Fine-tuning sequence to function deep learning models on large-scale proteomic data improves the accuracy of variant...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Sequence-to-function models show promise for variant effect prediction.
• Challenges remain in model accuracy and generalization to new genes and individuals.

Purpose of the Study:

• To fine-tune the Borzoi model using a large dataset from the UK Biobank Plasma Proteomic Project.
• To evaluate the impact of sample size and variant frequency on prediction accuracy.

Main Methods:

• Fine-tuning Borzoi on 54,219 individuals and 2,923 plasma proteins.
• Comparing the fine-tuned model against an Elastic Net baseline across 150 single-gene models.
• Analyzing the influence of rare (MAF < 0.01) and common (MAF > 0.05) variants on model performance.

Main Results:

• The fine-tuned Borzoi model improved variant effect prediction for 86% of genes compared to the baseline.
• Increased sample size and inclusion of rare variants were key drivers of improved performance.
• Fine-tuned Borzoi prioritized rare variants, while the baseline prioritized common variants in regulatory regions.

Conclusions:

• Larger sample sizes and the inclusion of rare variants are crucial for enhancing sequence-to-function models in variant effect prediction.
• Fine-tuned Borzoi demonstrates the capability for highly accurate variant effect prediction, outperforming pre-trained models and single-gene models trained jointly.
• This study highlights the feasibility and importance of incorporating diverse genetic data for robust predictive modeling.