Working Paper · Nutritional Biochemistry
Where the LDL hypothesis holds, where it fails as a decision rule, and what else HMG-CoA reductase inhibition switches off.
The popular heterodox position is that LDL cholesterol is a harmless bystander and statins are a profitable fraud. That position is wrong, and this paper does not take it. The genetic evidence that apoB-containing lipoproteins cause atherosclerosis is among the strongest causal evidence in all of chronic-disease epidemiology, and statins genuinely prevent heart attacks in people who have already had one. Any argument that requires denying those two facts is building on sand.
The interesting critique is a different one, and it survives the concession. A variable can be causal and still be a poor decision rule. LDL-C is causal at the population level, weakly informative at the individual level, and routinely treated as though it were the disease rather than one input to it. The paper develops four consequences of that gap. First, the benefit of statins is almost always quoted as a relative risk reduction — roughly 21–29% — while the absolute risk reduction in pooled trial data is on the order of 0.8% for all-cause mortality and 1.3% for myocardial infarction, a distinction that decides the drug for most people in primary prevention. Second, the relationship between total cholesterol and all-cause mortality is U-shaped, so “lower is better” is not the shape of the underlying curve. Third, statins do not inhibit cholesterol synthesis; they inhibit HMG-CoA reductase, which sits above a branching pathway that also produces coenzyme Q10, dolichol, and the farnesyl and geranylgeranyl groups that anchor a large class of signalling proteins to membranes — so the drug's effects are broader than its name suggests, in both directions. Fourth, the documented harms divide sharply on inspection: the diabetes signal is real, dose-dependent, and mechanistically on-target, while the muscle-symptom signal is mostly nocebo, which blinded n-of-1 trials establish cleanly and which the heterodox literature almost never concedes.
The paper then argues the positive case: that serum lipids are substantially a readout of metabolic terrain rather than an independent lever, which is why the same LDL number carries different risk in an insulin-resistant patient and a lean athlete. That framing is developed into the interventions with the best mechanistic and trial support — reducing linoleic-acid load, restoring insulin sensitivity, dietary nitrate for blood pressure, and above all exercise, where a randomised trial using serial intravascular ultrasound found coronary atheroma regressed under supervised interval training while it progressed under guideline care, with no lipid target anywhere in the intervention. That last result is the paper's strongest evidence that terrain moves the disease and not merely its markers — and, because it was delivered on top of standard therapy rather than instead of it, also a caution against reading any of this as a case for stopping treatment. Evidence is graded inline; the strongest counterargument is stated at full strength before it is answered. This is not medical advice, and nothing here is a reason to stop a prescribed medication.
Eight steps, one per part
Not “LDL doesn't matter” — that LDL-C is a real cause and a bad dashboard, and that treating the dashboard has costs the dashboard doesn't show.
This paper is an argument about evidence and mechanism. It is not medical advice, and it is not a reason to stop a prescribed statin. The distinction that matters most is the one between secondary prevention — people who have already had a heart attack, stroke, or revascularisation — and primary prevention in people who have not.
In secondary prevention the absolute benefit is large, replicated, and includes all-cause mortality;4 the critique developed below does not apply there with anything like the same force, and stopping abruptly after an acute coronary event is associated with harm. Familial hypercholesterolaemia is a second case where the argument does not transfer. If any of this changes how you think about a medication you are taking, the next step is a conversation with the person who prescribed it, not a unilateral change.
The heterodox side of this literature is a small, interconnected, and commercially interested network, and it is named rather than laundered. The immediate prompt for this paper was a post by Ashley Armstrong (@farmer_ash_), who farms and sells low-PUFA animal foods and who cited two real papers accurately: the CTT Collaboration's 2024 individual-participant analysis of statins and new-onset diabetes,17 and a NHANES analysis finding a U-shaped total-cholesterol/mortality curve.6 Both are engaged directly below. Uffe Ravnskov and THINCS, Malcolm Kendrick, Aseem Malhotra, and Nicholas Norwitz supply much of the rest of the sceptical case; Norwitz's KETO-CTA work is funded through the low-carb research ecosystem and one of its papers has been retracted, which is stated at point of use rather than buried. On the other side, the CTT Collaboration holds the trial-level data and has been criticised — fairly, in my view — for not releasing individual participant data for independent re-analysis, and most large statin trials were industry-funded. Neither side here is disinterested. Where a claim rests on advocacy synthesis rather than a study that demonstrates it end-to-end, it is flagged.
The argument runs in the main text and needs no biochemistry background. Blocks marked ▷ In the weeds (blue left-border) are optional enzyme-level detail; each is preceded by a plain-language takeaway that carries the essential point. Evidence is graded inline with the same fixed vocabulary used across this series:
A grade attaches to a specific claim, never to a whole section.
Almost every argument in this field goes wrong in the first ten seconds, because the phrase “LDL cholesterol” is doing three jobs at once and none of them well. Ten minutes here makes the rest of the paper follow almost mechanically.
Cholesterol is a rigid four-ring lipid that the body makes — roughly 70–80% of the cholesterol in your body was synthesised by your own cells, mostly the liver, and only the remainder came from food. It is not soluble in blood. To move it, the liver packages it with triglycerides, phospholipids, and a structural protein into a lipoprotein particle: a boat.
