ChromBPNet

ChromBPNet factorizes assay-specific enzyme biases from regulatory sequence determinants of chromatin accessibility using base-resolution deep learning DNA sequence models.

Key Features:

• Bias factorization: ChromBPNet detects, learns and deconvolves assay-specific enzyme biases from regulatory sequence determinants of accessibility.
• Base-resolution accessibility prediction: The model predicts base-resolution chromatin accessibility profiles from DNA sequence.
• Lightweight architecture: ChromBPNet achieves competitive performance with much larger contemporary models despite its lightweight design.
• Variant effect prediction: The model predicts variant effects on chromatin accessibility, pioneer transcription factor binding and reporter activity across assays, cell contexts and ancestry.
• Regulatory syntax interpretation: ChromBPNet provides interpretation of disrupted regulatory syntax and helps prioritize and interpret regulatory variants that influence complex traits and rare diseases.
• Motif and footprint discovery: The model enables robust discovery of compact transcription factor motif lexicons, cooperative motif syntax and precision footprints across assays and sequencing depths.

Scientific Applications:

• Cis-regulatory element analysis: ChromBPNet reveals cis-regulatory sequence syntax at context-specific cis-regulatory elements mapped by chromatin accessibility assays.
• Transcription factor footprinting: The model identifies precision transcription factor footprints across assays and sequencing depths.
• Cooperative motif syntax discovery: ChromBPNet discovers cooperative motif syntax governing transcription factor binding and chromatin accessibility.
• Regulatory variant prioritization: The tool prioritizes and interprets regulatory variants that influence complex traits and rare diseases.
• Variant effect benchmarking: ChromBPNet predicts effects of genetic variants on chromatin accessibility, pioneer transcription factor binding and reporter activity.

Methodology:

ChromBPNet employs deep learning DNA sequence models trained on base-resolution chromatin accessibility profiles. The architecture factorizes assay-specific enzyme biases from regulatory sequence determinants through deconvolution during model training and interpretation.