Identifying genes that are capable of inducing tumorigenesis has been a major challenge in cancer research. In many cases, such genes frequently show somatic mutations in tumor samples; thus various computational methods for predicting cancer genes have been developed based on “significantly mutated genes.” However, this approach is intrinsically limited by the fact that there are many cancer genes infrequently mutated in cancer genomes. Therefore, we recently developed MUFFINN (Mutations For Functional Impact on Network Neighbors), a method for cancer gene prediction based not only on mutation occurrences in each gene but also those of neighbors in functional networks. This enables the identification of cancer genes with infrequent mutation occurrence. We demonstrated that MUFFINN could retrieve known cancer genes more efficiently than gene-based methods and predicted cancer genes with low mutation occurrences in tumor samples. Users can freely access a web server (http://www.inetbio.org/muffinn ) and run predictions with either public or private data of cancer somatic mutations. For given information of mutation occurrence profiles, the MUFFINN server returns lists of candidate cancer genes by four distinct predictions with different combinations between gene networks and scoring algorithms. Stand-alone software is also available, which allows MUFFINN to be run on local machines with a custom gene network. Here, we present an overall guideline for using the MUFFINN web server and stand-alone software for the discovery of novel cancer genes.