Model: predicting who responds to a menin inhibitor - no patient-level data is public, so this cannot be trained either
also called response predictor scoping · biomarker model feasibility · AUGMENT-101 data availability
We were asked whether we could build a model that predicts which patients will respond to a menin inhibitor. Hundreds of patients have now been treated, but the trials publish only group averages. The individual patient data that a model needs is held by the companies and is not released, so the answer is no.
The same thing, for a scientist
Roughly 330 patients have received a menin inhibitor in the four published cohorts - revumenib phase 1 (n=68), AUGMENT-101 phase 2 KMT2A-rearranged (n=94, 57 efficacy-evaluable), AUGMENT-101 phase 2 NPM1-mutant (n=84, 64 evaluable) and KOMET-001 phase 2 NPM1-mutant (n=92). Neither NCT04065399 (enrollment 447) nor NCT04067336 (enrollment 263) has results posted on ClinicalTrials.gov. Every publication reports aggregate response rates. The Nature 2023 data-availability statement makes sponsor review a precondition of access. The only paired response-plus-molecular data at patient granularity that is genuinely public is the Perner resistance cohort supplement (31 sequenced patients) and four cell-free DNA cases - on the order of 35 patients, not 330. A supervised response predictor cannot be trained on 35 patients with heterogeneous disease, two different molecular subgroups and a roughly 23% response rate, which yields about eight responders.
What exists, counted
| Cohort | Treated | Efficacy-evaluable | Reported response | Source |
|---|---|---|---|---|
| Revumenib phase 1 | 68 | 60 | - | Nature 2023 1 |
| AUGMENT-101 ph2, KMT2A-r | 94 | 57 | CR+CRh 22.8% | JCO 2025 2 |
| AUGMENT-101 ph2, NPM1m | 84 | 64 | - | Blood 2025 3 |
| KOMET-001 ph2, NPM1m (ziftomenib) | 92 | 92 | CR/CRh 22% | JCO 2025 4 |
Registry status: NCT04065399 enrollment 447, recruiting, primary completion 2027-12-15, has_results: false. NCT04067336 enrollment 263, active not recruiting, completion 2028-10-16, has_results: false 5.
So the aggregate evidence base is large and growing, and the patient-level evidence base is approximately 35.
Why this cannot be trained
Take the most generous reading: 35 patients with paired baseline molecular data and outcome. At the observed 22-23% response rate that is roughly eight responders. Feature space: mutations, fusion partner, prior therapy lines, age, blast burden - dozens of candidate features against eight positive events. This is the classical regime in which a model reports excellent cross-validated performance and generalises to nothing.
The subgroup problem compounds it. KMT2A-rearranged and NPM1-mutant disease respond to the same drug through the same target but are different diseases; pooling them to reach n is exactly the error that makes a model look trainable when it is not. And the patients this project cares about - infants with KMT2A::AFF1 - are a small minority of even the KMT2A-rearranged arm.
What is feasible instead, and what it is worth
Feasible: a cell-line sensitivity model. DepMap and PRISM-style drug-response data across roughly 1,250 lines supports a supervised model that predicts sensitivity from baseline expression. The training data exists, the labels are quantitative, and the baseline to beat is well established (elastic net or random forest on expression, which in published benchmarks is hard to beat with deeper architectures).
What it is worth: less than it appears. Such a model predicts cell lines. The clinically decisive variables in infant ALL - prior therapy exposure, cell-state plasticity, bridging to transplant - have no cell-line analogue at all. A cell-line response model is a hypothesis generator about mechanism, not a patient-selection tool, and this entry exists partly so that it is never presented as one.
The highest-value action is not a model at all: ask for the data. Sponsor-held patient-level data from AUGMENT-101 becomes shareable after approval in the US and EU per the published data-availability statement. A formal request, and a collaboration approach through the landscape entries this graph already holds, is cheaper than any modelling work and is the only route to the dataset that would make the model possible.
| task | Predict which patients respond to a menin inhibitor from baseline molecular features. |
|---|---|
| verdict | CANNOT BE TRAINED at patient level. Aggregate data only; roughly 35 patients with paired molecular data and outcome in the public record, implying about 8 responders. |
| public patient level data | approximately 35 patients: the Perner resistance cohort supplement (31 sequenced) plus 4 cell-free DNA cases |
| highest value action | Request sponsor-held patient-level data through the published data-availability route and through a collaboration approach, rather than building a model on what is public. Cheaper than modelling and the only route to the dataset that would make a model possible. |
| status | planned |
| compute | the feasible cell-line substitute trains on a laptop; no GPU needed |
published cohorts — the numbers behind the chart
[
{
"cohort": "revumenib phase 1",
"treated": 68,
"evaluable": 60,
"pmid": "36922593",
"title": "The menin inhibitor revumenib in KMT2A-rearranged or NPM1-mutant leukaemia.",
"authors": "Issa GC, Aldoss I, DiPersio J, et al.",
"venue": "Nature",
"year": 2023,
"title_provenance": "Title/authors/venue backfilled from PubMed esummary on 2026-09-20 because this citation stored a bare PMID with no title, which made it unscoreable by verify-citations.py. BACKFILLING IS NOT VERIFICATION: it confirms the PMID resolves to a real paper, not that this is the right paper for the claim. That still requires reading."
