PUNNHMD A: Microbe-Disease Association prediction based on PU learning and optimized K -Nearest Neighbor algorithm

By deepening our understanding of the relationship between microorganisms and disease, we can study the formation and pathogenesis of various complex diseases. This, in turn, enables the development of more effective treatment methods and prevention measures to improve human health. However, traditional prediction methods are often faced with challenges in accurately predicting associations and can be time-consuming. As a result, computational approaches have emerged as a powerful tool for identifying and predicting potential associations between microorganisms and diseases. Here we develop a prediction algorithm called PUNNHMDA to predict associations. First, the PU learning (Positive-Unlabeled learning) method was used to select reliable negative microbe-disease association samples. Second, the Mahalanobis distance function was combined with the KNN (K-Nearest Neighbor) algorithm and utilized innovatively for microbe-disease association prediction. To validate the effectiveness of PUNNHMDA, we benchmarked it against five existing approaches on the task of predicting microbe-disease associations. Through evaluating PUNNHMDA on the HMDAD dataset using LOOCV and 10-fold cross-validation, the results demonstrated that PUNNHMDA achieved superior predictive performance over other advanced prediction models, attaining the highest accuracy scores. Specifically, it obtained accuracies of 0.9308 and 0.9351 in LOOCV and 10-fold CV respectively, which were markedly higher than those of other benchmarked methods.