Each such boat carries exactly one molecule of apolipoprotein B (apoB), and that protein stays with the particle from the moment it leaves the liver until it is cleared. VLDL, IDL, LDL and Lp(a) are all apoB particles — different stages and variants of the same vessel. So a measurement of apoB is a count of particles. A measurement of LDL-C is a measure of how much cholesterol those particles happen to be carrying.
The mass of cholesterol carried inside LDL particles, per decilitre of blood. On most standard panels it is not even measured — it is calculated by the Friedewald equation from total cholesterol, HDL-C and triglycerides, an estimate that degrades badly when triglycerides are high or very low.
A direct count of atherogenic particles, one protein per particle. Measured, not inferred. When apoB and LDL-C disagree — discordance — risk tracks apoB.8
Everything in one bucket: LDL-C plus HDL-C plus roughly a fifth of triglycerides. It is the crudest of the three, and it is the number nearly all of the classical epidemiology — including Keys's — was actually built on.
The mental model most people carry — cholesterol as grease coating the inside of a pipe — is wrong in a way that matters. Atherosclerosis begins inside the arterial wall, not on its surface. An apoB particle crosses the endothelium, and if it binds to the proteoglycan matrix underneath it is retained there. Retained particles are chemically modified, principally by oxidation. Modified particles are recognised by macrophages through scavenger receptors that — crucially — are not down-regulated by intracellular cholesterol load, so the macrophage keeps eating until it becomes a foam cell and dies. That debris is the plaque core.
Two independent things govern the rate of this chain: how many particles arrive, and how readily they are modified once retained. The mainstream model optimises the first term. The argument in the linoleic acid paper is that the second term is modifiable too — the oxidisability of a particle is set by the fatty acids esterified into it, and those come from the diet.
StrongRetention of apoB particles in the subendothelial matrix is the initiating step in atherogenesis; this is not in dispute in either camp.
InferredThat dietary fatty-acid composition meaningfully shifts the oxidisability term in vivo to a degree that changes clinical events. Each link is documented; the chain has not been demonstrated end-to-end in an outcome trial.
No figure in nutrition is more caricatured, in both directions. Getting him right matters here for a specific reason: the modern LDL hypothesis is routinely attacked through Keys, and if that attack is factually wrong — which it largely is — then leaning on it damages the credibility of the better arguments that follow.
Keys's position, built through the 1950s and formalised in the Seven Countries Study, was a chain: dietary saturated fat raises serum total cholesterol, and populations with higher mean serum cholesterol have higher coronary mortality. Note what is and is not in that sentence. It is a claim about total cholesterol, at the level of populations, and it predates the ability to measure LDL at all — the Friedewald equation that made routine LDL-C estimation possible was not published until 1972, by which time Keys was near the end of his career.
This is the single most-repeated claim in the anti-Keys literature, and it does not survive contact with the primary sources. The 22-country graph comes from Yerushalmy and Hilleboe's 1957 critique of an earlier Keys paper — it is the critics' graph, built from UN food-balance-sheet data, not Keys's.11 The Seven Countries Study never had 22 countries to discard: it was a prospective cohort study, designed from the outset around seven countries chosen for feasibility and for spread along the exposure gradient, with investigators on the ground and actual blood draws in sixteen cohorts.1012 You cannot drop what you never enrolled.
Three real problems remain, and they are enough. (1) Seven Countries was ecological in its headline correlations — comparing population means — and ecological correlations do not establish individual-level causation. (2) Keys attributed the Cretan advantage to fat composition, but the Cretan cohort was examined partly during Lent, when a substantial share of the population was fasting from animal foods; the dietary characterisation is on weaker ground than the mortality data. (3) Most importantly, the framework confounded saturated fat with everything else that distinguished mid-century Finland from mid-century Crete — and the diet-heart hypothesis it launched was operationalised as replace saturated fat with polyunsaturated vegetable oil, a recommendation whose own controlled trials — the Minnesota Coronary Experiment and the Sydney Diet Heart Study — lowered cholesterol and did not reduce mortality, with recovered data suggesting the opposite.13
StrongThe 22-countries cherry-picking story is a misattribution; the graph is Yerushalmy & Hilleboe's.
ModerateThat the recovered MCE and Sydney data undermine the specific saturated-fat→PUFA substitution recommendation. The trials are real and the recovered analyses are published, but both were conducted with hydrogenated-oil-era formulations and have their own methodological critics.
This is the part of the paper that concedes. It is placed before the critique deliberately, because an argument that has to hide this evidence is not worth making.
Observational epidemiology cannot separate LDL-C from the dozens of things that travel with it. Mendelian randomization can, and this is its best application. Alleles are assigned at conception, before any confounder can act, and they are fixed for life. So if you compare people who inherited variants that give them lifelong slightly-lower LDL-C against those who did not, you have something close to a randomised trial of lifetime exposure.
The answer is consistent across more than a dozen independent loci — PCSK9, HMGCR, NPC1L1, LDLR, ABCG5/8 — and it does not depend on which gene does the lowering. Lower lifetime apoB, lower lifetime coronary disease, proportional to the magnitude and the duration of exposure.1 The effect per unit of LDL-C in the genetic data is roughly threefold larger than in statin trials, which is exactly what a cumulative-exposure model predicts: five years of drug in middle age cannot do what fifty years of genetics does.
