Machine Learning Maps How Live and Heat-Killed Probiotics Shape Immune

Live probiotic bacteria caused much stronger immune responses than bacteria killed by heat in a new early-stage study. But the study also found something unexpected: the heat-killed bacteria produced clearer patterns showing how immune responses changed with dose. Published in Frontiers in Microbiology, the study was carried out by researchers at Probiotical Research in Novara, Italy. It questions three common beliefs in probiotic research: that probiotic products need at least one billion colony-forming units to work, that bacteria must settle and grow in the gut to be useful, and that live bacteria are always better than heat-treated bacteria. The researchers did not try to prove that probiotics improve health. Instead, they developed a method to test whether changes in immune activity at different doses can be measured reliably and whether live and heat-killed versions of the same bacterial strain behave differently when given in exactly matched cell amounts.

The researchers tested ten Gram-positive probiotic strains, labeled A through L, in both living and heat-treated forms. They used three concentrations: 100,000, 1 million, and 10 million cells per well. Each preparation was tested on human peripheral blood mononuclear cells (PBMCs), which are immune cells found in blood, from one healthy 58-year-old male donor. Each test condition was repeated three times, producing 180 treatment observations. The researchers measured 17 cytokines, which are proteins that help immune cells communicate, in the liquid surrounding the cells. Instead of using the usual method of counting colony-forming units, they counted cells using flow cytometry, following the ISO 19344:2015 standard. They used two fluorescent stains, thiazole orange and propidium iodide, to distinguish the types of cells being counted. They measured living bacteria in Active Fluorescent Units and heat-treated bacteria in Total Fluorescent Units. This allowed them to compare both preparations using the same unit: cells per well, whether or not the bacteria could still grow into colonies on a laboratory plate.

The difference between living and heat-treated bacteria was very large. For 16 of the 17 cytokines measured, living bacteria caused much stronger responses. The largest difference was in tumor necrosis factor alpha (TNF-alpha), where the average response to live bacteria was 310.8 times higher than the response to heat-treated bacteria. However, the 95 percent confidence interval ranged from 36.4 to 2,653 because many results for live bacteria exceeded the test's upper measurement limit. MIP-1beta showed a 69.4-fold difference, IL-8 a 25.8-fold difference, G-CSF a 15.5-fold difference, and IL-1beta a 13.3-fold difference. These results suggest that living bacteria broadly increased the activity of the body's early immune defenses, especially the activity of inflammation-related cytokines, rather than activating just one immune pathway. Responses linked to other parts of the immune system also increased. IFN-gamma, IL-2, and IL-12p70 rose about three- to four-fold, suggesting activation of Th1 immune responses. IL-4, IL-5, and IL-13 also increased, suggesting that Th2 responses were activated at the same time rather than the immune system shifting in only one direction.

One cytokine did not follow this pattern. MCP-1, a protein that attracts monocytes, a type of immune cell, showed no meaningful difference between living and heat-treated bacteria in the paired statistical analysis. The ratio of their average responses was 0.99, with a confidence interval from 0.89 to 1.12. This narrow range rules out differences greater than about 11 percent in either direction within the statistical model. The researchers treated this consistent result across all ten strains and three concentrations as an internal check on their method. MCP-1 can be triggered by parts of bacterial structures, such as peptidoglycan and lipoteichoic acid, even when bacteria are no longer alive or active. Its similar levels therefore suggest that both preparations contained comparable amounts of bacterial structural material. However, the researchers warned that one unchanged measurement cannot rule out other possible influences, such as differences in endotoxin levels or the breakdown of host cells. They did not measure these factors.

The clearest methodological finding concerned how immune responses changed as the dose increased. Heat-treated bacteria showed a stronger relationship between concentration and immune response than living bacteria for 13 of the 17 cytokines. The average Spearman correlation across strains reached 1.00 for IL-4 and 0.97 for both MIP-1beta and G-CSF. The researchers suggested that a hidden factor might explain this difference: living bacteria can multiply during the 24-hour experiment because the culture medium contains nutrients. Even the smallest starting dose may have grown considerably before the experiment ended, making the three starting doses less different in practice. A cautious estimate, assuming the bacteria doubled every 60 to 90 minutes, suggests that the largest starting dose could have increased by 16 to 256 times. The researchers did not count the bacteria again at the end of the experiment, and such measurements are rarely included in published studies of probiotic bacteria grown with immune cells. Therefore, this explanation remains a hypothesis. However, it suggests that tests comparing living bacteria at different starting doses may not actually compare bacterial amounts that remain 100 times apart.

