AI molecular property prediction for Parkinson's Disease reveals potential repurposing drug candidates based on the increase of the expression of PINK1

Background and objective: Parkinson's Disease (PD), a common neurodegenerative disorder and one of the major current challenges in neuroscience and pharmacology, may potentially be tackled by the modern AI techniques employed in drug discovery based on molecular property prediction. The aim of our study was to explore the application of a machine learning setup for the identification of the best potential drug candidates among FDA approved drugs, based on their predicted PINK1 expression-enhancing activity.

Methods: Our study relies on supervised machine learning paradigm exploiting in vitro data and utilizing the scaffold splits methodology in order to assess model's capability to extract molecular patterns and generalize from them to new, unseen molecular representations. Models' predictions are combined in a meta-ensemble setup for finding new pharmacotherapies based on the predicted expression of PINK1.

Results: The proposed machine learning setup can be used for discovering new drugs for PD based on the predicted increase of expression of PINK1. Our study identified nitazoxanide as well as representatives of imidazolidines, trifluoromethylbenzenes, anilides, nitriles, stilbenes and steroid esters as the best potential drug candidates for PD with PINK1 expression-enhancing activity on or inside the cell's mitochondria.

Conclusions: The applied methodology allows to reveal new potential drug candidates against PD. Next to novel indications, it allows also to confirm the utility of already known antiparkinson drugs, in the new context of PINK1 expression, and indicates the potential for simultaneous utilization of different mechanisms of action.