MEGPNM as a multiscale edge-aware GAT network with hybrid pooling predicting permeability of non-peptidic macrocycles

Non-peptidic macrocycles are attractive therapeutic candidates for targets that are difficult to drug, while still offering a path toward oral exposure. Predicting their membrane permeability, however, remains difficult because these molecules are conformationally flexible and can display molecular chameleon behavior. These properties make permeability prediction particularly challenging, so macrocycle-specific models are needed. In this work, we propose MEGPNM for permeability prediction. Our approach leverages multilayer edge-aware graph attention, incorporates Jumping Knowledge. We evaluate MEGPNM on PAMPA dataset from the Non-peptidic Macrocycle Membrane Permeability Database and benchmark it against fingerprint-based machine learning baselines and representative deep learning models, achieving the best performance. In addition, attention analyses highlight recurring structural motifs associated with permeability, providing practical clues for permeability-guided macrocycle optimization. We also release a web server for rapid permeability prediction and structure visualization to help early-stage screening. Non-peptidic macrocycles present a promising avenue for drug development, yet their conformational flexibility complicates membrane permeability prediction. Here, the authors introduce MEGPNM, a model that utilizes multilayer edge-aware graph attention, outperforms existing methods in predicting effective permeability, and reveals structural motifs linked to permeability, with a web server facilitating early-stage drug screening.