By accurately predicting inactive compounds, Cerella identified over 100 molecules which did not need to be synthesised. Our collaborators at Constellation Pharmaceuticals estimated that to be the work of roughly 12 chemists working full time for two months on this single project.
The Challenge
Measuring all endpoints for all compounds is impossible in drug discovery, and making accurate predictions is hard. Complex endpoints with non-linear relationships; experimental variation and error; missing data. These factors make it difficult to successfully apply most AI and machine learning methods to drug discovery. Constellation Pharmaceuticals wanted to see how Cerella could help.
The Solution
Cerella was applied to two of Constellation Pharmaceuticals data sets, for Project A (completed), and B (an ongoing project). Cerella’s models were compared to a range of other modelling approaches, and significantly outperformed them when making predictions. This was because Cerella could successfully harness incomplete experimental data to learn transferable assay-assay correlations.
Based on Cerella’s predictions and uncertainty estimates, a large number of molecules were confidently and accurately predicted to be inactive. This would have avoided the synthesis of over 100 molecules. Constellation estimated that this would have saved 12 chemists work over two months, corresponding to a saving of over $600k, based on a conservative estimate of $300k per FTE year.