As an important class of porous materials, MOFs have gained increasingly more interest due to their tunability and to their exceptional gas adsorption properties for gas storage and separation applications. Numerous high-throughput virtual screening (HTVS) studies have been performed that rely on atomistic Monte Carlo simulations to predict each material’s gas adsorption properties. The quality of these predictions relies on the partial atomic charge parameters used for the MOF framework atoms. For this purpose, DFT derived electrostatic potential fitted charges are best, but these require a full periodic DFT calculation of the MOF, which comes at a heavy computational cost. To overcome this bottleneck, we have developed rapid machine learning models to predict partial atomic charges in MOFs. Importantly the models were trained on a large and diverse database of approximately 300K MOFs for which periodic DFT calculations have been performed providing charges on tens of millions of atoms. Different descriptors and the combinations of them as the input for the machine learning models were explored. These descriptors describe the atomic environments for each atom within the MOF, including atomic-property-weighted radial distribution functions, atom-centered symmetry functions, chemical-bonding-based average properties, accessible area of the atom, etc. Several machine learning methods have been explored, including graph neural networks, multilayer perceptron neutral networks, and decision-tree based models. The models can produce charges with a MAD of 0.05 e- compared to the DFT charges for most elements and can do so for a whole MOF in mere seconds.