How to design a phase 2b MASH trial: lessons from 219 trials and their protocols

Suppose you are planning a phase 2b trial of a new drug for metabolic dysfunction-associated steatohepatitis (MASH, formerly NASH) in patients with moderate to advanced fibrosis (stages F2 to F3). Nearly every number in your sample size depends on things earlier trials have already measured: how often placebo patients “respond” on biopsy, how many people you must screen to randomise one, and how many will have a readable biopsy at the end. So we asked ClinLit AI to find those numbers in the registry, in the protocols and statistical analysis plans (SAPs) sponsors have posted, and in the published results.

The assumptions sponsors wrote into their protocols were often far from what their trials then saw. Six phase 2 trials we followed from protocol to paper assumed a placebo response of 10% to 20%; the placebo arms then responded at rates from 2% to 33%. One protocol planned for 25% of screened patients to fail screening; 72% did. And how the biopsies were read changed the answer: in the phase 3 trial behind resmetirom’s approval, two pathologists reading the same biopsies put the placebo response at 13% and 9%.

Key findings

• Two kinds of phase 2 MASH trial. Of 219 phase 2 trials, 77 have a liver biopsy primary endpoint and 61 measure liver fat by imaging. Recent biopsy trials enrol a median of 201 patients at 53 sites over 29 months; liver-fat trials enrol 103 over 21 months.
• Placebo response is the assumption that matters most, and it varies the most. Meta-analyses put it at about 11% for MASH resolution and 19% for fibrosis improvement, but individual trials ranged from 2% (pegozafermin) to 33% (lanifibranor). Both trials still met their primary endpoint.
• Screen failure is routinely underestimated. Lanifibranor’s protocol assumed 25% and saw 72%; tirzepatide’s trial screened 1,583 people to randomise 190 (87%). A pooled analysis of 58 trials puts the average at 56%.
• Plan for missing end-of-treatment biopsies. Between 12% and 25% of randomised patients in these trials had no usable final biopsy; lanifibranor had planned for at most 5%.
• How the biopsy is read is a design decision. The protocols range from one central pathologist to a panel of three; pegozafermin’s panel removed 27 randomised patients (12%) whose eligibility a single reader had confirmed.
• Regulators still require the biopsy. The FDA’s 2018 guidance is still a draft and accepts either histology endpoint; the EMA asks for both, as co-primaries. Neither yet accepts non-invasive tests in place of the biopsy.

The two kinds of phase 2 MASH trial

We defined MASH phase 2 trials as interventional phase 2 or phase 2/3 trials with a condition naming NASH, steatohepatitis or MASH, and sorted them by what their registered primary outcomes measure. That gives 219 trials: 135 completed, 26 terminated, 7 withdrawn and 51 still running or of unknown status.

Biopsy trials peaked in 2019–2021, when many sponsors were chasing the first approval; among trials starting in 2025–2026, short liver-fat studies outnumber biopsy studies 11 to 4. The two designs are different trials, not variants of one:

A biopsy trial needs about twice the patients, two and a half times the sites and eight more months than a liver-fat trial, and in practice it is multinational. The rest of this post is about the biopsy trial, since that is what a phase 2b in F2–F3 MASH almost always is.

CASE
  WHEN txt ~* 'histolog|biops|[[:<:]]NAS[[:>:]]|NAFLD activity score|fibrosis stage|resolution' THEN 1
  WHEN txt ~* 'PDFF|liver fat|hepatic fat|fat fraction|[[:<:]]MRI[[:>:]]|[[:<:]]MRS[[:>:]]|spectroscop|[[:<:]]CAP[[:>:]]' THEN 2
  WHEN txt ~* 'elastograph|stiffness|fibroscan|[[:<:]]MRE[[:>:]]' THEN 3
  WHEN txt ~* '[[:<:]]ALT[[:>:]]|alanine aminotransferase|[[:<:]]AST[[:>:]]' THEN 4
  WHEN txt ~* 'adverse|safety|pharmacokinetic|tolerab' THEN 5
  ELSE 6 END

Placebo response: the assumption that decides the sample size

For a binary biopsy endpoint, the sample size follows from two numbers: the response you expect on placebo and on the drug. We asked ClinLit AI to read the sample size section of six phase 2 trials that posted both a protocol and an SAP, and then to find each trial’s primary results paper and compare.

Three things stand out.

Placebo varied more than the drugs did. Pegozafermin’s placebo arm responded at a seventh of the assumed rate; lanifibranor’s at three times it. Both trials met their primary endpoint, but for different reasons: pegozafermin’s treatment effect (19 to 24 points) was smaller than the 30 points it was powered for, and it succeeded because so few placebo patients responded. Lanifibranor’s high dose beat placebo by 22 points, close to the 20 it planned for, but against 33% rather than 10% that was a risk ratio of 1.7, not the 3 the protocol had in mind, and its lower dose missed.

