Protein folding is the process in which a sequence of amino acids converge to form a three-dimensional shape. The shape of a protein is extremely important for determining its function. The field of bioinformatics has developed algorithms to accurately predict the structure of a protein. Quality assessment of protein models is a subset of problems associated with these algorithms.
The goal of QA is to score a computer-generated structure of a protein model based on its divergence from the native structure. This, however, is not trivial, as QA is under the assumption that the correct 3D structure is unknown. The main tool used to tackle this problem is Deep Learning. Energy-Based Models are a subset of Deep Learning where instead of predicting the target given some input, an EBM predicts a compatibility score (energy) between the feature and target vectors. On a given input, the goal is to now find a combination of input/target vectors that maximizes the probability they are compatible. This is done by minimizing the energy (model output) between the current input and target vectors.
In this way, EBMs learn a distribution of choices over an input as opposed to a single output prediction. It is known that proteins fold to their lowest energy state. Because of this, EBMs may perform better when predicting QA scores, as they follow more in line with the nature of how a protein folds.
