Digital Twin Approaches for Interpretable Side Effect Prediction in Drug Discovery

Artificial intelligence plays an ever-greater role in preclinical drug development, ranging from target identification and molecule design to ADME-Tox prediction; however, predicting side effects before performing clinical trials is still lagging behind. The best performing side effect predictors in the literature use either ATC codes, which are expert-derived features not even available at early stages, or graph neural networks based on chemical similarity, which - although use readily available features - are "black boxes" that do not deliver actionable insights. We argue that a paradigm shift is needed. Instead of using the latest neural network architectures that have proved worthy in other domains with a plethora of available data, one could use the off-targets of the compounds and build simple and interpretable predictors of side effects. To add another layer of biological realism, intricate biophysical mechanisms within the cells could also be simulated and used as features for training. Although not outperforming the current methods by a great margin, this digital twin-based model has the benefit of being interpretable, i.e., it puts biology behind the predictions. We showcase, with real-world examples, how the side effects predicted by this model can be interpreted and traced back to off-target proteins, and the complexes and signaling pathways in which they partake. In this way, the proposed model not only provides actionable insights, but in the future, may contribute to the amendment of secondary pharmacology assays.