Meta-Path-Based Multiple Instance Learning with Heterogeneous Graph Neural Network for lncRNA-disease Association Prediction

Long non-coding RNAs (IncRNAs) are involved in the development and progression of numerous diseases. Predicting associations between IncRNAs and diseases is crucial for uncovering their molecular mechanisms and devising new diagnostic and therapeutic approaches. However, conventional biological methods are time-consuming, expensive, and frequently produce inadequate outcomes. Thus, there is an urgent requirement for a predictive approach that comprehensively captures associations between entities and delves into network topology for IncRNA-disease prediction. We present a method called HGCNMRD, utilizing a heterogeneous graph neural network and meta-pathcentered deep multi-instance learning. This method builds a heterogeneous network by incorporating similarities between IncRNAs and diseases along with established associations. Instances are generated through meta-path creation, and attention mechanisms are applied to assess various meta-paths, yielding interpretable association probabilities. We performed 5 -fold and $\mathbf{1 0}$-fold cross-validation experiments on the IncRNA-Disease 2.0 dataset, illustrating HGCNMRD’s superior performance over alternative models and its ability to identify potential IncRNA-disease associations. We conducted case studies on breast and lung cancers on benchmark datasets, and as a result observed that authoritative databases all confirmed the top 10 ranked IncRNAs.