Currently scientists apply fundamental principles of chemical reactivity, learnings from previous/published experiments, and trial and error processes to predict how molecules will react. These reactions are an essential part of the manufacturing process of pharmaceuticals, but the trial-and-error process is often time consuming and expensive. However, through Pfizer's collaboration with the University of Cambridge, the adoption of AI and machine learning is seeking to rapidly improve and speed up this process.
The researchers at the University of Cambridge used a technique called transfer learning, where the AI programme was trained on a different task before being used to help predict chemical reactions.1 Researchers trained the AI programme using a large spectroscopic dataset (Carbon-13 Nuclear Magnetic Resonance), which identifies different chemical environments in a given molecule.1 Having trained the programme on this data alongside Pfizer’s internal chemical reaction data, it was then given the task of trying to predict chemical reaction outcomes.
The results generated by the AI programme were a significant step forward in the development of a predictive tool.
"The critical breakthrough was realising that we could leverage chemical knowledge from other sources, not just Pfizer's internal data, to guide our AI's learning. In the same way that we often learn more thoroughly through a variety of information sources, so to do these machine learning models. We hope to apply this technique of transfer learning towards a 'foundational model' - an AI tool that can be readily deployed towards a plethora of chemical reactivity-based tasks." said Dr Emma King-Smith, lead author of the publication.
During the research project, it was found that including negative data (data about reactions that didn’t work) was important for improving the performance of the AI tool.1
In another aspect of the collaboration, thousands of Pfizer’s chemical high-throughput experimentation datapoints were shared with the scientific community alongside a computational framework to uncover hidden chemical insights within large datasets.2
So, how will this lead to pharmaceutical advancement? These findings could fundamentally improve the accuracy and cost of medicine development, by potentially speeding up the drug design process through predictive modelling.
This research is a significant advancement for medicine design, however, there is still a need for further exploration. Researchers are hoping this breakthrough will inspire more research into the area. Pfizer is committed to working with researchers to optimise medicine design and to harness the potential power of AI.