How can we determine whether a mutation – a change of a few atoms in a much larger molecule - can modify the native function of a protein? In human genetics, extremely rare mutations that are found in less than 5 patients cannot be directly associated with the disease, unless extensive experimental validation is performed, and thus are classified as VUS (Variants of Unknown Significance). A similar issue is faced in tumor screens regarding the interpretation of low-frequency “passenger” mutations, especially in or genes do not fit Tier-1/Tier-2 cancer gene criteria. Since protein structure dictates motion, and motions determine function, to explain how a mutation triggers e.g. oncogenic activation needs first, knowing relevant conformations, and second, their inter-conversion pathways. Variant pathogenicity scoring is thus primarily based on sequence- features like evolutionary conservation, population frequency, etc. and simple assessments of residue-residue interactions, chemical properties or protein stability, but still, most approaches are built on a “static” view of structures and do not attempt to predict, from first principles, the conformational effect of mutations. To efficiently search for mutations with impact on conformation and function, we are developing an extremely advanced mutational screening workflow (NAISS 2024/1-7 Large Compute allocation), and in this project, we aim to develop Machine Learning models to evaluate structural-dynamical features and predict the pathogenicity of mutations.