ABIET: An explainable transformer for identifying functional groups in biological active molecules.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Molecular modeling

Background:

• Deep learning, particularly Transformer models, shows promise in drug discovery but lacks explainability.
• Identifying critical molecular regions like functional groups is vital for understanding drug-target interactions.
• Current explainability methods for AI models in chemistry are often insufficient.

Purpose of the Study:

• To introduce ABIET (Attention-Based Importance Estimation Tool), an explainable AI model for identifying key functional groups in drug molecules.
• To enhance the transparency of Transformer-based models in predicting drug-target interactions.
• To provide insights for improved molecular design and structure-activity relationship (SAR) analysis.

Main Methods:

• Utilized Transformer-encoder architectures trained on SMILES representations of molecules.
• Leveraged attention weights to assess the importance of molecular subregions.
• Developed a specific strategy for processing attention scores, including bidirectional interactions and layer-based extraction.

Main Results:

• ABIET successfully distinguished functional groups from non-functional group atoms by assigning them higher importance scores.
• Experimental validation on multiple drug-target datasets (VEGFR2, AA2A, GSK3, JNK3, DRD2) demonstrated model robustness and interpretability.
• Comparative analysis showed ABIET outperforms existing gradient-based and perturbation-based explainability methods.

Conclusions:

• ABIET significantly enhances the interpretability of deep learning models in drug discovery.
• The tool provides valuable insights into structure-activity relationships, aiding targeted drug development.
• This work paves the way for more transparent and reliable AI-driven molecular design.