},
{
"cohort": "AUGMENT-101 ph2 KMT2A-r",
"treated": 94,
"evaluable": 57,
"response": "CR+CRh 22.8%",
"pmid": "39121437",
"title": "Menin Inhibition With Revumenib for KMT2A-Rearranged Relapsed or Refractory Acute Leukemia (AUGMENT-101).",
"authors": "Issa GC, Aldoss I, Thirman MJ, et al.",
"venue": "J Clin Oncol",
"year": 2025,
"title_provenance": "Title/authors/venue backfilled from PubMed esummary on 2026-09-20 because this citation stored a bare PMID with no title, which made it unscoreable by verify-citations.py. BACKFILLING IS NOT VERIFICATION: it confirms the PMID resolves to a real paper, not that this is the right paper for the claim. That still requires reading."
},
{
"cohort": "AUGMENT-101 ph2 NPM1m",
"treated": 84,
"evaluable": 64,
"pmid": "40332046",
"title": "Menin inhibition with revumenib for NPM1-mutated relapsed or refractory acute myeloid leukemia: the AUGMENT-101 study.",
"authors": "Arellano ML, Thirman MJ, DiPersio JF, et al.",
"venue": "Blood",
"year": 2025,
"title_provenance": "Title/authors/venue backfilled from PubMed esummary on 2026-09-20 because this citation stored a bare PMID with no title, which made it unscoreable by verify-citations.py. BACKFILLING IS NOT VERIFICATION: it confirms the PMID resolves to a real paper, not that this is the right paper for the claim. That still requires reading."
},
{
"cohort": "KOMET-001 ph2 NPM1m",
"treated": 92,
"evaluable": 92,
"response": "CR/CRh 22%",
"pmid": "40997296",
"title": "Ziftomenib in Relapsed or Refractory NPM1-Mutated AML.",
"authors": "Wang ES, Montesinos P, Foran J, et al.",
"venue": "J Clin Oncol",
"year": 2025,
"title_provenance": "Title/authors/venue backfilled from PubMed esummary on 2026-09-20 because this citation stored a bare PMID with no title, which made it unscoreable by verify-citations.py. BACKFILLING IS NOT VERIFICATION: it confirms the PMID resolves to a real paper, not that this is the right paper for the claim. That still requires reading."
}
]registry status
[
{
"nct": "NCT04065399",
"enrollment": 447,
"status": "RECRUITING",
"has_results": false
},
{
"nct": "NCT04067336",
"enrollment": 263,
"status": "ACTIVE_NOT_RECRUITING",
"has_results": false
}
]why not trainable
[ "about 8 responders against dozens of candidate features", "KMT2A-rearranged and NPM1-mutant are different diseases and must not be pooled to reach n", "infants are a small minority of even the KMT2A-rearranged arm", "sponsor data-availability terms condition release on approval plus review" ]
feasible substitute
{
"task": "cell-line drug-sensitivity prediction from baseline expression",
"training_data": "DepMap and PRISM-style response across roughly 1,250 lines",
"baseline_to_beat": "elastic net or random forest on expression features",
"worth": "a mechanism hypothesis generator, not a patient-selection tool; predicts cell lines, and the clinically decisive variables in infant ALL have no cell-line analogue"
}- A response model built on cell lines and presented next to clinical response rates in a dashboard will be read as a clinical prediction. The framing risk here is larger than the technical risk.
- Pooling KMT2A-rearranged and NPM1-mutant patients to reach a trainable n would produce a model that has learned the difference between two diseases and calls it a response biomarker.
- If this project's own hypothesis that prior therapy installs resistance is correct, then the strongest predictor of response is a treatment-history variable that no molecular model would include, and a molecular model would attribute its effect to whatever mutation happens to correlate with heavy pretreatment.
- Do any AUGMENT-101 or KOMET-001 supplements contain a per-patient response-by-genotype table that this review did not find?
- Now that a menin inhibitor is approved, has the data-availability condition been met, and would a formal request succeed?
- Is there an academic consortium assembling menin-inhibitor outcome data across centres that this lab could join rather than duplicate?
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
The revumenib phase 1 study treated 68 patients with KMT2A-rearranged or NPM1-mutant leukaemia, with 60 efficacy-evaluable, and its data-availability statement conditions access on product approval or discontinuation with sponsor review.
