Long-lived conformational states and their interconversion rates critically determine protein function and regulation. When these states have distinct chemical shifts, the measurement of relaxation by NMR may provide us with useful information about their structure, kinetics, and thermodynamics at atomic resolution. However, as these experimental data are sensitive to many structural and dynamic effects, their interpretation with phenomenological models is challenging, even if only a few metastable states are involved. Consequently, approximations and simplifications must often be used which increase the risk of missing important microscopic features hidden in the data. Here, we show how molecular dynamics simulations analyzed through Markov state models and the related hidden Markov state models may be used to establish mechanistic models that provide a microscopic interpretation of NMR relaxation data. Using ubiquitin and BPTI as examples, we demonstrate how the approach allows us to dissect experimental data into a number of dynamic processes between metastable states. Such a microscopic view may greatly facilitate the mechanistic interpretation of experimental data and serve as a next-generation method for the validation of molecular mechanics force fields and chemical shift prediction algorithms.