Predicting metabolism at an early stage is important in maximising the chance of a drug’s success. However, accurate, useful models can be computationally expensive. To make good models more accessible, our team, in collaboration with the University of Cambridge, have been investigating new machine learning methods, which can be applied to solve our chemistry questions at a fraction of the computational cost of traditional methods.

This peer-reviewed paper in the Journal of Chemical Theory and Computation, describes a MACE interatomic potential, a new machine learning method, which increases the computational efficiency of predicting cytochrome P450 sites of metabolism.

The article builds on our metabolism prediction research, which underpins the StarDrop Metabolism module.

Graphical abstract for transferable machine leraning interatomic potential paper. It depicts the potential energy curves associated with geometry optimisation for bond dissociation energy for CYP metabolism


Empirical force fields for drug-like compounds often aren’t transferable, due to the large number of potential structures within this category. It is difficult to train force fields on enough different compounds to make them widely applicable. Parameters often have to be individually fine-tuned for each molecule for accuracy, or ab initio methods used. However, these can be very computationally expensive, meaning they aren’t currently routinely used on most drug discovery tasks.

Machine learning interatomic potentials (MLIPs) are one way in which computational costs can be reduced, compared to traditional ab initio methods. They can speed predictions up from taking minutes to taking mere milliseconds. However, MLIPs have so far only found applicability to near-equilibrium reactions and closed-shell molecular structures.

These parameters are not realistic for many potential scenarios that we want to model. For example, drug metabolism by cytochrome P450 (CYP) enzymes involves open-shell structures, as it proceeds via radical intermediates.

In this paper, we present a new transferable type of MLIP, a MACE interatomic potential, that is applicable to open- and closed-shell drug-like molecules containing hydrogen, carbon, and oxygen atoms. We use this new MACE method to predict bond dissociation energy (BDE) for CYP metabolism prediction. MACE, which is a type of message-passing neural network, shows exceptional speed, accuracy and transferability compared to other existing methods.

The paper demonstrates the potential of this new method, which can in future be extended to cover molecules with more chemical elements, other (CYP-mediated) reaction classes and modelling of full reaction paths, not only BDEs.

More metabolism resources