Heterogeneous biomedical entity representation learning for gene-disease association prediction

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Understanding gene-disease associations (GDA) is crucial for medical research, diagnosis, and drug development.
• Current GDA prediction methods, including machine learning and graph neural networks, often fail to capture deep semantic information and rely heavily on training data.
• Predicting GDA experimentally is time-consuming and costly.

Purpose of the Study:

• To propose a novel Gene-Disease Association (GDA) prediction model named FusionGDA.
• To improve the accuracy and efficiency of GDA prediction by enriching semantic representations of genes and diseases.
• To address the limitations of existing methods in capturing deep semantic information and data dependency.

Main Methods:

• FusionGDA utilizes a pre-training phase with a fusion module to enrich gene and disease semantic representations from pre-trained language models.
• Multi-modal representations are generated by integrating protein sequences and disease descriptions.
• A pooling aggregation strategy compresses multi-modal representation dimensions, and contrastive learning loss is employed for feature extraction.

Main Results:

• FusionGDA demonstrated superior performance in GDA prediction across five datasets compared to five baseline models.
• The model effectively captures deep semantic information from heterogeneous biomedical entities.
• A case study highlighted FusionGDA's ability to discover hidden gene-disease associations.

Conclusions:

• FusionGDA offers a powerful and effective approach for predicting gene-disease associations.
• The model's ability to leverage multi-modal data and pre-training enhances semantic understanding for improved GDA prediction.
• FusionGDA has the potential to accelerate biomedical research and drug discovery by identifying novel gene-disease links.