The strength of the design is that it is a naturally randomised instrument with a known mechanism. The standard objections to MR are pleiotropy (the variant affects the outcome through some other path) and linkage (the variant is a marker for a nearby causal one). Both are addressed here by the convergence across unlinked loci acting through different proteins: PCSK9 governs LDL-receptor recycling, NPC1L1 governs intestinal sterol absorption, HMGCR governs synthesis. They share nothing except the downstream particle concentration, and they give the same answer. That convergence is the reason this evidence is hard to dismiss, and it is why the strong form of “LDL is a bystander” is untenable.
Heterozygous FH — roughly 1 in 250 people — means a lifetime of LDL-C around 200–400 mg/dL from a single defective LDL-receptor allele, with no accompanying insulin resistance, hypertension or inflammation. Untreated, it produces coronary disease decades early. Homozygous FH, with LDL-C above 500 mg/dL, produces it in childhood. This is isolated apoB elevation without metabolic confounding, and it is unambiguously atherogenic.
Any model in which LDL/apoB is merely a passive marker has to explain FH, and none of them do. ApoB is causal. The rest of this paper is written on that assumption, not against it.
The CTT Collaboration's pooled analyses, covering roughly 170,000 participants across 26 trials, find a consistent ~21–22% proportional reduction in major vascular events per 1.0 mmol/L (≈39 mg/dL) of LDL-C lowering, and that proportionality holds across baseline risk strata, sexes, and starting LDL levels.23 In 4S — secondary prevention, high baseline risk — simvastatin cut all-cause mortality by 30% over 5.4 years.4 That is a real, large, replicated effect on death, and it is not explained away.
StrongApoB-containing lipoproteins are causal in atherosclerosis; statins reduce major vascular events proportionally to LDL-C lowering; secondary-prevention mortality benefit is real.
Everything in Part II is true. None of it establishes that LDL-C is the right thing to manage in an individual patient, and the gap between those two propositions is where this paper lives.
“Statins cut your risk of a heart attack by a third” is true and almost useless, because a third of a small number is a smaller number. Byrne and colleagues pooled 21 trials and reported both metrics side by side; the contrast is the single most decision-relevant table in the field.5
The same data, two presentations. A 1.3% absolute reduction in myocardial infarction over the trial durations means roughly 77 people take the drug for one to avoid a heart attack; the other 76 get the side-effect profile and none of the benefit. Whether that trade is worth making is a genuine value judgement that depends on baseline risk — which is the point. It is a judgement the relative figure quietly makes on the patient's behalf.
The CTT investigators would say Byrne's pooled ARR is an artefact of averaging over heterogeneous baseline risk and short follow-up. Both objections are legitimate. Proportional risk reduction is the stable quantity across trials; absolute reduction is derived from it by multiplying against the patient's own baseline risk, and pooling ARR across populations with wildly different baselines produces a number that describes no actual patient. And because atherosclerosis is cumulative, five-year trial ARRs systematically understate the benefit of thirty years of treatment — the same cumulative-exposure logic that makes the MR estimates larger.
The reply to the reply: both points are correct and neither rescues the clinical practice. If ARR must be computed per-patient from baseline risk, then quoting a population RRR to an individual is precisely the error, and guidelines built on 10-year risk calculators should be producing individualised absolute numbers — which, in practice, consultations rarely do. And the extrapolation from 5-year trials to 30-year treatment is an assumption, not a finding; it is also the extrapolation that makes the cumulative harms in Part V matter more, not less. You cannot claim the long horizon for benefits and the short one for risks.
“Lower is better” is a claim about the shape of a curve, and for all-cause mortality the shape is not monotonic. A NHANES analysis of roughly 30,000 adults using restricted cubic splines found U-shaped associations between total cholesterol and all-cause, cardiovascular, and cancer mortality. The lower limb was steep: total cholesterol below ~120 mg/dL strongly predicted elevated all-cause mortality, while values at or above 280 mg/dL predicted elevated mortality only for the cardiovascular cause specifically.6
Reverse causation is the first thing to raise here and it is real: advanced cancer, liver failure, malabsorption, frailty and chronic infection all lower cholesterol, so sick people have low numbers because they are sick. Serious analyses address this by excluding early deaths and adjusting for baseline illness, and a residual U generally survives — but it is attenuated, and how much survives is genuinely Contested. What the U-curve licenses is a modest conclusion: that the population-level dose-response for all-cause mortality is not monotonic, so “as low as possible” is an extrapolation past the data rather than a reading of it. What it does not license is the inference that lowering LDL pharmacologically moves you up the left-hand limb — the trials show no such mortality signal.
The related finding in the elderly is on similar footing: a systematic review of cohorts covering 68,000 people over 60 found LDL-C inversely associated with all-cause mortality in most of them.7 That paper's authors are the sceptical camp's core group and the review has been criticised for selective inclusion; it is cited here as Contested and load-bearing on nothing.
This is the deepest version of the critique. LDL-C is one variable in a risk function that also contains insulin resistance, triglyceride/HDL ratio, blood pressure, inflammation, and visceral adiposity — and those covary. In the general population, high LDL-C usually arrives bundled with the rest of the metabolic syndrome, and the bundle is what the epidemiology is actually measuring.
