IBD and Precision Medicine: Not Yet

Tulisokibart works: the ARTEMIS-UC Phase 2 data

ARTEMIS-UC, the Phase 2 trial published in the New England Journal of Medicine in 2024, was a clear win. In the all-comers cohort, clinical remission at week 12 was 26% on tulisokibart versus 1% on placebo (a 25-point difference, P<0.001), and the drug beat placebo on every ranked secondary endpoint, from endoscopic improvement to mucosal healing. This was a refractory population, roughly half of them already failed by an advanced therapy. A 25-point remission gap in that setting is the kind of result most IBD programs never see.

So tulisokibart works. The harder question is for whom, and whether that can be known in advance.

Prometheus built a TL1A companion diagnostic

What Merck bought, first and foremost, was the antibody. Its acquisition announcement called PRA-023 (now MK-7240) "Prometheus' leading clinical candidate," a humanized monoclonal antibody, and dwelt on the positive ARTEMIS-UC and APOLLO-CD Phase 2 data. And the drug was scarce: in early 2023 only one other anti-TL1A antibody (Roivant/Pfizer's RVT-3101, now Roche's afimkibart) had Phase 2 IBD data.

But the quieter, more ambitious part of the story was the diagnostic: a companion test meant to flag, in advance, the patients whose disease is genuinely TL1A-driven, and baked into ARTEMIS-UC's design. The promise: a smaller, enriched trial; a cleaner signal; eventually a labeled test that tells a gastroenterologist who to give the drug to.

The trial was built to prove it. ARTEMIS-UC ran two cohorts (Cohort 1 enrolled patients regardless of test status, Cohort 2 enrolled only biomarker-positive patients) with prespecified, hierarchically-tested analyses in the pooled biomarker-positive group. It was a serious, well-designed attempt to do precision immunology properly.

The companion diagnostic reads genotype, not expression

The deployed test is, in fact, well documented. Per the trial's supplementary appendix, the companion diagnostic is a proprietary, noninvasive PCR genotyping assay (a Clinical Trial Assay) run on a buccal swab and scored in a CLIA-certified, CAP-accredited lab into a single binary positive-or-negative call. The genotypes it reads, and the machine-learning model that selected them, are laid out in US Patent 11,136,386 (Cedars-Sinai and Prometheus). The test is genetic, but the biology it is chasing is TL1A expression, and our reading below is how the two connect.

TNFSF15, the gene that encodes TL1A, is one of the strongest inherited risk loci in IBD: certain germline versions of it raise how much TL1A a person's immune cells produce (the TL1A target brief covers that biology). The patents go further, mapping what that elevated TL1A does: in cultured human Th17 cells, TL1A switches on a 166-gene program (76 genes up, 90 down, by microarray); patients with TL1A-driven disease carry elevated TL1A messenger RNA in blood monocytes and gut biopsies (by PCR); and the downstream result is a Th17 cytokine output of IL-17, IL-22, and IL-9. That cluster, high TL1A plus a switched-on Th17 program, is what a "TL1A-driven" patient looks like on paper, and it is the patient an anti-TL1A drug should help.

The test reads genotype rather than expression, because genotype is far easier to measure. Per US 11,136,386, a machine-learning model selected polymorphisms across TL1A-pathway and immune genes (TNFSF15, JAK2, ETS1, ICOSLG, and others); the assay then turns a patient's genotypes into one positive-or-negative call. In ARTEMIS-UC, 32 of the 135 Cohort 1 (all-comers) patients tested positive; pooled with the entirely test-positive Cohort 2, the biomarker-positive analysis group came to 75.

Our interpretation, reading the test against the biology, is this: the genotype is the readout; expression is the target. None of the selected polymorphisms measures TL1A directly; together they are a cheap, one-time, buccal-swab stand-in for the expression phenotype the trial actually cares about. Two patent families bracket that logic: "Signature of the TL1A Signaling Pathway" (WO2015010108, Cedars-Sinai) covers the expression side, and US 11,136,386 (Cedars-Sinai and Prometheus) covers the genotypes and the algorithm. What is not public is the trained model and the exact cutoff, the proprietary step that turns the genotypes into a yes or no.

