Professor Anna Panorska and her former master’s student David Kweku (Statistics and Data Science), along with an interdisciplinary and international team of researchers, recently authored an article in which they were able to classify bacteria as virulent, avirulent or dead using an incredibly sensitive motion detector and a machine learning model Kweku trained.
, outlines the process researchers took to develop a method to classify whether bacteria are virulent, avirulent or dead based on their nanomotion signals. Avirulence, a state in which bacteria are unable to cause disease or spread in their host, can be characterized by distinct metabolic activity from that in virulent bacteria.
The need for new fast classification methods has become increasingly important. When a patient arrives at a medical center and it is determined that they have a bacterial infection, an effective antibiotic needs to be determined.
The chemical tests that are the standard way of testing whether a given bacterium is going to respond to a given antibiotic can take days, because the bacteria need to be cultured to get a large enough sample to determine which antibiotics are effective. Meanwhile, the bacterial infection in the patient could be spreading, so doctors typically go straight toward using a broad-spectrum antibiotic rather than one that targets bacteria more specifically. The samples needed for the proposed nanomotion testing of the response of bacteria to different antibiotics are very small and the testing is rapid.
Tiny motions can be a big deal in medicine
In 2013, one of the coauthors on the article, Sandor Kasas from the Laboratory of Biological Electron Microscopy in Switzerland, found a way to show that detection of nanoscale vibrations could indicate life. Researchers wanted to expand on this technique to see if motions could help identify bacteria as being virulent, avirulent or dead.
The technology used to measure oscillations involves placing an organism, like bacteria, onto a cantilever so small it is invisible to the naked eye and recording movement over time as new chemicals are introduced.
“The motion of the bacteria causes the motion of the cantilever,” Panorska said.
The movements are detected by atomic force microscopy, which is so precise it can detect opening and closing of ion channels in the organelle of a cell, just nanometers in size, which could indicate metabolic activity. Because it is so sensitive, it can be tricky to determine whether the motion of the cantilever is due to the organisms on it or whether it is from external environmental factors. Something as seemingly distant as a door closing in another part of a building can be picked up by the sensors. To address this, the measuring device is placed in an isolated environment immune to any motions from external environment.
The sample size needed to conduct the tests are so small, there isn’t a need for culturing the bacteria. The method described in the research paper allows medical experts to determine whether bacteria are reacting to a given antibiotic quickly, providing results for patients in a matter of hours rather than days and likely improving patient outcomes significantly.
Panorska and her colleagues, including her master’s student and the first author on the paper David Kweku, who has since graduated, developed a trained classification algorithm using two statistical methods, Random Forest and Model Based Clustering, to automatically classify the bacteria as avirulent, virulent or dead. The researchers collected five-minute nanomotion observations of the Bordetella pertussis, which causes whooping cough.
One of the important variables selected for training was the length of the time analyzed by the algorithm. Analyzing five minutes of data at a time is far too memory-intensive for most computers, so the researchers broke up the five-minute observation data into 60 five-second segments. They randomly selected 90 segments to train the model, 30 from dead bacteria observations, 30 from the avirulent bacteria and 30 from the virulent bacteria. The models were then checked for validity and consistency, and then tested on five-minute segments 200 times.
Using the trained classifier, the researchers achieved perfect classification results using both the Random Forest and Model Based Clustering methods. Panorska noted that the most errors in earlier approaches arose with the avirulent state where the bacteria are alive but not as active, which is expected. Nonetheless, the model ended up classifying the bacterial states correctly.
A field of opportunity
“This is a really fast-growing area of research with commercial applications,” Panorska said.
There are already researchers considering how this technology can be applied in other ways, including cancer care. Panorska added that the researchers are working with medical doctors who will be implementing these techniques and said that the closeness of collaboration is rewarding and informative for the scientists. She is also excited to be working on a project with direct applications.
“I am a statistician, so I mostly work on abstract things,” Panorska said. “Having the opportunity to work on such a project is really cool.”
Panorska also said she hopes to continue working on this project and maybe visit the locations where the experiments are being conducted to better understand the experimental setup.
“Sometimes it’s important for us to see the experiments to look for the sources of variability,” Panorska said. “Talking on Zoom is not the same as actually looking at how the experiment is performed.”
Panorska hopes that as machine learning becomes more useful in a wide range of disciplines that more students will consider taking data science classes. She and her colleague Emily Hand in the College of Engineering are working on developing new courses for applied data science.
“I think it’s important to showcase what people can do when they know some coding and they know some statistics,” Panorska said. “You can do really cool stuff! And the jobs are all over.”
