Gene panel and exome sequencing have revealed a high rate of molecular diagnoses among diseases where the genetic architecture has proven suitable for sequencing approaches, with a large number of distinct and highly penetrant causal variants identified among a growing list of disease genes. The challenge is, given the DNA sequence of a new patient, to distinguish disease-causing from benign variants. Large samples of human standing variation data highlight regional variation in the tolerance to missense variation within the protein-coding sequence of genes. This information is not well captured by existing bioinformatic tools, but is effective in improving variant interpretation. To address this limitation in existing tools, we introduce the missense tolerance ratio (MTR), which summarizes available human standing variation data within genes to encapsulate population level genetic variation. We find that patient-ascertained pathogenic variants preferentially cluster in low MTR regions (P < 0.005) of well-informed genes. By evaluating 20 publicly available predictive tools across genes linked to epilepsy, we also highlight the importance of understanding the empirical null distribution of existing prediction tools, as these vary across genes. Subsequently integrating the MTR with the empirically selected bioinformatic tools in a gene-specific approach demonstrates a clear improvement in the ability to predict pathogenic missense variants from background missense variation in disease genes. Among an independent test sample of case and control missense variants, case variants (0.83 median score) consistently achieve higher pathogenicity prediction probabilities than control variants (0.02 median score; Mann-Whitney U test, P < 1 × 10−16). We focus on the application to epilepsy genes; however, the framework is applicable to disease genes beyond epilepsy.