DescriptionSynonymous single nucleotide variants (sSNVs), a common type of genomic variant, does not alter the protein sequence but can have a variety of functional impacts. Integrating sSNVs into complex disease prediction and precision medicine is very difficult—mostly due to the lack of a reliable computational tool to evaluate the effects of sSNVs. Here, to bypass the bottlenecks that most relevant predictors suffer from (i.e. limited size and reliability of training data), we inferred groups of neutral and effect sSNVs based on large-scale variant data from population sequencing cohorts. We then built a novel machine learning-based predictor, synVep (synonymous Variant effect predictor), to evaluate whether a given sSNV has molecular functional effect. Validation on multiple experimental datasets demonstrated synVep’s good performance and suggested the promising utility of synVep in disease prediction. In a separate investigation, we found that incorporating synVep with conservation and variant frequency allowed better identification of cancer genes and cancer driver variants. We then, using synVep and other annotations, proposed a list of sSNVs that are potentially cancer driving variants. These results may help interpretation and prioritization on sSNVs in the context of cancer research.