Refining sequence-to-activity models by increasing model resolution.

Area of Science:

• Genomics
• Computational Biology
• Immunology

Background:

• Chromatin accessibility, measured by ATAC-seq, reveals gene regulatory regions.
• Sequence-to-function (S2F) models link genomic sequences to chromatin states.
• Previous AI-TAC model predicted immune cell chromatin accessibility but missed ATAC-seq profile details.

Purpose of the Study:

• To develop a novel model, bpAI-TAC, that incorporates base-pair resolution ATAC-seq profiles.
• To enhance predictions of differential chromatin accessibility and regulatory element function.
• To investigate the impact of ATAC-profile information on model learning and motif identification.

Main Methods:

• Developed bpAI-TAC, a multi-task neural network modeling ATAC-seq at base-pair resolution.
• Applied the model across 90 immune cell types.
• Utilized sequence attributions to analyze learned motifs and their effect sizes.

Main Results:

• Incorporating ATAC-profile information consistently improved differential chromatin accessibility predictions.
• Multi-task learning on related cell types outperformed single-task models.
• bpAI-TAC identified motifs with differential effect sizes when trained on profiles, revealing regulatory syntax.

Conclusions:

• Modeling ATAC-seq at base-pair resolution provides a more sensitive representation of regulatory syntax in immune cells.
• bpAI-TAC enhances understanding of regulatory mechanisms driving cell type differences.
• This approach improves predictions of variant effects on chromatin accessibility.