mwHIT: accelerated and accurate histone modification imputation using multi-scale window attention

Area of Science:

• Genomics
• Epigenetics
• Computational Biology

Background:

• Histone modifications are crucial for chromatin organization and gene activity.
• Aberrant histone modification patterns are linked to transcriptional dysregulation in diseases.
• Current methods like ChIP-seq for profiling are expensive and time-consuming.

Purpose of the Study:

• To develop an efficient computational tool for predicting histone modification signals.
• To reduce the cost and time associated with genome-wide histone modification profiling.
• To identify disease-associated regulatory signals and prioritize candidate regulatory regions.

Main Methods:

• Proposed mwHIT (multi-scale window attention-based Histone Modification Imputation Transformer).
• Utilized run-on sequencing (RO-seq) and DNA sequence features as input.
• Implemented a multi-scale window attention mechanism to capture local and long-range information efficiently.

Main Results:

• mwHIT achieved a high average Pearson correlation (0.7086), outperforming existing methods like dHIT and Enformer.
• Demonstrated significant reduction in mean squared error (MSE) on the GM12878 benchmark compared to dHIT and Enformer.
• The multi-scale window attention design reduced computational runtime by 33% while maintaining predictive accuracy.
• Successfully recovered histone modification patterns across various genomic regions and highlighted disease-associated signals.

Conclusions:

• mwHIT offers a computationally efficient and accurate method for histone modification imputation.
• The tool can accelerate the discovery of disease-associated regulatory markers.
• Provides valuable computational evidence for prioritizing regulatory regions and targets in genomic studies.