AI is accelerating the discovery of new medicines, agricultural chemicals and safer industrial compounds — but these systems can fail in silent, dangerous ways. A UQ team is pinpointing where AI predictions break down, and building safeguards to make them trustworthy.
What's the problem?
Artificial intelligence is increasingly used to predict how chemicals will behave — whether a drug candidate is likely to be toxic, how a compound moves through the body, or whether an agricultural or industrial chemical is safe for people and the environment.
These predictions can dramatically speed up discovery, but they can also fail silently: a model may return a confident answer that is simply wrong, particularly for molecules unlike those on which it was trained. Because many of these systems are opaque, it can be hard to tell when a prediction should not be trusted — a wrong call can send a flawed candidate forward into costly testing, or wrongly set aside a promising one.
What's UQ's innovation?
Dr Alex de Sa and Professor David Ascher, in UQ’s School of Chemistry and Molecular Biosciences, build and optimise AI tools that predict chemical properties such as toxicity and pharmacokinetics — the systems used to screen compounds before they are made or tested. Much of de Sa’s work targets the reliability of these tools: using automated machine learning (AutoML) to construct prediction models, then opening those models up to understand how they reach their answers and, crucially, where and why they fail.
By making the automated pipelines interpretable rather than treating them as black boxes, the team can identify the conditions under which a prediction is unreliable and design safeguards that flag them — helping ensure the AI shaping our medicines, crops and environment can be trusted.
What's the impact?
More reliable predictions mean fewer wrong turns in discovery. Catching an untrustworthy result early — before a flawed drug candidate advances to expensive testing, or an unsafe chemical reaches use — saves time, money and effort.
It also builds the confidence that researchers, regulators and industry need to rely on AI for decisions that matter, from developing new medicines to assessing agricultural and industrial chemicals. Because the team’s prediction tools are freely available and already widely used, improvements in their reliability reach a broad community of researchers worldwide.
Did you know?
The free prediction tools built by Professor Ascher’s group — the kind of AI systems this project works to make more reliable — are used more than 12 million times a year by researchers in over 120 countries.
Interested in working with this team?
Contact the research partnerships team
Led by
Dr Alex de Sa
Project team
Professor David Ascher, School of Chemistry and Molecular Biosciences
Key partners
- The University of Melbourne
- Baker Heart and Diabetes Institute
- The University of Tasmania
AI Research Strength | Data-Centric AI · Human-Centred AI . Automation of AI |
|---|---|
Industry portfolio | Advancing AI |
Project hub | |
Key publications | de Sa, Pappa, Freitas & Ascher (2025). Interpreting machine learning pipelines produced by evolutionary AutoML for biochemical property prediction. GECCO Companion ’25. de Sa, & Ascher (2024). Towards Evolutionary-based Automated Machine Learning for Small Molecule Pharmacokinetic Prediction GECCO Companion ’24. Myung, de Sa & Ascher (2024). Deep-PK: deep learning for small molecule pharmacokinetic and toxicity prediction. Nucleic Acids Research. de Sa, Long, Portelli, Pires & Ascher (2022). toxCSM: comprehensive prediction of small molecule toxicity profiles. Briefings in Bioinformatics. |