Deciphering lncRNA-disease associations based on multi-representation fusion and boosting with Gaussian process

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Long noncoding RNA-disease association (LDA) identification is crucial for understanding disease pathogenesis.
• Current deep learning models struggle with effective feature fusion and accurate classification of unknown lncRNA-disease pairs (LDPs).

Purpose of the Study:

• To develop a novel deep learning framework, LDA-RMGPB, for enhanced LDA prediction.
• To improve the fusion of multi-representation features for lncRNAs and diseases.
• To accurately classify unknown LDPs.

Main Methods:

• Utilized randomized singular value decomposition for extracting linear LDP features.
• Employed a masked graph autoencoder to learn nonlinear LDP features.
• Applied a boosting algorithm with Gaussian process for classifying unlabeled LDPs using fused features.

Main Results:

• LDA-RMGPB significantly outperformed seven state-of-the-art methods across six evaluation metrics and four cross-validation strategies on two LDA datasets.
• Further analyses confirmed LDA-RMGPB's superior LDA identification capabilities.
• Predicted potential LDA linkages for lncRNAs ATP6V1G2-DDX39B and PSORS1C3 with breast cancer and prostatic neoplasms, respectively.

Conclusions:

• LDA-RMGPB offers a robust approach for identifying lncRNA-disease associations, contributing to disease mechanism understanding.
• The framework facilitates the discovery of novel therapeutic molecular targets.
• LDA-RMGPB is publicly available for research use.