A representation learning model based on variational inference and graph autoencoder for predicting lncRNA-disease...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Long non-coding RNAs (lncRNAs) are implicated in numerous human diseases.
• Accurate prediction of lncRNA-disease associations is crucial for disease prognosis, diagnosis, and therapy.
• Existing machine and deep learning methods struggle to learn effective low-dimensional representations from high-dimensional lncRNA and disease data.

Purpose of the Study:

• To develop an advanced computational model for predicting lncRNA-disease associations.
• To improve the accuracy and robustness of predicting unknown associations.
• To address the challenge of learning efficient low-dimensional representations from complex biological data.

Main Methods:

• Proposed VGAELDA, an end-to-end model integrating variational inference and graph autoencoders.
• Employed Variational Graph Autoencoders (VGAE) for inferring lncRNA and disease representations.
• Utilized graph autoencoders for label propagation based on known associations.
• Implemented an alternating training strategy using the variational expectation-maximization algorithm.

Main Results:

• VGAELDA effectively captures efficient low-dimensional representations from high-dimensional features.
• The model demonstrates enhanced robustness and precision in predicting lncRNA-disease associations.
• VGAELDA addresses a geometric matrix completion problem using a deep learning approach.
• Cross-validation and numerical experiments confirm superior performance compared to state-of-the-art methods.

Conclusions:

• VGAELDA significantly outperforms existing methods in lncRNA-disease association prediction.
• The model successfully identifies potential lncRNA-disease associations through case studies.
• Source code and data are publicly available for reproducibility and further research.