This page provides an independent statistical analysis and educational interpretation of publicly reported results. ClinicalTrials.gov provides the official trial registry record.
1. Trial at a Glance
ATLAS was a randomized, double-blind, placebo-controlled phase 2 trial with seven parallel arms. It asked whether three investigational agents with different mechanisms, given alone or in pairs for 48 weeks, could improve liver fibrosis without worsening steatohepatitis in people with advanced NASH-related fibrosis, and how safe these regimens were.
| Feature | ATLAS |
|---|---|
| Phase | Phase 2 |
| Condition | Nonalcoholic steatohepatitis (NASH) with bridging fibrosis or compensated cirrhosis |
| Design | Randomized, parallel-group, double-masked, placebo-controlled |
| Arms | Selonsertib; Firsocostat; Cilofexor; SEL + FIR; SEL + CILO; FIR + CILO; Placebo |
| Enrollment | 395 |
| Primary endpoints | Treatment-emergent adverse events; treatment-emergent laboratory abnormalities; ≥ 1-stage fibrosis improvement without worsening of NASH at Week 48 |
| Primary efficacy analysis | Stratum-adjusted Mantel-Haenszel difference in percentages, stratified by baseline diabetes status and baseline cirrhosis status |
| Hypothesis framework | Superiority, two-sided 95% confidence intervals |
| Dates | Start 2018-03-21; primary completion 2019-10-30 |
| ClinicalTrials.gov | NCT03449446 |
| Lead sponsor | Gilead Sciences (industry) |
2. Clinical Question
NASH with advanced fibrosis is a setting in which several biological pathways are thought to drive disease, and ATLAS was built to screen single agents and pairwise combinations in parallel rather than to confirm one regimen. The central efficacy question was whether any of the six active regimens increased the proportion of participants achieving at least a one-stage improvement in fibrosis without worsening of NASH at Week 48 compared with placebo, alongside a formal description of adverse events and laboratory abnormalities.
Population
Participants with bridging fibrosis or compensated cirrhosis due to nonalcoholic steatohepatitis.
Intervention
Selonsertib, firsocostat or cilofexor alone, or the pairs SEL + FIR, SEL + CILO and FIR + CILO, for 48 weeks.
Comparator
Matching placebos for each study drug (placebo to match SEL, FIR and CILO), which allowed all arms to be blinded.
Primary question
Is the percentage of participants with ≥ 1-stage fibrosis improvement without NASH worsening higher on an active regimen than on placebo, and is each regimen acceptably safe?
3. Trial Design
The registry lists three active drugs and three matching placebos. Using a separate placebo for each drug is the standard way to blind a multi-arm trial in which some participants take one active drug and others take two: every participant can receive the same set of dosing units, with active or placebo content varying by arm. The numbers of participants at risk in the serious adverse event reporting show that the combination arms were roughly twice the size of the monotherapy and placebo arms.
| Arm | Type | Participants at risk (safety reporting) |
|---|---|---|
| Selonsertib (SEL) | Monotherapy | 39 |
| Firsocostat (FIR) | Monotherapy | 40 |
| Cilofexor (CILO) | Monotherapy | 40 |
| SEL + FIR | Combination | 79 |
| SEL + CILO | Combination | 77 |
| FIR + CILO | Combination | 78 |
| Placebo | Control | 39 |
4. Stratification and Analysis Populations
The efficacy comparisons were stratum-adjusted for baseline diabetes mellitus status and baseline cirrhosis status. Both factors are plausibly related to the chance of fibrosis regression, so accounting for them protects the treatment comparison against chance imbalances and aligns the analysis with the way the population was likely balanced at randomization.
| Analysis population | Definition / role |
|---|---|
| Full Analysis Set (FAS) | All participants who were randomized and took at least 1 dose of study drug. Used for the fibrosis endpoint, restricted to participants with available Week 48 data. |
| Safety reporting | Participants at risk in each arm for adverse event reporting, from first dose up to 48 weeks plus 30 days. |
The FAS is a modified intention-to-treat population: participants are analyzed in the arm to which they were randomized, but those who never received a dose are excluded. Because the fibrosis analysis also required an available Week 48 assessment, participants without an evaluable end-of-treatment result did not contribute to the estimate. The ClinicalTrials.gov record does not describe an imputation strategy for these participants.
