A team of chemists and computer scientists from the University of Münster has joined forces and developed an AI tool to predict reaction results.
 

Model Based Directly on Molecular Structures

A chemical reaction is a highly complex system. In contrast to the prediction of properties of individual compounds, a reaction is the interaction of many molecules and thus a multidimensional problem. Moreover, there are no clearly defined “rules of the game” which, as in the case of modern chess computers, simplify the development of AI models. For this reason, previous approaches to accurately predicting reaction results such as yields or products are mostly based on a previously gained understanding of molecular properties. The development of such models involves a great deal of effort.
 

The focus of the work presented was therefore on a general applicability of the program, so that other chemists can easily use it for their own work. To ensure this, the model is based directly on molecular structures. Every organic compound can be represented as a graph, in principle as an image. On such graphs, simple structural queries – comparable to the question of colors or shapes in photo – can be made in order to capture the so-called chemical environment as accurately as possible.
 

Molecular Fingerprint to Represent Chemical Structure

The combination of many such successive queries results in a molecular fingerprint. These simple number sequences have long been used in chemoinformatics to find structural similarities and are well suited for computer-aided applications. In their approach, the authors use a large number of such fingerprints to represent the chemical structure of each molecule as accurately as possible.

“In this way, we have been able to develop a robust system that can be used to predict completely different reaction results. The same model can be used to predict both yields and stereoselectivities, which is unique,” adds Marius Kühnemund from the field of computer science.

The authors demonstrated that their program can be applied easily and allows accurate predictions, especially in combination with modern robotics, by using a data set that was not originally created for machine learning. This data set contains only relative sales of the starting materials and no exact yields. For exact yields, calibrations have to be created. However, due to the high effort involved, this is rarely done in reality.

The team will continue to develop their program further and equip it with new functions in the future. Prof. Frank Glorius is confident: “When it comes to evaluating large amounts of complex data, computers are fundamentally superior to us. However, our goal is not to replace synthetic chemists with machines, but to support them as effectively as possible. Models based on artificial intelligence can significantly change the way we approach chemical syntheses. But we are still at the very beginning.”