The natural test is to find people with isolated high LDL-C and an otherwise pristine metabolic profile. The lean mass hyper-responder phenotype — LDL-C ≥200, HDL-C ≥80, triglycerides ≤70, typically lean and on a ketogenic diet — is exactly that population, and the KETO-CTA study followed 100 of them with serial coronary CT angiography for a year. In that cohort, neither apoB nor cumulative LDL-C exposure predicted plaque progression; baseline plaque did.9
Handle with real caution. (1) A JACC: Advances paper reporting these results was retracted — that fact belongs in any honest citation of this work. (2) There was no control group; the design cannot separate “LDL doesn't matter here” from “one year is too short to detect it.” (3) The headline finding cuts both ways: median non-calcified plaque volume rose by about 37% relatively over the year, which is not a reassuring absolute result. (4) A null within a cohort selected for a narrow, uniformly high LDL range has little power to detect a dose-response — restriction of range produces nulls mechanically. (5) The funding and the investigators are openly aligned with the conclusion.
What it does establish, weakly: that in this phenotype, one year of apoB around 180 mg/dL did not produce plaque progression proportional to that exposure, and that baseline disease dominated. That is a hypothesis worth testing properly, not a refutation of Part II.
SpeculativeThat metabolically healthy isolated LDL-C elevation carries materially lower risk than the same number in an insulin-resistant person. Mechanistically plausible, consistent with discordance analyses, and not established by outcome data.
Set the terrain argument aside entirely; LDL-C still underperforms. Where apoB and LDL-C disagree, risk follows apoB — across meta-analyses and discordance analyses, consistently.8 Discordance is not rare: it is concentrated precisely in insulin-resistant and hypertriglyceridaemic people, who carry many small cholesterol-poor particles and whose LDL-C therefore understates their particle burden. The people most likely to be falsely reassured by a normal LDL-C are the ones at highest metabolic risk.
Statins are described as cholesterol-lowering drugs. That is a description of an effect, not of a mechanism, and the difference explains nearly everything about their side-effect profile.
A statin inhibits HMG-CoA reductase, the rate-limiting enzyme converting HMG-CoA to mevalonate. Cholesterol is downstream of mevalonate — but so is a whole isoprenoid tree. Clamping the trunk reduces every branch.
Every branch below mevalonate is reduced by the same inhibition.14 This is not a fringe claim — it is standard pharmacology, and it is also the accepted explanation for statins' beneficial pleiotropic effects. Reduced geranylgeranylation of Rho and Rac is why statins lower CRP and stabilise endothelium independently of lipid lowering. The same mechanism produces both columns of the ledger.
Coenzyme Q10 carries electrons between Complex I/II and Complex III of the respiratory chain, and is a lipid-phase antioxidant. A meta-analysis of placebo-controlled statin arms found plasma CoQ10 falls by a weighted mean of −0.44 µmol/L.15 That depletion is not in question.
Two honest gaps. First, plasma CoQ10 travels on lipoproteins — so some of the measured fall is simply fewer carrier particles, not less tissue CoQ10. Muscle-biopsy data are sparser and less consistent than the plasma data. Second, if CoQ10 depletion caused statin myopathy, supplementation should fix it, and the trials are genuinely split: several meta-analyses report significant reduction in statin-associated muscle symptoms, others report no benefit over placebo.16 Given Part V's nocebo finding, that split is unsurprising — a supplement cannot fix a symptom the drug isn't causing.
StrongStatins reduce circulating CoQ10; the mevalonate pathway makes this obligatory.
ContestedThat CoQ10 depletion causes clinically meaningful mitochondrial dysfunction, and that supplementation relieves statin muscle symptoms.
The sceptical literature treats statin side effects as a single undifferentiated mass. They are not. On close inspection one major claim gets stronger than its advocates usually argue, and another collapses almost entirely — and conceding the second is what earns the right to press the first.
This is the finding from the post that prompted this paper, and it holds up. The CTT Collaboration's individual-participant meta-analysis pooled 19 placebo-controlled trials (123,940 participants) and four intensity-comparison trials (30,724 participants).17 Among participants without diabetes at baseline:
Two features make this more than a statistical curiosity. It is dose-dependent — a 10% excess at moderate intensity, 36% at high intensity — which is the signature of a causal drug effect rather than confounding. And it is on-target: Mendelian randomization studies of HMGCR variants find the same association between genetically lower HMG-CoA reductase activity and higher type 2 diabetes risk, meaning the diabetes effect flows through the drug's intended target, not some incidental property of a particular molecule. There is no statin that escapes it.
The average glycaemic effect was small, and most of the new diagnoses occurred in people whose baseline glycaemia was already close to the diagnostic threshold. In large part the drug is nudging people over a line they were already standing on, rather than producing diabetes in the metabolically healthy. Whether that is reassuring depends on your view of thresholds; the mechanism — impaired β-cell insulin secretion and reduced peripheral insulin sensitivity — is the same either way, and crossing the line still carries the clinical consequences of the diagnosis. The CTT authors' own conclusion is that the cardiovascular benefit outweighs this risk; that conclusion is a value judgement laid over the data, and it will be right for high-risk patients and wrong for some low-risk ones.
StrongStatins increase new-onset diabetes incidence, dose-dependently, through on-target HMGCR inhibition.
