Predicting metabolite-disease associations based on auto-encoder and non-negative matrix factorization.

Area of Science:

• Biochemistry and Bioinformatics
• Computational Biology
• Disease Diagnostics

Background:

• Metabolism involves chemical reactions essential for life, with deviations linked to diseases.
• Abnormal metabolite levels can indicate specific diseases, but traditional detection is labor-intensive.
• Accurate identification of metabolite-disease relationships is crucial for diagnostic advancements.

Purpose of the Study:

• To develop an efficient deep learning model for predicting potential associations between metabolites and diseases.
• To overcome the time and labor constraints of traditional experimental methods for metabolite analysis.
• To enhance disease diagnosis through reliable prediction of metabolite-disease links.

Main Methods:

• Developed MDA-AENMF, a deep learning model integrating auto-encoder and non-negative matrix factorization.
• Utilized multiple similarity networks and three modules (auto-encoder, NMF, graph attention auto-encoder) to extract metabolite and disease features.
• Merged extracted features into comprehensive vectors for training a multi-layer perceptron classifier.

Main Results:

• Achieved an Area Under the Receiver Operating Characteristic Curve (AUC) of 0.975 and Area Under the Precision-Recall Curve (AUPRC) of 0.973 via 5-fold cross-validation.
• Demonstrated superior performance compared to existing state-of-the-art predictive methods.
• Validated most novel associations predicted by MDA-AENMF through case studies, confirming its reliability.

Conclusions:

• MDA-AENMF effectively predicts potential metabolite-disease associations with high accuracy.
• The model offers a reliable and efficient alternative to traditional experimental methods for disease-related metabolite analysis.
• This approach holds significant promise for advancing disease diagnosis and understanding metabolic pathways.