Explainable Artificial Intelligence for Structure-Activity Relationships in Real-World Drug Design Datasets

Deep Neural Networks have been successfully used for predicting molecular properties and activities. However, due to their complexity, their reasoning is inherently hard to follow, leading to the so-called black-box character. Unfortunately, this can often hinder Drug Development, as it becomes impossible to identify the components of a molecule that have positive contributions to its activity. This problem is addressed by Explainable Artificial Intelligence, which untangles the complicated Machine Learning models and provides clear connections between the input features and outputs. We applied such methods to a lead optimization dataset, with a well-established Structure Activity Relationship and visualized the results in easy-to-understand heatmaps. We compared some simple baseline methods with novel methods, in particular SHAP-based explanations. Our analysis focused on how well we could reproduce the known Structure-Activity Relationship and how the model quality influence the explanations. We showed that SHAP-based methods are very powerful for explaining model predictions, but a good understanding of theunderlying machine learning model and data is required to prevent misunderstandings.