With their legendary sense of smell, dogs are adept at identifying the characteristic scents of cancers from breath, urine, and poop. But with trained cancer-sniffing pups in short supply, animals are unlikely to become widely available for routine diagnostics. Instead, Andreas Mershin wants man’s best friend to teach machine learning algorithms to sniff out diseases, and he plans to put this technology into your pocket. Mershin, a research scientist at the MIT Center for Bits and Atoms, says his eventual goal is to build electronic nose capability into smartphones.
The detection of a cancer signal by electronic noses isn’t a new concept, but those that have been developed so far still can’t match the accuracy of dog’s, says Mershin. To get closer to that ability, Mershin and his interdisciplinary team establish a proof-of-concept method for the integration of canine olfaction with machine odor analysis of prostate cancer in a study published February 17 in PLOS ONE.
Prostate cancer is the second most common cancer in males, affecting an estimated one in nine men at some point in their lives. A widely used tool for disease detection is prostate-specific antigen testing, but the test often fails to detect the disease or leads to incorrect diagnoses. In the hunt for better diagnostic options, researchers have searched for olfactory biomarkers of prostate cancer in the chemical bouquet of urine samples. One team was able to detect prostate cancer by analyzing urine odors with about 86 percent accuracy. The idea of using dogs to detect cancers was first proposed for melanomas in 1989, and since then, canines’ cancer-detecting skills have often outshone machine-based odor analysis—in one 2015 study, disease-sniffing pups detected prostate cancer from urine samples with 98–99 percent accuracy.
Mershin tells The Scientist that he was struck not only by dogs’ disease-sniffing prowess but by the fact that some pups, trained to detect a certain type of cancer, are able to detect other malignancies, despite low similarity in odors among various cancers. Some untrained pets have even detected cancer in their owners. “[Dogs] don’t go by the list of molecules. . . . They go by the scent character, which means they somehow figure out the cancer essence,” says Mershin. “That blew my mind. No analytical tool to this day can do this because it’s looking at the list of ingredients. Knowing what something is made of isn’t the same as knowing what it smells of.”
Inspired by canines, Mershin and his colleagues sought to develop artificial intelligence that emulated doggie decisions. “The specific question we try to answer [in the study] is, what are the hurdles and challenges in taking the dog’s nose and its functionality and plugging it into your smartphone?” says Mershin, who leads MIT’s Label-Free Research Group, named for its disregard of boundaries between scientific disciplines.
For their study, the researchers obtained urine from 12 men with biopsy-confirmed high-grade Gleason 9 prostate cancer and 38 men who had negative biopsies. Part of the urine specimens were sent to Medical Detection Dogs in the UK for diagnoses by Florin, a four-year-old female Labrador, and Midas, a seven-year-old female Wirehaired Vizsla. After training the animals with 5 cancer and 15 noncancer samples, the researchers used the remaining samples to test Midas’s and Florin’s skills. At each testing run, the dog examined a carousel containing three cancer-negative samples and one cancer-positive sample. After getting a whiff of each container of urine, the dog made a selection—Florin indicated a positive sample by standing and staring, whereas Midas sat in front of her choice. A correct choice earned the pup a well-deserved treat.
Both dogs accurately identified five out of seven prostate cancer samples. Out of 21 cancer-negative samples, Florin made the right decision 16 times and Midas 14 times. Overall, the dogs showed 71 percent sensitivity and 70–76 percent specificity. Mershin says the main reason for Midas’s and Florin’s moderate accuracy was because they received limited training, due to the limited number of urine samples available. With such a small number of test samples, Santiago Marco, a physicist who studies data analysis of sensors at the Institute for Bioengineering of Catalonia and the Barcelona Institute for Science and Technology, is not convinced that the dogs’ choices were made on the basis of cancer detection. “It’s not clear that they have very strong support for the claim that the canine olfaction is sufficiently sensitive or even specific,” says Marco, who was not involved in the study.
Mershin says that with additional training, the animals’ skills would have improved. “We weren’t trying to make these dogs go to 99 percent—which we can. Many dogs have been trained to 99 and even 99.8 percent accuracy with COVID and malaria and Parkinson’s and various cancers.” Given the study’s goal of identifying the feasibility of the group’s machine learning approach, Mershin says that the dog’s level of precision was adequate.
From canine to computer
The researchers collected the volatile compounds that make up each of the urine specimens’ aroma and analyzed these chemicals using gas chromatography–mass spectrometry (GC-MS). They found that the amounts of several volatile compounds were elevated or reduced in the cancer samples, but the distinguishing chemicals were different from those identified by other studies. Mershin says that this apparent lack of consistent biomarkers for prostate cancer highlights the problem of disease diagnoses based on particular compounds or sets of compounds. Rather than basing their decisions on specific chemicals, dogs identify something that’s “cancery” about a sample, he says.
As a first step toward developing electronic noses with similar capability, the team used the dogs’ diagnoses to train a type of artificial intelligence called an artificial neural network (ANN) to evaluate the volatile chemicals detected from the urine by GC-MS. The ANN detected cancer samples with high accuracy, says Stephen Thaler, the president and CEO of Imagination Engines and a coauthor of the study. But due to the small sample size, the researchers say their results need to be validated with a larger experiment.
“It’s really interdisciplinary what they’re doing and that’s kind of cool,” says Marc Aubreville, a professor of image understanding and medical application of artificial intelligence at Technische Hochschule Ingolstadt in Germany who didn’t contribute to the study. “I think the method is sound in what they did, but I would really have loved to see more samples and better proof that this really works with machine learning.”
Mershin says the team’s eventual goal is to apply its canine-trained machine algorithm to an electronic nose that contains synthetic analogs of animal olfactory receptors that they have patented. But before this tool is ready for smartphones, they need to use many more samples to boost the dogs’ cancer-detecting accuracy and then train the ANN to match this performance.
Amanda Siegel, an analytical and biophysical chemist at Integrated Nanosystems Development Institute (INDI) who did not participate in the work, says the research is promising but notes that the group only looked at very sick subjects. “They’re all Gleason 9, so that’s significant amounts of disease in the prostate. I think that’s a good first step. But ultimately, they’re going to need to look at samples with lower Gleason scores [for the tool] to be more useful.”
C. Guest et al., “Feasibility of integrating canine olfaction with chemical and microbial profiling of urine to detect lethal prostate cancer,” PLOS ONE, doi:10.1371/journal.pone.0245530, 2021.