Answers are grounded in this paper and cite their sources.
Histone modifications regulate chromatin organization and gene activity, and abnormal modification patterns are closely associated with disease-related transcriptional dysregulation. Although ChIP-seq provides accurate measurements of histone modification profiles, genome-wide experimental profiling remains costly and time-consuming. Here, we propose mwHIT, a multi-scale window attention-based Histone Modification Imputation Transformer that predicts histone modification signals from run-on sequencing (RO-seq) and DNA sequence features. mwHIT captures both local and long-range regulatory information while reducing the computational cost of transformer-based inference. Across cell lines and histone marks, mwHIT achieved an average Pearson correlation of 0.7086 in matched method comparisons, outperforming dHIT and Enformer by 13.7% and 33.4%, respectively. On the GM12878 benchmark, mwHIT reduced average mean squared error (MSE) by 69.7% relative to dHIT and 34.6% relative to Enformer. Compared with a full-attention transformer baseline, the multi-scale window attention design reduced runtime from 0.05 h to 0.03 h while maintaining comparable or slightly higher predictive accuracy. mwHIT also recovered histone modification patterns in promoters, enhancers, gene bodies, and regions centered on transcription start sites (TSSs), and highlighted disease-associated regulatory signals near the IL2RA locus. These results demonstrate that mwHIT can support the discovery of disease-associated regulatory markers and provide computational evidence for prioritizing candidate regulatory regions and potential regulatory targets. mwHIT’s data processing methods and model code are freely available in the mwHIT GitHub repository at https://github.com/zhichunlizzx/mwHIT .
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: