Refining sequence-to-activity models by increasing model resolution

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Understanding gene expression regulation is key to cell differentiation and disease.
• Deep learning sequence-to-activity (S2A) models predict regulatory elements by modeling chromatin accessibility.
• AI-TAC previously decoded immune cell differentiation syntax using multi-task learning.

Purpose of the Study:

• Introduce bpAI-TAC, a base-pair resolution model for ATAC-seq data.
• Enhance predictions of differential chromatin accessibility by incorporating ATAC-seq profiles.
• Identify novel cis-regulatory motifs and improve understanding of immune cell chromatin landscapes.

Main Methods:

• Developed bpAI-TAC, a novel S2A model operating at base-pair resolution.
• Modeled both chromatin accessibility and ATAC-seq profiles (Tn5 insertion site distribution).
• Employed multi-task learning across related immune cell types and compared sequence attributions.

Main Results:

• Modeling ATAC-seq profiles alongside accessibility significantly improved differential chromatin accessibility predictions.
• Multi-task learning across immune cell types consistently outperformed single-task models.
• Inclusion of ATAC-seq profiles revealed novel motifs with strong effect sizes.

Conclusions:

• Base-pair resolution ATAC-seq modeling provides a more nuanced understanding of cis-regulatory syntax.
• bpAI-TAC enhances the prediction of gene regulatory elements in immune cells.
• The findings advance the study of cell differentiation and immune system regulation.