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

Area of Science:

• Computational chemistry and cheminformatics.
• Artificial intelligence in drug discovery.
• Explainable AI (XAI) for molecular modeling.

Background:

• Graph Neural Networks (GNNs) excel at compound property prediction, especially with limited data.
• Current GNN interpretability methods struggle with accurate atomic attribution, hindering lead compound optimization.
• The growing AI-generated chemical space demands efficient and reliable XAI methods.

Purpose of the Study:

• To propose a novel two-stage framework for calculating atomic attribution values in GNNs.
• To enhance the accuracy and reliability of GNN interpretability for compound property prediction.
• To accelerate the drug development process by providing deeper insights into GNN predictions.

Main Methods:

• A two-stage framework involving model training via structural pooling.
• Atomic attribution value calculation using substructure mapping.
• Quantification of task-specific atomic attribution values for GNNs.

Main Results:

• Achieved 30%/20%/15% enhancement in atomic attribution accuracy for GCNs on aromaticity/LogP/TPSA datasets.
• Demonstrated high Pearson correlation coefficients (0.93/0.63/0.88), significantly outperforming existing methods (0-0.3).
• The framework shows robustness to model parameter changes and structural variations.

Conclusions:

• The proposed interpretable framework significantly improves GNN atomic attribution accuracy.
• This advancement facilitates more efficient and reliable lead compound optimization in drug discovery.
• The method offers a valuable tool for accelerating AI-driven research in chemistry and pharmacology.