Breast Milk Biology: Cytokines, HMOs and Infant Growth

The question comes to me at least once a semester, usually from a clinician who has just read a splashy paper: "So which milk components actually drive infant growth?" Sometimes it's phrased more confidently — "we know HMOs program the microbiome, right?" — and sometimes it arrives as a grant aim already written, with a causal arrow drawn between a single oligosaccharide and weight-for-length z-score at twelve months.

My usual answer used to be some version of: milk is a complex, dynamic biological fluid, and the components co-vary, so single-analyte stories are fragile. That's true. It's also incomplete, and I've come to think the incompleteness does real damage, because it leaves people with the impression that the field is hopeless rather than that it's hard in specific, tractable ways.

So let me try to give the fuller answer I wish I'd given a second-year PhD student who emailed me last spring, asking whether her milk cytokine panel was "worth salvaging" after her supervisor told her the between-mother variance was too large.

What we're actually asking

The question underneath most of this work is whether human milk composition, beyond gross macronutrient delivery, has independent effects on how infants grow and how their immune systems develop. That splits into at least three questions that get conflated constantly:

• Does variation in milk composition within breastfed infants predict variation in growth and health outcomes?

• Do specific components have mechanistic, dose-dependent biological effects demonstrable outside the context of breastfeeding?

• Would supplying those components to non-breastfed infants reproduce any part of the benefit?

The first is an epidemiological question with an enormous confounding structure. The second is a bench and model-system question. The third is a trial question, and it's the one with the cleanest answers and the narrowest scope. People slide between them in a single paragraph, and reviewers let them.

The mechanisms, briefly and honestly

Human milk oligosaccharides

HMOs are unconjugated glycans, largely indigestible by the infant, structurally diverse — on the order of two hundred structures characterized, with a handful dominating by mass. Their principal proposed mechanisms are reasonably well supported: selective substrate for specific Bifidobacterium taxa (notably B. infantis and related strains with the gene clusters to import and cleave them), soluble decoy receptors for pathogens and toxins, and direct signaling at the intestinal epithelium and on immune cells.

Maternal secretor status, driven by FUT2, is the single largest structural determinant: non-secretor mothers produce little or no 2'-fucosyllactose. That genetic switch is the closest thing the field has to a natural experiment, and it's been used well — and also over-interpreted, because FUT2 genotype affects the infant's own mucosal glycosylation and susceptibility to several enteric pathogens independently of milk.

Cytokines and immune factors

Milk contains secretory IgA, lactoferrin, lysozyme, growth factors including TGF-β and EGF, and measurable concentrations of cytokines and chemokines. The plausible mechanisms are local: immune education at the gut mucosa, modulation of epithelial barrier maturation, tolerance induction toward food and commensal antigens. TGF-β in particular has a coherent story around IgA class switching and oral tolerance.

What I'd push back on is the implicit systemic-signaling model that creeps into discussion sections — the idea that maternal cytokines circulate in the infant and tune growth centrally. Proteolysis in the infant gut is substantial, absorption of intact cytokines is limited and variable with gestational age, and the quantitative case for meaningful systemic exposure is not made. Milk hormones such as leptin and adiponectin are a more defensible route to systemic effects, and even there the concentrations are low relative to endogenous production.

Evidence, graded

Strongest: randomized trials of formula supplementation

The firmest ground in this whole literature is the set of randomized controlled trials adding synthesized HMOs — typically 2'-FL alone or with lacto-N-neotetraose, more recently five- and six-HMO blends — to infant formula. These trials consistently demonstrate safety and tolerance, and growth equivalence to control formula. Several report shifts in fecal microbiota composition and metabolites toward a breastfed-like profile, and some report reductions in reported infections, antibiotic use or respiratory morbidity.

Read those secondary outcomes carefully. They're frequently parent-reported, frequently not the pre-specified primary endpoint, and the trials are powered for growth equivalence rather than morbidity. I've reviewed manuscripts where a nominally significant reduction in one of eleven reported infection categories carried the abstract. The honest summary is: supplementation is safe, it moves microbial intermediates in the expected direction, and the clinical outcome evidence is suggestive and not yet decisive. Meta-analyses of these trials have reached broadly that conclusion, and a PubMed search for HMO formula randomized trials will show you how quickly the base is growing.

There is also the cluster-randomized breastfeeding promotion literature — PROBIT and its long follow-ups — which remains the best causal evidence that breastfeeding itself affects some outcomes. Note what it shows: effects on gastrointestinal infection and atopic eczema in infancy, and on cognitive measures, with weak or null effects on adiposity and blood pressure in later childhood. That null-for-adiposity finding deserves more weight than it gets in grant introductions.

Intermediate: large observational cohorts

Mother–infant cohorts with serial milk sampling have reported associations between individual HMOs and infant growth trajectories, body composition, and risk of allergic outcomes. The recurring findings worth taking seriously are that HMO profiles — secretor status, Lewis status, clustered composition — associate with infant microbiome assembly more robustly than any single structure associates with growth, and that associations with anthropometry are small, inconsistent in direction across cohorts, and sensitive to covariate choice.

I've watched this play out in person. A postdoc I worked with found a clean inverse association between a fucosylated HMO and weight-for-age z-score at six months. Beautiful figure. Then we stratified by secretor status, as we should have from the start, and the association collapsed into the secretor/non-secretor contrast — which in that cohort also tracked with maternal BMI and with site. She published it anyway, properly framed as a profile-level observation, and it's a better paper for it. The temptation to report the analyte-level result is strong because analyte-level results are what people cite.

