Refining sequence-to-expression modelling with chromatin accessibility

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Gene regulation is complex and influenced by DNA sequence and chromatin accessibility.
• Current sequence-to-expression models often overlook chromatin accessibility, a critical regulatory factor.
• This study investigates the impact of integrating chromatin accessibility into predictive models.

Purpose of the Study:

• To develop and evaluate sequence-to-expression models augmented with chromatin accessibility data.
• To determine if incorporating accessibility improves the prediction of gene expression, especially for highly variable genes and across different cell types.
• To analyze how accessibility influences the model's identification of important DNA sequence patterns.

Main Methods:

• Developed sequence-to-expression models incorporating chromatin accessibility as an input feature.
• Compared the performance of augmented models against sequence-only and accessibility-only models.
• Utilized attribution scores to interpret the model's focus on DNA sequences.
• Investigated the effect of fine-tuning pre-trained models with both sequence and accessibility data.
• Assessed the influence of sequencing depth on prediction performance.

Main Results:

• Augmented models significantly outperformed sequence-only or accessibility-only models.
• Improved prediction accuracy for highly variable genes and gene expression in different cell types.
• Attribution scores in augmented models correlated with chromatin accessibility, revealing cell type-specific sequence patterns.
• Fine-tuning pre-trained models with accessibility data further enhanced performance.
• Sequencing depth was identified as a crucial factor for accurate prediction.

Conclusions:

• Chromatin accessibility is a vital feature for improving sequence-to-expression model performance.
• Augmented models provide a more accurate and nuanced understanding of gene regulation.
• The findings highlight the potential of integrating multi-modal genomic data for predictive modeling.