Model: resistance-mutation prediction - cannot be trained. There are six patient-observed labels in the world, and a base-editor screen already does the job better
also called resistance predictor scoping · MEN1 mutation model · why we are not training this
We were asked whether we could build a model that predicts which mutations make menin inhibitors stop working, and we counted the real examples: six mutations across about a dozen patients. That is not a training set, and an experiment that makes every possible mutation and sees which survive is both cheaper and more informative.
The same thing, for a scientist
This is the clearest negative feasibility finding among the three models, and stating it is the point of a readiness review. The union of MEN1 substitutions observed in patients with acquired resistance to menin inhibition is **six substitutions at four residues** - S160T, M327I, M327V, G331D, G331R, T349M - across roughly twelve to sixteen patients in the entire public literature. Perner et al. sequenced 31 revumenib-treated patients with material available and found MEN1 mutations in 12 (38.7%); a later cell-free DNA series adds G331D across four illustrative cases. No supervised model can be trained on six positive labels. The alternative is not a better model, it is the experiment that already exists: a CRISPR base-editor screen across MEN1 has been published and identifies both shared and inhibitor-specific resistance substitutions prospectively. The correct phase-2 action is to run or extend that screen for the compounds this lab cares about, and to use the six patient observations as the validation set they can legitimately be.
The labels that exist, counted
| Residue | Substitutions | Source |
|---|---|---|
| S160 | S160T | Perner 2023, 1 patient |
| M327 | M327I, M327V | Perner 2023 (3 of 4 affected patients); also in cfDNA series |
| G331 | G331R, G331D | Perner 2023 (R); cfDNA series (D) |
| T349 | T349M | Perner 2023, 1 patient; also cfDNA |
Six substitutions. Four residues. Twelve to sixteen patients, with overlap between the two series likely. All revumenib; ziftomenib-specific patient counts are UNVERIFIED (the only figure this review located is a conference poster whose text could not be extracted).
Screen-derived, not patient-observed 3: a CRISPR base-editor screen reports shared substitutions M327I/V/T and G331D, and inhibitor-specific C334R, E368K/V and V372A. These are predictions about what could arise, generated experimentally, and they are a far richer source than the clinic has produced.
Why no model can be trained on this
Six positive labels at four positions, against a background of every other possible MEN1 substitution as an unlabelled negative. There is no split that is not degenerate, no cross-validation that is not leave-one-out on a handful of points, and no way to distinguish a model that has learned resistance biology from one that has learned to point at residues 327 and 331.
The seductive move - train a general protein-stability or drug-binding model and apply it to MEN1 - is the same out-of-domain problem the hotspot model has, with the added difficulty that resistance is not only about binding: a substitution must abolish drug binding and preserve enough menin-KMT2A function for the cell to survive. That is a two-constraint problem and nothing in the public data teaches the second constraint.
What to do instead
Run the experiment, not the model. A base-editor or saturation-mutagenesis screen across the MEN1 coding sequence under drug pressure generates the full map in one experiment, with the two constraints coupled automatically because only variants that both resist the drug and support growth come through. The published screen demonstrates the method works for this exact gene and these exact compounds.
The extension worth doing here is compound-specific: the screen reports that some substitutions are shared across inhibitors and others are inhibitor-specific. Whether a compound's resistance spectrum is narrow or broad is a real differentiator between menin inhibitors, and it is directly relevant to hyp-partner-selective-menin-ligand-separates-mandates and to the question of which agent to take into infants.
Use the six patient observations as a validation set. If a screen's top hits do not include M327 and G331, the screen is wrong. That is the only legitimate role six labels can play.
| task | Predict which MEN1 substitutions confer resistance to a given menin inhibitor. |
|---|---|
| verdict | CANNOT BE TRAINED. Six patient-observed positive labels at four residues is a validation set, not a training set. |
| n patients public | approximately 12-16, with likely overlap between the two series; all revumenib |
| ziftomenib patient counts | UNVERIFIED - the only located figure is a conference poster whose text could not be extracted |
| recommended alternative | Run or extend a CRISPR base-editor saturation screen across MEN1 under pressure from the specific compounds this lab cares about. The two constraints are coupled automatically because only variants that resist the drug and support growth survive. Use the six patient observations as the validation set: a screen whose top hits omit M327 and G331 is wrong. |
| differentiating question the screen answers | Is a given menin inhibitor's resistance spectrum narrow or broad? That is a real basis for choosing which agent goes into infants, and it connects directly to hyp-partner-selective-menin-ligand-separates-mandates. |
| status | planned |
| compute | none for modelling, because there is no model. The screen is wet-lab plus standard sequencing analysis. |
patient observed substitutions
[ "S160T", "M327I", "M327V", "G331D", "G331R", "T349M" ]
patient observed residues
[ "S160", "M327", "G331", "T349" ]
screen derived additional
[ "M327T", "C334R", "E368K", "E368V", "V372A" ]
why not trainable
[ "6 positive labels", "no non-degenerate train/test split", "resistance is a two-constraint problem (abolish drug binding AND preserve menin-KMT2A function) and nothing public teaches the second constraint", "a general stability or affinity model applied here is out of domain in the same way the hotspot model is" ]
- Publishing a resistance-prediction model built on six labels would be actively harmful: clinicians would read a per-variant score as calibrated, and it would not be.
