By using Cerella to guide generative chemistry algorithms, we found success in the Open Source Malaria (OSM) competition to create new antimalarial candidates. We did not enter the compound with the highest predicted activity, but the one which Cerella predicted would be most likely to be active.
The Challenge
In lead optimisation, the OSM consortium had followed several routes which led to less potent antimalarial compounds, wasting valuable time, money and resources. They wanted better potency predictions at an early stage, so launched a public competition to develop a predictive model for the identification of PfATP4 inhibitors. The resulting models were used alongside generative chemistry algorithms to explore different optimisation strategies.
The Solution
Cerella’s deep learning algorithm was trained on the OSM’s publicly available data set, developing a model to predict missing experimental assay values for potency. It was one of the top ranked models within the competition. By assessing the uncertainty estimates provided by Cerella on each potency value, we could focus on the compounds predicted most likely to be active.
New compound ideas were generated with the Nova™ module of Optibrium’s StarDrop™ platform. Cerella predicted anti-malarial activities for the compounds generated by Nova, and the results were screened to prioritise the compounds most likely to achieve anti-malarial activity and good physicochemical properties.
Approximately 1 million compounds were whittled down to just the most promising, a tert-butyl derivative. This was submitted to the OSM competition, synthesised and assessed experimentally. It was found to have the best activity out of all the compounds suggested by the different competitors. The structure of the compound, which would not have been considered by most medicinal chemists, opened up new directions for exploration for future OSM studies.