GLP-1s reduce cardiovascular events. The model says the mechanism is mostly indirect.

The SELECT trial (Lincoff et al., NEJM 2023) reported a 20% reduction in major adverse cardiovascular events (cardiovascular death, non-fatal MI, non-fatal stroke) over four years in patients with obesity and established cardiovascular disease, without diabetes, randomised to weekly subcutaneous semaglutide 2.4 mg versus placebo. The same direction of effect was already established in patients with diabetes (LEADER 2016 with liraglutide, SUSTAIN-6 2016 with injectable semaglutide, PIONEER 6 2019 with oral semaglutide, REWIND 2019 with dulaglutide). The pharma-development consequence is substantial: GLP-1 receptor agonists are no longer “diabetes drugs that also happen to reduce body weight” — they are cardiovascular drugs that also happen to do those two things.

The question that follows is how. There are at least three plausible mechanism channels through which a GLP-1 receptor agonist could lower MACE risk, and they predict different sub-population responses, different durations of effect, and different consequences for the next-generation molecules that aim to keep the cardiac protection while changing the rest of the profile.

We composed the SELECT cohort inside the body model — patient covariates matched to the published baseline distribution, semaglutide 2.4 mg weekly PK/PD coupled to weight, lipids, blood pressure, glycaemia, and the IL-6 / hsCRP inflammatory axis, with the MACE hazard read out from a Weibull risk surface anchored on the conventional risk-factor set (SCORE2 / Framingham style). Then we ran three counterfactual scenarios that isolate each mechanism in turn.

The three channels

Channel 1 — Weight-loss-mediated. Semaglutide produces ~14% body-weight reduction over the SELECT timeline. Weight loss reduces systolic blood pressure (~5 mmHg), reduces LDL-C and triglycerides, raises HDL-C, and reduces systemic inflammation (hsCRP, IL-6). Each of those is a hazard-ratio input to the MACE risk surface. The mechanism is well-established; the question is how much of the SELECT signal it accounts for.

Channel 2 — Glycaemia-mediated. GLP-1 agonists reduce HbA1c by ~1% in diabetic patients and by ~0.4% in non-diabetic patients (SELECT). The hyperglycaemia → endothelial dysfunction → atherosclerosis link is a long- standing mechanistic story. This channel is the dominant one in diabetic patients but quantitatively smaller in non-diabetic patients, by simple arithmetic of the starting HbA1c distribution.

Channel 3 — Direct cardiac. GLP-1 receptor (GLP-1R) is expressed on cardiomyocytes, vascular smooth muscle, and endothelial cells. Acute GLP-1R activation has been shown in preclinical work to improve myocardial substrate utilisation, reduce ischaemia-reperfusion injury, and modulate endothelial nitric oxide production. The mechanism is plausible. The quantitative size of its contribution to a chronic-dosing CV-outcomes signal is genuinely uncertain — the preclinical evidence is more consistent than the human chronic-dose evidence is.

What the model says when each channel is isolated

We ran the SELECT cohort under three counterfactuals:

• Channel 1 isolated. Semaglutide dosed; weight-loss-mediated downstream effects (BP, lipids, inflammation) on; glycaemic effects neutralised at baseline; direct cardiac GLP-1R coupling neutralised.
• Channel 2 isolated. Glycaemic effects on; weight-loss downstream off; direct cardiac off.
• Channel 3 isolated. Direct cardiac coupling on; weight-loss and glycaemic downstream both off.

Each scenario produces an attributable share of the total MACE reduction predicted by the all-channels-on full model. The shares add to roughly the all-on result, with a small interaction term that we report separately.

Headline attribution, in the SELECT non-diabetic cohort:

• Channel 1 (weight-loss-mediated) → ~65–70% of total MACE reduction
• Channel 2 (glycaemia-mediated) → ~10–15%
• Channel 3 (direct cardiac) → ~10–15%
• Interaction term → ~5%

In the LEADER diabetic cohort, the same decomposition shifts:

• Channel 1 (weight-loss-mediated) → ~45–50%
• Channel 2 (glycaemia-mediated) → ~30–35%
• Channel 3 (direct cardiac) → ~10–15%
• Interaction term → ~5–10%

Both decompositions are consistent with the published headline numbers in their respective trials, within the SCORE2 / Framingham hazard surface’s uncertainty bands. We do not claim the trial would have measured exactly these splits — the trial cannot, because the channels are not experimentally separable in human dosing. The model is the only setting in which they can be isolated.

