Machine learning and computational hydrogenation: an experiment

For half a century, rhodium catalysts have been converting alkenes enantioselectively with hydrogen (hydrogenation). Countless articles have already been published on this type of catalyst and the underlying reaction mechanism, so one might reasonably expect a well-developed understanding of these catalytic reactions to exist. Yet there is still no streamlined method for quickly selecting the right ligands for a homogeneous catalyst when switching substrates. Adarsh Kalikadien, Evgeny Pidko, and colleagues from TU Delft and Janssen Pharmaceutica wanted to explore whether they could develop a predictive model using machine learning — but the project took an unexpected turn.

‘The idea wasn’t complicated,’ says Kalikadien, a PhD student in Pidko’s group. ‘We set up a simple model reaction with a very well-known rhodium catalyst. The aim was to use it to build statistical models so we could predict which catalysts and ligands to use, reducing the need for trial and error.’ They applied various machine learning models to a combination of a computational dataset and high-throughput experiments from Janssen.

Random

The team compared, amongst other things, the performance of those models. Kalikadien: ‘We calculated all manner of properties based on quantum chemistry — the most intensive and expensive calculations — as well as 2D cheminformatics and 1D representations.’ These properties are different representations of the catalyst as a model would perceive it. As a test, they also added a random set containing 34 random numbers between -100 and 100. ‘The bizarre thing was that all the simpler models — including the random model — performed identically to the expensive variant; it turned out to be completely uninformative.’

‘We made everything open source’

Something that doesn’t appear in the paper, but which did influence the project, was a small oversight within the team. ‘On the computer, you draw the 3D structure of the catalyst that you’ve tested under certain conditions. You then run DFT calculations on it and extract properties,’ explains Kalikadien. ‘We were using the CAS numbers of the ligands to do this. But what we hadn’t realised was that our CAS numbers and the drawings on the bottles in the lab didn’t match our 3D structures.’

‘For months the team discussed the properties, made improvements, and we ultimately had genuinely good calculations at a high computational level,’ the PhD student continues. ‘But during a particular meeting, it turned out that the computational structures weren’t pointing to the correct identifiers for the experimental data! We had to go through all those structures one by one to find where things didn’t add up. When we’d processed the correct molecules and built a new statistical model, we got almost exactly the same results — much to our astonishment.’ One of the conclusions was therefore that, for this out-of-domain modelling approach, it doesn’t matter what you feed into it. It was an indication that the model wasn’t learning much from the given representation. ‘In hindsight we can laugh about it, but during the project it did take something of a toll on my mental wellbeing,’ he says with a smile.

Valuable

The intention had been for this to be a straightforward side project, but it didn’t go as planned. ‘I found quite a few of the results a bit disappointing, if I’m honest,’ Kalikadien admits. Nevertheless, the research — and in particular the data generated — was genuinely valuable, especially in light of the growing prominence of machine learning. ‘That’s why we made everything open source. Not only is all the data publicly accessible, but we’re also providing the code, complete with packages and documentation, so that anyone who wishes to can carry out the same kind of research.’

In doing so, they have published one of the largest datasets for a particular type of hydrogenation reaction. ‘Publication was still quite a challenge. It was a deeply probing investigation into how machine learning functions within chemistry, and not all the conclusions were positive. That led one high-profile journal to reject the article on the grounds that “it doesn’t belong here”. Fortunately, Chemical Science was rather more receptive, so we were able to publish our data, code, and even interactive figures there.’

Meaningful

What comes next? ‘Our representation wasn’t as meaningful as we’d hoped, so we’re now looking for a representation of the catalyst that is perhaps a little less simplified, yet still as simple as possible,’ says Kalikadien. ‘You also want to avoid the costs becoming too high, so we’re trying to incorporate more information about the reaction mechanism into the model without making it overly elaborate. A dynamic version of the representation, in other words.’

Kalikadien, A.V. et al. (2024) Chem. Sci., DOI: 10.1039/D4SC03647F