What impact will AI have on antibiotic research?

Antimicrobial resistance (AMR) is sometimes described as a future crisis, but for Australian clinicians it is already part of everyday practice. Resistant infections narrow treatment options, complicate care and increase the risk that previously reliable antimicrobials will fail. 

The Australian Commission on Safety and Quality in Health Care states that hundreds of Australians die each year as a result of AMR, while national surveillance continues to identify critical resistance to last-line agents across hospitals, aged care and the community. Australia also recorded more than 557,000 hospitalisations due to infectious diseases in 2024.

The long-term response must include stewardship, infection prevention and surveillance. But it must also include new antibiotics. That is difficult because drug discovery (including antibiotics) is expensive, slow and scientifically challenging. Artificial intelligence (AI) offers a way to search large chemical spaces and prioritise molecules for laboratory testing. The problem is that many high-performing AI systems remain “black boxes.” They may tell a scientist that a molecule is promising without reliably showing which chemical features drove the prediction. In drug discovery, a plausible-looking but incorrect explanation can send chemists in the wrong direction and waste scarce experimental resources.

Evaluating explainable AI in antimicrobial drug discovery

Our study therefore focused on a question that is often overlooked: not simply “How accurate is the AI?” but “How do we evaluate whether its explanation is useful and trustworthy?”

In our research we designed a framework that tests whether AI can provide reliable chemical reasoning during antibiotic development. The framework aims to overcome the “black box” problem of AI — where AI provides an answer but there’s no explanation of how it got there. This ensures scientists have a strong understanding of the scientific explanations behind its predictions. Our goal is the new tool will help speed up the development of antibiotics long-term. 

To develop the framework, we used a curated dataset of 43,777 molecules previously tested against the superbug bacterium Staphylococcus aureus and built three different AI approaches. The first learned from calculated chemical properties and used SHAP (SHapley Additive exPlanations) to identify which properties were most important for each prediction. The second learned directly from text-based representations of molecular structures and tested which parts of the representation influenced its predictions. The third learned from molecules represented as networks of connected atoms and bonds and examined which chemical substructures were driving its predictions.

Rather than judging explanations by whether they looked chemically convincing, we created a common evaluation framework that tests the credibility of AI’s chemical reasoning in antibiotic development. We asked whether each model could recognise established chemical structures of antibiotics (also known as antibiotic scaffolds); whether it could distinguish very similar molecules with very different activity (which we call “activity cliffs”); whether the explanation agreed with the model’s prediction; and whether explanations remained consistent when the model was retrained under different conditions.

Activity cliffs are especially important. In medicinal chemistry, a small structural change can turn an active compound into an inactive one. An explanation method that simply highlights the familiar antibiotic scaffold may look sensible while missing the small modification that actually changed activity. That is exactly the kind of failure that matters when a chemist is deciding what analogue to make next.

Similar predictions but different explanations

All three approaches were reasonably good at predicting antibacterial activity, and all could recognise important antibiotic scaffolds. But their explanations were not equivalent. The clearest differences appeared when the models were challenged with activity cliffs.

The feature-based Random Forest model method gave strong scaffold recognition, but its fragment contributions were less useful for distinguishing the specific structural changes responsible for different activity. The sequence-based model provided useful explanations, although some were inconsistent with known experimental structure–activity relationships. The graph-based model generally gave the most chemically informative discrimination of activity-cliff fragments.

This matters because predictive accuracy alone would not have revealed these differences. Two models can achieve similar headline performance while offering very different value to the scientist who must decide what to synthesise, test or discard. The framework is therefore not a stamp of approval for one particular AI architecture; it is a way to expose strengths and weaknesses before an explanation is used to guide experiments especially in a high-stake field such as antibiotic discovery.

From better predictions to better decisions

For clinicians, the AI framework is not a diagnostic tool and it does not decide which antibiotic a patient should receive. Its potential value is earlier in the pipeline, helping discovery teams make better-informed choices about which molecules deserve precious laboratory time.

That distinction is important as AI becomes more common across health and biomedical research. We should not equate an explanation with evidence merely because it is visually persuasive. Explanations need to be challenged against known biology or chemistry, tested for stability and examined in the difficult cases where small changes matter most.

Australia’s national AMR response rightly emphasises surveillance, stewardship, infection prevention and responsible antimicrobial use. Drug discovery should complement those efforts. We would like to see evaluation of AI reasoning become routine in preclinical discovery, particularly when AI is used to prioritise costly experiments. Funders, research organisations and developers should ask not only for benchmark accuracy, but also for evidence that explanations are faithful, reproducible and useful for the decisions scientists actually make.

AI may help us search the chemical space faster, but speed is only valuable if it improves decisions. In antibiotic discovery, where the need is urgent and experimental resources are finite, the goal should not be to replace medicinal chemists with an algorithm. It should be to give them AI systems whose predictions — and limitations — can be interrogated before the next experiment begins.

Dr Abdulmujeeb T. Onawole is an Associate Research Assistant at The University of Queensland’s Institute for Molecular Bioscience. His research focuses on building trust-aware AI systems for molecular and biological applications.