DeepBend: An interpretable model of DNA bendability.

Area of Science:

• Genomics
• Computational Biology
• Biophysics

Background:

• DNA bendability influences critical genomic processes like chromatin packaging and protein binding.
• A comprehensive understanding of DNA bendability-associated motifs is lacking.
• Existing machine learning models lack accuracy and interpretability for DNA bendability prediction.

Purpose of the Study:

• To develop an accurate and interpretable machine learning model for predicting DNA bendability.
• To identify sequence motifs and their spatial arrangements that govern DNA bendability.
• To explore the relationship between DNA bendability, chromatin conformation, and genome organization.

Main Methods:

• Development of DeepBend, a convolutional neural network (CNN) model.
• CNN architecture designed to capture DNA sequence motifs and their periodic occurrences.
• Genome-wide prediction of DNA bendability using the DeepBend model.

Main Results:

• DeepBend achieves performance comparable to existing models while offering mechanistic interpretations.
• Identification of known and novel DNA sequence motifs influencing bendability.
• Demonstration of how the spatial patterns of motif occurrences modulate DNA bendability.
• Genome-wide predictions link DNA bendability to chromatin conformation and topologically associated domain organization.

Conclusions:

• DeepBend provides a powerful tool for deciphering the sequence determinants of DNA bendability.
• The model reveals novel insights into how DNA sequence and motif organization control genome structure and function.
• Understanding DNA bendability is key to unraveling chromatin dynamics and gene regulation.