A human milk oligosaccharide alters the microbiome, circulating hormones, and metabolites in a randomized controlled trial of older adults

Summary

Aging-related immune dysfunction is linked to cancer, atherosclerosis, and neurodegenerative diseases. This 6-week randomized controlled trial evaluated whether 2′-fucosyllactose (2′-FL), a human breast milk oligosaccharide with established benefits in infants and animal models, could improve gut microbiota and immune function in 89 healthy older adults (mean age 67.3 years). While the primary endpoint of cytokine response change was not met, 2′-FL supplementation increased gut Bifidobacterium levels and elevated serum insulin, high-density lipoprotein (HDL) cholesterol, and FGF21 hormone. Bifidobacterium “responders” experienced additional metabolic and proteomic changes and also performed better on a cognitive test of visual memory. Nonresponders were more likely to lack Bifidobacterium in their gut microbiota at the start of the intervention. Multi-omics analysis indicated a systemic response to 2′-FL, which could be detected in blood and urine, showcasing the potential of this prebiotic to provide diverse benefits for healthy aging. This trial was registered at ClinicalTrials.gov (NCT03690999).

Graphical abstract

Highlights

• •2′-fucosyllactose (2′-FL) stimulates Bifidobacterium growth in older adults
• •2′-FL increases levels of HDL cholesterol, insulin levels, and FGF21 in this cohort
• •Bifidobacterium responders show greater metabolic and inflammatory changes

Carter et al. show that 2′-fucosyllactose, a human milk oligosaccharide, increases Bifidobacterium in older adults' gut microbiomes, associated with elevated HDL cholesterol, insulin, and FGF21 hormone. Microbiome “responders” showed improved visual memory, suggesting that milk oligosaccharides offer benefits beyond infancy by supporting healthy aging through microbiome modulation.

Introduction

Aging is associated with the broad decline of biological and cognitive function and the progression of diseases such as atherosclerosis and cancer.^1,2,3 Central to this process of deterioration is immunosenescence, which is characterized by both chronic inflammation and hyporesponsiveness to stimuli.⁴ Combating the aging process will require the identification of therapeutic modalities that can counter immune system deterioration.

It is increasingly recognized that an important modifier of metabolic and immune system function is the trillions of microorganisms that comprise the human gut microbiota, or microbiome.⁵ Furthermore, microbiota-accessible carbohydrates (MACs) have long been realized as potent modulators of the gut microbiota. Consumption of these complex carbohydrates by the microbiota not only results in the production of beneficial metabolites but also may serve to properly tune the systemic immune system.⁶ Recent clinical trials investigating this premise have shown that dietary fiber supplementation in human cohorts results in heterogeneous, personalized responses that often but do not always result in a benefit to the host.^7,8,9,10 Previous work has also demonstrated that our microbiome changes as we age and that these changes are potentially implicated in immunosenescence.^11,12

A type of MAC that remains under-explored with regards to interventional studies in adult humans is the human milk oligosaccharide (HMO). HMOs are the third most abundant constituent of human breast milk after lactose and lipids. They act as a key source of nutrients for key members of the developing infant gut microbiota, namely species of the genus Bifidobacteria.¹³ While HMOs cannot be directly metabolized by humans, they have recently been shown to have direct interactions with host cells independent of the gut microbiota.^14,15,16 Clinical and pre-clinical studies suggest that HMOs have pleiotropic benefits for the host that include fostering a healthy infant gut microbiota that protects against enteropathogens,^17,18 training the immune system^19,20 and enhancing cognitive development.^21,22 When combined with the potential for microbiota-dependent (prebiotic) and microbiota-independent (direct impact on host cells) effects, we hypothesized that HMOs could provide a benefit to older adults. In sum, our study tests whether human biology and gut microbes possess extant pathways that would enable response to these molecules long after weaning. The capacity for HMOs to act as immune modulators in adults, especially aging adults who suffer from the sequelae of immunosenescence, requires detailed study.

Results

A randomized, placebo-controlled clinical trial to study the effects of 2′-FL on the microbiota and host immune system

We performed the RAMP Study (Rejuvenating the Aging Microbiota with Prebiotics; ClinicalTrials.gov ID: NCT03690999) to investigate the effects of the HMO 2′-fucosyllactose (2′-FL) (Figure 1A) on the gut microbiota, immune system, and metabolism of older individuals. We recruited a cohort of 89 individuals (mean age = 67.4 years, minimum age = 60 years, maximum age = 83.9 years) and randomized them to one of three arms: a High Dose 2′-FL arm where subjects consumed 5 g 2′-FL per day (n = 29), a Low Dose 2′-FL arm where subjects consumed 1 g 2′-FL per day (n = 30), and a placebo arm where subjects consumed a glucose product free of 2′-FL (n = 30; Figures 1B and 1C; Table 1; STAR Methods). The treatment period lasted for 6 weeks; 32 participants began the intervention after the start of the COVID-19 pandemic, resulting in modest deviations from the study protocol—primarily extending the duration of the protocol over the time that blood sampling facilities were locked down during the pandemic (these protocol deviations did not substantively affect the results of our study, see Methods S1 and Table S1). We observed subjects starting 2 weeks before the intervention (pre-baseline) as well as for 4 weeks after the intervention (washout). Stool and blood samples were collected at weeks −2, 0, 3, 6, and 10 (Figure 1D). We also collected anthropometrics and cognitive testing data at each of these time points. Urine was collected at week 6.

Figure 1.

Design of a randomized controlled trial in healthy, older human subjects

(A) The chemical structure of 2′-FL. 2′-FL is composed of lactose (purple) covalently linked to a fucose (orange) via a glycosidic linkage (green).

(B) A Consolidated Standards of Reporting Trials (CONSORT) flow diagram of the individuals recruited and screened for this study.

(C) The treatment arm allocation of the 89 individuals who were randomized to a treatment arm and received the product. The three arms were placebo, Low Dose 2′-FL (consuming 1 gram of 2′-FL per day; 0.5 g morning and evening), and High Dose 2′-FL (consuming 5 grams of 2′-FL per day; 2.5 g morning and evening).

(D) Trial timeline (top) and sample collection schema (bottom). Each subject was tracked for a 2 weeks of baseline period; supplementation began on week 0 and continued for 6 weeks; participants were observed for a 4 week washout period after cessation of supplementation. Blood and stool were collected at each time point; urine was collected at the week 6 time point. Subjects also took cognitive tests at each time point.

Table 1.

Demographic information of participants

		Placebo	Low Dose 2FL	High Dose 2FL	Combined
Sex, no. (%)	women	20 (66.7)	17 (56.7)	12 (41.3)	49 (55.1)
	men	10 (33.3)	13 (43.3)	17 (58.6)	40 (44.9)
Age, mean (SD), y		68.8 (7.4)	66.7 (3.8)	66.5 (4.3)	67.3 (5.4)
Highest level of education achieved, no. (%)	some college	4 (13.3)	2 (6.7)	0 (0)	6 (6.7)
	college graduate	11 (36.7)	6 (20.0)	8 (27.6)	25 (28.1)
	some post-graduate school	3 (10.0)	3 (10.0)	5 (17.2)	11 (12.4)
	post-graduate degree	12 (40.0)	19 (63.3)	15 (51.7)	46 (51.7)
	not reported	0 (0)	0 (0)	1 (3.4)	1 (1.1)
Race/ethnicity, no.	White	26 (86.7)	24 (80.0)	21 (72.4)	71 (79.7)
	Hispanic/Latinx	1 (3.3)	2 (6.7)	1 (3.4)	4 (4.5)
	Asian	3 (10.0)	4 (13.3)	4 (13.8)	11 (12.3)
	Black/AA	0 (0)	0 (0)	1 (3.4)	1 (1.1)
	other	0 (0)	0 (0)	2 (6.9)	2 (2.2)
Weight, mean (SD), kg	women	71.8 (10.9)	68.5 (16.8)	73.4 (14.2)	71.2 (14.0)
	men	69.7 (11.5)	71.8 (9.8)	72.8 (14.5)	71.4 (11.9)
	both sexes	71.1 (10.9)	69.9 (14.1)	73.0 (14.2)	71.3 (13.0)
Body mass index, mean (SD)	women	25.6 (3.2)	24.1 (4.3)	25.0 (2.8)	24.9 (3.4)
	men	24.8 (2.3)	24.4 (2.4)	24.4 (4.0)	24.5 (2.9)
	both sexes	25.3 (2.9)	24.2 (3.5)	24.6 (3.5)	24.7 (3.3)
Blood pressure, mean (SD), mm Hg	systolic	127.4 (15.5)	124.1 (17.6)	127.3 (14.3)	126.2 (15.8)
	diastolic	72.5 (8.7)	73.5 (11.6)	72.3 (11.1)	72.8 (10.5)
Blood lipid level, mean (SD), mg/dL	total cholesterol	209 (41)	199 (40)	194 (40)	201 (40)
	HDL cholesterol	64 (17)	65 (19)	63 (19)	64 (18)
	LDL cholesterol	127 (35)	116 (36)	114 (33)	119 (35)
	triglycerides	88 (29)	92 (44)	89 (42)	90 (38)
Fasting glucose, mean (SD), mg/dL		95 (9)	96 (17)	92 (10)	94 (12)
Fasting insulin, mean (SD), μIU/mL		8.9 (3.2)	8.7 (4.4)	9.7 (3.7)	9.1 (3.8)