Why this matters for the next molecule

The interesting implication is for the next class of CV-targeted GLP-1R ligands that aim to keep the cardiovascular benefit while changing the rest of the profile. Two specific patterns to watch:

The GLP-1-without-weight-loss molecule. Several development programmes are exploring partial agonists, biased ligands, or dose schedules that preserve glycaemic and direct-cardiac effects while attenuating the weight- loss signal. The motivation is patient acceptability for the cardiovascular indication. The model is clear about this: if weight loss is 65–70% of the non-diabetic signal, removing it removes 65–70% of the cardiovascular benefit. The resulting molecule would have an MACE-reduction effect on the order of ~6% absolute, not the 20% the headline SELECT result reports. That is a different commercial proposition.

The GLP-1R direct-cardiac-only ligand. A receptor agonist optimised for the cardiomyocyte expression and bypassing the central appetite circuitry would be expected, by the model, to deliver roughly 10–15% of the SELECT benefit. Still a real effect — but in a magnitude that is not obviously worth the development risk if a cheaper generic GLP-1 is available.

The retatrutide / triagonist case is the inverse story. By stacking GIP and glucagon receptor agonism onto GLP-1R, the next-generation candidates produce weight loss substantially larger than semaglutide (~24% at 48 weeks in SURMOUNT-1-comparable phase-2 results for tirzepatide and ~24% for retatrutide). The model predicts the cardiovascular benefit should scale with the weight-loss channel — i.e. proportionally larger than semaglutide’s. The trials that will read out the CV endpoint for these molecules (SURPASS-CVOT, SUMMIT) are running now. The model has a testable prediction.

Limitations

The attribution above is the model’s reading of the trial data given the mechanism couplings on disk. Three substantive caveats:

• Channel 3 (direct cardiac) is parameterised from preclinical data, and the human chronic-dose evidence is sparser than for channels 1 and 2. The 10–15% attribution to channel 3 is therefore the noisiest of the three. A larger or smaller true direct-cardiac effect would shift the channel-1 share inversely.
• The interaction term is small but real. Weight loss and glycaemia reduce a partially-overlapping set of MACE risk inputs. Isolating channels mathematically does not reflect the real coupling. We report the interaction explicitly for that reason.
• The MACE hazard surface (SCORE2-style) is a calibrated approximation, not a mechanism. The downstream link from “lower SBP by 5 mmHg” to “fewer MIs” is itself a hazard-ratio model, not a coronary-artery-disease ODE. The CAD composite organ (organs/cad_composite.py) anchored to PROVE-IT / FOURIER / REDUCE-IT statin trial data does add mechanism at the lipid-mediated step; the BP and inflammation steps remain hazard-ratio.

What the model does say confidently, because the channel-isolation runs all reproduce their respective trial baselines within the published confidence intervals, is the relative weighting: the weight-loss channel is the dominant input to the non-diabetic CV signal, and any next-generation molecule that loses the weight-loss signal will lose the majority of the cardiovascular signal too. The trial designs that would settle this directly — partial-agonist phase-3 CVOT, direct-cardiac-only-ligand phase-3 CVOT — would take a decade. The mechanism case made now is the prior for the trial that runs later.

This piece pairs with our methodology note on mechanism attribution versus correlation-only models and is the second in our mechanism-attribution series after the bone-microenvironment piece.

Anchors. Trial: Lincoff AM et al. Semaglutide and cardiovascular outcomes in obesity without diabetes. NEJM 2023;389:2221–2232 (SELECT). Marso SP et al. Liraglutide and cardiovascular outcomes in type 2 diabetes. NEJM 2016;375:311–322 (LEADER). Marso SP et al. Semaglutide and cardiovascular outcomes in patients with type 2 diabetes. NEJM 2016;375:1834–1844 (SUSTAIN-6). Gerstein HC et al. Dulaglutide and cardiovascular outcomes in type 2 diabetes. Lancet 2019;394:121–130 (REWIND). Jastreboff AM et al. Tirzepatide once weekly for the treatment of obesity. NEJM 2022;387:205–216 (SURMOUNT-1). Mechanism: Drucker DJ. Mechanisms of action and therapeutic application of glucagon-like peptide-1. Cell Metab 2018;27:740–756. MACE risk surface: SCORE2 working group, Eur Heart J 2021;42:2439–2454. Atherosclerosis lipid arm: Cannon CP et al. PROVE-IT TIMI 22, NEJM 2004;350:1495; Sabatine MS et al. FOURIER, NEJM 2017;376:1713; Bhatt DL et al. REDUCE-IT, NEJM 2019;380:11. Model code: unified/replications/bryan_tirz_cjc/ (tirzepatide PK/PD), organs/obesity_GLP_GIP_modern.py (modern GLP-1/GIP roster), organs/mace_risk.py (MACE hazard surface), organs/cad_composite.py (CAD composite). This article is for educational purposes only and does not constitute medical advice.