Here the sceptical case largely fails, and the failure is instructive. Muscle pain is the most-cited reason people stop statins, and the obvious causal story — mevalonate depletion in myocytes — is mechanistically tidy. The problem is that blinded n-of-1 trials test it directly.
SAMSON gave 60 patients who had previously abandoned statins for side effects a randomised sequence of twelve one-month periods: four on atorvastatin, four on placebo, four on nothing. Symptom burden on placebo was statistically indistinguishable from symptom burden on the statin, and both were markedly higher than during the no-tablet months. Roughly 90% of the symptom burden attributed to the drug was reproduced by placebo.18 When shown their own data, half the participants successfully restarted a statin. StatinWISE, a larger independent series of n-of-1 trials in UK primary care, reached the same conclusion.19
Three caveats are legitimate. (1) Nocebo is not imaginary — the pain is real; what the trials establish is that the drug molecule is not producing it. (2) These trials enrolled people who had already stopped statins for symptoms, a population enriched for expectation effects; they do not exclude a smaller group with genuine pharmacological myopathy. (3) Rhabdomyolysis is unambiguously drug-caused, and objectively-measured myopathy with raised creatine kinase is a real, if uncommon, entity — CTT estimates roughly 1 excess case of myopathy per 10,000 treated per year.
But the honest summary is that the bulk of what is attributed to statins in popular discussion is expectation, not pharmacology. Any critique that leads with muscle pain is leading with its weakest card, and sophisticated readers will discount everything after it.
StrongMost statin-attributed muscle symptoms are nocebo; blinded n-of-1 designs establish this directly.
Because every steroid hormone is built from cholesterol (Part VI), the inference that statins lower testosterone is natural. Meta-analyses do find a statistically significant reduction, but a small one — and steroidogenic tissue is not supply-limited by circulating cholesterol under ordinary conditions; the adrenal and gonad can up-regulate LDL-receptor uptake and local synthesis. The signal is real and probably clinically minor for most people. Moderate
On cognition, the FDA added a label warning in 2012 on the strength of case reports, but randomised data — including PROSPER and the Heart Protection Study, both with formal cognitive endpoints — have been essentially null. Brain cholesterol is synthesised locally behind the blood-brain barrier and turns over slowly; lipophilic statins (simvastatin, lovastatin) penetrate more than hydrophilic ones (pravastatin, rosuvastatin), which is a plausible basis for individual variation but not for a population effect. Contested, and weaker than the sceptical literature presents it.
Two questions get conflated constantly: what does cholesterol do, and why is mine what it is. They have different answers, and III.3 issued an instruction — ask why a number is elevated before asking how to lower it — that this paper owes an actual answer to. VI.1 covers the function; VI.2 covers the regulation.
The framing of cholesterol as a pathogen to be minimised obscures that it is one of the most functionally central molecules in the body. None of what follows argues that lowering LDL-C harms these functions — in a healthy person, it mostly doesn't, because cells synthesise their own. It argues something narrower and more useful: that the body's insistence on making 70–80% of its own supply, under tight feedback control, is information about how much it needs.
| Function | Tier | What cholesterol does there |
|---|---|---|
| Membrane architecture | Universal | Intercalates between phospholipid tails, setting membrane fluidity and thickness. Concentrated with sphingolipids into lipid rafts — ordered platforms where receptors cluster to signal. Insulin-receptor signalling, immune-synapse formation and neurotransmitter receptor function are all raft-dependent. |
| Steroid hormones | Obligate | The sole precursor. Cholesterol is imported into mitochondria by StAR and cleaved by CYP11A1 to pregnenolone — the parent of progesterone, cortisol, aldosterone, DHEA, testosterone and oestradiol. There is no alternative substrate. |
| Bile acids | Obligate | CYP7A1 converts cholesterol to cholic and chenodeoxycholic acid. This is the body's main disposal route for cholesterol — and those bile acids are what emulsify dietary fat and enable absorption of vitamins A, D, E and K. They are also signalling molecules via FXR and TGR5. |
| Vitamin D | Obligate | 7-dehydrocholesterol in the skin is cleaved by UVB to cholecalciferol — the direct link to the sunlight paper. The precursor is a cholesterol intermediate one step from the end of the synthesis pathway. |
| Myelin & the brain | Critical | The brain is ~2% of body mass and holds ~20–25% of the body's cholesterol, most of it in myelin. CNS cholesterol is synthesised locally — lipoproteins do not cross the blood-brain barrier — which is the main reason the cognitive-harm hypothesis is weaker than it looks. |
| Cell division | Obligate | A dividing cell must build two membranes' worth of material, and cholesterol synthesis is required to complete the cell cycle. Separately, cytokinesis requires geranylgeranylated RhoA — a mevalonate-pathway product, not cholesterol itself. This is the basis for the (entirely separate) research programme on statins as antiproliferative agents in oncology. |
| Hedgehog signalling | Developmental | Sonic hedgehog is covalently modified with cholesterol; the pathway is central to embryonic patterning. Severe inherited defects in cholesterol synthesis (Smith-Lemli-Opitz syndrome) cause profound developmental abnormalities — the clearest natural evidence that cholesterol synthesis is not optional. |
It would be easy, and wrong, to run from this list to “so lowering cholesterol must damage all of these.” It generally does not, for a concrete reason: these are intracellular functions served by local synthesis and receptor-mediated uptake, both of which are homeostatically defended. Statin-treated patients do not develop adrenal insufficiency or demyelination. What the table actually supports is narrower: that a molecule under this much regulatory control is not plausibly something the body wants as little of as possible, and that the burden of proof on “lower indefinitely” is higher than it is usually made to carry. Smith-Lemli-Opitz is the proof of principle for catastrophic deficiency, not a model of statin therapy.
