MEMDA: A Multi-similarity Integration Pre-completion Algorithm with Error Correction for Predicting Microbe-Drug Associations

Exploring the relationship between microbes and drugs can enhance our understanding of the mechanisms underlying their interactions. Traditional wet lab experimental methods are often time-consuming and laborious, prompting a shift towards computational approaches to address this challenge. While some similarity-based computational models have been developed to predict microbe-drug associations, they only use the integration of multiple similarities as input and ignore the unique contribution of a single similarity, reducing the predictive performance of the model. To address these limitations, we develop MEMDA, a method designed to pre-complete the association matrix using different similarity combinations and subsequently employ a label propagation algorithm with error correction to predict microbe-drug associations. In MEMDA, we collect three type of microbe similarities and drug similarities, and calculate their integrated similarity by Similarity Network Fusion (SNF) method. The collected similarities and integrated similarities are taken together as input. Then, MEMDA pre-completes the association matrix using different combinations of microbe and drug similarities. Subsequently, we use an error correction method to correct the predictive scores obtained from the label propagation algorithm. Experimental results on three benchmark datasets demonstrate that our proposed method outperforms state-of-the-art approaches in both the 5-fold cross-validation and \(de\) \(novo\) test. Furthermore, case studies on drugs and microbes underscore the considerable potential of our method in uncovering novel microbe-drug associations. The MEMDA code is publicly available at https://github.com/lyx8527/MEMDA .