Yes, Sarah was designed to give interpretable predictions.
Predictions are derived from the self-organising hypothesis network. With this approach, the model organises the training data into hypotheses (clusters) based upon the presence of key structural fragments – which can be activating or deactivating. A typical prediction will be derived by considering several of these hypotheses that are matched to fragments of the query molecule. The signal from each of these hypotheses is shown – being generated by considering the training compounds held within that cluster. This then gives the user a high-level view to all the activating and deactivating fragments found by the model. So provides all the fragments that the model has learnt to be deactivating. For a more complete analysis, the user can look at the supporting compounds within each hypothesis which are ordered by relevance to the model. This then provides the user with the ability to challenge each step the model takes and use expert analysis to come to a judgement of the model’s trustworthiness in addition to the confidence metric that is provided alongside the overall call.