Here is the fact that reorganises the whole question: cholesterol synthesis is under negative feedback, so the body makes more precisely when cells sense they have too little. The sensor is the SCAP–Insig complex in the endoplasmic reticulum membrane. When ER cholesterol falls, SCAP escorts SREBP-2 to the Golgi, where two proteases release its transcription-factor domain; that fragment enters the nucleus and switches on both HMG-CoA reductase (make more) and the LDL receptor (import more). When cholesterol is replete, Insig retains the complex in the ER and the programme shuts off.32
Why this explains the drug: a statin lowers serum LDL only indirectly. By blocking HMG-CoA reductase it depletes hepatocyte cholesterol, which trips this very sensor, which upregulates LDL receptors, which pull LDL out of the blood. The drug works through the feedback loop, not around it — which is also why the effect plateaus, and why adding ezetimibe or a PCSK9 inhibitor (both of which act on other points in the same loop) stacks on top of it.
Once the loop is clear, “high cholesterol” stops being one condition. It is at least four, with genuinely different meanings:
| Why the number is up | Tier | What is actually happening |
|---|---|---|
| Impaired clearance | Most common | LDL-C is a concentration — production minus clearance — and receptor activity is usually the dominant term. Hypothyroidism is the textbook case: T3 upregulates LDL-receptor transcription, so low thyroid means fewer receptors and a higher number with no change in synthesis.33 PCSK9 activity does the same by degrading the receptor, and familial hypercholesterolaemia is the genetic version. Here the number is a clearance readout, and the useful response is to find out why clearance is impaired. |
| Genuine demand | Physiological | Proliferation and tissue repair consume membrane, and membrane requires cholesterol; steroidogenic tissue under load draws on it. The biggest lever here is bile acid loss: excrete bile acids and the liver must synthesise replacements from cholesterol, depleting hepatic stores, tripping SREBP-2, raising LDL-receptor expression — which is why the number falls. This is the mechanism of every soluble fibre and every bile-acid sequestrant. |
| Energy trafficking | Context-dependent | Under carbohydrate restriction, VLDL turnover rises to shuttle fat as the dominant fuel, and LDL-C can rise as a transport consequence rather than a disease state — the lean-mass hyper-responder phenotype of III.3. Whether this carries the same risk as the same number arrived at by impaired clearance is exactly the open question that paper section flags, and nobody knows. |
| Immune / acute phase | Points the other way | Lipoproteins are genuinely part of innate immunity — LDL, VLDL, HDL and chylomicrons all bind and neutralise bacterial LPS, and LPS-binding protein circulates with apoB particles to hand endotoxin off for hepatic clearance.34 But the direction matters: in human endotoxaemia and sepsis the observed pattern is rising triglycerides with falling LDL-C and HDL-C.35 Infection lowers cholesterol; it does not raise it. |
A popular version of this argument runs: raw carrot lowers cholesterol; carrot has a mild antibiotic effect on gut flora; therefore endotoxin drives cholesterol up, and cholesterol is the body's defence against toxins (a framing associated with Ray Peat). The underlying study is real — Robertson et al. 1979, 200 g raw carrot at breakfast for three weeks, serum cholesterol down about 11%.36 The reasoning on top of it is not.
The study measured its own mechanism. The same abstract reports a 50% increase in faecal bile acid and fat excretion — that is row two of the table above, the ordinary soluble-fibre pathway shared by psyllium, oat beta-glucan and cholestyramine. The "change in bacterial flora or metabolism" line the claim quotes is the authors speculating about something else entirely: why the effect persisted three weeks after stopping. The popular version promotes that aside to primary mechanism and discards the measured one.
And it names the wrong molecule. In the innate-immunity literature, LPS-neutralising capacity tracks the phospholipid content of lipoproteins — cholesterol and triglyceride content specifically do not correlate with it.34 Whatever protection lipoproteins provide belongs to the particle's phospholipid surface, not to the cholesterol it is carrying. "Cholesterol protects against toxins" credits the cargo with the work of the boat.
What survives is better than what was claimed, and worth stating in its own right: the reason a very low number can be a bad sign is not that cholesterol is a toxin-sponge. It is that cholesterol is largely a downstream readout, and the states that drive it very low — inflammation, infection, liver dysfunction, malabsorption, hyperthyroidism, frailty — are themselves harmful. That is the honest mechanism behind the U-shaped curve in III.2, and it is the same terrain logic as the rest of this paper: the gauge reads low for bad reasons as well as good ones, and neither the high end nor the low end interprets itself.
StrongThe SREBP-2/SCAP–Insig feedback loop and its control of HMGCR and LDLR; thyroid hormone's regulation of LDL-receptor expression; lipoprotein binding of LPS.
ModerateThat the four categories above are cleanly separable in an individual patient. In practice they overlap, and the common case — metabolic syndrome — involves several at once.
