Bi-SeqCNN: A Novel Light-Weight Bi-Directional CNN Architecture for Protein Function Prediction.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning

Background:

• Deep learning, including recurrent neural networks (RNNs) and convolution neural networks (CNNs), excels at predicting protein function.
• RNNs offer strong sequential processing, capturing both short- and long-range dependencies, while CNNs focus on short-term information.
• Existing CNNs are limited in their ability to process sequential data compared to RNNs.

Purpose of the Study:

• To introduce a novel bi-directional CNN architecture, Bi-SeqCNN, for protein function prediction.
• To develop an ensemble-based framework that leverages the strengths of bi-directional CNNs for improved prediction accuracy.
• To apply bi-directional CNNs to general temporal data analysis beyond protein sequences.

Main Methods:

• Developed Bi-SeqCNN, a sub-sequence-based framework employing a novel bi-directional CNN architecture.
• Implemented an ensemble approach within Bi-SeqCNN to enhance prediction performance.
• Designed the bi-directional CNN to mimic the sequential processing capabilities of RNNs.

Main Results:

• Achieved improvements of up to +5.5% over contemporary state-of-the-art (SOTA) methods on three benchmark protein sequence datasets.
• Demonstrated that Bi-SeqCNN is substantially lighter, utilizing 0.50-0.70 times fewer parameters than SOTA methods.
• Showcased the first application of bi-directional CNNs for general temporal data analysis.

Conclusions:

• Bi-SeqCNN offers a significant advancement in protein function prediction accuracy and efficiency.
• The proposed bi-directional CNN architecture effectively captures sequential dependencies, outperforming existing methods.
• This work highlights the potential of bi-directional CNNs for both specialized biological sequence analysis and broader temporal data modeling.