Increasing the dose did not always increase immune activity. Six of the ten living strains showed a steady rise in cytokine production as the dose increased. Two strains showed a different pattern: their responses peaked at 1 million cells per well and then fell significantly at 10 million cells. In one of these strains, G-CSF production fell by 83 percent and IL-6 production fell by 25 percent at the highest dose compared with the peak. In the other strain, the decreases were 36 percent and 50 percent, respectively. The authors suggested several possible explanations that could work together. These included IL-10, which might act as a feedback signal to reduce NF-kappaB activity; immune receptors becoming less responsive after prolonged exposure to large numbers of bacteria; damage to host cells at high bacterial concentrations; and a drop in the acidity level of the culture medium caused by active bacteria. The researchers did not measure host-cell survival, cell damage, or pH. As a result, they could not tell which explanation, if any, was responsible. They reported the bell-shaped response patterns as observations, not as proven biological control mechanisms.

To make comparisons between living and heat-treated bacteria more useful, the researchers introduced a measure called the Functional Equivalence Dose, or FED70. It is the lowest concentration of heat-treated bacteria that produces at least 70 percent of the immune activity caused by the living version at the reference dose of 1 million cells. The researchers chose the 70 percent threshold because PBMC tests typically vary by 20 to 35 percent. They also adapted the 80 to 125 percent range used to assess bioequivalence in regulatory pharmacology. Among the 140 strain-and-cytokine combinations that could be evaluated, heat-treated bacteria met the equivalence threshold at the highest tested concentration in 80 cases, or 57.1 percent. Some individual strains showed much closer results. One heat-treated preparation produced an IL-6 response equal to 266 percent of the response to its living counterpart at the reference dose. However, one strain showed much weaker activity after heat treatment at every concentration tested. The authors cautioned that they had not included a positive control stimulus, which would have helped confirm that the donor's PBMCs were functioning properly on the day of the experiment.

Machine learning revealed patterns that were not clear from examining each cytokine separately. The researchers used unsupervised hierarchical clustering, a method that groups samples with similar patterns, to identify three groups of strains based on their cytokine responses. One group contained two strains that produced strong, broad inflammatory responses. A larger group had moderate responses with varied patterns over time. A third group contained two strains that produced weaker responses when alive. A Random Forest classifier, a machine-learning method, reproduced these groups with 83.3 percent consistency, compared with the 33.3 percent expected by chance. However, the authors stressed that this result measured how well the model recovered groups created from the same data, not how well it would work on new data. When they used a stricter test that left out one strain at a time, accuracy fell to 66.7 percent, which was still twice the chance level. An especially interesting result came from the ranking of important features. IL-12p70, IL-13, and IFN-gamma were the strongest markers for distinguishing strains, rather than the cytokines with the largest responses, such as IL-6 or IL-1beta. IL-13 ranked second even though its response differed relatively little between living and heat-treated bacteria. This shows that a measurement can be relatively unaffected by bacterial viability but still help identify a strain. Detecting such differences is one advantage of using machine learning to study many immune markers at once.

The authors clearly stated that all findings came from one donor and should be treated as starting points for further research, not as firm conclusions. All p-values describe results within this single experimental setting. The three groups of immune responses may reflect this donor's particular immune system. Before the FED70 thresholds or classification accuracy could be used for wider populations or regulatory decisions, the methods would need to be tested with at least five donors who differ in sex, age, and genetic background. Earlier research suggests that the relative ranking of bacterial strains may remain similar across donors even when the actual response levels differ. However, the authors also acknowledged several other limitations: no positive control, no measurements of host-cell survival, many results outside the test's measurement range that had to be estimated, and only three dose levels, which were not enough to fit formal drug-response curves. Despite these limits, the study provides a computational method that can be repeated and checked by other researchers. It combines standardized cell counting by flow cytometry, measurement of many immune markers, and machine-learning classification. This approach could help researchers measure dose responses more reliably and assess whether heat-treated bacteria have immune effects similar to those of live bacteria. However, studies involving multiple donors and, eventually, living organisms will be needed to determine whether these laboratory results predict meaningful health effects.

Subject of Research: Measuring immune responses to different doses of live and heat-treated Gram-positive probiotic strains in human peripheral blood mononuclear cells

Article Title: Beyond the billion: measuring immune responses at different doses and using machine learning to classify live and heat-treated Gram-positive probiotic strains in human peripheral blood mononuclear cells—a proof-of-concept study using one donor

Article References: Deusebio, G., Visciglia, A., Amoruso, A., & Pane, M. (2026). Beyond the billion: dose-response immunophenotyping and machine learning classification of live vs. heat-treated gram-positive probiotic strains in human peripheral blood mononuclear cells—a single-donor proof-of-concept study. Frontiers in Microbiology, 17, Article 1943891. https://doi.org/10.3389/fmicb.2026.1943891

Image Credits: AI Generated

DOI: 10.3389/fmicb.2026.1943891

Keywords: probiotics, postbiotics, PBMC, cytokines, dose-response, flow cytometry, machine learning, heat-treated bacteria, immunomodulation, TNF-alpha, Random Forest, Frontiers in Microbiology