TradePool: A Novel Interpretable Framework for Quantifying Atomic Attribution Values in Molecular Property Prediction

The interpretability of GNNs (Graph Neural Network Models) has always been a focal point in the field of compound property prediction. GNNs perform well in modeling small-sample compound data sets, but current interpretable methods struggle to accurately interpret atomic attribution values, quantitative measures of individual atom attribution to model predictions. This makes the optimization of lead compounds reliant on the experience of senior chemists, slowing the drug development process. The rapid expansion of the AI-generated chemical space necessitates efficient interpretation of AI (XAI) methods. These tools can uncover insights beyond human intuition, complementing expert knowledge and significantly accelerating optimization cycles. To tackle these challenges, we propose a novel two-stage atomic attribution value calculation framework for GNNs, which includes model training based on structural pooling and atomic attribution value calculation based on substructure mapping. Our interpretable framework quantifies task-specific atomic attribution values─enhancing atomic attribution accuracy (agreement between computed and ground-truth atomic attribution values) by 30%/20%/15% with 0.93/0.63/0.88 Pearson correlation coefficient (Pearson's <i>r</i>) for GCNs on aromaticity/LogP/TPSA data sets, surpassing commonly used interpretable methods that achieve only 0-0.3 Pearson's <i>r</i>. Additionally, our method is not sensitive to changes in model parameters and provides relatively stable prediction results for changes in the compound structures.