When Dr. Abdulmujeeb Onawole looks at a petri dish growing drug-resistant bacteria, he sees a race against time. Antimicrobial resistance is edging closer to rendering many antibiotics useless, leaving doctors with fewer tools to fight life-threatening infections. Now, a team at the University of Queensland has developed a framework that could help change that trajectory—not by replacing scientists, but by making AI a more trustworthy partner in the hunt for new medicines.
The problem, Onawole explains, is that AI can suggest promising drug candidates, but it often can't explain why. Researchers call this the "black box" of artificial intelligence—a system that provides answers without showing its work. For antibiotic development, where a misleading recommendation can waste months of laboratory effort and millions of dollars, that lack of transparency is a significant barrier.
"AI is revolutionizing drug development, but scientists struggle to trust its recommendations because the technology often can't explain its reasoning," Onawole said. "This lack of transparency is dangerous during antibiotic development, as misleading AI explanations can lead to incorrect decisions and wasted resources in the lab."
Working with colleagues at UQ's Center for Superbug Solutions, including Dr. Johannes Zuegg, Onawole designed a framework to evaluate whether AI systems can provide reliable chemical explanations during drug discovery. The team built three AI models using datasets of chemical compounds previously tested against Staphylococcus aureus—a bacterium responsible for difficult-to-treat infections in hospitals and communities. They then assessed how well each model could identify important drug structures and interpret what researchers call "activity cliffs," situations where tiny chemical modifications dramatically alter a drug's effectiveness.
The findings, published in the Journal of Cheminformatics, revealed that while all three models excelled at recognizing known antibiotic structures, they varied considerably in their ability to explain what made certain molecules work against superbugs. That variance matters. By providing a way to measure whether an AI system's reasoning is sound, the framework gives researchers a tool to separate genuinely useful recommendations from statistical noise.
"We have shown our framework can successfully assess whether AI systems can provide trustworthy chemical explanations, which is critical to medical chemists in drug development," Zuegg said. "This is an important step toward speeding up the integration of AI into antibiotic research."
For patients and clinicians confronting drug-resistant infections, this work represents cautious optimism. Antibiotic development has long been slow and expensive, but AI offers a path to accelerate the process—if scientists can trust what they're seeing. The Queensland team's framework doesn't promise instant solutions, but it lays groundwork for AI to become a more reliable collaborator in one of medicine's most urgent challenges. As Onawole put it, "Longer term, this could contribute to the faster discovery of new antibiotics to combat drug-resistant superbugs."