Metal–organic frameworks (MOFs) are an emerging class of materials possessing significant potential in separation and storage applications. Identifying optimal candidates from tens of thousands of MOFs that have been reported is a challenging task. To this end, machine learning (ML) represents a promising approach to facilitate the selection of best-performing MOFs. In this study, we propose a scheme to develop chemistry-encoded convolutional neural network (CNN) models to predict gaseous adsorption properties, i.e., Henry’s constants of adsorption and adsorption selectivity, in chemically diverse MOFs. To train CNN models, the MOF structures are represented by their atomic locations coupled with associated chemical information of each framework atom including the 6–12 Lennard-Jones parameters (i.e., σ and ε) and point-charge values (i.e., q). Henry’s constants of CH4 and CO2 in approximately 10 000 MOF structures computed via molecular simulations are used for training and testing. Our developed CNN models show a superior prediction accuracy. Models for zeolites are also developed for comparative purposes. Various key aspects of the CNN models, such as data augmentation and spatial resolution, are also systematically investigated for achieving high accuracy.