Project: Machine learning that can explain its own chemical predictions
Challenge Space
Environment Challenge
Prinicipal Investigator(s) and team names
Dr Darren Fayne and Paul Clarke, with Prof Ole-Christoffer Granmo, University of Agder, Norway
Collaborating Institutions
Funding support
DCU, co-supervised doctoral programme
Project start date
September 2025
End date
Current
Problem
Decisions about chemicals, whether to restrict a substance, whether to prioritise it for testing, whether an exposure is tolerable, have to be explainable to the people affected by them and defensible to a regulator.
Machine learning is increasingly capable of making the underlying predictions, but the methods that perform best return a number with no account of how it was reached. Guidance for computational models used in a regulatory setting asks for mechanistic interpretation wherever it can be given, which is where most modern models stop.
Background
The group works with a different class of learner. Tsetlin Machines, developed by Prof Granmo's group at the University of Agder, learn in propositional logic: the trained model is a set of readable rules stating which combinations of features matter, so a chemist can inspect the reasoning, agree with it, or argue against it.
Working together, the two groups have benchmarked this approach against conventional methods for predicting chemical activity, published in 2026, establishing that the interpretability does not have to be paid for in accuracy.
Solution
The doctoral work continues jointly between DCU and Agder, extending the method to describe binding sites as three-dimensional grids of chemical properties and to learn explicit rules about which kinds of molecule interact with them.
The present application is to drug resistance, but the property that makes the approach useful there, producing a stated reason rather than a score, is the same property environmental risk assessment needs. The group would be glad to apply it to prediction problems arising in the Institute's environmental work.