Artificial intelligence (AI) has seen a significant surge in interest recently. However, similar to humans, understanding the decision-making process of AI models is challenging. This is where Explainable AI (XAI) comes into play, offering insights into how these models reach their conclusions. Researchers are beginning to leverage XAI to analyze predictive AI models in depth, potentially leading to advancements in antibiotic development.
AI has gained immense traction lately, being used in various applications such as autonomous vehicles, grammar checking for emails, and even the design of new drugs. Nevertheless, figuring out how AI comes to its conclusions can be quite complex. Explainable AI (XAI), a branch of this technology, aims to shed light on the rationale behind AI decisions. Researchers are currently applying XAI not only to examine predictive AI models more thoroughly but also to delve deeper into chemical research.
The findings will be presented at the upcoming fall meeting of the American Chemical Society (ACS).
AI’s widespread applications have made it an essential part of modern technology. Nonetheless, many AI models function as black boxes, leaving uncertainty about the processes behind their outputs. This lack of transparency can create doubt, especially when dealing with potential drug molecules. “As scientists, we value the need for justification,” says Rebecca Davis, a chemistry professor at the University of Manitoba. “If we can develop models that illuminate how AI makes its choices, it could trust what scientists and the public think about these methods.”
One effective way to provide this clarity is through XAI. The machine learning algorithms within XAI allow us to peek behind the curtain of AI decision-making. While XAI can be beneficial in numerous contexts, Davis’ research specifically investigates its use in AI models for drug discovery, particularly for predicting new antibiotic candidates. Given the rigorous screening process thousands of potential drug candidates undergo just to approve one new medication—and the ongoing challenge of antibiotic resistance—accurate and efficient prediction models are essential. “I want to utilize XAI to gain insight into what information we need to teach computers about chemistry,” says Hunter Sturm, a graduate student in Davis’ lab who will present the findings at the meeting.
The research team began by inputting databases of known drug molecules into an AI model designed to forecast the biological effects of different compounds. They then employed an XAI model created by collaborator Pascal Friederich from Karlsruhe Institute of Technology in Germany to analyze which specific aspects of the drug molecules influenced the AI’s predictions. This approach elucidated why certain molecules were considered active or inactive by the model, helping Davis and Sturm comprehend what factors were deemed significant by the AI and how it categorized various compounds.
Through their work, the researchers discovered that XAI can uncover factors that humans might overlook, as it can process far more variables and data points simultaneously than a human mind. For example, while evaluating a collection of penicillin molecules, XAI revealed some unexpected findings. “Many chemists traditionally view the core of penicillin as central to its antibiotic actions,” explains Davis. “But XAI indicated the structures linked to that core as the key elements in its classification, rather than the core itself. This might explain why some derivatives of penicillin with that core exhibit limited biological activity,” she says.
Beyond identifying significant molecular features, the team hopes to enhance predictive AI models using XAI. “XAI reveals what the algorithms consider important in terms of antibiotic activity,” clarifies Sturm. “This information can then train the AI model on what to focus on,” Davis adds.
Next, the research team plans to collaborate with a microbiology lab to synthesize and evaluate some of the compounds predicted by the refined AI models to have antibiotic properties. In the long run, they aim for XAI to assist chemists in developing improved or even novel antibiotic compounds, which could be crucial in combating antibiotic-resistant pathogens.
“AI can lead to a lot of skepticism and concern among people. However, if we prompt AI to clarify its processes, it increases the chances of this technology being embraced,” Davis states.
Sturm believes that the incorporation of AI in chemistry and drug discovery signifies the future direction of the field. “We need to establish a solid groundwork. That’s what I aspire to accomplish.”
This research received funding from the University of Manitoba, the Canadian Institutes of Health Research, and the Digital Research Alliance of Canada.
