MPEMDA: A multi-similarity integration approach with pre-completion and error correction for predicting microbe-drug associations

Abstract

Exploring the associations between microbes and drugs offers valuable insights into their underlying mechanisms. Traditional wet lab experiments, while reliable, are often time-consuming and labor-intensive, making computational approaches an attractive alternative. Existing similarity-based machine learning models for predicting microbe-drug associations typically rely on integrated similarities as input, neglecting the unique contributions of individual similarities, which can compromise predictive accuracy. To overcome these limitations, we develop MPEMDA, a novel method that pre-completes the microbe-drug association matrix using various similarity combinations and employs a label propagation algorithm with error correction to predict microbe-drug associations. Compared with existing methods, MPEMDA simultaneously utilizes the integrated and individual similarities obtained through the Similarity Network Fusion (SNF) method to pre-complete the known drug-microbe association matrix, followed by error correction to optimize the predictive scores generated by the label propagation algorithm. Experimental results on three benchmark datasets show that MPEMDA outperforms state-of-the-art methods in both the 5-fold cross-validation and de novo test. Additionally, case studies on drugs and microbes highlight the method's strong potential to identify novel microbe-drug associations. The MPEMDA code is available at https://github.com/lyx8527/MPEMDA.