DeepFormer: a hybrid network based on convolutional neural network and flow-attention mechanism for identifying the...

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Accurate DNA sequence function identification is crucial but challenging in genomics.
• Existing deep learning methods like DeepSEA, DanQ, DeepATT, and TBiNet face computational complexity and limitations in capturing distant feature interactions.

Purpose of the Study:

• To propose DeepFormer, a hybrid deep neural network model for improved DNA sequence function prediction.
• To address the limitations of existing methods by incorporating a convolutional neural network (CNN) and a flow-attention mechanism.

Main Methods:

• Developed DeepFormer, a hybrid model combining CNN for local feature and motif detection with a flow-attention mechanism for capturing distal interactions.
• The flow-attention mechanism operates with linear time complexity, inspired by flow network conservation laws.
• Evaluated DeepFormer against four classical methods on a dataset of 4.9 million noncoding DNA sequences with 919 chromatin features.

Main Results:

• DeepFormer demonstrated significantly superior performance, achieving an average recall rate at least 7.058% higher than existing methods.
• Effectiveness was confirmed in predicting functional variations related to Alzheimer's disease, alpha-thalassemia mutations, and CTCF activity.
• Validated generalization on maize chromatin accessibility across five tissues, showing at least 1.54% higher average recall than classical methods.

Conclusions:

• DeepFormer offers a robust and accurate solution for DNA sequence function prediction.
• The hybrid approach effectively captures both local and distant interactions, overcoming limitations of previous models.
• Demonstrates strong robustness and generalization capabilities across different biological contexts and species.