SSLDTI: A novel method for drug-target interaction prediction based on self-supervised learning.

Area of Science:

• Computational biology
• Drug discovery
• Machine learning

Background:

• Identifying drug-target interactions (DTIs) is crucial for accelerating drug development.
• Graph neural networks (GNNs) show promise but face limitations in feature extraction due to shallow aggregation or over-smoothing in deep architectures.
• A significant challenge is the scarcity of known DTIs, leading to severe class imbalance in predictive modeling.

Purpose of the Study:

• To develop an advanced computational model for accurate prediction of DTIs.
• To overcome the limitations of existing GNNs in capturing comprehensive graph features.
• To address the class imbalance problem inherent in DTI datasets.

Main Methods:

• A hybrid model combining graph autoencoder and self-supervised learning (SSL) was proposed.
• The model efficiently encodes multilevel graph features using limited labeled data.
• A positive sample compensation coefficient was integrated into the objective function to mitigate class imbalance.

Main Results:

• The proposed SSLDTI model demonstrated superior performance compared to four baseline methods across two datasets.
• Experimental results confirmed the model's effectiveness in accurately predicting DTIs.
• Newly predicted DTIs by the SSLDTI model were validated using the DrugBank database.

Conclusions:

• The developed model effectively captures complex graph features for DTI prediction.
• The approach successfully mitigates the class imbalance issue in drug-target interaction datasets.
• This method offers a promising computational strategy for enhancing drug discovery pipelines.