Consumption of 2′-FL by adult subjects is well tolerated and leads to detectable 2′-FL in urine and plasma

Consumption of 2′-FL was well tolerated and was not associated with any adverse events, such as gastrointestinal distress or nausea (Figure S1). To assess adherence to trial protocols and potential for systemic effects of 2′-FL, we performed targeted mass spectrometry of 2′-FL in urine (n = 66 samples) and plasma (n = 89 samples) of each participant at the week 6 time point (the lower number of urine samples available was due to COVID-related challenges of sample collection). Indeed, we detected 2′-FL in the plasma of 62% of the individuals in the High Dose 2′-FL arm (mean = 0.066 μg/mL, standard deviation [SD] = 0.067 μg/mL; lower limit of detection [LOD] = 0.02 μg/mL) and 13.3% of the individuals in the Low Dose 2′-FL arm (mean = 0.021 μg/mL, SD = 0.0083 μg/mL; LOD = 0.02 μg/mL) compared to 0% of the individuals in the placebo arm (mean = 0.02 μg/mL, SD = 0; LOD = 0.02 μg/mL). We also detected 2′-FL in urine in 90.4% of the individuals in the High Dose 2′-FL arm (mean = 5.05 μg/mL, SD = 6.55 μg/mL, LOD = 0.2 μg/mL) and 52.1% of the Low Dose 2′-FL arm (mean = 0.7 μg/mL, SD = 1.07 μg/mL, LOD = 0.2 μg/mL) compared to 31.8% of the placebo arm (mean = 0.27 μg/mL, SD = 0.20 μg/mL, LOD = 0.2 μg/mL). The relatively low levels of 2′-FL detection in the placebo groups in plasma and urine suggests that these subjects are not consuming 2′-FL in their diet at appreciable levels. Previous studies have noted an absence of fucosylated lactose in the milk of domesticated animals whose milk may be consumed in diet.²³ The High Dose 2′-FL group had significantly higher concentrations of 2′-FL compared to the placebo group for both urine and plasma (p = 0.003 and p = 9.6 × 10⁻⁴, respectively, Student’s t test) (Figures 2A and 2B). We also found a strong correlation between the 2′-FL concentrations found in urine and plasma (Figure 2C, Pearson’s r = 0.729, p = 1.9 × 10⁻⁸). Additionally, while systemic circulation of HMOs have been demonstrated in infants,²⁴ lactating mothers,²⁵ and animal models,²⁶ the mechanism of translocation from the lumen of the gastrointestinal tract to the bloodstream is unclear. In vitro data suggest that paracellular transport and/or receptor-mediated transcytosis may be involved.²⁷

Figure 2.

2′-FL is detectable in plasma and urine of the High Dose 2′-FL group and results in significant increases in fasting insulin and HDL cholesterol, although it does not affect cytokine response score

(A) 2′-FL is detected at significantly greater levels in plasma at the week 6 time point in the High Dose 2′-FL group compared to both the Low Dose 2′-FL group (p = 3.62 × 10⁻⁵, Wilcoxon rank-sum test) and the placebo group (p = 5.37 × 10⁻⁷, Wilcoxon rank-sum test). Lower limit of quantification is 0.02 μg/mL. p < 0.05 is indicated by ∗, p < 0.01 by ∗∗, and p < 0.001 by ∗∗∗.

(B) 2′-FL is detected at significantly greater levels in urine at the week 6 time point in the High Dose 2′-FL group compared to both the Low Dose 2′-FL group (p = 0.0002, Wilcoxon rank-sum) and the placebo group (p = 2.56 × 10⁻⁶, Wilcoxon rank-sum). The Low Dose 2′-FL group trended toward higher concentrations of 2′-FL compared to the placebo group (p = 0.076, Wilcoxon rank-sum). Lower limit of quantification is 0.2 μg/mL.

(C) A scatterplot showing the correlation between 2′-FL concentrations measured in plasma and urine, for participants that contributed both urine and plasma samples but excluding participants in the placebo arm. Pearson correlation coefficient is 0.729, and the p value of the association is 1.9 × 10⁻⁸.

(D) Cytokine response score does not change significantly from baseline (week 0) to end of intervention (week 6) for any of the three treatment groups.

(E) Percent change from baseline (week 0) to end of intervention (week 6) for metabolic markers. Fasting insulin and HDL cholesterol increased significantly from week 0 to week 6 in the High Dose 2′-FL group but not the Low Dose 2′-FL group or the placebo group. Dots are the mean change from baseline (week 0) to end of intervention (week 6), and lines are 95% confidence intervals of the means. p < 0.05 is indicated by ∗.

Consumption of 2′-FL had no effect on cytokine response score but is associated with increases in high-density lipoprotein cholesterol and fasting insulin

The pre-registered primary outcome of the study was change in the cytokine response score (CRS) from baseline to week 6. The CRS was defined previously as a set of features within immune profiles that discriminated younger from older individuals⁴; these features indicate hyporesponsiveness in peripheral blood mononuclear cells (PBMCs) of older individuals, and we reasoned that change in CRS may indicate a reversal of inflammation associated with aging. We did not observe significant changes in CRS between treatment groups or from baseline to end of intervention within any of the three groups (Figure 2C).

We used a standard clinical panel to measure the effect of 2′-FL on host metabolic status. We found that the High Dose 2′-FL group, but not the placebo or Low Dose groups, had significantly increased high-density lipoprotein (HDL) cholesterol at end of intervention (week 6) compared to baseline (week 0) (mean = 6.0% increase, 95% confidence interval [CI]: 0.8%–11.3%) while low-density lipoprotein (LDL) cholesterol was unchanged (mean = 0.7% increase, 95% CI: −3.7% decrease–5.2% increase). We also found that fasting insulin increased significantly and specifically in the High Dose 2′-FL group (mean = 23.6% increase, 95% CI: 8.1%–39.1% increase) (Figure 2D). While higher fasting insulin is sometimes associated with insulin insensitivity,²⁸ no subjects’ fasting insulin values exceeded reference ranges during week 6 for the High Dose 2′-FL group (Figure S2).

Bifidobacterium blooms in response to consumption of 2′-FL

We performed 16S rRNA sequencing on each stool sample collected in order to assess compositional changes in the microbiota during the course of the trial (Table S2). We did not find statistically significant changes in Shannon diversity over the course of the intervention (Figure 3A). A beta diversity analysis revealed a significant association between gut microbiota composition with treatment arm at week 3 (p = 0.03, Adonis test, Figure 3B) but not week 0 (p = 0.99, Adonis test) or week 6 (p = 0.35, Adonis test). Comparing individual treatment groups at week 3 revealed that the High Dose 2′-FL group was significantly different from the placebo group (p = 0.006, Adonis test). We performed metagenomic sequencing to determine if microbiota functional capacity was affected by 2′-FL supplementation and found that a small but statistically significant portion of variation in Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway beta diversity was explained by the treatment arm (p < 0.001, Adonis test; Figure S3B).

