Pelacarsen, Lp(a)HORIZON, 8,323 patients, and how a mechanistic model earns credibility

Nova In Silico announced the results of its prospective simulation of the Lp(a)HORIZON Phase III trial of pelacarsen, deposited on Zenodo under restricted access on June 2, 2026, before the topline results were known.

A manuscript describing the simulation results of the Lp(a)HORIZON trial is under peer review. The simulation anticipated the direction of the outcome. We are sharing it to make a case about sequence: when a mechanistic model is built early, it can inform which trials we commit to, and that is a question about patients before it is a question about cost.

What the trial asked of people

Lp(a)HORIZON enrolled 8,323 patients with elevated lipoprotein(a) and established cardiovascular disease. It ran across 902 sites in 43 countries. Enrollment opened on December 12, 2019, and the study completed on July 16, 2026. That is six years and seven months, and tens of thousands of patient-years.

The trial produced knowledge that no model can produce alone

Lp(a)HORIZON is the first Phase III outcomes trial designed to test directly whether lowering Lp(a) pharmacologically reduces cardiovascular events. Before the Lp(a)HORIZON trial, there was no clinical evidence tying Lp(a) reduction to event reduction. For LDL-C, decades of trials established that relationship quantitatively. For Lp(a), the question was open.

On September 4, 2026, Novartis reported that the trial did not meet its primary endpoint, a composite of cardiovascular death, non-fatal myocardial infarction, non-fatal stroke, and urgent coronary revascularization requiring hospitalization (4P-MACE). Lower Lp(a) levels were achieved with pelacarsen. Those lower levels did not translate into a reduction in the primary endpoint in the overall study population.

This is a real result, and the field is better for having it. Running the trial was a reasonable decision on the evidence available in 2019. Reporting a null result clearly is a service to every team working on Lp(a).

What the model estimated, and when

Nova simulated the trial independently, using a mechanistic quantitative systems pharmacology model of atherosclerotic cardiovascular disease. The model was calibrated on FOURIER data and evaluated against ODYSSEY-OUTCOMES and the Copenhagen General Population Study.

The simulation predicted only a modest difference in 4P-MACE between the pelacarsen and placebo arms, with no statistically significant treatment effect expected.

We deposited this on Zenodo under restricted access on June 2, 2026, which time-stamps the prediction almost 100 days before the topline announcement. The deposit provided a time-stamped record without making the underlying results publicly accessible. The deposit came six weeks before the trial was completed. It was a test of the method against an unknown answer. It was not, and could not have been, an input to any decision about this trial. The trial was almost over.

The topline release does not yet contain enough quantitative detail for a numerical comparison against our estimates. What we can say today is that the direction was consistent. The detailed results will be the real test, and we will report that comparison whichever way it falls.

Why a mechanistic model can speak to a question with no precedent

A statistical model learns the relationship between a biomarker and an outcome from trials that have already run. For Lp(a) before Lp(a)HORIZON, no such trial existed, so there was nothing to learn from.

A mechanistic model starts elsewhere. It represents the causal chain: how the drug changes the biomarker, how that biomarker enters plaque biology, how plaque burden accumulates, and how that accumulation produces events in a given population. Each link rests on physiology and pharmacology that was already known. That structure is what lets the model produce a quantitative expectation for an endpoint never measured for this target. Nova's quantitative simulation predicted that despite a substantial change in the biomarker, clinical events would change only modestly.

The two approaches answer different questions and are stronger together. Where they disagree, the disagreement is informative.

What this result changes for the next program

Lp(a)HORIZON could not reasonably have been designed differently. In 2019, there was no clinical evidence linking Lp(a) lowering to cardiovascular events, and no model of that link that had been tested against an outcome. The question was open, the hypothesis was well founded, and answering it required a trial of this size and duration. Nothing that follows suggests these patients should have been studied differently. Their participation produced the answer the field needed. It also produced the test that tells us what our model is worth.

What is different is what exists now that did not exist then. A mechanistic model of this disease has been compared against a Phase III outcome it had not seen. They do mean that a team facing a comparable question no longer has to begin where the pelacarsen team began.

That is where the value lands, and it lands forward. For the next program in this area, a model can be brought to the table while the design is still open, when there is room to ask whether a better design is available and what it would need to demonstrate. The useful point is: the question can now be examined quantitatively, and examined early.

The loss this can prevent is economic, and it is also ethical. A Phase III outcomes trial spends money, and it spends something scarcer. Patients with established cardiovascular disease give years of their remaining life to a study on the understanding that the question is worth asking. That trust is finite, and the field draws on it continuously. Reducing the number of programs that commit to a large outcomes trial without a tested quantitative expectation is not only a way to spend less. It is a way to be more careful with what patients give us.

Where simulation belongs in the sequence

Clinical trials establish efficacy and safety. Nothing replaces them, and we are not proposing anything of the kind.

