Predicting gene expression from histone modifications with self-attention based neural networks and transfer learning

Area of Science:

• Genomics and Bioinformatics
• Epigenetics and Gene Regulation

Background:

• Histone modifications are crucial for DNA replication, repair, and transcription.
• Computational models for predicting gene expression from histone modifications are widely studied but require improved accuracy, particularly for cross-cell line predictions.

Purpose of the Study:

• To develop a novel deep learning model, TransferChrome, for predicting gene expression from histone modifications.
• To enhance the accuracy of cross-cell line gene expression prediction using transfer learning.

Main Methods:

• TransferChrome utilizes a densely connected convolutional network for feature extraction from histone modification data.
• Self-attention layers are employed to aggregate global data features.
• Transfer learning is incorporated to boost prediction accuracy across different cell lines.

Main Results:

• The model achieved an average Area Under the Curve (AUC) score of 84.79% on 56 cell lines from the REMC database.
• TransferChrome demonstrated improved prediction performance compared to three state-of-the-art models on most cell lines.
• Experimental results confirmed TransferChrome's superior performance in cross-cell line gene expression prediction.

Conclusions:

• TransferChrome is an effective deep learning model for predicting gene expression from histone modifications.
• The model shows significant promise for accurate and efficient cross-cell line gene expression prediction, advancing epigenetic research.