CEMDA

CEMDA predicts potential associations between microRNAs (miRNAs) and diseases by integrating pair and node embeddings to capture fine-grained pairwise relational features for computational discovery of miRNA-disease interactions.

Key Features:

• Pair and node embedding integration: Integrates pair embedding and meta-path-based node embedding to capture fine-grained pairwise miRNA-disease relationships.
• Heterogeneous network construction: Constructs a heterogeneous network from supported miRNA-disease pairs, disease semantic similarity, and miRNA functional similarity.
• Meta-path extraction: Identifies associated context paths (meta-paths) for each confirmed miRNA-disease pair within the heterogeneous network.
• GRU-based representation learning: Processes meta-paths using a gated recurrent unit (GRU) to learn similarity representations between miRNAs and diseases.
• Multi-head attention weighting: Applies multi-head attention to assign weights to hidden states of each meta-path across network layers.
• Pair embedding refinement: Processes pair embeddings with a multi-layer perceptron (MLP) to emphasize critical segments of pairwise relationships.
• Embedding fusion via cost function: Combines meta-path-based node embeddings and pair embeddings using a cost function to predict associations.
• Cross-validation performance: Validated by leave-one-out cross-validation and fivefold cross-validation with reported accuracies of 93.16% and 92.03%, respectively.

Scientific Applications:

• miRNA-disease association prediction: Predicts and prioritizes potential miRNA-disease associations, including unconfirmed interactions relevant to disease research.
• Cancer case studies: Applied to identify miRNAs associated with lung cancer, breast cancer, prostate cancer, and pancreatic cancer, with confirmations reported in the HDMM V2.0 database.
• Biomarker and therapeutic target discovery: Supports prioritization of candidate miRNA biomarkers and therapeutic targets by ranking predicted miRNA-disease associations.

Methodology:

Constructs a heterogeneous network from supported miRNA-disease pairs, disease semantic similarity, and miRNA functional similarity; extracts meta-paths for each confirmed miRNA-disease pair; processes meta-path sequences with a gated recurrent unit (GRU) and applies multi-head attention to weight hidden states; processes pair embeddings via a multi-layer perceptron (MLP) and fuses meta-path-based node embeddings with pair embeddings using a cost function; evaluates performance using leave-one-out cross-validation and fivefold cross-validation (reported accuracies 93.16% and 92.03%).