We all know that algorithms are better at board games than humans beating the world’s best, first at chess and more recently Go. We wanted to explore whether similar results would play out in areas of drug discovery.

So we devised a challenge that embraced the complexity of real-life compound optimisation towards key discovery goals.

Ten experienced volunteers and one Exscientia algorithm were given the same fixed chemistry task and after 20 rounds of optimisation the results brought together. In line with gameplay, the Exscientia algorithm outperformed all but one of the human chemists, clearly demonstrating that a computational learning approach can achieve the necessary performance for primetime use. 

Further algorithmic improvements have improved performance further and now specific flavours of the systems we refer to as Active Learning can be employed depending on the task at hand.  

The graph shown here summarises Exscientia original system performance (reported by Nature in 2016, light orange circles) and subsequent improvements (orange circles) in the context of the mean human performance (black circles) and overall variation (grey shading). For more information see the original report 