DeepChrome: deep-learning for predicting gene expression from histone modifications.

Area of Science:

• Computational biology
• Genomics
• Epigenetics

Background:

• Histone modifications are crucial regulators of gene expression.
• Understanding combinatorial effects of histone modifications is key for gene regulation insights and epigenetic drug development.
• Previous computational methods often fail to capture these combinatorial effects or use fragmented approaches.

Purpose of the Study:

• To develop a unified computational framework for predicting gene expression from histone modification data.
• To automatically extract and visualize complex interactions among histone modifications.
• To provide intuitive insights into epigenetic mechanisms regulating genes.

Main Methods:

• A deep convolutional neural network (DeepChrome) was developed for gene expression classification using histone modification signals.
• A novel optimization-based technique was employed to generate feature pattern maps for visualizing combinatorial interactions.
• The model was trained and tested on data from the REMC database across 56 cell types.

Main Results:

• DeepChrome significantly outperforms traditional methods like Support Vector Machines and Random Forests in gene expression classification.
• The visualization technique successfully validated existing knowledge and uncovered novel insights into histone modification interactions.
• These findings offer a deeper understanding of epigenetic regulation relevant to diseases like cancer.

Conclusions:

• DeepChrome provides a powerful, unified approach for analyzing histone modifications and gene expression.
• The visualization tool offers valuable insights into complex epigenetic mechanisms.
• This work facilitates the development of targeted epigenetic therapies.