Figure 3.

2′-FL induces a bloom of Bifidobacterium that is most prominent in week 3 and subsides to baseline levels during the washout period

(A) Shannon (alpha) diversity of each treatment group at each time point during the intervention.

(B) Beta diversity of each treatment group at baseline (week 0), midpoint (week 3), and end of intervention (week 6), depicted via principal coordinate analyses (PCoA). Treatment group explained a significant portion of variation at the midpoint (week 3) time point only (p = 0.03, Adonis test). Ellipses represent 95% confidence intervals around the centroids of each group.

(C) Relative abundance of the genus Bifidobacterium for each treatment group at each time point during the intervention. The High Dose 2′-FL group had higher Bifidobacterium relative abundance at midpoint (week 3) compared to the placebo group (p = 0.0016, Wilcoxon rank-sum test) and also compared to the High Dose 2′-FL group at the pre-baseline week −2 (p = 0.01, Wilcoxon rank-sum test), baseline week 0 (p = 0.037, Wilcoxon rank-sum test), and washout week 10 (p = 0.037, Wilcoxon rank-sum test) time points. p < 0.05 is indicated by ∗, and p < 0.01 by ∗∗.

We hypothesized that the taxon responsible for driving this change in composition was the genus Bifidobacterium, whose member species are known to be specialist degraders of HMOs in the developing infant microbiota^29,30 but also inhabit the adult gastrointestinal tract. Relative abundance of Bifidobacterium was significantly higher in the High Dose 2′-FL group (mean relative abundance = 9.7%, standard error [SE] = 1.4%) compared to the placebo group (mean relative abundance = 3.5%, SE = 0.96%) at the study midpoint (week 3) (Figure 3C). Indeed, Bifidobacterium was the only genus that showed a statistically significant fold change in abundance in any of the three treatment groups from baseline to study midpoint (p = 0.009, Student’s t test, Benjamini-Hochberg correction, Figure S3A). The relative abundance of Bifidobacterium in the Low Dose 2′-FL group was not significantly different from the placebo or the High Dose 2′-FL groups at week 3 but was at an intermediate abundance between both groups (mean relative abundance = 6.0%, SE = 1.3%). Bifidobacterium relative abundance was also significantly higher for the High Dose 2′-FL group at the midpoint of the study (week 3) compared to the same group at baseline (week −2 and week 0) and at the end of the study (week 10) (p = 0.015, p = 0.037, and p = 0.035, respectively, paired Wilcoxon rank-sum tests).

Analysis of the relative abundance of individual species of Bifidobacterium, which we could detect with metagenomics (B. adolescentis, B. pseudocatenulatum, B. infantis, and B. bifidum) over time across groups, revealed two species that followed a similar trend as the genus: B. adolescentis and B. pseudocatenulatum. B. adolescentis had greater relative abundance at the week 3 and week 6 time points in the High Dose 2′-FL group compared to the placebo groups, though these differences were not statistically significant (p = 0.32 and p = 0.09, for weeks 3 and 6, respectively, Wilcoxon rank-sum tests, Figures S3C and S3D). B. pseudocatenulatum also showed greater relative abundance in the High Dose 2′-FL group compared to the placebo group at the week 3 and week 6 time points (p = 0.02 and p = 0.40 for weeks 3 and 6, respectively, Wilcoxon rank-sum tests). While B. pseudocatenulatum isolates have been shown to be able to degrade 2′-FL,³¹ the finding of increased B. adolescentis abundance is surprising given evidence from genomic analysis from publicly available genomes showing that B. adolescentis has limited HMO degradation capacity (Sakanaka et al.³¹). Additional studies are needed to determine whether presently available genomes (some of which are derived from oral rather than gut microbiomes) capture the full extent of B. adolescentis metabolism, particularly related to HMO degradation, or if cross-feeding mechanisms, in which bacteria within the gut rely on other species’ glycan degrading machinery, may be at play.^32,33,34 The data, taken together, demonstrate the strongest microbiome effect in the week 3 time point, which subsided to baseline levels during the post-intervention washout period, a resilience effect consistent with other studies examining dietary microbiota perturbations.^35,36

We next sought to determine whether consumption of 2′-FL led to increases in the production of short-chain fatty acids (SCFAs) by the microbiota. We performed targeted metabolomics on stool samples collected from each subject at week 0 and week 6 to determine the stool concentrations of eight SCFAs (acetate, propionate, butyrate, 2-methylpropionate, pentanoate, 3-methylbutyric, 2-methylbutyric, and hexanoate). We specified a linear mixed effects model (LMM) for each of these SCFAs based on treatment arm and time point and did not observe any statistically significant effects. We next correlated stool SCFA concentrations with the relative abundance of each genera detected in these stool samples. We found that Akkermansia was negatively associated with acetate and propionate (p = 0.0056 and p = 0.0073, respectively, LMM, Benjamini-Hochberg correction) and Dorea was negatively associated with 2-methylbutyric, 3-methylbutyric, and 2-methylpropionate (p = 0.032, p = 0.048, and p = 0.037, respectively, LMM, Benjamini-Hochberg correction). We also found that Dialister was positively associated with acetate, propionate, and butyrate (p = 1.72 × 10⁻⁵, p = 3.64 × 10⁻⁴, and p = 0.023, respectively, LMM, Benjamini-Hochberg correction) and that Holdemanella was positively associated with acetate and propionate (p = 0.0066 and p = 0.0015, respectively, LMM, Benjamini-Hochberg correction) (Figure S4A). We next specified LMMs to determine if components of diet (total energy intake, total carbohydrates, dietary fiber, or total protein) correlated with each stool SCFA. While dietary fiber did not demonstrate statistically significant correlations with SCFAs, we found that total energy intake was significantly associated with acetate, propionate, and butyrate (p = 0.017, p = 0.0089, and p = 0.026, respectively, LMM, Benjamini-Hochberg correction; Figure S4B). Acetate, propionate, and butyrate were also statistically significantly correlated with total carbohydrate intake (p = 0.0089, p = 0.0089, and p = 0.017, respectively, LMM, Benjamini-Hochberg correction). Propionate was also statistically significantly correlated with total protein intake (p = 0.024, LMM, Benjamini-Hochberg correction).

Lastly, we also performed targeted metabolomics on stool samples collected at week 0 and week 3 to determine the stool concentrations of 15 primary and secondary bile acids. As with the SCFAs, we specified a LMM for each of these bile acids based on treatment arm and time point and did not observe any statistically significant effects. We then correlated stool bile acid concentrations with the relative abundance of each genus detected in these stool samples (Figure S4C). We found four pairwise correlations that were statistically significant after multiple hypothesis testing correction with the Benjamini-Hochberg method (Prevotella and glycolithocholic acid, p = 1.67 × 10⁻⁶; Roseburia and chenodeoxycholic acid, p = 7.25 × 10⁻⁶; Roseburia and taurodeoxycholic acid, p = 2.29 × 10⁻⁴; and Agathobacter and taurocholic acid, p = 1.32 × 10⁻³).

Metabolites and circulating inflammatory proteins are altered in 2′-FL-consuming individuals

We next sought to understand how 2′-FL consumption might affect the host metabolic and immune status. To this end, we analyzed serum samples from each subject with both untargeted metabolomics on 127 circulating metabolites (Table S3) and targeted proteomics on 92 circulating cytokines, chemokines, and inflammation-associated proteins (Table S4). We specified linear mixed effects models for each metabolite and immune protein in order to discover metabolites and proteins that are enriched or depleted during the intervention. More specifically, we used these LMMs to determine which features had significant interaction effects between treatment arm and time (that is, significantly different slopes of normalized feature abundance between the three treatment arms over time). We then performed post hoc tests to determine in which treatment arm each feature was enriched or depleted.

