Multitask joint learning with graph autoencoders for predicting potential MiRNA-drug associations.

Area of Science:

• Biomedical Informatics
• Computational Biology
• Genomics

Background:

• MicroRNA (miRNA) abnormalities are linked to numerous diseases.
• Accurate prediction of drug-miRNA associations is crucial for disease treatment and novel drug discovery.
• Existing computational methods often neglect valuable information from related tasks.

Purpose of the Study:

• To develop a multitask joint learning framework (MTJL) for predicting drug-miRNA associations.
• To leverage multitask learning to enhance prediction accuracy by utilizing information from related tasks.
• To improve the embedding representations of drugs and miRNAs for better association prediction.

Main Methods:

• Constructed high-quality drug and miRNA similarity networks using integrated information.
• Employed a graph autoencoder (GAE) to learn separate embedding representations for drugs and miRNAs.
• Incorporated an auxiliary drug classification task to refine drug embeddings.
• Utilized linear transformation of embeddings to generate predictive association scores.

Main Results:

• MTJL achieved superior prediction performance compared to state-of-the-art models.
• Ablation experiments confirmed that the auxiliary task enhances embedding quality and model robustness.
• Case studies demonstrated MTJL's utility in predicting potential drug-miRNA associations.

Conclusions:

• The proposed MTJL framework effectively predicts drug-miRNA associations.
• Multitask learning and auxiliary tasks significantly improve model performance and robustness.
• MTJL offers a valuable tool for advancing disease treatment and drug discovery through accurate association prediction.