Positional interpretation of cis-regulatory code and nucleosome organization with deep learning models

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Sequence-to-function neural networks model genomic data but are difficult to interpret.
• Complex readouts like MNase-seq present challenges due to experimental biases.
• Understanding cis-regulatory sequence rules is crucial for deciphering biological processes.

Purpose of the Study:

• Introduce Pairwise Influence by Sequence Attribution (PISA) for interpreting sequence-to-function models.
• Develop PISA to decode combinatorial sequence contributions and identify motifs.
• Enable improved modeling of genomic data, including nucleosome occupancy.

Main Methods:

• Developed PISA using attribution methods to analyze sequence contributions at specific genomic coordinates.
• Applied PISA to MNase-seq data to learn and correct for experimental biases.
• Integrated PISA with neural networks for nucleosome prediction and motif discovery.

Main Results:

• PISA visualizes transcription factor motif effects and detects novel motifs with complex patterns.
• The method successfully identified and corrected for MNase-seq experimental biases.
• PISA-enabled models achieved unprecedented nucleosome prediction accuracy and facilitated motif discovery.
• Systematic motif perturbations revealed insights into Micro-C chromatin domain boundaries.

Conclusions:

• PISA is a versatile tool for training and interpreting sequence-to-function neural networks in genomics.
• The approach enhances the understanding of cis-regulatory codes and biological mechanisms.
• PISA facilitates the design of synthetic sequences with specific nucleosome configurations.