Molecular dynamics simulations provide a wealth of data whose in-depth analysis can be computationally demanding and, sometimes, even unnecessary. Dimensionality reduction techniques are thus routinely employed to simplify and improve the interpretation of trajectories focusing on specific subsets of the system's atoms; a key issue, in this context, is to determine the optimal resolution level, i.e. the smallest number of atoms needed to preserve the largest information content from the full atomistic trajectory. Here, we introduce the protein optimal resolution identification method (PROPRE), an unsupervised approach built on information theory principles that determines the smallest number of atoms that need to be retained to attain a synthetic yet informative description of a protein. By applying the method to a protein dataset and two particular case studies, we show that this number is typically between 1.5 and 2 times the number of residues in a protein; nonetheless, the degree of conformational variability of the system influences the specific number importantly, in that a broader range of large-scale conformations correlates with fewer retained sites. The PROPRE method is implemented in efficient and user-friendly python scripts, which are made available for download on a github repository. Here, the raw data employed for the preparation of the associated manuscript are made available.