We found three metabolites that had statistically significant interaction effects in our LMM modeling: octanoylcarnitine, glutamate, and taurine (p = 0.0011, p = 0.042, and p = 0.049, respectively, LMM, Benjamini-Hochberg corrected, Table S5). Octanoylcarnitine is a medium-chain fatty acid covalently linked to carnitine, which undergoes fatty acid oxidation in the mitochondria. Post hoc analysis revealed that octanoylcarnitine decreased significantly in the High Dose 2′-FL group during the intervention (p = 0.015, Wilcoxon rank-sum test, Figure 4A), but not in the Low Dose 2′-FL group or placebo group (p = 0.27 and p = 0.39, respectively, Wilcoxon rank-sum test). Glutamate and taurine significantly increased in abundance in the placebo group during the intervention (p = 0.028 for glutamate and p = 0.003 for taurine, Wilcoxon rank-sum tests). Levels of these metabolites also increased in the Low Dose 2′-FL (p = 0.14 for glutamate and p = 0.10 for taurine, Wilcoxon rank-sum tests) and High Dose 2′-FL groups (p = 0.12 for glutamate and p = 0.13 for taurine, Wilcoxon rank-sum tests), albeit not in a statistically significant manner.

Figure 4.

Serum metabolites and proteins that vary significantly by treatment arm during the intervention

(A) Boxplots showing the log2 abundance of octanoylcarnitine (top), glutamate (middle), and taurine (bottom) from untargeted metabolomics analysis of serum samples. Three panels from left to right show metabolite abundance for the placebo group, Low Dose 2′-FL group, and High Dose 2′-FL group, respectively, in the baseline and intervention time points. P_int values next to each metabolite name are the p values of the linear mixed effect modeling, adjusted across all metabolites using the Benjamini-Hochberg method. p values within each facet are the results of Wilcoxon rank-sum post hoc tests between baseline and intervention time points. The baseline time points are weeks −2 and 0, and the intervention time points are weeks 3 and 6.

(B) Boxplots showing the normalized protein expression values of FGF21 (top) and CD40 (bottom) from targeted proteomics analysis of serum samples. Three panels from left to right show metabolite abundance for the placebo group, Low Dose 2′-FL group, and High Dose 2′-FL group, respectively, in the baseline and intervention time points. P_int values next to each protein name are the p values of the linear mixed effect modeling, adjusted across all metabolites using the Benjamini-Hochberg method. p values within each facet are the results of Wilcoxon rank-sum post hoc tests between baseline and intervention time points. The baseline time points are weeks −2 and 0, and the intervention time points are weeks 3 and 6.

With regards to circulating proteins, we found two proteins that had statistically significant interaction effects in our LMM modeling: fibroblast growth factor 21 (FGF21) and CD40 (p = 0.0032 and p = 0.036, respectively, LMM, Benjamini-Hochberg correction, Table S5). FGF21 is a liver-secreted hormone that regulates lipid metabolism.³⁷ Post hoc analysis revealed that FGF21 increased significantly in the High Dose 2′-FL group during the intervention (p = 0.047, Wilcoxon rank-sum test, Figure 4B), but not in the Low Dose 2′-FL group or the placebo group (p = 0.40, p = 0.087, respectively, Wilcoxon rank-sum test). In human clinical trials, FGF21 analogs have been shown to increase HDL cholesterol levels while down-regulating lipolysis,^38,39 consistent with our finding of increased HDL cholesterol (Figure 2B) and decreased serum octanoylcarnitine (Figure 4A) in the High Dose 2′-FL group. CD40 is expressed by a wide range of antigen-presenting cells and is involved with B cell activation and immunoglobulin production. Post hoc analysis did not reveal any significant changes during the intervention in any of the treatment groups (p = 0.18 for High Dose 2′-FL group, p = 0.24 for the Low Dose 2′-FL group, and p = 0.18 for the Placebo group, Wilcoxon rank-sum tests).

Increased Bifidobacterium abundance corresponds to more Bifidobacterium at baseline and circulating 2′-FL at end of intervention

We observed a heterogeneous response to 2′-FL consumption with regards to changes in Bifidobacterium relative abundance during the intervention. This heterogeneity may be due to (1) lack of Bifidobacterium in the gut microbiota at baseline; (2) lack of Bifidobacterium that is capable of degrading 2′-FL; or (3) other ecological factors, including other microbes, which may be consuming 2′-FL.⁴⁰ To provide insight into whether the metabolomic and proteomic changes observed in the 2′-FL treatment groups may depend on a bloom of Bifidobacterium in the gut microbiota, we performed a post hoc analysis that subsetted participants from the High Dose 2′-FL and Low Dose 2′-FL groups based on their change in Bifidobacterium relative abundance from week 0 to week 3. Participants above the median Bifidobacterium change were categorized as “responders” (n = 30 participants), and those below the median were categorized as “nonresponders” (n = 29 participants) (Figure 5A). Of the 30 responders, 18 were in the High Dose 2′-FL group and 12 were in the Low Dose 2′-FL group. While responders had greater Bifidobacterium relative abundance at week 0 (baseline) on average compared to nonresponders, this difference was not statistically significant (p = 0.055, Wilcoxon rank-sum test, Figure 5B). We next performed a presence-absence analysis to determine whether or not responders and nonresponders have Bifidobacterium in their microbiome at baseline (week 0). Indeed, 93.3% of responders have Bifidobacterium in their microbiome at baseline, compared to only 58.6% of nonresponders (p = 0.0046, chi-squared test). Lastly, neither age (p = 0.81, Wilcoxon rank-sum test) nor gender (p = 0.76, chi-squared test) was associated with responder status.

Figure 5.

Subsetting subjects into responders and nonresponders based on change in Bifidobacterium abundance reveals widespread changes to levels of circulating cytokines

(A) Subsetting of subjects into responders and nonresponders based on change in Bifidobacterium relative abundance between baseline (week 0) and midpoint (week 3). Participants were classified as responders if they were in the top tertile of Bifidobacterium increase from week 0 to week 3.

(B) Bifidobacterium relative abundance at baseline (week 0) for responders and nonresponders is not significantly different (p = 0.055, Wilcoxon rank-sum test).

(C) Plasma concentrations of 2′-FL at end of intervention (week 6) for responders and nonresponders. Responders have significantly greater plasma concentrations of 2′-FL (p = 0.038, Wilcoxon rank-sum test). The dashed line represents the lower limit of quantification (0.02 micrograms/mL).

(D) Scatterplot showing the relationship between week 0 to week 3 change in Bifidobacterium abundance and plasma concentration of 2′-FL. Points are colored according to responder status; gray segments indicate standard error of linear trend. Linear regression analysis revealed that change in Bifidobacterium abundance from week 0 to week 3 was associated with higher levels of 2′-FL in plasma (p = 0.016, linear regression). Horizontal dashed line indicates limit of 2′-FL detection.

(E) Scatterplot showing the relationship between week 0 to week 3 change in Bifidobacterium abundance and urine concentration of 2′-FL. Points are colored according to responder status; gray segments indicate standard error of linear trend. Linear regression analysis revealed that change in Bifidobacterium abundance from week 0 to week 3 was not associated with higher levels of urine 2′-FL concentration (p = 0.60, linear regression). Horizontal dashed line indicates limit of 2′-FL detection.

(F) Volcano plot of untargeted serum metabolomics showing the metabolites enriched in responders and nonresponders. p values were false discovery rate corrected with the Benjamini-Hochberg method. Horizontal dashed line represents a q value threshold of 0.1. ACC, 1-aminocyclopropane-1-carboxylic acid.

(G) Volcano plot of targeted serum proteomics showing the proteins enriched in responders and nonresponders. p values were false discovery rates corrected with the Benjamini-Hochberg method. Horizontal dashed line represents a q value threshold of 0.1.

Interestingly, we found that responders had higher levels of 2′-FL in plasma at week 6 compared to nonresponders (p = 0.038, Wilcoxon rank-sum test, Figure 5C). Moreover, this relationship appears to be linear. A linear regression analysis revealed a significant association between the magnitude of the change in Bifidobacterium abundance from week 0 to week 3 and the concentration of 2′-FL in plasma at week 6 (Figure 5D, p = 0.016, linear regression). Such a relationship was not found between the magnitude of the change of Bifidobacterium abundance and 2′-FL concentration in urine (Figure 5E, p = 0.60, linear regression).

