Predicting Pore Limiting Diameter in Metal-Organic Frameworks Using Only Metal-Linker Building Blocks

Metal-organic frameworks (MOFs) show important potential in applications such as gas storage, catalysis, and drug delivery. However, the vast chemical space of MOFs makes experimental exploration impractical. While existing computational models often require the full MOF structure to predict key properties, we introduce a deep learning approach that predicts the pore limiting diameter (PLD) using only the metal-linker building blocks. Our method employs a 3D convolutional neural network (CNN) trained on spatial features derived from the 3D coordination of atoms within metal-linker complexes. We constructed a dataset of 25,529 MOFs from CoREMOF-2019 and Tobacco databases, deconstructed them into metal-linker units, and applied voxelization to generate 3D features. These features were then used to train a 3D CNN model, achieving an R² value of 0.86 on an unseen test set. By enabling accurate PLD prediction without requiring the full MOF structure, our approach facilitates early-stage screening and rational design of MOFs. The data and source code for our model are available at https://github.com/ClinicalAI/MOF_LPD_Prediction.