Differentially expressed genes prediction by multiple self-attention on epigenetics data.

Area of Science:

• Genomics
• Computational Biology
• Epigenetics

Background:

• Predicting differentially expressed genes (DEGs) from epigenetics data is crucial for understanding gene regulation and cell heterogeneity.
• Current machine learning methods often lack accuracy, interpretability, or speed.
• This knowledge gap hinders the development of targeted epigenetics drugs for complex diseases like cancer.

Purpose of the Study:

• To develop an accurate, interpretable, and fast machine learning model for predicting DEGs from epigenetics signals.
• To address the limitations of existing methods in prediction accuracy, interpretability, and training speed.

Main Methods:

• Proposed Epi-MSA (Epigenetics Multiple Self-Attention) model.
• Utilizes convolutional neural networks for embedding neighborhood bin information.
• Employs multiple self-attention encoders for epigenetics factor analysis and a soft attention module for factor significance.
• Leverages pure matrix operations for parallel computation and faster training.

Main Results:

• Epi-MSA demonstrated superior performance compared to existing methods on Roadmap Epigenomics and BDAP datasets.
• The model exhibited a smaller standard deviation, indicating high effectiveness and stability.
• Attention weights provided biological interpretability, aligning with existing knowledge.

Conclusions:

• Epi-MSA offers an effective and stable solution for predicting DEGs from epigenetics data.
• The model's interpretability facilitates biological insights into gene regulation.
• Epi-MSA advancements can accelerate the development of novel epigenetics-based therapies for diseases such as cancer.