Abstract
Drug repositioning is a valuable and efficient strategy to discover new applications for traditional medications. In
contrast to experimental methods, developing accurate and effective computational methods is crucial. The identification
of potential drug-disease associations is a vital aspect of drug repositioning. In the paper, we proposed a new
computational model called DDNMFNN to identify potential drug-disease associations, combining nonnegative matrix
factorization and neural networks. The sparsity of validated drug-disease associations leads to subpar model
generalization performance. To address this issue, a novel dual multi-graph regularization nonnegative matrix
factorization algorithm with adaptive weights is proposed to reconstruct the association matrix. An efficient
optimization algorithm is designed and convergence proof is provided. Furthermore, a multi-kernel neural network is
utilized to predict potential associations based on the multiple similarity matrices and the reconstructed association
matrix. This network effectively combines the nonparametric flexibility of the multi-kernel method with the structural
characteristics of deep learning. The experimental results of 10-fold cross-validation demonstrate the proposed model
achieved the best performance by comparing it with state-of-the-art models on three datasets. Case studies of three
diseases and prediction results of five real-world network datasets further indicate that the proposed model as a
precise prediction tool that can facilitate drug repositioning efforts effectively.