Two caveats. First, what stays proprietary is not the markers but the math: the polymorphisms and the selection method are published, but the trained weights and the positive-or-negative cutoff are not. Second, as the next section shows, when the test was run, it did not cleanly separate responders from everyone else.

Did the biomarker actually work?

In biomarker-positive patients, clinical remission was 32% on tulisokibart versus 11% on placebo, the number that looked like validation. Two other lines in the same table complicate it.

First, the treatment effect was smaller in the biomarker-positive group, not larger. All-comers: a 25-point gap. Biomarker-positive: a 21-point gap (95% CI 2–38, P=0.02). The reason is the placebo arm: biomarker-positive patients on placebo remitted at 11%, versus 1% in the unselected cohort. The test selected patients who fare well overall, not patients for whom the drug beats placebo by more, and only the latter makes a companion diagnostic useful.

Second, in the biomarker-positive group the key endoscopic improvement endpoint missed significance (37% vs 19%, P=0.06). Under the trial's own hierarchical testing rule, that stopped the sequence, leaving every biomarker-positive endpoint below it formally exploratory, as the authors noted.

None of this makes the diagnostic wrong; it leaves it unproven. To the trial's credit, the biomarker analyses were prespecified and reported in full, the smaller subgroup effect and the borderline P=0.06 included, not buried.

Tulisokibart's Phase 3 enrolls all-comers

Tulisokibart's Phase 3 trials run all-comers. The UC trial (NCT06052059, ~1,020 patients; "ATLAS-UC") and the Crohn's trial (NCT06430801, ~1,200 patients) do not use the diagnostic to select patients. The UC trial's eligibility criteria carry no genetic or biomarker gate, and its primary endpoints are clinical remission in the unselected population. The primary hypothesis is the plainest possible: at least one tulisokibart dose level superior to placebo on clinical remission, in everyone.

On June 22, 2026, that bet delivered: Merck reported that the ATLAS-UC induction study met its primary endpoint (clinical remission by Modified Mayo Score at week 12) along with key secondary endpoints including endoscopic improvement and histologic-endoscopic mucosal improvement; Merck's Eliav Barr called them "the first positive Phase 3 induction results for an anti-TL1A biologic." The topline announcement reported the win, with the full remission rates to follow at an upcoming congress (standard practice). It did not reference the companion diagnostic.

That is a rational call. Merck did not set the diagnostic aside; it moved it from an entry criterion to a measured-within endpoint. The test survives as a prespecified secondary endpoint: among the UC trial's registered secondary outcomes are four that track "Diagnostic Assay Positive (Dx+)" participants (clinical remission and endoscopic improvement at weeks 12 and 52), measured alongside, not instead of, the all-comers primary. Even there, the registry describes the test only as "protocol-specific diagnostic assay criteria" applied at screening; as in the Phase 2 paper, its actual composition is undisclosed. So Merck is treating everyone, powering its registration on the unselected population, and keeping the precision question as a readout it can exercise later, a possible label-expansion argument rather than the foundation it was first designed to be. A broad label is worth more than a narrow enriched one, and you only restrict your market with a test if that test cleanly separates responders from non-responders. ARTEMIS-UC suggested it doesn't, so the diagnostic went from deciding who enters the trial to being measured inside it.

Merck has not spelled out why it set the diagnostic aside, but the reasoning is not hard to guess. For a biotech, a companion diagnostic is de-risking: a smaller, cheaper trial with a cleaner signal and a sharper story for investors and acquirers. For a big pharma that has just paid $10.8 billion, the same strategy looks mostly like downside: gating on a biomarker that did not clearly separate responders threatens the pivotal trial, narrows the market, and ties approval to a diagnostic that must clear the FDA on its own.

And it would have cost optionality. Merck did not buy a UC drug; it bought a TL1A franchise, and a companion test validated for one indication would have to be re-proven for the next. A genotype tuned to gut inflammation need not predict response in joint or fibrotic disease. All-comers keeps every indication open; the enrichment that helped Prometheus would mostly have boxed Merck in.

There is a clever detail here. Merck still runs the test on every Phase 3 patient; it just no longer uses the result to choose who gets in. That makes the Dx+ endpoint a low-risk experiment: test everyone, gate no one, and check afterward whether the test-positive patients did better. If they did, those results let Merck sharpen the test and bring it back for the Crohn's program or its next drug.