The menin inhibitor revumenib in KMT2A-rearranged or NPM1-mutant leukaemia · Issa GC, Aldoss I, DiPersio J, et al. · Nature 2023What this evidence cannot show
Phase 1 dose escalation; response rates from a dose-finding population are not comparable with phase 2 rates at the recommended dose.
Evidence that argues against this (1)
- A response predictor trained on these trials is at high risk of data leakage in the technical sense. Kapoor and Narayanan surveyed ML-based science across 17 fields, found leakage affecting 294 papers, some with wildly over-optimistic conclusions, and set out a taxonomy of eight leakage types with model info sheets for testing. With 60 efficacy-evaluable patients in the phase 1 study, feature-selection or preprocessing performed before the train/test split would be enough to manufacture apparent predictive performance. Kapoor S et al. 2023
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
In the AUGMENT-101 phase 2 KMT2A-rearranged cohort, 94 patients were treated with a median age of 37 years, and among 57 efficacy-evaluable patients the CR plus CRh rate was 22.8%.
Menin Inhibition With Revumenib for KMT2A-Rearranged Relapsed or Refractory Acute Leukemia (AUGMENT-101) · Issa GC, et al. · Journal of Clinical Oncology 2025What this evidence cannot show
Median age 37 with a range from 1.3 years: this is predominantly an adult cohort, and the infant subgroup is small. Response rates here should not be transferred to infants without that caveat attached.
Evidence that argues against this (1)
- The label is not comparable across the studies this entry pools. AUGMENT-101's KMT2A-r cohort reports CR+CRh of 22.8% among 57 efficacy-evaluable patients out of 94 treated, median age 37 (range 1.3-75), with the efficacy set restricted to centrally confirmed rearrangement. KOMET-001 reports CR/CRh of 22% in 92 NPM1-mutant patients, median age 69. The two numbers are nearly equal and describe different diseases, different age distributions, different drugs and different analysis sets; treating them as one label column would be a leakage of study identity into the outcome. Wang ES et al. 2025
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
The AUGMENT-101 NPM1-mutant cohort treated 84 patients with 64 adults in the primary efficacy analysis, reported separately from the KMT2A-rearranged cohort.
Menin inhibition with revumenib for NPM1-mutated relapsed or refractory acute myeloid leukemia: the AUGMENT-101 study · Blood 2025What this evidence cannot show
NPM1-mutant AML is a different disease from KMT2A-rearranged infant ALL; these patients cannot be pooled with the KMT2A-rearranged cohort to increase n for a model.
Evidence that argues against this (1)
- The NPM1-mutant and KMT2A-rearranged cohorts are reported separately for a reason and should not be merged as one response label. Ziftomenib's KOMET-001 NPM1-mutant registration cohort - a different sponsor, drug and protocol - lands at essentially the same CR/CRh rate in a population with median age 69, which shows how little a headline response rate constrains the underlying biology. Wang ES et al. 2025
- E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
The KOMET-001 phase 2 study treated 92 patients with NPM1-mutant AML, median age 69, with a CR/CRh rate of 22%.
Ziftomenib in Relapsed or Refractory NPM1-Mutated AML · Journal of Clinical Oncology 2025What this evidence cannot show
Median age 69; an adult AML population with no paediatric read-across.
Evidence that argues against this (1)
- Response to menin inhibition is lineage-dependent in a way a single pooled label erases. Adriaanse et al. showed KMT2A-r ALL undergoes rapid dose-dependent cell death under revumenib while KMT2A-r AML responds slowly through myeloid differentiation, and that acquired resistance in ALL arises both with and without MEN1 mutations. A model trained on pooled AML-dominated response data will not transfer to the infant ALL population this project exists to serve. Adriaanse FRS et al. 2024
- E5strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.trial_registry
Neither AUGMENT-101 (NCT04065399, enrollment 447, recruiting) nor KOMET-001 (NCT04067336, enrollment 263, active not recruiting) has results posted on ClinicalTrials.gov.
ClinicalTrials.gov records NCT04065399 and NCT04067336 · ClinicalTrials.gov 2026What this evidence cannot show
Registry status fetched this cycle. Results posting is not required until after primary completion, so absence now is expected and is not a criticism of the sponsors.
Evidence that argues against this (1)
- Neither pivotal trial collecting the covariate a response model would need is the finding, and the entry should state what follows: such a model cannot be trained on registrational data and would require either prospective collection or a biobank with linked outcomes. That is a resourcing conclusion, not a modelling one. Cancer Discovery 2021
Counter-evidence: menin-inhibitor response is lineage-dependent (ALL dies, AML differentiates; PMID 38892207), and the trial labels available differ in disease, drug, age and analysis set - so a pooled response label leaks study identity (PMID 37720327).
aggregate response data exists at scale; patient-level data does not
if prior therapy installs resistance, the strongest predictor is a treatment-history variable no molecular model would include
the cohort whose patient-level data would be required and is not public