The experiment was fully factorial. Ten Gram-positive probiotic strains, coded A through L, were tested in both viable and heat-treated forms across three concentrations — one hundred thousand, one million, and ten million cells per well — against human peripheral blood mononuclear cells from a single healthy 58-year-old male donor. Each condition was run in technical triplicate, yielding 180 treatment observations, and the culture supernatants were screened for 17 cytokines using a multiplex suspension array. Critically, the researchers abandoned conventional colony-forming-unit counting in favor of flow cytometric enumeration under the ISO 19344:2015 standard. Using thiazole orange and propidium iodide dual staining, they quantified viable bacteria as Active Fluorescent Units and heat-treated bacteria as Total Fluorescent Units, allowing both preparation types to be compared in identical physical units of cells per well, independent of whether the cells could still form colonies on a plate.

The viability effect was enormous. Across 16 of the 17 cytokines measured, viable preparations induced substantially greater responses than heat-treated ones. Tumor necrosis factor alpha showed the largest difference, with live bacteria producing a geometric-mean fold-change 310.8 times higher than heat-treated preparations, though the 95 percent confidence interval stretched from 36.4 to 2,653 because the analyte exceeded the assay’s upper quantification limit in many viable conditions. MIP-1beta followed at 69.4-fold, IL-8 at 25.8-fold, G-CSF at 15.5-fold, and IL-1beta at 13.3-fold. The pattern suggested broad amplification of innate immune signaling rather than selective pathway activation, with pro-inflammatory cytokines disproportionately affected. Adaptive responses shifted too: IFN-gamma, IL-2, and IL-12p70 rose roughly three- to four-fold, indicating Th1 activation, while IL-4, IL-5, and IL-13 also climbed, pointing to parallel Th2 activation rather than polarization in a single direction.

One cytokine broke the pattern. MCP-1, a monocyte chemoattractant, showed no significant difference between viability states under the paired statistical analysis, with a geometric-mean ratio of 0.99 and a confidence interval of 0.89 to 1.12 — a bounded null excluding any difference larger than roughly 11 percent in either direction. The researchers interpreted this invariance across all ten strains and three concentrations as an internal methodological control. MCP-1 can be triggered by bacterial structural components such as peptidoglycan and lipoteichoic acid regardless of whether the bacteria are metabolically active, so its consistent secretion suggests both preparation types delivered comparable structural material. The team cautioned, however, that a single invariant analyte cannot formally rule out confounders such as endotoxin content or host-cell lysis, which were not measured.

The most methodologically striking finding concerned dose-response shape. Heat-treated preparations showed stronger concentration-response correlations than viable ones for 13 of the 17 cytokines, with mean within-strain Spearman correlations reaching 1.00 for IL-4 and 0.97 for MIP-1beta and G-CSF. The researchers attribute this asymmetry to a hidden variable: live bacteria proliferate during the 24-hour co-culture in nutrient-rich medium. Even the lowest inoculum may have expanded substantially before the experiment ended, compressing the effective difference between the three nominal dose levels. A conservative estimate based on one bacterial doubling every 60 to 90 minutes suggests the highest inoculum could have expanded 16- to 256-fold. Because endpoint bacterial enumeration was not performed — and is rarely performed in any published probiotic co-culture study — this explanation remains a hypothesis, but it implies that comparing live bacteria at supposedly different doses may not be comparing a 100-fold range at all.

Dose escalation also did not uniformly increase immune activity. Six of the ten viable strains showed monotonically increasing cytokine output with concentration, but two strains displayed bell-shaped kinetics, peaking at one million cells per well and then attenuating significantly at ten million. For one of these strains, G-CSF output dropped 83 percent and IL-6 dropped 25 percent at the highest dose relative to the peak; the other showed reductions of 36 and 50 percent. The authors propose several non-exclusive mechanisms, including IL-10-mediated autocrine feedback suppressing NF-kappaB signaling, receptor desensitization under sustained high-density stimulation, host-cell cytotoxicity at high bacterial density, and medium acidification by metabolically active bacteria. Because no host-cell viability, cytotoxicity, or pH measurements were collected, the study cannot distinguish among these explanations, and the bell-shaped profiles are reported as observed dose-response shapes rather than established regulatory phenomena.

To make live-versus-killed comparisons operationally useful, the team introduced the Functional Equivalence Dose, or FED70 — the minimum heat-treated concentration at which a preparation achieves at least 70 percent of the immunological activity of its viable counterpart at the one-million-cell reference dose. The 70 percent threshold reflects typical PBMC assay variability of 20 to 35 percent and adapts the 80 to 125 percent bioequivalence range used in regulatory pharmacology. Across the 140 evaluable strain-by-analyte combinations, heat-treated preparations reached this equivalence criterion at the highest tested concentration in 80 cases, or 57.1 percent. Some individual strains converged dramatically: one heat-treated preparation reached 266 percent of its viable counterpart’s IL-6 response at the reference dose. One strain, however, showed strongly attenuated activity in heat-treated form across all concentrations, and the authors note this observation carries the caveat that no positive control stimulus was included to formally verify PBMC functional competence on the experimental day.

Machine learning added a layer that univariate analysis could not reveal. Unsupervised hierarchical clustering of strain-level cytokine fingerprints identified three immunological phenotype clusters: one pair of strains producing high-magnitude, broad-spectrum pro-inflammatory responses; a larger group with moderate, kinetically diverse profiles; and a pair with low-magnitude viable responses. A Random Forest classifier recovered this internally derived partition with 83.3 percent consistency against a 33.3 percent chance level, though the authors stress this measures internal recovery, not generalization, since the clusters came from the same data. Under the stricter leave-one-strain-out cross-validation, accuracy fell to 66.7 percent — still double chance. More intriguing was the feature importance ranking: IL-12p70, IL-13, and IFN-gamma, not the high-amplitude cytokines IL-6 or IL-1beta, were the primary discriminators of strain identity. IL-13 ranked second despite carrying the second-smallest viability effect in the panel, showing that an analyte can be viability-insensitive yet strain-informative — precisely the kind of dissociation multiplex machine-learning analysis is designed to detect.

The authors are explicit that every finding rests on a single donor and must be treated as hypothesis-generating. All p-values are descriptive within this one experimental context, the three phenotype clusters reflect one immune background, and the FED70 thresholds and classification accuracies would require validation across at least five donors of diverse sex, age, and genetic background before any population-level or regulatory use. Prior work suggests relative strain rankings tend to be preserved across donors even when absolute values differ, but the team also acknowledges the missing positive control, the absence of host-cell viability readouts, substantial out-of-range values requiring imputation, and the three-point dose range that precludes formal pharmacological curve fitting. Within those limits, the study delivers a computationally reproducible framework — flow cytometric standardization, multiplex immunophenotyping, and machine-learning classification — that could underpin evidence-based dose characterization and postbiotic equivalence assessment, provided multi-donor and eventually in vivo validation confirms that these laboratory signatures predict anything clinically meaningful.

Subject of Research: Dose-response immunophenotyping of live versus heat-treated Gram-positive probiotic strains in human peripheral blood mononuclear cells

Article Title: 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

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