Self-Supervised Learning for Label Sparsity in Computational Drug Repositioning.

Area of Science:

• Computational biology
• Pharmacology
• Machine learning in drug discovery

Background:

• Computational drug repositioning accelerates drug discovery but is hindered by scarce validated drug-disease associations.
• Limited labeled data impairs classification models, leading to poor generalization in predicting drug efficacy for new diseases.

Purpose of the Study:

• To develop a multi-task self-supervised learning framework to enhance drug representation for computational drug repositioning.
• To address the challenge of label sparsity in drug-disease association prediction.

Main Methods:

• A multi-task learning framework with drug-disease association prediction as the main task.
• An auxiliary self-supervised task utilizing data augmentation and contrast learning to mine intrinsic drug feature relationships without supervised labels.
• A multi-input decoding network to enhance autoencoder model reconstruction capabilities.

Main Results:

• The proposed framework effectively learns improved drug representations by leveraging self-supervised learning.
• Joint training of main and auxiliary tasks demonstrably enhances the prediction accuracy and generalization of drug-disease associations.
• Experimental validation on three real-world datasets confirms the framework's superior predictive performance compared to state-of-the-art models.

Conclusions:

• The multi-task self-supervised learning framework offers a robust solution to label sparsity in computational drug repositioning.
• The approach successfully improves drug representation learning, leading to enhanced predictive accuracy and model generalization.
• This method represents a significant advancement in accelerating the drug discovery pipeline through improved computational strategies.