Predicting gene expression from histone marks using chromatin deep learning models depends on histone mark function,...

Area of Science:

• Genomics
• Molecular Biology
• Computational Biology

Background:

• Predicting gene expression from histone marks is crucial for understanding epigenetics.
• Previous machine learning models overlooked cell state, histone mark function, and distal regulatory elements.
• There's a need for more comprehensive models that integrate these factors for deeper biological insight.

Purpose of the Study:

• To comprehensively investigate the relationship between seven histone marks and gene expression across eleven cell types and diverse cell states.
• To develop and apply advanced deep learning models (convolutional and attention-based) for predicting transcription from histone mark activity at both promoters and distal regulatory elements.
• To explore the utility of these models for uncovering new biological insights, including functional and disease-related loci, through in silico perturbation assays.

Main Methods:

• Utilized large-scale datasets encompassing seven histone marks across eleven distinct cell types.
• Employed convolutional neural networks and attention-based deep learning models to predict transcriptional activity.
• Investigated histone mark activity at both proximal (promoters) and distal regulatory elements.
• Performed in silico histone mark perturbation assays to assess functional impacts.

Main Results:

• Demonstrated that histone mark function, genomic distance, and cellular states collectively influence the relationship between histone marks and transcription.
• Found that no single histone mark consistently predicts gene expression across all genomic and cellular contexts.
• Identified specific functional and disease-related genomic loci through in silico perturbation experiments.
• Established a framework for leveraging chromatin deep learning models to generate novel biological discoveries.

Conclusions:

• The relationship between histone mark activity and gene expression is complex and context-dependent, influenced by cellular state and genomic location.
• A holistic approach considering multiple factors is necessary for accurate prediction of transcriptional state from epigenetic data.
• Deep learning models offer a powerful platform for dissecting epigenetic regulation and uncovering new biological insights, including potential therapeutic targets.