When it comes to translating texts or generating images, artificial intelligence already outperforms humans in many areas. When it comes to predicting chemical reactions, however, the use of artificial intelligence is still in its infancy.
Theoretical chemist Esther Heid from the Institute of Materials Chemistry at TU Wien wants to change this: in recent years, she has developed completely new ideas on how neural networks can be used to find better chemical processes. She has now been awarded an ERC Starting Grant by the European Research Council (ERC), endowed with 1.5 million euros over a project period of five years. Esther Heid now wants to use the money to expand her research group and make neural networks better and more efficiently usable for chemical research.
No ready-made solutions are available
Using artificial intelligence (AI) for scientific research is nothing new. There have long been ready-made AI tools that can be downloaded and then trained using your own data. An AI that is designed to assign images to different categories can then be used, for example, to assign medical images to different diseases.In chemistry, however, this is more complicated, explains Esther Heid: "You can’t simply use existing AI tools, train them on chemical data sets and then use them to predict chemical reactions. In recent years, it has become clear that in order to be successful here, you have to develop very specific new AI methods that are customised for chemistry."
Chemical properties and chemical reactions
What already works to a certain extent today is predicting the properties of different chemicals with AI: "You can train a neural network on many different molecules and their known properties, and then in many cases the AI can also correctly estimate the properties of other molecules - for example, the solubility of the molecule."Some researchers therefore hoped to be able to expand such neural networks in order to use them to correctly predict chemical reactions too - but until now, there has been no resounding success. "That’s not surprising when you look a little deeper into these methods," says Esther Heid. "You quickly realise that some phenomena cannot be described correctly in this way for very fundamental reasons."
This includes stereochemistry, for example: many molecules exist in different variants. They consist of the same atoms, but are mirrored versions of each other, like a right and a left glove. This is particularly important in biochemistry: in biochemical reactions, one of the two variants is often produced to a much greater extent than its mirror image. "Previous AI methods cannot cope with this at all," says Esther Heid. "The data structure with which these methods work simply cannot map such phenomena correctly."
New foundations for AI
The first step is to find a suitable data format in which not just individual molecules but entire chemical reactions can be coded so that neural networks can handle them correctly and efficiently. Basic research in AI is needed to be able to apply it to chemistry in a meaningful way. "We have made significant progress in this area in recent years. Benchmark tests show that our approaches are already making much more progress than other methods available today. But there is still a lot to do here, and we have a long list of ideas that we now want to realise as part of the ERC project."If the neural networks do their job as desired, they will be applied specifically to organic and biocatalytic chemical reactions. Such processes are difficult to describe but are particularly important for chemical research. This should make it possible to develop better, more efficient and more environmentally friendly processes for the production of important organic molecules.
"Today, such developments are often simply based on experience or trial and error," says Esther Heid. "If artificial intelligence can deliver suggestions in a short space of time in the future that humans might not even have thought of, that would be a game changer in chemistry."
Esther Heid
Esther Heid studied chemistry at the University of Vienna, where she also completed her doctorate in 2019. She spent some time abroad as a visiting scientist (at the University of Maryland, USA, and at Imperial College, UK) as a doctoral candidate, and after completing her dissertation, she conducted research as a postdoc at MIT in the USA. She finally moved to TU Wien in 2022.Esther Heid has already been honoured with numerous awards, including the Loschmidt Prize from the Chemisch-Physikalische Gesellschaft (Vienna) and a Schrödinger Fellowship from the FWF. Only recently, in June 2024, she received the FWF’s START Award. With the ERC Starting Grant, her research work is now being honoured with one of the most highly endowed and prestigious prizes in the European research landscape.