The positive claim of this paper: for most people outside FH and secondary prevention, serum lipids are substantially a readout of metabolic state, and the readout moves when the state moves. Most of this part is the most speculative material in the paper and is labelled accordingly — with one exception, VII.3, which carries the only direct anatomical evidence anywhere in it.
Part of the Primer's chain that mainstream practice leaves alone: an LDL particle's vulnerability to oxidative modification is set by the polyunsaturated fatty acids esterified into it, and those come from the diet. Controlled feeding studies show that substituting oleate for linoleate makes LDL measurably more resistant to oxidation.20 The full argument, with its counterarguments, is the subject of the linoleic acid paper and is not repeated here.
InferredThat reducing dietary LA lowers clinical cardiovascular events via reduced LDL oxidisability. The oxidation-resistance step is demonstrated in humans; the event step is not.
Restoring insulin sensitivity improves nearly every term in the risk function simultaneously — triglycerides, HDL, particle size and number, blood pressure, inflammation, visceral fat. The interventions are the ones in the metabolic health paper: carbohydrate quality and load, resistance training and muscle mass as a glucose sink, sleep, and visceral fat reduction. Unlike the lipid-lowering lever, this one moves the whole system rather than one gauge on it.
StrongThat improving insulin sensitivity improves the full metabolic risk profile including lipid subfractions. Moderate that this translates into event reduction comparable to pharmacological LDL lowering in a matched population — the head-to-head trial has not been done.
Everything else in this part rests on surrogate endpoints: oxidation resistance, blood pressure, insulin sensitivity. This section does not, and that makes it the single strongest piece of evidence the terrain argument has.
The CERT trial randomised 60 patients with stable coronary disease following PCI to either six months of supervised high-intensity interval training — 4 × 4-minute intervals at 85–95% of peak heart rate, twice weekly — or to contemporary preventive guidelines, and measured coronary plaque directly by intravascular ultrasound at baseline and follow-up.22 The plaque regressed.
Read the last two columns together: the control group's plaque drifted upward as untreated atherosclerosis does, and the exercising group's went backwards. Normalised total atheroma volume fell by 9 mm³ in the HIIT arm against a 3 mm³ rise in controls. This is disease regression on a direct anatomical measurement, in a randomised design.
(1) It is not exercise instead of statins — it is exercise on top of them. This is the most important caveat and the one most likely to be dropped in transmission. Both arms were post-PCI patients under contemporary secondary-prevention guidelines, which means both arms were, in the main, statin-treated. CERT demonstrates an additive effect on a background of optimal medical therapy. It says nothing whatsoever about substituting one for the other. (2) n = 60, six months. Small and short; plaque volume, not events. Regression of atheroma is strongly associated with reduced coronary events but is still a surrogate. (3) The contrast is attenuated by design — the control group received guideline-based care including cardiac rehabilitation, so this is high-intensity training versus some training, not versus nothing. That makes the observed difference more impressive, not less, but it also means the effect size does not transfer to a sedentary comparison. (4) It was supervised, and the authors say so plainly — they doubt the achieved intensity is reachable without continuous supervision. The intervention that worked is not the intervention most people will actually do.
Note also what was tested: 85–95% of peak heart rate. This is not Zone 2 steady-state work, and the distinction matters when citing this trial. Zone 2 training has its own strong rationale — mitochondrial density, fat oxidation, sustainability — but it is not what produced this result.
StrongSupervised HIIT produced coronary atheroma regression versus guideline care in post-PCI patients, by IVUS, in a randomised trial.
InferredThat this plaque regression translates into reduced events, and that it generalises to primary prevention or to unsupervised training at lower intensity. Neither was tested.
Blood pressure is an independent and arguably larger contributor to cardiovascular risk than LDL-C, and two food interventions have enough randomised data to be worth stating precisely. They are usually lumped together as “antioxidant juices.” They should not be — they act on different nodes of the same axis, which is the only reason it is worth discussing both.
Beetroot supplies substrate. Nitrate from beetroot and leafy greens is reduced to nitrite by oral commensal bacteria, then to nitric oxide in tissue — an endothelium-independent route to vasodilation that bypasses the eNOS pathway damaged in metabolic disease.
The practical corollary: antibacterial mouthwash abolishes the effect by killing the nitrate-reducing commensals. This is one of the cleanest demonstrations that an oral-microbiome function has a direct cardiovascular readout.
Meta-analyses of randomised trials find inorganic nitrate and beetroot juice lower systolic blood pressure by roughly 3–5 mmHg, with larger effects in hypertensive populations (−5.3 mmHg in one analysis restricted to them).2123 That is a modest effect per person and a large one per population — roughly the magnitude at which epidemiological models project meaningful reductions in stroke and ischaemic heart disease mortality. Dose and practical notes are in the beet juice entry.
Pomegranate restrains the opposing system. Where beetroot adds vasodilator substrate, pomegranate acts on the vasoconstrictor arm — the renin–angiotensin system — and on eNOS itself. Pomegranate's ellagitannins inhibit angiotensin-converting enzyme, the same target as the “-pril” drug class. An early human study reported a 36% reduction in serum ACE activity alongside a 5% systolic fall,26 and the specific inhibitory compounds have since been isolated and characterised — pedunculagin, punicalin and gallagic acid, with IC50 values around 0.9–1.8 µM, apparently binding the catalytic zinc.27 Punicalagin additionally raises NO output by activating eNOS, so the two mechanisms push the same direction.
