Understanding protein mutations requires learning which specific mutations disrupt a protein's function and offers the opportunity to further understand how to restore the normal function by introducing additional mutations. The approach here builds upon analyzing the huge protein sequence and structure data with two novel statistical frameworks, including high-dimensional Potts model inference with structure information, and inference on matrix-valued partial correlations, to develop quantitative predictions of which mutants disrupt function with a new uncertainty measure and models the mutation finesses by integrating the sequence and structure data. Together these innovative approaches with the uncertainty quantification will enable the prediction of compensatory mutations and the final construction of a protein mutant atlas that broadly disseminates the collective mutation information. The project will learn which mutants disrupt protein function, and what additional mutants will restore function. This project will demystify the interdependencies within the sequences to yield a deeper understanding of how protein mutations can change phenotypes. Preliminary results demonstrate how mutations that are intrinsically destabilizing, and destructive can persist but be neutralized by the introduction of additional compensatory mutations. The major aims of this project are to reliably distinguish between the neutral and deleterious mutations and learn how to repair these problems by introducing additional compensating mutations, by applying the new statistical inference frameworks. The tools developed in this project will have the power to make direct connections between gene mutations and changes in phenotypes. This is a highly interdisciplinary collaboration essential for establishing meaningful assessments of protein mutations and that will develop an important tool for informed protein editing.