GraphUnet-SS: A novel deep learning model for protein secondary structure prediction based on U-Net architecture

Area of Science:

• Bioinformatics
• Computational Biology
• Structural Bioinformatics

Background:

• Protein 3D structure prediction is vital for understanding protein function.
• Protein secondary structure prediction (PSSP) is a critical intermediate step.
• Deep learning methods have shown promise in enhancing PSSP accuracy.

Purpose of the Study:

• To propose a novel deep learning model, GraphUnet-SS, for improved protein secondary structure prediction.
• To leverage a hybrid architecture combining U-Net, CNNs, GCNs, and BiLSTMs.
• To explore the utility of diverse feature sets including PSSMs, profiles, and amino acid properties.

Main Methods:

• Developed GraphUnet-SS, a deep learning model based on U-Net architecture.
• Integrated convolutional neural networks (CNNs), graph convolutional networks (GCNs), and bidirectional long short-term memories (BiLSTMs).
• Utilized PSI-BLAST PSSMs, HHBlits profiles, amino acid properties, and contact map predictions for graph generation.

Main Results:

• GraphUnet-SS demonstrated superior performance compared to existing PSSP methods.
• The model's hyperparameters were optimized using Bayesian optimization.
• The optimal configuration involved using all layers with a depth of four.

Conclusions:

• GraphUnet-SS offers a significant advancement in protein secondary structure prediction.
• The hybrid deep learning approach effectively integrates diverse biological data.
• The proposed model contributes to more accurate protein 3D structure prediction and functional analysis.