BRMDA: prediction model for potential microbe-drug associations based on bilinear attention networks and random forest.

INTRODUCTION

Uncharted microbe-drug relationships constitute an under-exploited reservoir of therapeutic leads. In this manuscript, we introduced a hybrid framework named BRMDA by coupling a bilinear attention network with a random-forest classifier to systematically expose latent microbe-drug associations.

METHODS

Firstly, BRMDA integrated multiple drug-centric, microbe-centric, and disease-centric similarity profiles, along with experimentally validated microbe-drug associations, to construct a unified heterogeneous graph. And then, the bilinear attention network and random-forest classifier were employed to compute the predicted scores for potential microbe-drug associations based on the newly constructed unified heterogeneous graph. Next, benchmarking experiments were conducted under a rigorous five-fold cross-validation protocol using the MDAD dataset to validate the prediction performance of BRMDA. Additionally, case studies were further performed, focusing on front-line antibiotics including amoxicillin and ciprofloxacin as well as clinically relevant pathogens including Bacillus cereus and Mycobacterium tuberculosis, to evaluate the translational validity of the proposed model.

CONCLUSION

Intensive experimental results demonstrated that BRMDA outperformed seven state-of-the-art competitors in terms of both AUC and AUPR, and 9 out of the top 10 associations predicted by the model were corroborated by independent literature evidence. These findings underscored the accuracy and translational potential of BRMDA, offering a data-driven compass for antimicrobial discovery and microbe-oriented therapeutic design.