Predicting potential microbe-disease associations based on multi-source features and deep learning.

Area of Science:

• Microbiology
• Computational Biology
• Bioinformatics

Background:

• Complex diseases are linked to microbial communities, influencing tumorigenesis and metastasis.
• Clinical observation of microbiota in disease presents significant gaps.
• Current biological experiments for identifying disease-associated microbes are accurate but time-consuming and expensive.

Purpose of the Study:

• To develop a computational model for efficient prediction of microbe-disease associations.
• To reduce the capital and time costs associated with identifying disease-related microbes.
• To address the limitations of traditional experimental methods in microbiota research.

Main Methods:

• A novel model, DSAE_RF, was developed combining multi-source features and deep learning.
• Calculated four similarities between microbes and diseases to create feature vectors.
• Employed k-means clustering for reliable negative sample screening and a deep sparse autoencoder for feature extraction.
• Utilized a random forest classifier for predicting microbe-disease associations.

Main Results:

• The DSAE_RF model achieved high performance with an AUC of 0.9448 and AUPR of 0.9431.
• Extensive validation through 10-fold cross-validation demonstrated model reliability.
• Comparative analyses, including negative sample selection methods, model comparisons, and case studies (Covid-19, colorectal cancer), confirmed the model's effectiveness.

Conclusions:

• The DSAE_RF model offers a reliable and efficient computational approach for predicting microbe-disease associations.
• This method significantly reduces the costs and time required for identifying disease-associated microbes.
• The findings support the potential of computational models in advancing clinical microbiota research and disease understanding.