Employing bimodal representations to predict DNA bendability within a self-supervised pre-trained framework.

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• DNA bendability, a measure of DNA looping rate, is vital for biological processes.
• Loop-seq enables measurement of DNA fragment cyclizability but is resource-intensive for large-scale DNA.
• Understanding DNA bendability is crucial for deciphering regulatory elements like super-enhancers.

Purpose of the Study:

• To develop an innovative computational model, MIXBend, for forecasting DNA bendability.
• To integrate nucleotide sequences and physicochemical properties for enhanced prediction accuracy.
• To elucidate the impact of DNA bendability on human genome regulatory motifs.

Main Methods:

• Utilized DNABERT (a pre-trained language model) and a convolutional neural network with an attention mechanism.
• Developed sequence- and physicochemical-based extractors for DNA representation.
• Employed a k-mer matching module and a self-attention fusion layer for prediction.

Main Results:

• MIXBend demonstrated superior performance compared to existing state-of-the-art methods.
• Identified novel and known motifs in yeast DNA.
• Discovered significant bendability variations in human super-enhancer regions and transcription factor binding sites.

Conclusions:

• MIXBend offers a powerful computational approach for predicting DNA bendability.
• The model provides valuable insights into DNA structure-function relationships.
• Findings highlight the importance of DNA bendability in gene regulation.