) models predict molecular phenotypes from DNA sequences.
• Current models capture gene variation but struggle with inter-individual differences.

Purpose of the Study:

• To improve prediction of inter-individual differences in molecular phenotypes using pre-trained AI embeddings.
• To compare a novel transformer model against the state-of-the-art Enformer model.

Main Methods:

• Utilized pre-trained embeddings from the Nucleotide Transformer, a foundation model trained on over 3,000 genomes.
• Trained a transformer model and compared its predictive performance to Enformer using genotype and expression data from 290 individuals.

Main Results:

• The novel model significantly outperformed Enformer in predicting correlations across individuals.
• An elastic net regression approach using genetic variants narrowed the performance gap.

Conclusions:

• Large pre-trained AI/DL models, like the Nucleotide Transformer, show great potential for advancing functional genomics.
• Foundation models offer unique strengths in flexibility and interpretability for molecular phenotype prediction.
• Future improvements may lead to AI models surpassing regression-based approaches in accuracy.