Explainable AI for Antibiotic Discovery

UQ researchers have developed a framework that tests whether AI can provide trustworthy explanations during antibiotic discovery, helping scientists design novel antibiotics for new treatments for drug-resistant superbugs.  

What's the problem? 

Antimicrobial resistance is one of the world's most pressing health challenges, creating an urgent need for new antibiotics to combat drug-resistant bacteria. While artificial intelligence is increasingly being used to accelerate drug discovery, many AI models operate as "black boxes", providing predictions without explaining the scientific reasoning behind them. This makes it difficult for researchers to trust AI recommendations when designing and optimising new antibiotic candidates.  

Understanding how an AI model predicts that a molecule may be effective is critical in medicinal chemistry, where small structural changes can have a significant impact on a drug's performance.  

What's UQ's innovation? 

Researchers from UQ's Centre for Superbug Solutions at the Institute for Molecular Bioscience, including Dr Abdulmujeeb Onawole, Dr Johannes Zuegg and Professor Mark Blaskovich, have developed a framework for evaluating the quality of eXplainable Artificial Intelligence (XAI) in drug discovery.  

The framework tests whether AI systems can correctly identify the most effective parts of a drug structure and explain complex cases, like "activity cliffs" where small chemical changes dramatically alter a compound's effectiveness. Rather than simply generating prediction scores, the approach helps determine whether AI interprets a chemical structure properly and how AI can find patterns in part of the structures, to provide meaningful explanations that support medicinal chemists in the design of novel molecules with improved activity.  decision-making.  

Using datasets of chemical compounds previously tested against the bacterial pathogen Staphylococcus aureus, the team evaluated three different AI models with fundamentally different approaches for molecular representation, machine learning algorithms and explanation methods. These included models designed to predict antimicrobial activity while also identifying molecular patterns and features influence those predictions.  

What's the impact? 

The research helps make AI models more transparent, adding interpretable outcome that help the chemists to evaluate if the prediction for a specific molecule is based on some chemical reasoning or is just a statistical apparition, make an AI model more trustworthy and useful for general drug development. 

The team's evaluation showed that while all three AI models were effective at identifying known antibiotic structures and achieved similar predictive performance, they differed significantly in their ability to explain what makes a molecule active or inactive. The framework provides a practical way to assess whether AI-generated explanations are reliable enough to support real-world drug development decisions.  

By helping scientists identify which explainable AI approaches can provide the most useful and translatable chemical insights, the framework will support a faster development and better optimisation of future drug candidates. In the longer term, it may contribute specifically to the discovery of new antibiotics needed to combat antimicrobial resistance and drug-resistant superbugs.  

Key researchers 

• Dr Johannes Zuegg 
  Centre for Superbug Solutions, Institute for Molecular Bioscience 
• Dr Abdulmujeeb Onawole 
  Centre for Superbug Solutions, Institute for Molecular Bioscience 
• Professor Mark Blaskovich 
  Centre for Superbug Solutions, Institute for Molecular Bioscience 

Published 15 September 2026

99 more AI innovations at UQ