Applying negative sample denoising and multi-view feature for lncRNA-disease association prediction

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Long non-coding RNAs (lncRNAs) are increasingly implicated in human disease pathogenesis and prognosis.
• Accurate prediction of lncRNA-disease associations (LDA) is crucial for understanding complex diseases.
• Existing computational methods struggle with reliable negative sample generation and multi-perspective LDA information integration.

Purpose of the Study:

• To develop an advanced computational method for predicting lncRNA-disease associations.
• To address the challenges of negative sample selection and multi-view feature integration in LDA prediction.
• To enhance the accuracy and reliability of computational LDA prediction models.

Main Methods:

• Proposed NDMLDA method integrating multi-view feature extraction, unsupervised negative sample denoising, and a stacking ensemble classifier.
• Utilized K-means for unsupervised negative sample denoising to mitigate sample imbalance and noise.
• Employed graph attention networks for extracting multi-view features of lncRNAs and diseases.

Main Results:

• NDMLDA achieved high predictive performance with an AUC of 0.9907 and AUPR of 0.9927.
• Demonstrated superior performance compared to baseline methods in predicting lncRNA-disease associations.
• Exhibited low 5-fold cross-validation variance (<0.1%), indicating model stability.

Conclusions:

• NDMLDA effectively predicts lncRNA-disease associations by leveraging multi-view features and robust negative sample handling.
• The model shows significant potential for applications in cancer diagnosis and precision medicine.
• The developed method offers a promising advancement in computational approaches for disease-related lncRNA identification.