Fine-tuning sequence-to-expression models on personal genome and transcriptome data.

Area of Science:

• Genomics
• Computational Biology
• Machine Learning

Background:

• Deep learning models for genomic sequence-to-expression prediction struggle with individual genetic variation.
• Pre-trained models show limited accuracy when applied to personal genome sequences without adaptation.

Purpose of the Study:

• To investigate if fine-tuning sequence-to-expression models with paired personal genome and transcriptome data enhances predictive performance.
• To assess the impact of fine-tuning on model generalizability across individuals and genes.

Main Methods:

• Utilized Enformer, a pre-trained sequence-to-expression model, for fine-tuning experiments.
• Employed paired personal genome and transcriptome data for training.
• Evaluated fine-tuning strategies on held-out individuals and populations.

Main Results:

• Fine-tuning significantly improved gene expression prediction accuracy for genes included in the fine-tuning dataset.
• Performance on held-out individuals, including diverse populations, matched established linear models.
• Model generalizability did not improve for genes not seen during the fine-tuning process.

Conclusions:

• Incorporating individual genetic variation and expression data during training enhances model utility for known variants and individuals.
• The challenge of predicting expression for unseen genes remains unresolved.
• Fine-tuning offers a path to improved персонализированная genomics but requires further advancements for novel gene prediction.