- A base-editor screen accesses only the substitutions its editing window permits. The resulting map looks complete and is systematically missing whole classes of substitution, so its negatives must not be read as safe.
- The genetic route may not be the dominant escape route in infants at all. This graph already holds a hypothesis that genetic and non-genetic escape are substitutes, in which case a perfect MEN1 mutation map would predict the wrong failure mode for the patients this project cares about.
- Are the base-editor screen's inhibitor-specific hits (C334R, E368K/V, V372A) reachable in patients, or are they fitness-costly variants that a screen sees and a clinic never will?
- How many menin-inhibitor-treated patients worldwide have been sequenced at relapse but not published, and could a consortium request assemble a real training set?
- Does the very low variant allele fraction seen in cell-free DNA mean resistance clones are present far earlier than tissue sequencing suggests, and could that be an early-switching biomarker rather than a modelling target?
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
Somatic MEN1 mutations at the drug-menin interface mediate clinical resistance to menin inhibition; among 31 patients treated with revumenib for more than two cycles with DNA available, 12 (38.7%) carried one or more MEN1 mutations, with M327V or M327I in three of four affected patients and T349M, G331R and S160T each in one patient.
MEN1 mutations mediate clinical resistance to menin inhibition · Perner F, Stein EM, Wenge DV, et al. · Nature 2023What this evidence cannot show
A single programme, one drug, 31 sequenced patients selected by material availability. The 38.7% figure is not a population prevalence.
Evidence that argues against this (1)
- Interface MEN1 mutations are real and are not the whole resistance picture, which matters for a model trained to predict them: this graph records resistance arising with NO menin-inhibitor exposure and with on-target repression preserved, plus TP53 and lineage-switch routes that never touch the interface. A model predicting interface mutations would miss those entirely. Nature 2023
- E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.paper
A cell-free DNA series adds G331D to the patient-observed set and detects these substitutions at variant allele fractions below 2%, implying tissue sequencing under-counts both the prevalence and the diversity of resistance substitutions.
Detection of MEN1 resistance mutations in cell-free DNA from acute leukemia patients treated with menin inhibitors · Shukla N, et al. · Blood Cancer Journal 2026What this evidence cannot show
A letter reporting four illustrative patient cases from a single institution, with the authors noting the modest sample size of longitudinally banked plasma; cases may overlap with the earlier cohort.
Evidence that argues against this (1)
- Detection at low variant allele fraction in cell-free DNA is an assay achievement and not yet a clinical strategy: no study shows that cfDNA-guided switching improves outcome, and with cross-resistance being mutation- and drug-specific, detecting a substitution does not say which drug to switch to. Blood Cancer Journal 2026
- E3moderate confidence 3 of 5: One strong study, or several smaller consistent ones.paper
A CRISPR base-editor screen across MEN1 identifies the spectrum of substitutions that impair clinically used menin inhibitors, reporting shared resistance substitutions M327I/V/T and G331D and inhibitor-specific C334R, E368K/V and V372A.
CRISPR base editor screening identifies spectrum of MEN1 mutations impacting menin inhibitors in clinical trials · Nature Communications 2026What this evidence cannot show
Screen-derived, in cell lines under drug pressure. Base editing accesses only the substitutions its editing window permits, so the map is systematically incomplete in a way that is predictable from the codon table. Inhibitor-specific calls depend on the exposures chosen.
Evidence that argues against this (1)
- Screen hits that have never appeared in a patient may be unreachable in vivo because they cost fitness in the absence of drug; a screen over-calls relative to the clinic, just as the clinic under-calls relative to the screen. source 2023
- ← related toETV6::NTRK3 (ETV6-TRKC), t(12;15)(p13;q25)
Counter-evidence: Drilon 2016 (PMID 26884591) documents acquired NTRK3 G623R solvent-front resistance after a durable entrectinib response in an ETV6-NTRK3 tumour
- related to →MEN1 drug-interface mutations — the first proof that a chromatin drug can drive Darwinian escape in patients
six patient-observed substitutions at four residues - a validation set, not a training set
a compound-specific base-editor screen answers whether a ligand's resistance spectrum is narrow or broad
if genetic and non-genetic escape are substitutes, a perfect mutation map predicts the wrong failure mode in infants
the base-editor screen is the prospective alternative that entry recommends in place of a model trained on six patient labels