From PubMed to Protocol: How Research Becomes a Recommendation

Filter One: What Kind of Study Is This

The first question is the category, since the categories are not interchangeable.

In vitro. Cells in a dish, at a concentration the experimenter chose. Establishes that a molecule can do something to a cell. Says nothing about a person.

Animal. Establishes an effect in a living organism, with the caveat that translation to humans has historically been poor across most fields, and that rodent metabolic rates and lifespans differ from ours in ways that make dose and duration non-transferable.

Human observational. Associations at population scale, vulnerable to confounding and reverse causation. Useful for generating hypotheses.

Human randomised. The category that supports a recommendation, with strength depending on size, duration, blinding, endpoint and replication.

Meta-analysis and systematic review. Aggregates randomised evidence, and its quality depends entirely on the trials included.

The failure mode here is counting a large number of in vitro papers as cumulative evidence. They are cumulative evidence of plausibility, which is a different claim.

Filter Two: Is the Concentration Achievable

This filter eliminates more supplement candidates than any other, and it is the one most often skipped.

An in vitro study reports the concentration used. Human pharmacokinetic studies report the plasma concentration achievable at an oral dose. If the second is orders of magnitude below the first, the mechanism does not transfer.

The recurring examples: resveratrol, extensively conjugated so free plasma levels sit far below experimental concentrations; quercetin and EGCG, similarly; plain curcumin, very poorly absorbed; oral butyrate, which largely does not reach the colon; and NAD+ itself, which does not cross cell membranes intact at all.

The exceptions that pass: creatine, which reaches muscle at achievable doses; omega-3, which incorporates into membranes measurably; bioavailability-enhanced curcumin formulations; and collagen peptides, where specific di- and tripeptides appear in plasma.

Applying this filter honestly removes most of the polyphenol category, and it does so without any judgement about whether the underlying science was good. The science was often excellent. The concentration was unreachable.

Filter Three: Does the Endpoint Matter

The validated-surrogate distinction is the substantive one. A surrogate is validated when changing it by a given means has been shown to change the outcome. Blood pressure and apoB qualify for cardiovascular endpoints. Homocysteine and HDL do not, since lowering and raising them respectively did not change events.

An unvalidated biomarker moving is therefore weak evidence, and it is the most common form of supplement trial evidence. A compound that lowered an inflammatory marker in a twelve-week trial has demonstrated something real and not that anyone will be healthier.

Filters Four Through Seven

Four: was the trial adequate? Size, duration, blinding, allocation concealment, pre-registration, intention-to-treat analysis, and whether the primary endpoint was the one reported. Outcome switching, reporting a secondary endpoint as if it were primary, is common and detectable by comparing publication to registration.

Five: who funded it, and has it replicated? Industry funding does not invalidate a trial and is associated with more favourable results across fields. Independent replication is what converts a finding into knowledge, and its absence across decades for a heavily marketed compound is itself informative.

Six: does the population match? A trial in people with low baseline status does not support use in the replete. A trial in patients with established disease does not support use in healthy adults. A trial in older adults does not transfer to younger ones. This filter is where most extrapolation goes wrong.

Seven: does the product match the trial? Formulation, dose, strain, standardisation and stability. Curcumin trials used enhanced forms, omega-3 trials specify EPA and DHA content, probiotic trials used named strains. A product using a different form is not the studied intervention.

A candidate surviving all seven is a defensible recommendation. Most do not reach filter four.

What Comes Out the Other End

Applied to the whole supplement field, this process produces a short list and a long set of exclusions, and the shape of the result is worth stating.

Compounds that survive with outcome or strong functional data: creatine, omega-3 at adequate dose, and correction of genuine shortfalls in vitamin D, iron, B12, zinc, magnesium and iodine. Fibre, which is barely a supplement. A handful of situational items including CoQ10 on a statin, inositol in PCOS, and specific probiotic strains for specific indications.

Compounds with reasonable but incomplete cases: collagen peptides, bioavailable curcumin, NAD+ precursors, urolithin A, ashwagandha, berberine with its interaction caveats.

Compounds tested and found wanting: ginkgo for cognitive decline, glucosamine and chondroitin, B vitamins for cardiovascular prevention, vitamin E, beta-carotene, tribulus, D-aspartic acid, high-dose antioxidant combinations.

Compounds where mechanism does not survive the bioavailability filter: most of the polyphenol category as sold.

That distribution is the honest output of an evidence process, and it explains why a rigorous system recommends less than a permissive one. The shortness of the first list is the finding, not a failure of the method.

Using This as a Reader

The full process requires access and time, and a compressed version is available to anyone.

When you see a claim, ask what kind of study it rests on. If the answer is a cell line or a mouse, the claim is about plausibility.

Ask whether the dose in the study matches the product. This alone resolves a great deal, particularly for omega-3 and curcumin.

Ask what was measured. A marker, a symptom score or an outcome, and whether the marker is validated.

Ask whether it replicated. A single positive trial is a reason to watch, not to act.

Ask whether the population resembles you. Baseline status and disease state are the usual mismatches.

Notice what a source excludes. Any source that has never concluded against a compound is not applying a filter, and that tells you more about its reliability than any individual claim it makes.

That last habit is the most useful. A reliable source has a list of things it recommends against, and the presence of that list is the signal.