Serum metabolites and cytokines are enriched in responders

Next, we sought to determine whether any physiological consequences corresponded to the observed bloom in Bifidobacterium abundance. We re-analyzed the untargeted metabolomics data, this time comparing responders versus nonresponders, and discovered five metabolites enriched in responders (glutamate, taurine, ornithine, asparagine, and 1-aminocyclopropane-1-carboxylic acid) and three metabolites enriched in nonresponders (octanoylcarnitine, inosine, and arginine; Figure 5F). Pathway analysis of these eight metabolites reveals a significant impact on the arginine biosynthesis pathway (p = 9.71 × 10⁻⁴, Benjamini-Hochberg corrected).

We found that responders showed significant enrichment for 24 circulating cytokines compared to nonresponders (Figure 5G). Network analysis of these cytokines with the Search Tool for the Retrieval of Interacting Genes (STRING) database revealed a local clustering coefficient 0.591 (P < 1 × 10⁻¹⁶). Gene Ontology pathway analysis revealed significant pathway enrichment of “cytokine-mediated signaling pathway,” “cellular response to organic substance,” and “neutrophil chemotaxis” (p = 1.00 × 10⁻¹³, 7.36 × 10⁻¹⁰, and 8.38 × 10⁻⁹, respectively, Benjamini-Hochberg false discovery rate [FDR] correction). Notably, one significantly up-regulated protein, sirtuin 2 (SIRT2), has been strongly implicated in aging and longevity.⁴¹ We also re-analyzed the CRS using this responder subsetting but found no significant difference between responders and nonresponders at the end of the intervention (p = 0.61, Wilcoxon rank-sum test).

In order to ensure that we were not arbitrarily subsetting our data in a way that inflated our FDR, we simulated the responder/nonresponder enrichment analysis 500 times, randomly permuting the responder and nonresponder labels each time. For the metabolomics analysis the “true” responder/nonresponder labels yielded more significant features than >94% of simulated analyses, and for the proteomics analysis the “true” responder/nonresponder labels yielded more significant features than >84% of simulated analyses (Figure S5). This sensitivity analysis underscores the utility of our responder subsetting scheme, showing that Bifidobacterium-based subsetting of participants coincides with a broader biological impact of the intervention.

Multi-omics associations between serum metabolomics and cytokines

In addition to understanding the impact of a dietary intervention on human subjects, multi-omics studies such as this one can be used to discover novel interactions between different facets of host biology. To this end, we next sought to understand how changes in circulating levels of serum metabolites and serum cytokines might influence one another. We calculated the baseline-to-end fold change of each metabolite and cytokine for each participant and performed the all-versus-all correlations across the combined datasets (Figure 6A). We discovered extensive, significant associations both within and between these datasets. Eleven metabolites and 18 serum proteins were implicated in statistically significant cross-dataset correlations after multiple hypothesis testing correction (Figure 6B). Six of these 11 metabolites are purines (or purine derivatives), and four of these 11 metabolites are amino acids (or amino acid derivatives). Notably, five of these metabolites (asparagine, inosine, glutamate, taurine, and ornithine) and 10 of these cytokines (AXIN1, CD40, ENRAGE, HGF, OSM, SIRT2, ST1A1, STAMPB, TGF-alpha, and TNFSF14) were also significantly enriched in Bifidobacterium responders or nonresponders, indicating a broad rewiring of the host metabolome and immuno-proteome in response to the 2′-FL-mediated Bifidobacterium bloom.

Figure 6.

Multi-omic associations between serum metabolomics and serum cytokine proteomics

(A) A correlation network between serum metabolomics (green nodes), serum cytokine proteomics (orange nodes), and clinical markers (blue nodes). All-versus-all correlation network was based on centered and scaled baseline-to-end changes in individual metabolite/protein levels across all participants. Blue edges indicate positive correlation, and red edges indicate negative correlation. ACC = 1-Aminocyclopropane-1-carboxylic acid.

(B) Significant pairwise correlations between serum metabolites and serum cytokines using Pearson correlation (p value correction with Benjamini-Hochberg method, corrected p values ≤ 0.05). Larger circles indicate more significant corrected p values. Color map indicates the strength of the Pearson correlation with blue indicating positive correlation and red indicating negative correlation. Metabolites and cytokines with asterisks were significantly associated with responder status.

Responders perform better on assessments of cognitive abilities

During the course of the trial, participants took two forms of cognitive assessments from the Cambridge Neuropsychological Test Automated Battery of computer-based tests⁴²: the rapid visual information processing (RVP) test, which assesses sustained attention, and the paired associates learning (PAL) test, a measure of visual memory and learning. For both tests, there was no significant difference in performance between treatment groups (p = 0.81 and p = 0.94 for PAL and RVP, respectively, analysis of covariance analysis; Figures S6A and S6B). However, Bifidobacterium responders performed comparatively better than nonresponders on the PAL test (p = 0.016 analysis of covariance, Figure S6C). There were no significant differences between responders and nonresponders for the RVP test (p = 0.081, analysis of covariance; Figure S6D). These results, while speculative, are in line with research in infants and animal models regarding the effect of 2′-FL on cognition and memory^21,22,43 and suggest the need for future studies designed to test cognitive performance explicitly.

Discussion

The aging immune system is characterized by the impaired response to stimuli as a result of heightened baseline inflammation.⁴ This immune dysregulation is implicated in numerous non-communicable pathologies such as cancer, atherosclerosis, and neurodegeneration.^1,2 This study demonstrates the effects of the HMO 2′-FL on a cohort of older adults. Our work demonstrates that dosing older adult humans with 2′-FL results in a targeted modification of the microbiota, namely a bloom of Bifidobacterium in the High Dose 2′-FL group. The primary outcome of this study, the CRS,⁴ was not statistically significant and suggests that the 2′-FL intervention did not improve immune “responsiveness” according to this specific metric based on particular cell types and stimulating cytokines. However, secondary analysis showed that the bloom of Bifidobacterium was accompanied by widespread metabolic and immune alteration. Given that we show evidence of 2′-FL entering systemic circulation (Figures 2A and 2B) and that past work shows direct effects of 2′-FL on host cells,^14,15 future work will be needed to determine the extent to which these HMO-mediated effects are dependent upon activity of the microbiota.

We observed two significant changes in the clinical panels of participants in the High Dose 2′-FL group: significant increases in fasting insulin and HDL cholesterol. Previous work has shown that Bifidobacterium spp. can induce glucagon-like peptide-1 (GLP-1) secretion in the gastrointestinal tract.⁴⁴ GLP-1 in turn, stimulates the secretion of insulin.⁴⁵ Bifidobacterium spp. are also known to affect cholesterol and lipid homeostasis through the activity of bile salt hydrolases. These results suggest that the increase in Bifidobacterium that we observed in the High Dose 2′-FL group may be responsible for the changes to levels of circulating insulin, HDL cholesterol, and the lipid octanoylcarnitine. We also observed higher levels of the endocrine factor FGF21, which is known to regulate insulin sensitivity and lipid metabolism.⁴⁶ Directly measuring bacteria, metabolites, and hormones in the gastrointestinal tract and other tissues in situ is challenging in human studies, and it is often difficult to capture the dynamic interplay between these features. Future studies in animal models will help to uncover host-microbe interactions occurring in this part of the gastrointestinal tract. Future mechanistic studies will also be needed to understand how long-lasting some of these physiological adaptations are in response to a short-term nutritional intervention as long-term follow-ups are often challenging in human studies.

One speculative finding in our study was the observation that Bifidobacterium responders showed improved performance on a cognitive test relative to nonresponders. This “paired associates learning” test is a measure of visual memory. Notably, responders did not perform better on the “rapid visual processing” test, which is a measure of visual processing time. While this battery of two tests covers only a narrow range of human cognitive functions, we believe that tests relating to memory may be particularly relevant to an older cohort, such as the one used in this study. These results demonstrate a potential impact of 2′-FL supplementation on cognitive function in human adults. Previous work has shown that 2′-FL levels in breast milk are positively associated with cognitive development in human infants,²¹ and 2′-FL supplementation influences vagus-nerve-mediated memory adaptations in adult rodents.²² Future studies in humans should be performed using expanded batteries of tests as well as other ways of measuring cognitive function.

