With the rapid development of computational power over the past few decades, computational simulation has grown increasingly important in protein-related research due to its high resolution and predictability. However, simulating large proteins over extended time scales using all-atomic resolution models and explicit solvents remains infeasible due to the substantial computational burden. Consequently, researchers have turned to developing coarse-grained (CG) models, which reduce degrees of freedom to accelerate simulations. To develop a CG model, two tasks must be completed. One involves determining which degrees of freedom will be retained in the CG model, known as the CG mapping. The other task is accurately capturing equivalent interactions at the CG resolution using an appropriate force field.

In this dissertation, two data-driven methods are developed for the aforementioned tasks. The first method, Variational Approach for Markov Processes (VAMP)-based coarse-graining, provides a criterion to evaluate different CG mappings under a specific resolution directly from finer-grained simulation data. This evaluation is accomplished without constructing a complete CG model and running CG simulations. Additionally, based on this method, a Markov Chain Monte Carlo sampling method is developed to attain the optimal CG mapping without enumerating all possibilities. The second method optimizes the CG force field according to the native structure based on the energy landscape theory. Specifically, this method is used to extend an existing coarse-grained model, the Associative memory, Water-mediated, Structure, and Energy Model (AWSEM), to include protein-metal ion interactions. The use of these methods enables the construction of accurate coarse-grained protein models, making it possible to simulate larger protein systems over longer time scales.