Merck's wider program points the same way. Beyond UC, tulisokibart is in Phase 3 for Crohn's disease and in Phase 2 across rheumatoid arthritis, psoriatic arthritis, axial spondyloarthritis, hidradenitis suppurativa, and systemic sclerosis: seven immune-mediated diseases in all. A company that viewed its companion diagnostic as the key to TL1A would be using it to enrich those trials; instead, Merck is betting that TL1A blockade works broadly, across diseases and across patients, which is the opposite of a precision strategy.

The anti-TL1A class converged on all-comers

Three TL1A events landed within two weeks of each other in June 2026: Merck's Phase 3 win, a best-in-class-looking Phase 2 from Spyre, and Bionyra's $165 million launch. Merck did not arrive at all-comers alone; its rivals never made the precision bet to begin with.

Afimkibart, the former Roivant/Telavant asset Roche acquired for $7.1B, posted 36% clinical remission at week 56 in the Phase 2b TUSCANY-2, run in an unselected population; it also reported an enhanced biomarker-positive subgroup (43% remission, 64% endoscopic improvement), so the precision question is alive, not dead. But Roche made the same call Merck did: its Phase 3 program (three UC and three Crohn's trials) enrolls all-comers, biomarker and all left out of the gate. Sanofi and Teva's duvakitug posted a positive all-comers Phase 2 in 2024.

And the newest entrants make the point most plainly, because they never even raised the biomarker question. In June 2026 Spyre's SPY002 posted a positive all-comers Phase 2 (SKYLINE), with a 10.7-point reduction in histologic disease activity (P<0.0001), 33% clinical remission, and 42% endoscopic improvement at week 12, and pitched itself as best-in-class not on a test but on a ~75-day half-life that could allow dosing once every three to six months. Days earlier, Bionyra Pharma launched on a $165 million Series A to chase the same target with half-life-extended and bispecific anti-TL1A antibodies. Across the anti-TL1A class, the competition is potency, durability, and convenience, not patient selection.

Why is precision medicine so hard in immune disease?

In oncology, a single driver (HER2 amplification, an EGFR mutation, BRAF V600E) gives a clean binary: the patients with it respond, the patients without it largely don't. That is what makes a companion diagnostic work. Immune diseases don't offer that. Ulcerative colitis is polygenic and multi-pathway; block one cytokine and others compensate. A single biomarker, however well-chosen, rarely carves out a responder group cleanly distinct from the rest. TL1A is one instance of a problem that runs across immunology.

There is also a biological reason precision is hard here. TL1A is not only an inflammatory signal. On fibroblasts, TL1A-DR3 signaling drives fibrosis directly, and anti-TL1A antibodies reverse it in animal models; Merck and others describe this dual inflammatory-and-fibrotic role as "immuno-fibrosis," a mechanism our TL1A target brief covers in depth. A drug that works on both axes has a wider rationale for treating everyone, and its most differentiated benefit, slowing the fibrostenotic damage that drives surgery, unfolds over months to years, where a 12-week inflammatory-remission endpoint cannot detect it. Merck's Phase 2 program in systemic sclerosis, a fibrotic disease, shows it is pursuing that broader mechanism directly.

What would still validate the companion diagnostic?

The induction result is in; the precision question is not. Two readouts will help answer it. First, the detailed ATLAS-UC data at an upcoming congress: Merck's topline gave no remission rates and no "Diagnostic Assay Positive (Dx+)" numbers, and those prespecified secondary endpoints are where the precision story lives or dies. The specific number to watch is the Dx+ placebo arm: in Phase 2 it remitted at 11%, versus 1% in all-comers, which is what shrank the treatment difference. If that high placebo response recurs, the Dx+ subgroup will again fail to separate, however well the assay identifies "TL1A-driven" patients. Second, the week-52 maintenance results and the Phase 3 Crohn's program (primary completion 2028). If test-positive patients separate cleanly from the all-comers population, the diagnostic returns as a label-expansion argument; if they look like everyone else, the precision thesis stays where ARTEMIS-UC left it: unproven.

On paper, IBD had the makings of its first genuine precision-medicine approach: a validated genetic target, a drug that works, and a test to match them. Two of the three arrived; the test did not. The biology is real, but turning it into something that reliably picks patients is the part still missing, which is why IBD's first real precision-medicine bet has to wait.