Bifidobacterium spp. are widely used as probiotic agents and are associated with a panoply of host benefits,⁴⁷ which makes their enrichment in the gut microbiota via HMOs an attractive therapeutic target. Oral delivery of probiotics has demonstrated mixed but promising clinical value.^48,49,50 Furthermore, many probiotic formulations may fail to properly engraft in the host microbiota. Recent studies have shown improved engraftment of probiotic strains when they are delivered as “synbiotics” with prebiotic compounds.^51,52 We demonstrate a modulation of Bifidobacterium abundance in human subjects via consumption of an HMO without the use of a probiotic supplement.

Additionally, this 2′-FL-mediated modulation of Bifidobacterium was reversible: relative abundance subsided to baseline levels after cessation of prebiotic consumption. Similar findings have been reported with other prebiotics.^51,53,54 These personalized kinetics of “recovery” from dietary interventions with prebiotics are analogous to the microbiota resilience that follows cessation of a variety of perturbations such as seminal studies of recovery from antibiotics.⁵⁵ In this study, the high dose of 5 g/day was well tolerated and resulted in prominent changes to the gut microbiota as well as to the host metabolome and proteome. One recent study showed well-tolerated consumption of up to 20 g/day 2′-FL in younger adults,⁵⁶ and another showed 18 g/day of a mixture of HMOs derived directly from breastmilk.⁵¹ Longer-term intervention studies will be needed to understand the nature of resiliency in the gut microbiota during and after a dietary intervention. Future studies will also be needed to understand proper dosing regimens for HMOs and whether administration of live Bifidobacterium or other bacteria is required to achieve specific individualized versus generalized responses in adult subjects.

Limitations of the study

Despite the promising findings in this study, several limitations should be acknowledged. First, the sample size of 89 participants, while sufficient to detect significant changes in the primary microbiome outcomes, may have limited our ability to detect more subtle changes in host metabolism and immune parameters, particularly in subgroup analyses. Second, although the 6-week intervention period allowed us to observe significant changes in microbiota composition and metabolic markers, a longer intervention might reveal additional effects that develop over extended time frames.

While supplementation with 2′-FL was shown to have beneficial effects, a limitation of this HMO-only prebiotic approach is that broader metabolic and immunological benefits seem to be individualized. Studies of probiotic engraftment have shown that responses are contingent upon the baseline composition of an individual’s microbiota.⁵⁷

Our study focused on healthy older adults (mean age 67.3 years), which limits generalizability to other age groups or individuals with specific health conditions. The mechanisms linking 2′-FL, microbiome changes, and downstream metabolic effects warrant further investigation, particularly regarding the detection of 2′-FL in systemic circulation. Future studies should explore whether direct effects of 2′-FL on host cells contribute to observed outcomes independently of microbiome modulation.

Finally, while we observed improvements in visual memory performance in Bifidobacterium responders, these cognitive findings were exploratory and will require validation in studies specifically designed to assess cognitive outcomes. The reversibility of microbiome changes following cessation of supplementation suggests that continuous prebiotic intake might be necessary to maintain benefits, raising questions about optimal dosing regimens for long-term interventions.

Resource availability

Lead contact

All information and requests for further resources should be directed to and will be fulfilled by the lead contact, Justin L. Sonnenburg (jsonnenburg@stanford.edu).

Materials availability

This study did not generate new unique reagents.

Data and code availability

Metagenomic and 16S reads are publicly available at the NCBI Sequence Read Archive (SRA). Metabolomics data are publicly available at the National Metabolomics Data Repository. Accession numbers are listed in the key resources table. Code and analysis files for this project can be found at https://github.com/SonnenburgLab/RAMP_study_analysis. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

We wish to thank the participants for their engagement and effort to enable this study; Hannah Wastyk for providing guidance and mentorship; the University of Minnesota Genomics Center for 16S sequencing; Tala Khosroheidari and William Beckwith at Olink Proteomics, the Human Immune Monitoring Core, for staining and running PhosphoFlow and Olink assays; the Chan Zuckerberg Biohub for metagenomic sequencing; Metabolon for providing targeted mass spectrometry analysis; Yuqin Dai and the Metabolomics Knowledge Center for supporting mass spectrometry analysis; Will Van Treuren for aiding with metabolomics method development; and the Stanford Clinical and Translational Research Unit. This work was funded by Abbott Nutrition, United States. M.M.C. was supported by a Stanford Graduate Fellowship. J.L.S. is a Chan Zuckerberg Biohub investigator.

Author contributions

M.M.C., E.D.S., C.D.G., and J.L.S. performed the experiments and designed and performed data analysis. D.D. and D.P. provided patient interaction, dietary counseling, and analysis. M.S.O. and M.M.C. performed stool aliquoting, metagenomic sequencing, and analysis. K.M. and H.T.M. performed immune profiling assays. K.C. aided with statistical analyses. J.L.R. and D.D. oversaw patient recruitment and study design implementation. M.M.C., C.D.G., J.L.S., A.S.-D., J.M.C., and R.H.B. conceived of the study and wrote the manuscript.

Declaration of interests

J.M.C., A.S.-D., and R.H.B. are employees of Abbott Laboratories, which is a commercial manufacturer of nutritional products containing HMOs including infant formula. J.L.S. is a founder, a shareholder, and on the scientific advisory board of Novome Biotechnologies and Interface Biosciences. M.M.C., J.L.S., C.D.G., and Abbott Laboratories have filed a provisional patent application related to this work.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
CD3 Pacific Blue	BD Biosciences	Cat#558117; RRID: AB_397038
CD4 PerCP-Cy5.5	BD Biosciences	Cat#332772; RRID: AB_2868621
CD20 PerCP-Cy5.5	BD Biosciences	Cat#560735; RRID: AB_1727450
CD33 PE-Cy7	BD Biosciences	Cat#333946; RRID: AB_399961
CD45RA Qdot 605	BD Biosciences	Cat#560775; RRID: AB_1937333
Phospho-STAT1 (pY701) Alexa Fluor 488	BD Biosciences	Cat#612595; RRID: AB_399879
Phospho-STAT3 (pY705) Alexa Fluor 647	BD Biosciences	Cat#557815; RRID: AB_647144
Phospho-STAT5 (pY694) PE	BD Biosciences	Cat#612567; RRID: AB_399858
IFN-α	BD Biosciences	Cat#554618
IFN-γ	BD Biosciences	Cat#554617
IL-2	BD Biosciences	Cat#554603
IL-6	BD Biosciences	Cat#550071
IL-7	BD Biosciences	Cat#554608
IL-10	BD Biosciences	Cat#554611
IL-21	Fisher Scientific	Cat#PHC0214
Biological samples
Human stool samples	This study
Human blood samples	This study
Human urine samples	This study
Chemicals, peptides, and recombinant proteins
2′-fucosyllactose (2′-FL)	Abbott Nutrition
Ficoll-Paque PLUS	Sigma Aldrich	Cat#GE17-1440-02
Proteomic Stabilizer, Smart Tube	Fisher Scientific	Cat#501351692
Deposited data
16S rRNA sequencing data	This paper	NCBI BioProject: PRJNA1177009
Metagenomic sequencing data	This paper	NCBI BioProject: PRJNA1177009
Metabolomics	This paper	National Metabolomics Data Repository: PR002428 https://doi.org/10.21228/M8SN8M
Software and algorithms
QIIME pipeline version 1.8	Caporaso et al.58	http://qiime.org/
DADA2	Callahan et al.59	https://benjjneb.github.io/dada2/
Phyloseq	McMurdie and Holmes60	https://joey711.github.io/phyloseq/
HUMAnN	Beghini et al.61	https://huttenhower.sph.harvard.edu/humann/
BBtools suite	Bushnell62	https://sourceforge.net/projects/bbmap/
FastQC	Babraham Bioinformatics63	https://www.bioinformatics.babraham.ac.uk/projects/fastqc/
R	R Core Team	https://www.r-project.org/
Maaslin2	Mallick et al.64	https://github.com/biobakery/Maaslin2
MSDIAL software	Tsugawa et al.65	http://prime.psc.riken.jp/compms/msdial/main.html
DIVA v8	Becton Dickinson	https://www.bdbiosciences.com/en-us/products/software/instrument-software/bd-facsdiva-software
FlowJo v10	Becton Dickinson	https://www.flowjo.com/
Code for analysis	This paper	https://github.com/SonnenburgLab/RAMP_study_analysis

