Identification of drug-gene-disease interaction is important to find drugs effective toward disease. Nevertheless, it is not easy since we have to identify drugs effective to genes which are also critical for diseases. When starting from diseases (disease centric approach), we need to identify genes critical in diseases and find drugs effective to the selected genes. When we start from drugs (drug centric approach), we need to find genes that drugs target and then find diseases in which identified genes are critical. These are complicated processes. If we can identify genes which are effective to the disease and can be targeted by drugs, i.e. if we can start from genes (gene centric approach), it is much easier. Nevertheless, none knows how we can identify such sets of genes without specifying either target diseases or drugs. Hence, our novelty is attributed to an artificial intelligence-based approach, employing unsupervised methods and identifying such genes without specifying either diseases or drugs. To inspect the feasibility, we have applied tensor decomposition (TD) based unsupervised feature extraction (FE) to perform drug repositioning from protein-protein interaction (PPI) without using any other information. Proteins selected by TD based unsupervised FE included many genes related to cancers as well as drugs that target selected proteins. Thus, we could identify drugs for cancers from only PPI. Since the selected proteins have more interaction, we replaced the selected proteins with hub proteins and found that hub proteins themselves can be used for drug repositioning. In contrast to hub proteins that can identify only cancer drugs, TD based unsupervised FE enables us drugs for other diseases. In addition to this, TD based unsupervised FE can identify drugs effective inin vivoexperiments, which is difficult by hub proteins. In conclusion, TD based unsupervised FE is a useful tool to perform drug repositioning only using PPI without other information.