DeepDiff: DEEP-learning for predicting DIFFerential gene expression from histone modifications.

Area of Science:

• Computational biology
• Genomics
• Epigenetics

Background:

• Histone modifications play a crucial role in regulating gene expression and cellular heterogeneity.
• Existing computational methods often fail to capture combinatorial effects or focus solely on cell-type-specific analyses.

Purpose of the Study:

• To develop a unified, end-to-end deep learning model (DeepDiff) for predicting differential gene expression from histone modification signals.
• To interpret the dependencies among histone modifications that control differential gene regulation patterns.

Main Methods:

• Developed an attention-based deep learning architecture, DeepDiff, utilizing hierarchical Long Short-Term Memory (LSTM) modules.
• Implemented a multi-task learning framework with cell-type-specific gene expression predictions as auxiliary tasks.
• Incorporated two levels of attention to identify relevant modifications and genomic positions.

Main Results:

• DeepDiff significantly outperformed state-of-the-art baselines in differential gene expression prediction using Roadmap Epigenomics Project data.
• Validated learned attention weights against existing knowledge on epigenetic mechanisms and gene expression.
• Achieved superior performance across ten different cell type pairs.

Conclusions:

• DeepDiff offers a powerful and interpretable approach for modeling gene regulation through histone modifications.
• The model's ability to capture combinatorial effects and spatial structures advances the field of epigenomic analysis.
• Provides a unified solution for differential gene expression prediction, enhancing understanding of cellular heterogeneity.