5. Endpoints
All three registered primary endpoints are shown with the registry wording and time frames.
| Primary endpoint | Definition | Time frame |
|---|---|---|
| Percentage of participants experiencing treatment-emergent adverse events (TEAEs) | Proportion of participants with at least one adverse event emerging on treatment | First dose date up to 48 weeks plus 30 days |
| Percentage of participants experiencing treatment-emergent laboratory abnormalities | Values that increase at least one toxicity grade from baseline; participants with any laboratory abnormality were reported | First dose date up to 48 weeks plus 30 days |
| Percentage of participants who achieved a ≥ 1-stage improvement in fibrosis without worsening of NASH | Fibrosis improvement: ≥ 1-stage decrease from baseline by the NASH Clinical Research Network (CRN) classification. Worsening of NASH: ≥ 1-point increase from baseline in hepatocellular ballooning or lobular inflammation. Within-arm 95% CIs by the Clopper-Pearson method. | Week 48 |
The efficacy endpoint is a composite binary responder definition. A participant counts as a responder only if fibrosis improves by at least one stage and neither ballooning nor lobular inflammation increases by a point or more. The second condition prevents an apparent fibrosis gain from being credited while the underlying inflammatory activity deteriorates. Because fibrosis staging is ordinal and biopsy-based, the endpoint is also sensitive to sampling variability and reader variability, which adds noise to the observed response rates in every arm.
6. Primary Efficacy Results: Fibrosis Improvement Without NASH Worsening
The registry posts nine stratum-adjusted Mantel-Haenszel comparisons for the Week 48 fibrosis endpoint: each of the three monotherapies versus placebo, each of the three combinations versus placebo, and each combination versus its component monotherapies. All estimates are differences in percentages (percentage points) with two-sided 95% confidence intervals.
| Comparison | Difference in percentages | 95% CI | P-value |
|---|---|---|---|
| Firsocostat vs Placebo | 0.6 | −17.6 to 18.8 | 0.9449 |
| Cilofexor vs Placebo | 0.4 | −17.6 to 18.4 | 0.9646 |
| SEL + FIR vs Placebo | 3.7 | −10.9 to 18.3 | 0.6219 |
| SEL + CILO vs Placebo | 8.8 | −6.4 to 24.1 | 0.2554 |
| FIR + CILO vs Placebo | 10.8 | −4.5 to 26.1 | 0.1658 |
| Firsocostat vs SEL + FIR | 3.2 | −12.9 to 19.4 | 0.6963 |
| Firsocostat vs FIR + CILO | 10 | −6.8 to 26.8 | 0.2441 |
| Cilofexor vs SEL + CILO | 5.2 | −10.7 to 21.0 | 0.5229 |
| Cilofexor vs FIR + CILO | 10.4 | −6.0 to 26.9 | 0.2149 |
Every one of the nine intervals includes zero, and every p-value is above 0.05. The point estimates are ordered in a consistent way: the two monotherapy-versus-placebo estimates are close to zero (0.6 and 0.4 points), the combination-versus-placebo estimates are larger (3.7, 8.8 and 10.8 points), and the combinations containing FIR + CILO produce the largest contrasts against both placebo and the component monotherapies. For the four combination-versus-monotherapy contrasts, the registry labels the comparison with the monotherapy listed first; the sign convention should be read from the registry record rather than assumed from the label order.
Monotherapy versus placebo
Firsocostat and cilofexor alone vs placebo
FIR: 95% CI −17.6 to 18.8 · P = 0.9449
CILO: 95% CI −17.6 to 18.4 · P = 0.9646
What the estimate means. A stratum-adjusted difference of 0.6 percentage points means that, after weighting across the diabetes and cirrhosis strata, the responder percentage on firsocostat was estimated to be almost identical to that on placebo. The same holds for cilofexor at 0.4 points.
