An end-to-end method for predicting compound-protein interactions based on simplified homogeneous graph convolutional...

Area of Science:

• Computational biology
• Bioinformatics
• Drug discovery

Background:

• Identifying compound-protein interactions (CPIs) is vital for drug discovery and understanding protein functions.
• Deep learning models offer efficient, high-throughput CPI prediction, reducing reliance on costly experiments.
• Existing methods often struggle with the inherent data imbalance common in real-world biological datasets.

Purpose of the Study:

• To introduce SPVec-SGCN-CPI, an end-to-end deep learning approach for accurate compound-protein interaction prediction.
• To address the challenges of data imbalance and computational efficiency in CPI prediction.
• To validate the model's performance against existing state-of-the-art methods.

Main Methods:

• Utilized a simplified graph convolutional network (SGCN) model integrated with low-dimensional features from SPVec.
• Employed graph topology information alongside node features for interaction prediction.
• Incorporated layer-wise propagation for efficient aggregation of K-order neighbor information, mitigating neighbor explosion.

Main Results:

• SPVec-SGCN-CPI significantly outperformed four machine learning and six deep learning methods across three datasets.
• The model demonstrated superior performance, especially in scenarios with imbalanced data.
• Validated predictions on unlabeled ChEMBL data, with top-ranked interactions confirmed via molecular docking and existing evidence.

Conclusions:

• SPVec-SGCN-CPI reliably predicts compound-protein interactions, offering a powerful tool for drug re-profiling and discovery.
• The methodology effectively fuses heterogeneous information from compounds and proteins.
• This approach accelerates target identification and streamlines drug discovery by considering sample imbalance and computational efficiency.