The confounding structure here is brutal. Milk composition varies with lactation stage, time of day, within a single feed, maternal diet, BMI, parity, mode of delivery, geography and season. Infant growth varies with all the same maternal and social factors, plus feeding mode, which is itself only ever partly measured. Reverse causation is live: infant intake and suckling behavior influence milk composition, so a correlation between a milk component and a growing infant can run in either direction.

Cytokine measurements in cohorts carry an additional burden. Assay platforms differ, many analytes sit near or below the limit of detection, matrix effects in milk are nontrivial, and freeze–thaw history is rarely reported in enough detail. Between-study comparability is poor. If you want to see the spread of this literature, a search on milk cytokines and infant growth cohorts is instructive mostly for how heterogeneous the methods are.

Early and mechanistic work

Gnotobiotic mouse and piglet models, organoid systems, and in vitro fermentation have carried most of the mechanistic weight. Colonization of germ-free animals with infant-derived microbiota, with and without HMO exposure, has produced some of the most interesting results in the field, including effects on growth phenotypes in models of undernutrition. Microbiota-directed complementary food work in children with malnutrition grew out of that lineage, and it's a genuine translational success story — though it's about complementary feeding rather than milk composition per se.

Epithelial and immune-cell work showing direct HMO signaling — effects on barrier integrity, on dendritic cell maturation, on cytokine production — is real but mostly at concentrations and in isolation conditions that don't obviously map onto the infant gut lumen. I'd treat these as proof-of-possibility rather than as evidence of operating mechanism.

What I'd call established versus emerging

Established, in my reading: HMO composition is substantially genetically determined and shapes early Bifidobacterium colonization; added HMOs in formula are safe and support normal growth; breastfeeding reduces gastrointestinal infection in infancy; milk sIgA and lactoferrin have defensible local protective roles.

Emerging, and I'd hold loosely: that specific HMOs have independent, clinically meaningful effects on infant adiposity or linear growth; that milk cytokine concentrations predict atopic or metabolic outcomes; that HMO supplementation reduces clinically confirmed infection or allergy incidence; that any of this programs long-term metabolic trajectories.

Genuinely contested: whether HMO effects are mediated through the microbiome at all, or substantially through direct host signaling. Whether milk should be analyzed as a system of correlated components rather than as a set of candidate analytes. I'm firmly in the systems camp, and I think the analyte-by-analyte approach has absorbed an enormous amount of effort for a modest return.

Practical things I'd change in study design

Report your sampling protocol in painful detail — full expression or partial, foremilk or hindmilk or pooled, time since last feed, time of day, lactation day, storage temperature and freeze–thaw count. I've refereed papers where none of this was recoverable, and the measurements become uninterpretable.

Pre-register the secretor-status stratification. It's not a sensitivity analysis; it's the primary structure of the data.

Collect intake volume if you can, even approximately, by test weighing in a subsample. Concentration without volume is not dose, and almost every paper that discusses "exposure" is actually discussing concentration.

Treat growth outcomes with the same care you'd give an adult cardiometabolic endpoint. Weight-for-length z-score at a single time point is noisy; conditional growth models and body composition by a validated method are worth the extra cost. And decide in advance whether your hypothesis is about growth velocity, attained size, or body composition, because they behave differently.

Finally, build the measurement comparison into your protocol. If you're running a multiplex cytokine panel, run a bridging subset against a second platform. Reviewers increasingly ask, and the answer shapes how much your effect estimates can be pooled later. The EQUATOR Network reporting guidelines are worth checking before you write up, not after.

Keeping up with a literature that moves this fast

One practical frustration I hear constantly from students is the gap between reading and writing: you find forty relevant papers, build a mental model, and by the time you're drafting the discussion you've lost the thread of which finding came from which design. I've started pointing people toward tools that keep searching and reading connected to the writing rather than treating them as separate chores. PACR is an AI-native research workspace built around that continuity — literature discovery, summarizing, drafting and collaboration in one place, so the context you built while reading is still there when you're arguing a point. For a field like this one, where the whole game is keeping trial evidence, cohort evidence and animal work mentally separated, that matters more than it sounds.

What I'd watch next

Three things have my attention. Trials with microbiologically or clinically confirmed infection endpoints rather than caregiver report, properly powered — the field needs one or two of these more than it needs another compositional survey. Work on HMO utilization capacity in the infant microbiome, since the same HMO exposure in an infant without B. infantis-type glycan machinery is a different exposure entirely, and that may explain much of the cohort heterogeneity, particularly between high- and low-income settings. And donor milk and preterm populations, where the clinical stakes are highest and where milk processing — pasteurization effects on cytokines, bioactive proteins and enzyme activity — is itself an unresolved variable.

If you're starting a project in this space, my advice to that second-year student stands: her cytokine panel wasn't worthless. It just couldn't answer the question she'd written. She reframed it around within-mother longitudinal change rather than between-mother comparison, and the variance problem became the signal.

That reframing is available more often than people think. Before you decide the data are too noisy, check whether you've asked the noise to carry a question it was never going to carry. You can set up a workspace and start mapping the trial-versus-cohort evidence for your own question in an afternoon — and you'll likely find the design gap faster than you'd find it in a reviewer's report.