The literature gives a centre, not a range you can rely on. The largest meta-analysis, of 127 studies, puts the placebo rate of MASH resolution without fibrosis worsening at 11% (95% CI 8% to 14%) and found no patient or trial feature that explained the variation (PMID 41076041). An earlier one, of 43 trials, gives 11.7% for resolution and 18.8% for fibrosis improvement (PMID 34990037). Another found higher placebo responses in trials with a higher baseline NAS, in South America and where patients lost weight (PMID 29913275): lanifibranor’s trial required highly active disease. In the semaglutide phase 3 behind its MASH approval, the placebo resolution rate was 34%, according to the Wegovy label. Sizing a trial on one point estimate, without a sensitivity range, is the weakest step in these protocols.

The way the biopsies are read moves the placebo rate. Pegozafermin’s authors suggest their low placebo rates reflect their three-pathologist consensus panel, which may have lowered the drug’s response too. Resmetirom’s label reports each pathologist separately: the placebo rate of steatohepatitis resolution at month 12 was 13% by pathologist A and 9% by pathologist B, on the same 294 patients. A simulation study shows how sampling and reader variability alone can make 22% of placebo patients with F2–F3 appear to improve by a fibrosis stage (PMID 34091047).

A few caveats on the table: the observed rates are each paper’s primary analysis, which for lanifibranor used multiple imputation rather than the SAP’s “missing = non-responder”, and semaglutide’s 17% is for patients with F2–F3 fibrosis, while its sample size was planned on all randomised patients (F1–F3). The sponsor of efruxifermin blacked out the assumptions in the posted protocol: the sample size paragraph on PDF page 75 has black boxes where the rates, power and size would be.

Screen failure: plan for most patients not to qualify

The second assumption that sinks timelines is how many people must be screened to randomise one. The biopsy is the main filter: in a systematic review of 14 phase 2 trials, more than half of the people found ineligible were excluded on histology, most often because their disease was not active enough or their fibrosis not advanced enough (PMID 33216477).

The pooled analysis found higher screen failure in more recent trials, in global and industry-funded trials, and in trials with more sites (PMID 40323026): the profile of a modern phase 2b. Lanifibranor’s SAP records its own miss plainly: “the observed screening failure rate was of 72%” (SAP p. 35).

Every recent protocol therefore pre-screens with non-invasive tests before anyone is biopsied, but no two use the same tests or thresholds:

Pre-screening helps less than you might hope. In SYNERGY-NASH, adding the FAST score and a higher AST cut-off raised the share of biopsies that qualified only from 27.5% to 28.9%, with the gain at endocrinology sites rather than hepatology ones (PMID 38768298). In a screening cohort of 1,929 patients, the NIS2+ blood test brought biopsy failure down from 60% to 39% (PMID 38061448). Across eight non-invasive tests compared in 1,005 patients, none worked alone; only sequential combinations brought false positives down to about 30% (PMID 40212791).

The eligibility criteria of 31 industry biopsy trials started since 2019 agree on the rest: F2–F3 fibrosis in 16 (F1 to F3 in 6 more), NAS of at least 4 with at least 1 in each component, a qualifying biopsy no more than about six months old in 18, and type 2 diabetes allowed in 24, usually with HbA1c up to 9% or 9.5%. They differ on imaging cut-offs (MRI-PDFF from 7.5% to 10%) and on BMI, which 18 do not limit at all.

The biopsy at the end: who reads it, and how many will be missing

Two more decisions shape the primary endpoint: how the biopsies are read, and what happens to patients without a final one.

Reading. The six protocols take four approaches:

Sources: the posted protocols and SAPs of each trial, pages cited in the reproduce box below.

Readers disagree often. In 150 patients, local and central pathologists agreed on the diagnosis of NASH in only 50.7% of biopsies and on the total NAS in 26.7% (PMID 32514460); panels of three trained pathologists did much better (PMID 38126958). A panel costs something too: when pegozafermin’s panel re-read the baseline biopsies, 27 of 222 randomised patients no longer met the histology criteria and were left out of the main analysis (PMID 37356033). Whatever you choose, write the disagreement rule into the protocol: four of the six left it out or deferred it to a manual that was not posted.

Missing biopsies. Every one of these trials lost a meaningful share of end-of-treatment biopsies:

Sources: PMID 33185364, 38856224, 38847460, 37802088, 37356033, 34670042, and the protocols and SAPs above.

Where the drug causes the losses, they fall unevenly. A fifth of survodutide patients stopped treatment for adverse events, mostly during dose escalation, against 3% on placebo, and patients who stop early often get no final biopsy. In efruxifermin’s trial, counting the 15 missing biopsies as non-responders, as the SAP required for its sensitivity analysis, took the 50 mg arm from p = 0.036 to p = 0.123. How missing biopsies are handled (non-responder imputation, multiple imputation, or a “retrieved dropout” model as in tirzepatide’s SAP) is an estimand question to settle in the SAP, not afterwards: lanifibranor’s paper reported multiple imputation as its primary analysis, not the SAP’s non-responder rule.

