AI helps chemists consider more options when it comes to picking the best reactants for an experiment

Chemists have created a ‘scope score’ to help evaluate the breadth of a reaction with less bias.

Choosing which substrates to work with is an important part of developing a chemical reaction. Yet despite their importance, reaction scopes have traditionally been generated and evaluated subjectively, with chemists relying largely on their own judgement to decide which molecules to test and which to leave out. This means that useful chemistries may be missed.

To solve this problem, researchers created a new quantitative ‘scope score’ that they paired with a machine-learning algorithm called ScopeBO to standardise the selection of substrates to maximise both performance and structural diversity.

Existing approaches tend to favour one of two things: substrates that are likely to react or substrates that are structurally different from those already tested. ScopeBO combines both considerations, using machine learning to select substrates that are likely to perform well while also exploring new regions of chemical space. ‘ScopeBO is an open-source app that [combines] active learning with iterative chemical space pruning to deliver representative high-information scopes,’ says Abigail Doyle at the University of California, Los Angeles, US.

‘ScopeBO is trained on information about existing scope entries and selects the next ones based on their expected performance,’ adds Sven Roediger, who was a postdoctoral researcher in Doyle’s group at the time of the study. ‘At the same time, the algorithm also avoids the selection of substrates that are too similar to existing ones.’ He adds that ensures ‘the reaction space is explored broadly while also maintaining a focus on regions in which the reaction works well’.

Tested across multiple reaction datasets, ScopeBO statistically outperformed existing selection strategies, providing a more information-rich picture of what a reaction can and cannot do. The approach could help researchers design more informative reaction screens and reduce waste experiments, while also improving downstream applications such as synthesis planning, chemical discovery and machine-learning models.

That last point is significant because a reaction scope is more than just a list of examples but increasingly provides data for chemical databases and machine-learning models. How researchers choose their substrates can therefore influence how well those models understand what chemistry is possible.

Failed scope entries are also important for understanding a reaction’s behaviour but are often not reported. ‘Through the diversity enforcement in ScopeBO, the chemical space is explored more broadly so that poor performing regions are also highlighted and the information gain from the scope is maximised,’ says Roediger.

‘In reaction discovery, chemists naturally focus first on finding and documenting examples that work well,’ says Jean-Louis Reymond at the University of Bern, Switzerland, who was not involved in the study. ‘ScopeBO could help them explore a more diverse range of substrates, potentially helping them find those successful reactions faster.’

As AI and machine learning become increasingly embedded in scientific discovery, this raises a broader question around what happens when parts of the scientific process are handed over to algorithms.

‘The potential loss lies in over-relying on numerical featurisation … which might not fully capture structure–activity relationships or mechanistic discontinuities,’ says Doyle. ‘By keeping the chemist in the loop for search space definition and initial substrate choices, ScopeBO attempts to mitigate this trade-off to an extent.’

Reymond is optimistic that ScopeBO could make a useful contribution to chemistry. ‘I don’t think using ScopeBO means “handing things over to an algorithm”,’ he says. ‘It means using a clever method to choose test cases.’ Rather than replacing chemists’ expertise, ScopeBO could give them a way to explore chemical space more broadly, he adds.