Experimental model and subject details

Recruitment and selection of participants

Participants were recruited from the local community through online advertisement in different community groups as well as emails to past research participants that consented to being contacted for future studies. The current study assessed 335 participants for eligibility. They completed an online screening questionnaire and a clinic visit between March 2019 and December 2020. The primary inclusion criteria included age ≥60 years and general good health. Participants were excluded if they had a history of active uncontrolled inflammatory bowel disease (IBD) including ulcerative colitis, Crohn’s disease, or indeterminate colitis, irritable bowel syndrome (IBS) (moderate or severe), infectious gastroenteritis, colitis or gastritis, Clostridium difficile infection (recurrent) or Helicobacter pylori infection (untreated), malabsorption (such as celiac disease), major surgery of the gastroinstestinal tract, with the exception of cholecystectomy and appendectomy, in the past five years, or any major bowel resection at any time. Other exclusion criteria included a BMI ≥40, diabetes, renal disease, significant liver enzyme abnormality, smoking, a history of CVD, inflammatory disease, or malignant neoplasm. Consort flow diagram of participant recruitment shown in Figure 1B and demographics table shown in Table S1. 89 participants (49 female sex and gender identifying, 40 male sex and gender identifying) were used for full analysis with an average age of 67.3 ± 5.4 years. All study participants provided written informed consent. The objective of this study is to define the impact of a prebiotic supplement on microbiome, immune system, and metabolic status in older adults. The study was approved annually by the Stanford University Human Subjects Committee. Trial was registered at ClinicalTrials.gov, identifier: NCT03690999.

Specimen collection

Stool samples were collected at Weeks −2 (pre-baseline), 0 (baseline), 3 (midpoint of intervention), 6 (end of intervention), and 10 (washout). All stool samples were kept in participants’ home freezers (−20C) wrapped in ice packs, until they were transferred on ice to the research laboratory and stored at −80C. Blood samples were collected during research clinic visits at Weeks −2 (pre-baseline), 0 (baseline), 3 (midpoint of intervention), 6 (end of intervention), and 10 (washout). Blood for PBMC and whole blood aliquots were collected into heparinized tubes. Whole blood aliquots were incubated with Proteomic Stabilization Buffer (Smart tube, Fisher Scientific) for 12 min at room temperature and stored at −80C. PBMCs were isolated using Ficoll-Paque PLUS (Sigma-Aldrich), washed with PBS, frozen at −80C for 24 h then moved to liquid nitrogen for longer storage. Blood for serum was collected into an SST-tiger top tube, spun at 1,200xg for 10 min, aliquoted, and stored at −80C. Blood for plasma was collected into an EDTA tube, spun at 1,200xg for 10 min, aliquoted, and stored at −80C.

Method details

Randomization

Participants were randomized to the Placebo arm, the Low Dose 2′-FL arm or the High Dose 2′-FL arm. A simple randomization was done for the three groups using a random number generator (Excel), performed by a statistician not involved in the intervention or data collection.

Dietary logs and health questionnaires

Participants logged all their food and drink intake for 3 days (2 weekdays and 1 weekend) each week during the ramp phase and every other week for the rest of the study using the HealthWatch360 app. The dietitian reviewed the entries with participants to assess accuracy of entries and portions.

The following validated health surveys were used by participants: PROMIS v1.1 global health, PROMIS v1.0 - fatigue, WHO wellbeing index, PROMIS applied cognition short form, Perceived Stress Scale (Cohen et al., 1983), and the International Physical Activity Questionnaire (Craig et al., 2003).

16S amplicon sequencing

DNA was extracted from stool using the MoBio PowerSoil kit according to the Earth Microbiome Project’s protocol (Gilbert et al., 2014) and amplified at the V4 region of the 16S ribosomal RNA (rRNA) subunit gene and 250 nucleotides (nt) Illumina sequencing reads were generated. There was an average of 60,192 reads per sample (standard deviation of 22,043) and samples with less than 2,000 reads were filtered out (2 samples out of 419 removed). There was an average of 40,621 reads per sample recovered (standard deviation of 16,022) after filtering, denoising, and removing chimeras. 16S rRNA gene amplicon sequencing data from both stool samples and fermented food samples were demultiplexed using the QIIME pipeline version 1.8.⁵⁸ Amplicon sequence variants (ASVs) were identified with a learned sequencing error correction model (DADA2 method),⁵⁹ using the dada2 package in R. ASVs were assigned a taxonomy using the Silva database (version 138.1).⁶⁶ ɑ-diversity was quantified as the number of observed ASVs, Shannon diversity, or PD whole tree, in a rarefied sample using the phyloseq package in R (version 3.4.0). Data were rarefied to 6,727 reads per sample (lowest 5% of reads, 398 samples retained out of 419 total) also using the phyloseq package in R. Rarefied data were only used for ɑ-diversity measures. b-diversity was calculated using the ordinate function in the phyloseq package in R (version 3.4.0) for weighted and unweighted Unifrac. To determine if the treatment arms were significantly different at different timepoints, the samples were filtered to a given timepoint, beta diversity was calculated, and the analysis of variance using distance matrices was calculated with the adonis function (method = “bray”, permutations = 5000) in the vegan package in R (version 2.5.6).

Metagenomic sequencing

DNA extraction for shotgun metagenome sequencing was done using the MoBio PowerSoil kit as described in the 16S amplicon sequencing methods. For library preparation, the Nextera Flex kit was used with a minimum of 10ng of DNA as input and 6 or 8 PCR cycles depending on input concentration. A 12 base pair dual-indexed barcode (CZ Biohub) was added to each sample and libraries were quantified using an Agilent Fragment Analyzer. They were further size-selected using AMPure XP beads (Beckman) targeted at a fragment length of 450bp (350bp size insert). DNA paired-end sequencing (2x146bp) was performed on a NovaSeq 6000 using S4 flow cells (CZ Biohub). The average sequencing depth for each sample was 14.2 million paired-end reads (or 4.05 giga base pairs). Data quality analysis was performed by demultiplexing raw sequencing reads and concatenating data for samples that required multiple sequencing runs for target depth before further analysis. BBtools suite⁶² (https://sourceforge.net/projects/bbmap/) was used to process raw reads and mapped against the human genome (hg19) after trimming, with masks over regions broadly conserved in eukaryotes (http://seqanswers.com/forums/showthread.php?t=42552). Exact duplicate reads (subs = 0) were marked using clumpify and adapters and low-quality bases were trimmed using bbduk (trimq = 16, minlen = 55). Finally, reads were processed for sufficient quality using FastQC⁶³ (https://www.bioinformatics.babraham.ac.uk/projects/fastqc/).

Microbiome functional profiling

The abundance of KEGG orthologous groups and modules were determined using the HUMAnN pipeline⁶¹ with default parameters. The input to this pipeline were the processed reads described in the “metagenomic sequencing” section above.

Measuring stool short-chain fatty acids with LC-MS

Quantification of short-chain fatty acids was performed using an LC/MS-based technique described here,⁶⁷ using a method adapted from.⁶⁸ Briefly, stool from participants was homogenized in an extraction buffer containing isotopically labeled SCFA standards, followed by metabolite extraction through bead-beating, incubation at −20°C, and centrifugation. The supernatant was derivatized with 3-nitrophenylhydrazine and 1-ethyl-3-(3-dimethylaminopropyl)carbodiimide, then analyzed using an Agilent 6470 triple-quadrupole LC-MS with a BEH C18 column and gradient elution. Analyte quantification was performed via isotope dilution, using standard curves to ensure linearity. Optimization steps included full-mass scans, collision energy tuning, and signal-to-noise maximization for accurate SCFA measurement.

