With advancements in bioinformatics, an accumulating body of evidence highlights the substantial implication of piRNA in the progression of various diseases. piRNAs are acknowledged as potential therapeutic targets and biomarkers for disease diagnosis and therapy. Exploring the intricate relationship between piRNA and diseases through computational methods can streamline costs, improve efficiency, and mitigate experimental risks. However, existing computational methods face challenges related to the sparsity of heterogeneous adjacency matrix and inherent noise in similarity networks. Therefore, in this study, we propose a matrix completion-based multi-kernel learning approach for predicting the association between piRNA and diseases, termed MCMKL-PDA. The primary goal of MCMKL-PDA is to populate the sparse adjacency matrix and mitigate the impact of noise from similarity networks on model performance. Specifically, firstly, compute piRNA and disease similarity networks using data from multiple sources. Subsequently, employ bounded matrix completion (BMC) to prefill piRNA-disease association matrices. Next, employ the low-rank fast kernel learning (LRFKL) algorithm fusion similarity networks, and the Network Enhancement (NE) algorithm refines the fused similarity networks. Finally, the Random Forests (RF) methodology is applied to forecast potential associations. The experimental results unequivocally establish the superior performance of MCMKL-PDA over other state-of-the-art methods. Moreover, independent test sets further substantiate the practical effectiveness of MCMKL-PDA.