Abstract by Bryce Hedelius
Physics and Astronomy
Dennis Della Corte
Inference of Protein Structure via Multidimensional Scaling of Deep Learning Predictions
Proteins are biological macromolecules that perform life-sustaining functions. The structure of a protein gives rise to its function but it is very difficult and expensive to experimentally determine the structure. Therefore, techniques have been developed to predict protein features using deep learning. Current methods of determining the structure run molecular dynamics simulations with those features as restraints but require a lot of computational power and struggle to satisfy the predicted features. Rather than starting with an arbitrary full-atomistic model, we use multidimensional scaling (MDS) to calculate coordinates of certain atoms around which we fit the rest of the structure. Repeating this process yields an ensemble of potential structures from which we select the lowest energy structures to refine. This method requires orders of magnitude less computational power while creating structures that are both chemically sound and satisfy the predicted features.