Because they act on opposite arms of the same control loop, there is a reasonable mechanistic case that the two are additive rather than redundant. Speculative — I could find no trial that tested them in combination, so this is an inference from mechanism, not a result.
The pomegranate trial data are real but weaker than beetroot's. A meta-analysis of 8 RCTs found systolic pressure fell −4.96 mmHg (95% CI −7.67 to −2.25) and diastolic −2.01 mmHg (−3.71 to −0.31).24 A later analysis of 14 trials (n = 573) found a near-identical systolic effect of −5.02 mmHg — but with two findings that complicate the picture considerably: the effect was concentrated at doses of ≤300 mL/day (larger doses did less for systolic pressure), and the benefit was lost after two months of continued intake.25
(1) The attenuation finding is the important one and it is rarely quoted. If the effect genuinely disappears by month three, then pomegranate is an acute vasoactive intervention rather than a treatment for chronic hypertension, and every enthusiastic summary that stops at “−5 mmHg” is describing a transient. It may be a statistical artefact of fewer, smaller long-duration trials — but it is what the data currently say. (2) The two meta-analyses share a senior author (Sahebkar), so they are not the independent replication they superficially appear to be. (3) The provenance is unusually compromised. POM Wonderful funded a very large share of the pomegranate literature — reportedly over $35 million across a hundred-plus studies — including much of Michael Aviram's foundational work, and the FTC found the company's cardiovascular and cancer advertising claims deceptive, a ruling upheld on appeal in 2015. That does not make the ACE-inhibition biochemistry wrong; it does mean the clinical literature was assembled by an interested party, and should be discounted accordingly.
A 240 mL serving of pomegranate juice carries roughly 30 g of sugar and ~130 kcal; beet juice is lower but not trivial. This paper's central positive claim (VII.2) is that insulin sensitivity is the upstream variable. Recommending a daily glass of concentrated fruit sugar to an insulin-resistant person in order to capture a few mmHg is close to self-contradictory, and I would rather name that tension than route around it. The resolution, such as it is: whole pomegranate arils carry the ellagitannins with the fibre intact, and nitrate is better obtained from leafy greens and whole beets than from juice. Juice is the form the trials used, not necessarily the form that makes sense.
StrongDietary nitrate lowers blood pressure in randomised trials. Moderate that pomegranate juice lowers blood pressure acutely — real meta-analytic effect, compromised provenance, apparent loss of effect past two months. Inferred that either reduces cardiovascular events; no outcome trial exists for either.
The claim now circulating widely — that sodium alone does not cause high blood pressure, and only a lack of potassium and magnesium does — is an overcorrection of a real insight, and it is worth separating the two. What is right: the sodium/potassium ratio does predict blood pressure and cardiovascular mortality better than sodium in isolation; potassium intake in the modern diet is genuinely poor; and the mechanisms usually given (potassium promoting natriuresis and vascular relaxation, magnesium supporting vasodilation) are broadly correct. What is wrong: the strong form is refuted by controlled feeding data. DASH-Sodium fed 412 people three sodium levels — 50, 100 and 150 mmol/day — in random sequence, within each diet arm, so potassium was held constant while sodium moved. Blood pressure fell stepwise as sodium fell, on both the high-potassium DASH diet and the control diet, and the two effects were additive and independent.31 Sodium reduction works even when potassium is already high. The honest statement is that both matter, the ratio captures them jointly, and potassium is the more neglected half — not that sodium is exonerated. The magnesium claim is also oversold at the popular level: the measured effect is about 2 mmHg, not a fix for hypertension, and the blanket assertion that supplementation is “usually necessary” has no support in the trial data.
Intellectual honesty requires naming the cases where the argument of this paper does not apply, and where the drug simply wins:
Six claims, in descending confidence
The disagreement worth having is not whether LDL causes atherosclerosis. It is whether a population-level cause makes a good individual-level target — and on that, the mainstream case is much weaker than its confidence suggests.
Limitations of this paper
What would change my mind: a properly controlled trial of the LMHR phenotype with multi-year imaging endpoints showing plaque progression proportional to apoB; or conversely, a primary-prevention trial reporting individualised absolute benefit alongside new-onset diabetes and quality-of-life endpoints, showing net benefit at low baseline risk. Both are feasible. Neither has been done.
Outstanding citation-audit items for the next draft: the author list and exact volume of 6 (the NHANES nonlinear-cholesterol paper) need confirming against the published version; the HMGCR Mendelian-randomization/diabetes literature cited in V.1 is currently asserted from the consensus position and needs a primary citation (Swerdlow et al., Lancet 2015 is the likely anchor); the PROSPER and HPS cognitive-endpoint claims in V.3 need direct citation rather than the summary given; the statin/testosterone meta-analysis referenced in V.3 is asserted without a citation and should either be sourced or removed; and 33 (thyroid hormone and LDL-receptor expression) is standard endocrinology asserted without a primary citation and needs one pinned. The KETO-CTA preprint at 9 should be re-checked for peer-reviewed publication before this draft is promoted past v1.