Graph with Residue-Based Cross-Modal Framework Enhances Cell Function-Related Protein Properties Prediction.

Area of Science:

• Computational Biology
• Bioinformatics
• Protein Engineering

Background:

• Accurate protein property prediction is vital for drug design and biological research.
• Existing methods using sequence analysis or structure-aware features have limitations.
• Protein Language Models (PLMs) show promise but struggle with structure-dependent functions.

Purpose of the Study:

• To develop a novel cross-modal framework integrating sequence and structure for improved protein property prediction.
• To enhance the predictive performance beyond current PLM-based approaches.

Main Methods:

• Proposed a Graph with Residues (GwR) framework utilizing Layer-Aggregated Graph Convolutional Network (LA-GCN) and Geometric Vector Perceptron-Graph Neural Network (GVP-GNN).
• Employed two complementary residue graphs: one for semantic features (PLMs, self-attention) and another for structural characteristics (spatial topology).
• Applied GwR to predict subcellular localization, solubility, metal ion binding, and thermal stability.

Main Results:

• GwR consistently outperformed existing methods in predictive performance and training efficiency across four tasks.
• Demonstrated superior or comparable performance against state-of-the-art deep learning models and parameter-efficient fine-tuning strategies.
• Highlighted the effectiveness of integrating sequence and structure information via the GwR framework.

Conclusions:

• The GwR framework offers a robust and efficient approach for protein property prediction.
• Integrating sequence-derived semantic features with structure-aware topological features is key to improving predictive accuracy.
• GwR provides a valuable tool for advancing drug design, disease research, and biological experimentation.