Abstract
Quantitative structure-activity relationship (QSAR) models are essential for predicting endpoints that are otherwise challenging to estimate using other in silico approaches. Developing interpretable models for endpoint prediction is valuable as interpretable models may provide valuable insights into the relationship between molecular structure and observed biological or toxicological properties of compounds. In this study, we introduce a novel modification of counter-propagation artificial neural networks that aims to identify key molecular features responsible for classifying molecules into specific endpoint classes. The novel approach presented in this work dynamically adjusts molecular descriptor importance during model training, allowing different molecular descriptor importance values for structurally different molecules, which increases its adaptability to diverse sets of compounds. We applied the method to enzyme inhibition and hepatotoxicity classification datasets. Our findings show that the proposed approach improves the classification of molecules, reduces the number of neurons excited by molecules from different endpoint classes, and increases the number of acceptable models. The proposed approach may be useful in compound toxicity prediction and drug design studies.