What it does not mean. A near-zero estimate with a very wide interval is not evidence that the drugs have no effect. The intervals, running from −17.6 to 18.8 and from −17.6 to 18.4 percentage points, are compatible with a meaningful harm, no difference, or a meaningful benefit. With roughly 40 participants per arm, the trial was not able to distinguish among these possibilities for a single agent.
Why the p-value is not an effect size. P = 0.9449 says only that the observed difference is very small relative to its standard error. It would be the same order of magnitude for a tiny true effect measured precisely or for a moderate true effect measured imprecisely; the confidence interval, not the p-value, conveys which of these situations applies.
Combinations versus placebo
FIR + CILO vs placebo (largest contrast)
95% CI −4.5 to 26.1 · P = 0.1658
SEL + CILO: 8.8 (−6.4 to 24.1), P = 0.2554 · SEL + FIR: 3.7 (−10.9 to 18.3), P = 0.6219
What the estimate means. The FIR + CILO arm had an estimated responder percentage 10.8 points higher than placebo. This is an absolute difference in the proportion of responders, not a relative change: it does not say that response was 10.8% more likely in relative terms.
Precision. The interval from −4.5 to 26.1 points extends only slightly below zero and well above it. The data are more compatible with a benefit than with harm, but the lower limit still includes no difference. The combination intervals are somewhat narrower than the monotherapy intervals because the combination arms were larger, yet the placebo arm of 39 participants still limits precision.
Cautions. Nine comparisons were posted for the same endpoint, and the registry does not describe a multiplicity adjustment. When several contrasts are examined, the largest observed difference tends to overstate the true effect of the best-looking arm (a "winner's curse"). A p-value of 0.1658 does not count as a near-miss that can be rescued by selecting the most favorable contrast; the appropriate reading is that FIR + CILO produced the most promising signal worth testing in a larger trial.
Combinations versus their component monotherapies
What these comparisons address. Contrasts such as Firsocostat vs FIR + CILO (10.0; −6.8 to 26.8; P = 0.2441) and Cilofexor vs FIR + CILO (10.4; −6.0 to 26.9; P = 0.2149) address whether adding a second agent contributes beyond the first. This is the logic of a factorial-type question embedded in a multi-arm design, although ATLAS did not include every cell needed for a formal factorial interaction analysis on its own.
What they do not show. None of the four intervals excludes zero, so the trial does not establish that any combination outperforms its components. Because both FIR-versus-combination and CILO-versus-combination point in the same direction for FIR + CILO, the pattern is internally consistent, but consistency across correlated comparisons that share an arm is weaker evidence than it appears: the same FIR + CILO participants contribute to both contrasts.
Sign convention. Reading these estimates correctly requires knowing which group is subtracted from which. The label order alone ("Firsocostat vs FIR + CILO") should not be used to infer direction without checking the registry's group definitions.
Safety primary endpoints
The percentages of participants with treatment-emergent adverse events and with treatment-emergent laboratory abnormalities are registered as primary endpoints and have results posted on ClinicalTrials.gov. No between-arm statistical tests were posted for these two endpoints, which is typical: safety primary endpoints in phase 2 trials are usually descriptive, reported as counts and percentages by arm, and interpreted for patterns (for example, specific laboratory shifts clustered in arms containing a particular drug) rather than through hypothesis tests. If a formal comparison were desired, a risk difference with a confidence interval, or an exact test for small counts, would be the usual approach.
7. Serious Adverse Events by Arm
| Arm | Participants with serious adverse events / at risk |
|---|---|
| Selonsertib | 7/39 |
| Firsocostat | 3/40 |
| Cilofexor | 8/40 |
| SEL + FIR | 7/79 |
| SEL + CILO | 10/77 |
| FIR + CILO | 8/78 |
| Placebo | 2/39 |
Serious adverse events were numerically fewest in the placebo arm (2/39) and firsocostat arm (3/40), and more frequent in the selonsertib (7/39) and cilofexor (8/40) monotherapy arms. Raw counts must be read against the denominators: 8 events in the FIR + CILO arm come from 78 participants at risk, about twice the denominator of the monotherapy arms.
