New Paper: Interpretation of Structure–Activity Relationships in Real-World Drug Design Data Sets Using Explainable Artificial Intelligence
27 January 2022
Neural Networks have achieved high predictive performance in a variety of tasks related to Drug-Design. However their inherent black box character means that their decision making is opaque and little insight can be gained from predictions. In a recent collaboration with scientists from Sanofi in Frankfurt we explored a variety of explanation methods for Neural Networks. By using a novel visualization scheme, we are able to generate easily understandable explanations for model predictions. In addition we applied the methods to real world Drug-Design datasets and could demonstrate that model explanations are consistent with known SAR trends.
The full story can be found in the Journal of Chemical Information and Modeling here.