Duration, arms and multiplicity

The six trials read their primary biopsy at 24 weeks (efruxifermin, pegozafermin, lanifibranor), 48 (survodutide), 52 (tirzepatide) or 72 (semaglutide). Shorter trials can work for drugs with fast effects on liver fat and inflammation, but the FDA’s draft guidance asks phase 2 to run “at least 12–18 months” unless there is a clear justification, and to characterise effect size, variability and dose response well enough to plan phase 3.

All six had three or four arms. Their handling of multiple doses differed: tirzepatide, efruxifermin and pegozafermin tested each dose against placebo at 5% two-sided without adjustment; lanifibranor used a Hochberg procedure; semaglutide tested its doses and two endpoints in a fixed hierarchy; survodutide used a dose-response method (MCPMod), with a one-sided 5% error rate across candidate dose-response shapes. Tirzepatide’s paper later reported P values with a post hoc adjustment for its three doses.

What regulators ask for

The FDA’s guidance for non-cirrhotic NASH with fibrosis is still the December 2018 draft. It accepts, as endpoints reasonably likely to predict clinical benefit, resolution of steatohepatitis without fibrosis worsening, or at least one stage of fibrosis improvement without worsening of steatohepatitis, or both. Resmetirom (14 March 2024) and semaglutide (15 August 2025) were approved under accelerated approval on these histology endpoints, both with outcome trials still running (FDA).

The EMA’s reflection paper, in effect since 1 October 2024, asks for both endpoints, as co-primaries (EMA). A phase 2b meant to inform a global phase 3 should therefore measure both well, even if only one is primary.

Neither agency accepts non-invasive tests in place of the biopsy yet. The EMA says they are “still insufficiently validated”, while recommending them for screening. In August 2025 the FDA accepted a letter of intent to qualify liver stiffness by vibration-controlled transient elastography as a reasonably likely surrogate endpoint; that is the first step of the qualification process, not a change to the guidance (FDA).

Why phase 2 MASH trials stopped

Of the 219 trials, 26 were terminated and 7 withdrawn before enrolling anyone. Among the 12 biopsy trials that stopped, 7 gave a business or strategic reason, 2 slow enrolment (one citing “repeat-biopsy burden and COVID-19 pandemic”, NCT03198572), 1 lack of efficacy (NCT04267393) and 2 no reason. Few phase 2 MASH trials stop for the reasons a design can control; most stop because the sponsor changed course.

A design sketch, from the evidence

Putting the pieces together, this is what the evidence suggests for a phase 2b in non-cirrhotic F2–F3 MASH. It is a starting point for a team’s own judgement and for discussion with regulators, not a recommendation.

How we counted

How we ran the analysis step by step, with the questions to reuse for another disease and the traps we hit, is in How to benchmark a clinical trial design.

Trials. Interventional studies in the AACT snapshot of 9 November 2026 with phase 2 or phase 2/3, and at least one condition matching nash|steatohepatitis|[[:<:]]mash[[:>:]] (case-insensitive). Trials registered only as NAFLD, MASLD or steatosis are left out. Primary endpoint categories come from the registered primary outcomes’ measure and description text, first match wins, as in the SQL above. A trial whose primary outcomes mention both histology and liver fat counts as histology. Start periods use the registry start date, which is estimated for 5 of the 21 trials starting in 2025–2026. Sizes and durations use completed trials with actual enrolment; duration is start date to primary completion date; sites are ClinicalTrials.gov’s facility count.

Protocols and results. We chose six well-known phase 2 trials that posted both a protocol and an SAP, so the sample is not random: they are larger and better resourced than most. ClinLit AI read the posted documents through their extracted text; we opened every page this post cites in the PDFs. The observed rates are from each trial’s primary results paper as ClinLit AI read it; where it used a full-text copy rather than the PubMed abstract, we say so. Endpoints and analyses differ between trials, so the comparisons are within each trial.

A bug we hit, and what it means if you reuse the definition. We first wrote the word boundaries as \m and \M. ClinLit AI’s SQL doubled the backslashes (\\m), which in PostgreSQL matches a literal backslash, so terms such as NAS, MRI and ALT silently matched nothing: it counted 72 histology trials instead of 77, and 56 liver-fat trials instead of 61. Our own SQL caught the difference. The definition above uses [[:<:]] and [[:>:]], which need no backslashes and give the same counts as our SQL. If you adapt it, keep that form.

Cross-check. We ran the definition in our own SQL. The trial count, the status counts, every cell of the endpoint table, the size and duration table and the list of stopped trials matched ClinLit AI’s answers exactly. We also checked every PMID in this post against PubMed, and the key numbers against their abstracts.

Limits. The registry’s free text can misclassify a trial’s endpoint, and the “why stopped” field is optional. Screen failure and missing-biopsy figures come from papers that define them differently. The FDA and EMA positions are as published on 9 November 2026. Nothing here is regulatory or medical advice.