8. Statistical Methodology
Stratum-adjusted Mantel-Haenszel difference in percentages
The primary efficacy comparisons used the Mantel-Haenszel (MH) method to estimate a common risk difference across strata defined by baseline diabetes status and baseline cirrhosis status. Within each stratum, the difference in responder proportions is computed; the MH estimate is a weighted average of these stratum-specific differences, with weights that favor strata containing more participants in both arms.
where p1k and p0k are the responder proportions in the two arms within stratum k, and n1k, n0k are the corresponding numbers of participants.
The associated test is closely related to the Cochran-Mantel-Haenszel (CMH) test, which asks whether there is any association between treatment and response after conditioning on strata. The MH approach assumes that the treatment effect on the risk-difference scale is broadly similar across strata; if the effect were very different in cirrhotic and non-cirrhotic participants, a single pooled difference would summarize the strata less faithfully.
Clopper-Pearson confidence intervals
Within-arm 95% confidence intervals for the responder percentage were based on the Clopper-Pearson method. This "exact" interval inverts two one-sided binomial tests and is guaranteed to achieve at least the nominal coverage. It is conservative, meaning the actual coverage is usually above 95% and the interval is somewhat wider than alternatives such as the Wilson interval, a reasonable choice when arms contain only about 40 participants and response counts are small.
Superiority framework and two-sided testing
Each posted comparison is labelled as a superiority test with a two-sided 95% confidence interval. The null hypothesis is a difference of zero; a two-sided test treats an increase and a decrease in response rate symmetrically. There is no non-inferiority margin in this design.
Analysis population and missing biopsies
The efficacy analysis used participants in the Full Analysis Set with available Week 48 data. This is a complete-case analysis for the biopsy endpoint. If participants who lacked a Week 48 biopsy differed systematically from those who had one, for example through early discontinuation for adverse events or lack of benefit, the complete-case estimate may not reflect the effect in all randomized, treated participants. A common conservative alternative in liver-histology trials is to count participants with missing biopsies as non-responders; the registry does not describe such a sensitivity analysis.
Multiplicity
Nine comparisons were posted for a single primary efficacy endpoint, and the ClinicalTrials.gov record does not describe a multiplicity adjustment or hierarchical testing order. In a phase 2 screening trial this is often acceptable because the goal is to select regimens for further development rather than to confirm efficacy, but the familywise probability of at least one false-positive result across nine unadjusted 5%-level tests is substantially higher than 5%. In ATLAS the issue is academic for claims of benefit, since no comparison reached nominal significance, but it remains important when ranking arms by point estimate.
Safety analysis
Safety endpoints cover the period from the first dose up to 48 weeks plus 30 days, so that events emerging shortly after the last dose are captured. Treatment-emergent laboratory abnormalities are defined by a shift of at least one toxicity grade from baseline, which anchors each participant to their own starting values rather than to a fixed laboratory threshold.
9. Statistical Methods Explained
Why was a Mantel-Haenszel risk difference used rather than a simple difference in proportions?
Diabetes and cirrhosis status are both expected to influence the chance of fibrosis improvement. Adjusting for them through MH stratification removes the influence of any chance imbalance in these factors between arms and generally improves precision compared with a crude difference, while still reporting the result on the intuitive absolute scale of percentage points.
What does a difference in percentages of 10.8 mean?
It means the estimated proportion of responders was 10.8 percentage points higher in the FIR + CILO arm than in the placebo arm, averaged across strata. It is an absolute difference, not a relative one.
Why are the monotherapy intervals so wide?
The width of a confidence interval for a difference in proportions depends mainly on the arm sizes. Comparing two groups of about 40 participants yields a large standard error, so the interval from −17.6 to 18.8 is wide even though the point estimate is close to zero. This width reflects the design's limited sample size, not unusual variability in the data.
Why is a p-value of 0.9449 not proof that firsocostat does nothing?