Measuring stool bile acids with LC-MS

Bile acid concentrations are analyzed by LC-MS/MS according to Metabolon Method TAM178 (“LC-MS/MS Method for the Quantitation of Bile Acids”). The Human Feces Bile Acid Panel measures all of the major primary and secondary bile acids and their conjugates: Cholic Acid (CA), Chenodeoxycholic Acid (CDCA), Deoxycholic Acid (DCA), Lithocholic Acid (LCA), Ursodeoxycholic Acid (UDCA), Glycocholic Acid (GCA), Glycochenodeoxycholic Acid (GCDCA), Glycodeoxycholic Acid (GDCA), Glycoursodeoxycholic Acid (GUDCA), Taurocholic Acid (TCA), Taurochenodeoxycholic Acid (TCDCA), Taurodeoxycholic Acid (TDCA), Taurolithocholic Acid (TLCA), Tauroursodeoxycholic Acid (TUDCA), and Glycolithocholic Acid (GLCA). Calibration samples are prepared at eight different concentration levels by spiking an acidified methanol solution with corresponding calibration spiking solutions. Calibration samples, study samples, and quality control samples are spiked with a solution of labeled internal standards and subjected to protein precipitation with an organic solvent (acidified methanol). Following centrifugation, an aliquot of the organic supernatant is evaporated to dryness in a gentle stream of nitrogen. The dried extracts are reconstituted and injected onto an Agilent 1290 Infinity/SCIEX QTRAP 6500 or 6500+ LC-MS/MS system equipped with a C18 reverse phase UHPLC column. The mass spectrometer is operated in negative mode using electrospray ionization (ESI). The peak area of each bile acid parent (pseudo-MRM mode) or product ion is measured against the peak area of the respective internal standard parent (pseudo-MRM mode) or product ion. Quantitation is performed using a weighted linear least-squares regression analysis generated from fortified calibration standards prepared immediately prior to each run.

Serum cytokines

Cytokine data were generated from serum samples submitted to Olink Proteomics for analysis using their provided inflammation panel assay of 92 analytes (Olink Inflammation). Data are presented as normalized protein expression values (NPX, Olink Proteomics arbitrary unit on log2 scale). We removed 29 proteins from the dataset that were below the limit of detection in greater than 30% of samples.

PhosphoFlow

This assay was performed by the Human Immune Monitoring Center at Stanford University. Normal PBMCs were isolated by density-gradient centrifugation and cryopreserved. PBMC were thawed in warm media, washed twice and resuspended at 0.5x10⁶ viable cells/mL. 200 μl of cells were plated per well in 96-well deep-well plates. Cytokine stock solutions were prepared at a concentration of 100 g/mL for each cytokine (IFN, IL-6, IL-7, IL-10, IL-2, or IL-21), or 10⁶ U/ml for IFN. After resting for 2 h at 37°C, cells were stimulated by adding 50 μl of cytokine stock solution and incubated at 37°C for 15 min. The PBMCs were then fixed with paraformaldehyde, permeabilized with methanol, and kept at −80°C overnight. The cells were washed with FACS buffer (PBS supplemented with 2% FBS and 0.1% sodium azide), and stained with the following antibodies (all from BD Biosciences, San Jose, CA): CD3 Pacific Blue, CD4 PerCP-Cy5.5, CD20 PerCp-Cy5.5, CD33 PE-Cy7, CD45RA Qdot 605, pSTAT-1 Alexa Fluor 488, pSTAT-3 Alexa Fluor 647, pSTAT-5 PE. The samples were then washed and resuspended in FACS buffer. 100,000 cells per stimulation condition were collected using DIVA software on an LSRII flow cytometer (BD Biosciences). Data analysis was performed using FlowJo software (BD), by gating on live cells based on forward versus side scatter profiles, then on singlets using forward scatter area versus height, followed by cell subset-specific gating. The 90th percentile fluorescence for each pSTAT marker on each cell population was reported.

Untargeted serum metabolomics

Untargeted metabolomics analysis was performed using a previously published mass spectrometry pipeline.⁶⁹ Briefly, serum samples were extracted in LC-MS grade methanol (4:1 v/v). These serum samples were identical to the samples used for serum cytokine analysis detailed above. Protein precipitation was conducted by incubating the samples for 5 min at room temperature and centrifugation at 5,000xg for 10 min. Sample supernatants were then transferred, evaporated, and reconstituted in an internal standard mix (50% Methanol). Metabolite samples were analyzed on an LC-MS qTOF instrument using reverse phase C18 positive mode as described.^69,70 Compound annotation was carried out using the MSDIAL software⁶⁵ and an authentic standard reference library. To quantify metabolite levels, area under the curve for each annotated metabolite was normalized using the sum of internal standards in each sample.

Quantification and statistical analysis

Location of statistical details in the text

Results of each experiment can be found in the results and figure legends. Significant values of statistical tests are also indicated by asterisks in the figures according to the following scheme: p < 0.05 is indicated by ∗, p < 0.01 by ∗∗, and p < 0.001 by ∗∗∗.

Statistical analysis of primary outcome

The primary outcome as listed on ClinicalTrials.gov was the change in Cytokine Response Score (CRS) within each arm from baseline (week 0) to end of intervention (week 6). CRS was calculated using the PhosphoFlow data according to the methods described in (Shen-Orr et al., 2016). Briefly, we computed the CRS for each individual by summing the normalized values of 15 reproducible age-associated cytokine responses. Significant changes were evaluated using Wilcoxon rank-sum tests.

Quantification of alpha and beta diversity measures

Alpha and beta-diversity measures were calculated using the phyloseq package⁶⁰ in R. Alpha diversity was calculated at the ASV level. Bray-Curtis dissimilarity was computed at each timepoint and for each subject using the vegdist function in the R vegan package.⁷¹

Testing for significant changes in microbe and host features during the intervention

For each data type, we constructed linear mixed effect models with Maaslin2.⁶⁴ Each model was constructed using the treatment group and the interaction between the treatment group (“Placebo”, “Low Dose 2′-FL” or “High Dose 2′-FL”) and the intervention phase (“Baseline” or “Intervention”) as fixed effects. Both baseline timepoints (Weeks −2 and 0) were used for the “Baseline” phase and both Weeks 3 and Week 6 were used for the “Intervention” phase in order to include as many samples as possible in the regressions. Subject identity was used as a random effect to control for auto-correlation due to repeated measures of the same individual. Age and gender were included in the models as covariates. Features were deemed to vary significantly by treatment arm if the interaction term had a Benjamini-Hochberg-corrected p-value (q-value) less than 0.05. Individual post hoc tests for participants in each treatment arm were then performed separately for each significantly varying feature using Wilcoxon rank-sum tests between Baseline and Intervention samples. For 16S data, we used a centered-log ratio transformation on the input data matrix. For the metabolomics and proteomics data, we used a log2 transformation on the input data matrix.

Centering and scaling data

In analyses comparing different data types to each other, parameters were centered and scaled by column for the purposes of data regularization. All methods describing data as centered and scaled were done using the scale function in base R (scale function with parameters: center = TRUE, scale = TRUE).

Identifying significant correlations in serum proteomics and serum metabolomics datasets

First, centered and scaled each feature from the serum proteomics (Olink), serum untargeted metabolomics and clinical panel datasets. We imputed missing values in the data using the “impute.knn” function from the R package impute (v1.7). Then, we determined the change in Z score from baseline-to-end for each participant. We then calculated an all-versus-all Pearson correlation for all features using the function “rcorr” from the R package Hmisc (v4.7). We calculated the Benjamini-Hochberg corrected p-value for each correlation and filtered the correlation matrix to contain only correlations with adjusted p-values <0.05. We then generated a correlation network from this filtered correlation matrix using the R package igraph (v1.3.4).

Additional resources

Clinical trial registry #NCT03690999: https://www.clinicaltrials.gov/ct2/show/NCT03690999.

Published: July 29, 2025