Researchers from the University of Manitoba used Explainable AI (XAI) to analyze AI predictions of potential antibiotics.
Using AI to peer deeper into the field of chemistry
Watch a short Q&A video about this research. Video Credit: The American Chemical Society
XAI revealed insights previously missed by humans, like the importance of structures surrounding penicillin's core. XAI-guided improvement of AI models could lead to better antibiotic discovery and combat antibiotic resistance. Ultimately, XAI could foster trust in AI and revolutionize drug development. The researchers will present their findings at the American Chemical Society (ACS) fall meeting.
The popularity of artificial intelligence (AI) has skyrocketed. It fuels models that assist in driving cars, editing emails, and even creating novel drug compounds. It is difficult to read an AI's mind, though, just like a human can.
A subset of the technology known as explainable AI (XAI) may be able to assist humans in defending a model's choices. Furthermore, scientists are currently using XAI to explore the realm of chemistry and closely examine predictive AI models.
AI has so many applications that it is practically a given in today's technological environment. But many AI models are opaque, so it is unclear precisely what processes are involved in generating a result. Not knowing the steps could make experts and the general public skeptical, especially when that result is something like a possible medication molecule.
As scientists, we like justification. If we can come up with models that help provide some insight into how AI makes its decisions, it could potentially make scientists more comfortable with these methodologies.
Rebecca Davis, Professor, Department of Chemistry, University of Manitoba
Using XAI is one method of presenting that rationale. These machine learning algorithms can give insight into AI decision-making's inner workings. Davis' research focuses on using XAI to AI models for drug development, such as those that predict new antibiotic candidates, even though it can be applied in many different situations.
Accurate and effective prediction models are essential, as thousands of candidate molecules must be examined and rejected before one new treatment can be approved. Antibiotic resistance constantly jeopardizes the effectiveness of already-approved medications.
I want to use XAI to better understand what information we need to teach computers chemistry.
Hunter Sturm, Graduate Student, University of Manitoba
Strum is a student from Davis’ lab who is presenting the work at the meeting.
First, the researchers fed databases of well-known pharmacological compounds into an artificial intelligence (AI) algorithm designed to forecast a compound's likelihood of producing a biological effect.
Then, they examined the precise components of the medication compounds that contributed to the model's prediction using an XAI model created by partner Pascal Friederich at Germany's Karlsruhe Institute of Technology.
This made it easier for Davis and Sturm to comprehend what an AI model may consider significant and how it establishes categories after looking at many different chemicals. According to the model, it also helped explain why a given molecule had activity.
XAI can analyze many more variables and data points simultaneously than a human brain, and researchers discovered that it can identify things that humans might have overlooked. For instance, the XAI discovered something intriguing when examining a collection of penicillin compounds.
Many chemists think of penicillin’s core as the critical site for antibiotic activity. But that is not what the XAI saw.
Rebecca Davis, Professor, Department of Chemistry, University of Manitoba
Rather than the core itself, it determined that the structures connected to it made it essential for classification.
“This might be why some penicillin derivatives with that core show poor biological activity,” explained Davis.
Apart from pinpointing significant molecular structures, the researchers aim to leverage XAI to enhance predictive artificial intelligence models.
“XAI shows us what computer algorithms define as important for antibiotic activity,” explained Sturm.
Davis added, “We can then use this information to train an AI model on what it is supposed to be looking for.”
The group will then collaborate with a microbiology lab to create and evaluate a few substances the enhanced AI models indicate have potential as antibiotics. Ultimately, they anticipate that XAI will assist chemists in producing new or improved antibiotic molecules, which may help halt the spread of infections that are resistant to antibiotics.
“AI causes a lot of distrust and uncertainty in people. But if we can ask AI to explain what it is doing, there is a greater likelihood that this technology will be accepted,” said Davis.
Sturm continues, saying he believes AI applications in drug development and chemistry will shape the sector going forward. “Someone needs to lay the foundation. That is what I hope I am doing,” Sturm says.
This study was supported by the University of Manitoba, the Canadian Institutes of Health Research, and the Digital Research Alliance of Canada.