What we are proposing is that simulation belongs earlier in the sequence than it usually sits. A mechanistic model is most useful at the point where a program is still deciding what to test, because that is when its output can still change something. Used there, it supports target assessment, dose selection, endpoint choice, follow-up duration, and enrollment criteria.

Used late, as ours was here, a model can only be graded. Grading is useful, and it is how a method earns the right to be used early. But the value to patients comes from the early use.

A thorough credibility assessment

The credibility of the model has previously been assessed through multiple complementary lines of evidence, including model plausibility, calibration goodness-of-fit, population-based retrospective validation and robustness analyses, consistent with the framework outlined in the FDA guidance on assessing the credibility of computational modeling and simulation.

More recently, the model successfully predicted the effect of enlicitide on LDL-C observed in the Phase III CORALreef Lipids trial. The prediction was made and released before the clinical results were known, providing a prospective test of the model's predictive performance. The prospective validation results were subsequently published in Atherosclerosis in April 2026.

Hence, Lp(a)HORIZON is the second time Nova has put a prediction on the record before a Phase III readout in cardiovascular disease. The Lp(a)HORIZON simulation is at an earlier stage: the prediction is deposited and time-stamped on Zenodo, and a manuscript describing the simulation results is under peer review. The currently available topline results do not provide sufficient quantitative detail to establish a direct numerical comparison between the model prediction and the observed trial outcome. A complete assessment of predictive performance will therefore only be possible once the detailed trial results are released.

Putting the prediction on the record before the readout is the point. A model that is fitted after the fact can be made to agree with anything. A prediction with a time stamp cannot be revised, which makes it a fair test, including when it is wrong. We intend to keep working this way, and to report the misses with the same detail as the hits.

Building confidence in model-informed drug development

Regulatory agencies increasingly support model-informed drug development, including QSP models. The open question is how much confidence a given simulation deserves when it informs a development or regulatory decision.

Prospective validation is one honest answer. It tests a model against outcomes it could not have seen, and it exposes limitations as clearly as strengths. We would like to see it become a normal expectation for any model presented in support of a decision, including ours.

Building on its mechanistic modeling expertise, Nova is now using agentic AI through Jinkō's connections with Anthropic and OpenAI to expand the scope and speed of scientific investigation. Agents help scientists explore more hypotheses, execute simulation workflows and compare alternative trial scenarios. The underlying models remain explicit and inspectable, while scientists define the questions, approve assumptions and retain ownership of the scientific work and responsibility for its conclusions. This combination of greater scientific capacity, traceability and human accountability is central to Nova's strategy.

About Nova In Silico

Nova In Silico is a privately owned company founded in 2024, specializing in Digital Twins for clinical research. Nova has delivered more than 80 modeling and simulation projects across oncology, rare diseases, and cardiovascular conditions. Nova builds mechanistic models that make trial design decisions inspectable and traceable, so that development teams can examine the assumptions behind a projected outcome before committing to it.

About Jinkō

Jinkō is Nova's platform for mechanistic modeling and In Silico trials. It combines white-box mechanistic disease models, grounded in physiology and pharmacology, with virtual populations and simulated trial protocols. Every model, protocol, virtual population, and trial is a versioned object that a reviewer can open and inspect. Jinkō is agent-ready and headless, so modeling teams can drive it through a shared workspace or through the SDK.

Media contact

References

1. Peyronnet E, Wang Y, Béchet E. Prospective prediction of cardiovascular outcomes for HORIZON-Lp(a) trial using a mechanistic QSP model of ASCVD. Preprint, version 6, June 2, 2026.
2. Novartis. Novartis announces Lp(a)HORIZON Phase III topline results for pelacarsen in patients with elevated Lp(a) and established cardiovascular disease. September 4, 2026.
3. Assessing the Impact of Lipoprotein (a) Lowering With Pelacarsen (TQJ230) on Major Cardiovascular Events in Patients With CVD. ClinicalTrials.gov, NCT04023552.
4. Wang Y, Courcelles E, Peyronnet E, Porte S, Diatchenko A, Jacob E, Angoulvant D, Amarenco P, Boccara F, Cariou B, Mahé G, Steg PG, Bastien A, Portal L, Boissel JP, Granjeon-Noriot S, Bechet E. Credibility assessment of a mechanistic model of atherosclerosis to predict cardiovascular outcomes under lipid-lowering therapy. NPJ Digit Med. 2025;8(1):171. doi:10.1038/s41746-025-01557-7.
5. FDA. Assessing the Credibility of Computational Modeling and Simulation in Medical Device Submissions. November 2023.
6. Cariou B, Wang Y, Angoulvant D, Boccara F, Mahé G, Steg PG, Bechet E. In Silico prediction of the lipid-lowering efficacy of the oral PCSK9 inhibitor enlicitide using a mechanistic computational model. Atherosclerosis. 2026;415:120682. doi:10.1016/j.atherosclerosis.2026.120682.
7. Nova In Silico. Nova In Silico prospectively predicts LDL-C outcomes of Merck's Phase 3 CORALreef Lipids trial using an In Silico clinical trial. November 14, 2025.