Protein and blood cholesterol: what 32 trials show

Protein is bought and sold on muscle. Three meta-analyses published in 2026 looked instead at what it does to blood lipids, and between them they cover 58 randomised trials. The pooled numbers are real and statistically significant. They are also very small: about 3.45 mg/dL of LDL-cholesterol, roughly 0.09 mmol/L, in a population whose baseline LDL sits near 100 mg/dL.

The gap between significant and useful is the whole story for anyone specifying a protein system. It changes nothing in the purchase specification, and in the European Union it changes nothing on the label either: the authorised wording for protein does not mention cholesterol, and the thirteen authorised entries that do mention it belong to other substances, at daily doses between 0.8 and 10 grams. Below is what the three papers found, and what can and cannot be done with it.

What the three syntheses measured

Khosravi et al., Nutrition Reviews 2026. Thirty-two randomised controlled trials, 39 intervention arms, 1,357 participants (710 intervention, 647 control), all overweight or obese adults. The intervention node pools milk protein, casein, whey protein and soy protein into a single estimate. Outcomes are weighted mean differences in mg/dL for LDL-cholesterol, total cholesterol, triglycerides and HDL-cholesterol.

Beitsaeed et al., Diabetes & Metabolic Syndrome 2026. Twenty-one trials, a dose–response model for milk proteins including whey and casein, registered as PROSPERO CRD42024615328. It reports the change per 60 g/day increment rather than the effect of supplementation as such. That is a different quantity and it cannot be read against the first one unit for unit.

Zhao et al., Frontiers in Nutrition 2026. Five trials, 267 patients, whey protein only, after bariatric surgery. Doses ranged from 0.5 to 1.2 g/kg/day or 30 to 48 g/day; follow-up ran from one to six months across three surgical techniques.

Three populations, three questions. «Protein lowers cholesterol» is not one finding, and the chart below keeps the three apart on purpose.

How big the difference is

Converted into the units a clinician uses, the two significant estimates are −0.09 mmol/L of LDL and −0.15 mmol/L of total cholesterol. Both intervals exclude zero. Both sit close to the day-to-day variation of a routine lipid panel. Triglycerides and HDL did not move at all, and in the post-bariatric set no lipid moved.

The dose–response model is steeper per unit, but it asks for more material: −5.85 mg/dL of LDL (−8.87 to −2.83) and +3.39 mg/dL of HDL (1.46 to 5.33) for each additional 60 g/day of milk protein. Sixty grams a day is two full servings of a 30 g powder. The same paper reports suggested effective doses of 30 g for fat-free mass, 50 g for triglycerides and 40 g for HDL, and states that the relationship was non-linear for fat-free mass and triglycerides while remaining linear for the lipoproteins.

The post-bariatric trials are the useful contrast. They found nothing on lipids and something on body composition: fat-free mass +3.87 kg (1.35 to 6.39) and fat mass −2.10 kg (−3.92 to −0.27), both at low certainty. The authors’ own reading is that ingested whey is preferentially used for muscle protein accretion, so the lipid panel is the wrong place to look for its effect.

Where the effect stops

The most important line in the first paper is not a number. It is the sentence saying that the LDL and total cholesterol reductions «were attenuated when high-risk-of-bias studies were excluded». How far they were attenuated is not published in the abstract. An effect that starts at 3.45 mg/dL and shrinks when the weaker trials are removed is a signal worth following, not a property of the ingredient.

What it means for a formulator

A lipid result does not become a claim by raising the dose. The Union list of authorised health claims is a closed list. Cholesterol wording belongs to thirteen listed substances, each with its own conditions of use: 0.8 g/day of plant sterols or stanols, 3 g/day of beta-glucans from oats or barley, 10 mg/day of monacolin K from red yeast rice, 10 g/day of guar gum. If a brief says cholesterol, the answer is a listed claim carrier added to the system, with its own specification, its own maximum levels, its own allergen declaration and its own behaviour in the matrix. It is not more protein.

The grams in a dose–response curve are not the grams on a label. Sixty grams a day is a design target expressed in protein as the trials measured it. Kjeldahl and Dumas measure nitrogen, not protein, and the conversion factor is not the same for dairy (N × 6.38) and for soy (N × 6.25), while the EU requires the declaration to be made with N × 6.25 (Annex I, Reg. (EU) 1169/2011). A multi-source system specified with one factor and declared with another is not delivering the dose the trials used. What makes the figure defensible lot to lot is an amino acid profile by HPLC with a per-amino-acid acceptance criterion and free amino acids declared separately, not a line reading protein ≥ 80 %. That is the parameter a multi-fraction protein system has to carry in writing.

The claim threshold is a property of the finished food, not of the ingredient. All three authorised protein entries are conditional on the food being at least a source of protein: 12 % of its energy value from protein, and 20 % to be high in protein. That is an energy ratio, so the bulking agent, the flavour system and the sweetener move the claim exactly as the protein fraction does. The ingredient supplier answers for the declared content and the method that verifies it; the claim is the brand’s, as food business operator. Anyone sizing a first pallet against a claim should settle which of the two numbers the specification is being written to.

Limits of this evidence

Two of the three papers were read from the indexed abstract only. Neither Nutrition Reviews nor Diabetes & Metabolic Syndrome deposits these articles in PubMed Central, and both publishers serve the full text behind a paywall. Confidence intervals are published for every estimate plotted above, but I² values, risk-of-bias distributions, GRADE ratings and subgroup analyses by protein type are not. Only Zhao et al. was read in full.

One published interval cannot be read as printed. The triglyceride estimate in the post-bariatric set appears as 0.41 (14.52, 15.34) in both the results text and the GRADE table, a lower bound above the point estimate and missing its sign. It is reported here as published and left out of the chart.

Certainty is low throughout. Zhao et al. grade total cholesterol and HDL as low and LDL and triglycerides as very low, on five trials, one with a 52.9 % completion rate, and with some comparisons built from medians converted into means. Publication bias could not be assessed on five studies.

A pooled node is a category, not a formulation. Khosravi et al. combine milk protein, casein, whey and soy into one estimate, so the result describes a class of ingredients rather than any specification. Beitsaeed et al. report a slope per 60 g/day, which is a modelled quantity and not a dose any trial administered. None of the three papers reports a funding analysis of the included trials that could be read from the accessible text.