ABIET: An explainable transformer for identifying functional groups in biological active molecules
Tiago Pereira, Maryam Abbasi, Joel Perdiz Arrais · Computers in Biology and Medicine · 2025
Recent advancements in deep learning have revolutionized the field of drug discovery, with Transformer-based models emerging as powerful tools for molecular design and property prediction. However, the lack of explainability in such models remains a significant challenge. In this study, we introduce ABIET (Attention-Based Importance Estimation Tool), an explainable Transformer model designed to identify the most critical regions for drug-target interactions — functional groups (FGs) - in biologically active molecules. Functional groups play a pivotal role in determining chemical behavior and biological interactions. Our approach leverages attention weights from Transformer-encoder architectures trained on SMILES representations to assess the relative importance of molecular subregions. By processing attention scores using a specific strategy — considering bidirectional interactions, layer-based extraction, and activation transformations — we effectively distinguish FGs from non-FG atoms. Experimental validation on diverse datasets targeting pharmacological receptors, including VEGFR2, AA2A, GSK3, JNK3, and DRD2, demonstrates the model’s robustness and interpretability. Comparative analysis with state-of-the-art gradient-based and perturbation-based methods confirms ABIET’s superior performance, with functional groups receiving statistically higher importance scores. This work enhances the transparency of Transformer predictions, providing critical insights for molecular design, structure–activity analysis, and targeted drug development. • Attention-based tool that identifies the most critical molecular regions for biological interactions. • Uses Transformer attention weights to compute token importance scores in SMILES. • Strategy associates higher importance scores to atoms in functional groups. • Tool that contributes to the explainability of molecular property prediction models.