Optimizing Link Prediction for the CSD Cocrystal Network: A Demonstration Using Praziquantel

Abstract

The physicochemical properties of chemical compounds can be altered and optimized by cocrystallization with a suitable coformer. However, discovering suitable coformers is a difficult and expensive process. Link prediction is one of the several techniques developed to predict suitable new coformers computationally. Link prediction uses a network of known coformers extracted from, e.g., the Cambridge Structural Database (CSD) to predict new cocrystals. We have investigated link prediction methods and were able to improve the performance of these methods using a scoring function called “multi-steps resource allocation”. Further improvements were obtained by examining the local structure of the network to remove imperfections and by using an algorithm previously designed by us to bipartise the network, thus removing imperfections on a global scale. By repeatedly predicting and synthesizing new cocrystals and adding them to the network to predict more new cocrystals, we obtain more and better predictions, but saturation of the local network eventually leads to diminishing returns. We demonstrate this for praziquantel (PZQ), a drug used to treat schistosomiasis. We discovered 11 new cocrystals for this compound, one of which is a racemic conglomerate that could be used to improve the medical efficacy of PZQ, and present 6 new cocrystal structures.

Link prediction for the CSD cocrystal network is optimized and iteratively applied to praziquantel.