A large p-value means the data are compatible with no difference, but they are equally compatible with any value inside the confidence interval, including differences as large as 18.8 points. Absence of evidence of an effect is not evidence of absence, especially in a small arm; demonstrating equivalence would require a prespecified margin and a much narrower interval.
Why use Clopper-Pearson intervals for the within-arm percentages?
With small arms, the normal (Wald) approximation for a binomial proportion can produce intervals that undercover or even extend below 0%. The Clopper-Pearson interval is based directly on the binomial distribution and never falls outside 0–100%, at the cost of being slightly conservative.
Why were combination arms compared with their component monotherapies as well as placebo?
A combination that beats placebo might do so only because one of its components is active. Comparing, for example, FIR + CILO with firsocostat alone and with cilofexor alone asks whether each drug contributes to the combination. In ATLAS these comparisons were exploratory in precision terms, but they show how multi-arm designs can address additivity directly.
10. Limitations
- Small arms: about 40 participants in each monotherapy arm and in the placebo arm make every placebo comparison imprecise; confidence intervals such as −17.6 to 18.8 points around the monotherapy estimates cannot exclude clinically meaningful effects in either direction.
- Multiple comparisons: nine efficacy contrasts were posted without a described adjustment; the most favorable estimate is likely to be optimistic.
- Complete-case analysis: only FAS participants with available Week 48 data were analyzed; missing biopsies could bias estimates if missingness was related to treatment or outcome.
- Histologic endpoint: fibrosis staging from a single biopsy is subject to sampling and reader variability, which attenuates observed treatment differences.
- Shared comparator: all placebo comparisons reuse the same 39-participant placebo arm, so their errors are correlated; an unusually high or low placebo response would shift every contrast together.
- Duration: 48 weeks of treatment may be short relative to the pace of fibrosis regression in advanced disease.
- Descriptive safety: serious adverse event counts are small and were not formally compared between arms.
11. Why This Trial Matters Statistically
ATLAS is a useful teaching case for the analysis of binary responder endpoints in multi-arm phase 2 trials, and for reading a set of null results without over- or under-interpreting them.
| Concept | How it appears in ATLAS |
|---|---|
| Multi-arm design | Three monotherapies, three combinations and placebo in one randomized, blinded trial |
| Double-dummy style blinding | Separate matching placebos for SEL, FIR and CILO |
| Stratified risk difference | Mantel-Haenszel difference in percentages adjusted for diabetes and cirrhosis status |
| Exact binomial intervals | Clopper-Pearson 95% CIs for within-arm responder percentages |
| Confidence intervals vs p-values | Wide intervals around near-zero estimates show imprecision rather than proof of no effect |
| Multiplicity | Nine contrasts for one endpoint; ranking arms by point estimate invites optimism bias |
| Composite responder endpoint | Fibrosis improvement required together with no NASH worsening |
| Modified ITT / complete-case | FAS with available Week 48 data |
| Descriptive safety | Serious adverse events reported as affected/at-risk counts by arm |
12. Related Tutorials
Learn more about the methods used in this trial:
13. Related Calculators
14. Sources
- ClinicalTrials.gov: NCT03449446 — Study to Evaluate the Safety and Efficacy of Selonsertib, Firsocostat, Cilofexor, and Combinations in Participants With Bridging Fibrosis or Compensated Cirrhosis Due to Nonalcoholic Steatohepatitis (NASH).
- PubMed: PMID 33169409.
Continue exploring
Connect this trial's endpoints and methods to in-depth biostatistics tutorials and to calculators for working through similar analyses.
15. Record Summary
ATLAS randomized 395 participants with advanced NASH fibrosis across seven blinded arms. For the Week 48 endpoint of ≥ 1-stage fibrosis improvement without NASH worsening, stratum-adjusted Mantel-Haenszel differences versus placebo ranged from 0.4 points (cilofexor) to 10.8 points (FIR + CILO), with every 95% confidence interval including zero. The most informative reading combines the direction and ordering of the point estimates, the width of the confidence intervals imposed by small arms, and the multiplicity of the comparisons: the trial identified the FIR + CILO pairing as the most promising signal without establishing efficacy for any regimen.