Artificial intelligence accelerates discovery of next-generation disinfectants

Chemists and computer scientists have used AI to find new disinfectants to combat the growing threat of dangerous “superbugs.”

The Journal of Chemical Information and Modeling published a computational experimental framework for developing quaternary ammonium compounds (QACs) to kill bacteria.

This method yielded 11 new QACs that showed activity against antimicrobial-resistant bacteria.

We believe this is the first example of using AI to generate molecules for disinfectants. As an experimental chemist, I think it’s great to see machines help design new chemicals. ”

Bill Wuest, Emory chemistry professor and senior author of the study

“We have created an effective feedback loop between AI research, computational biochemistry, and experimental chemistry,” said Liang Zhao, Emory associate professor of computer science. “While we have proven that our concept works to generate QAC, we also believe that a wide range of scientific fields could benefit from this concept.”

Additional senior authors whose labs also contributed to the study include: Amalda Sheff, a computational biochemist at George Mason University in Virginia. and Kevin Minbior, an experimental chemist at Villanova University in Pennsylvania.

Arms race with microorganisms

Check the labels of any antibacterial cleaning products, from homes to hospitals, and you’ll likely see QACs listed among the ingredients. QAC is inexpensive, easy to create, and generally effective. As a result, they have been pioneers in disinfecting everything from kitchen counters to operating room floors for more than a century.

But while QACs remain relatively unchanged, bacteria continue to evolve, and in some cases evolve in ways that allow them to survive the onslaught of these cleaning agents. On a microscopic scale, it resembles an arms race. And recently, some microbes have begun to win, becoming dangerous “superbugs” that are resistant to QAC.

West and Minbiore, leading experts on the issue of rising antimicrobial resistance to disinfectants, are developing pioneering methods to modify QACs to ensure their efficacy. They fine-tune the structure of QAC molecules, synthesize these new designs, and test their ability to kill pathogens.

This is a painstaking and time-consuming process.

Zhao wondered if AI could help speed things up. He develops machine learning and artificial intelligence methods to advance scientific discovery and medical diagnostics.

“The design of new molecules has traditionally been done one at a time by humans in chemistry labs,” Zhao says. “But with AI models, we can create thousands of new designs at once.”

Mr. Zhao “knocked on the door of the building.”

After some back and forth, they decided to form a team that included Shehu and Minbiole’s labs to develop a computational experimental framework for QAC discovery.

The National Science Foundation funded this project.

Building a database

Most QACs consist of a nitrogen atom at the center of a four-carbon chain. In its simplest form, the positively charged head of the nitrogen center is attracted to the negatively charged phosphates of the fatty acids that surround the bacterial cell.

When QAC is anchored to a bacterial cell, the tip of the carbon chain acts like the tip of a spear, penetrating both the protective fatty membrane and the intracellular membrane, causing the bacteria to collapse.

Improper use of cleaning agents may be one factor in how bacteria evolve ways to evade the killing power of this QAC, West theorizes. He added that the increased use of cleaning products during the coronavirus pandemic may have created more opportunities for pathogens to develop resistance that are difficult to kill.

Over the past decade, Wuest and Minbiole have built a library of hundreds of new QAC molecules that their lab has designed, synthesized, and tested for toxicity against mammalian cells and antimicrobial activity against a variety of pathogens.

“This is the strongest dataset for QAC available anywhere,” West says.

A key part of what makes a dataset so powerful is its standardization, he added. Chemists followed consistent procedures for testing and classifying QACs and compiled results in a uniform manner.

graph problems

Developing an effective algorithm for this project was “challenging,” says Liang. “We had to customize the architecture of the model so that it could be mathematically designed to follow specific chemical rules.”

It was essentially a graph problem, he explains. The geometry of a molecule can be mapped onto a graph where atoms are treated as nodes and chemical bonds as edges.

The researchers divided the problem into a hierarchical two-stage generative process. One step concerns the nitrogen center of the QAC, and the other concerns the multiple tails of the molecule. I was then able to assemble these two pieces.

They extracted 603 molecules in the QAC dataset to train their algorithm.

This model generated approximately 300 molecular structures. The computer scientists sent these results to chemists for review. Team members decided to limit this review time to 4 hours to ensure that the approach being developed was practical.

Chemists applied systematic decision criteria such as QAC geometric compatibility, synthetic feasibility, structural novelty, and predictive ability of antimicrobial activity to access AI-generated molecules.

Analysis by their experts found that 9 percent of the molecules produced were suitable for synthesis and testing. More than half of the compounds, or 65%, were either not new or showed only incremental changes from existing compounds. 5% considered synthesis impractical.

Also, 21 percent of the molecules produced were invalid, of which 18 percent were not QACs and 3 percent were not any category of compounds.

Improvements in the method

In the second experimental workflow, the researchers enhanced the process.

They handpicked a library of 603 QAC compounds and retained only those that showed activity against four dangerous bacterial strains: Staphylococcus aureus, Enteroccus faecalis, Escherichia coli and Pseudomonas aeruginosa.

The curation yielded 421 compounds, which were used to retrain the AI-generated model.

The resulting 2,000 candidates underwent an automated structural validity check by a computer program to retain only chemically valid molecules. This filter reduced the number of candidates to 800.

These 800 molecules were further filtered using a computational classifier that predicted activity against each of the four bacterial strains. The classifier rated each molecule from 1 to 4 depending on the number of strains in which the molecule was predicted to be active.

This filtration process yielded 300 top-ranked compounds, which the lab’s chemists focused on evaluating within a four-hour time limit.

This second experimental workflow significantly improved the results. Invalid output decreased from 21 percent to zero, and compounds deemed worthy of synthesis increased from 9 percent to 38 percent.

Testing the results

Chemists in the lab selected and synthesized 29 molecules generated from both workflows and tested their efficacy. The results yielded 11 novel QACs whose ability to inhibit bacterial pathogens was experimentally verified.

“One of these QACs stands out in particular in that it exhibits broad-spectrum activity against all seven bacterial strains we tested,” West says. “This includes Gram-negative bacteria, which are the most difficult to kill.”

Gram-negative bacteria cells have two membranes that make it difficult for molecules to penetrate, he explains.

The research has already attracted interest from the private sector as a potential model to accelerate the discovery of new, more effective disinfectants, West said.

The framework also serves as a model for scientists in other fields on how to collect and standardize datasets for potential AI applications.

“On the one hand, this study gives us a huge list of lead compounds to study,” West says. “We are having more undergraduate students synthesize and test the resulting compounds. This is good training for them and could lead to more discoveries.”

The paper’s lead author is Shiva Ghaemi, a doctoral student at George Mason University. Co-authors include Bo Pan, an Emory doctoral student in the Liang lab. and Alice Wu and Elise Bezold from the Wuest lab. Additional authors are Amanda Consylman, Ashley Petersen, Gabe Chang, Alice Wu, and Diana McDonough, all from Villanova University, and Mark Forman from St. Joseph’s University in Philadelphia.