Prototypical Explanations in an AI Method for Protein Pocket Detection

Prototype-based explanations are a well-established technique in eXplainable Artificial Intelligence (XAI), commonly used for classification problems like image classification. This study presents a novel approach to find prototypical explanations for a protein pocket detection algorithm previously developed by the authors. The method aims to identify pockets that were predicted in a way similar to a specific instance of interest, thereby providing insights into relevant information for medicinal chemists. To validate our approach, we tested it as a binary classification problem, distinguishing between similar and dissimilar pocket pairs using the ProSPECCTs benchmark. The results showed that our method outperformed other state-of-the-art methods, taking into account the uncertainty of the predictions due to variations in the training data. The proposed approach uses k-nearest neighbors in a d-dimensional latent space of pocket descriptors to identify prototypes linked to the instance being explained.