Portfolio theory — a hedge only works against uncorrelated risk, and nobody has estimated the correlation matrix of resistance routes
also called covariance matrix of resistance · diversification analogy
Combining two drugs only helps if the ways the leukemia escapes them are genuinely different. Investors have a precise way to check that, and it needs a table of how each escape route relates to every other one. That table has never been built for this disease.
The same thing, for a scientist
Diversification reduces variance only to the extent that constituent risks are uncorrelated; correlated holdings are one holding with extra fees. The graph currently names four resistance routes to menin inhibition — MEN1 interface mutation, adaptive non-genetic rewiring, PRC1.1 loss with MYC reactivation, and signalling co-mutation — and reasons about combinations as though they were independent. Portfolio theory says the object that determines the right combination is the covariance matrix, and that its off-diagonal terms can be negative as well as positive. A negative correlation between routes is not a technicality: it means the routes are *substitutes*, so the observed frequency of one is an inverse readout of the availability of another. That gives a sharp prediction about infants, whose leukemia has an exceptionally silent mutational landscape [E3] and who should therefore show markedly less genetic and markedly more non-genetic escape than the adults in whom the 39% MEN1-mutation figure was measured [E1].
What the existing orthogonality argument does and does not cover
The graph already contains hyp-menin-plus-engager-orthogonal-pre-emption, which argues qualitatively that partners outside the transcriptional layer will extend durability and partners inside it will not. That is a statement about the sign of the correlation between mechanisms, made by inspection.
This analogy adds two things that inspection cannot supply. First, a magnitude: how correlated, estimated from data rather than asserted. Second, and more interesting, the possibility of negative correlation, which qualitative layer-reasoning does not naturally generate and which has a concrete clinical consequence — a genotype that rarely mutates is not a safer genotype, it is one whose escape is invisible to sequencing.
The substitution insight, stated as plainly as possible1 deeper
If a leukemic cell can escape a menin inhibitor by rewiring its transcriptional state, it does not need to wait for a MEN1 mutation. If it cannot rewire, a mutation is its only route. Selection therefore does not sample the routes independently — it takes whichever is cheapest, and the cheapest route pre-empts the others.
The testable consequence is an anti-correlation between genetic and non-genetic escape across populations and across models. Adults with AML, carrying substantial mutational burden and clonal diversity, show MEN1 mutations in 39% of post-revumenib relapses 1. Infant KMT2A-rearranged ALL has an exceptionally silent mutational landscape consistent with its prenatal origin 3 and, in the one serial-sample study available, becomes menin-resistant without any menin exposure and with sustained on-target drug effect 2. If the substitution logic holds, MEN1 mutation should be rare in children and the field's principal resistance biomarker will not work in the population this project exists to serve.
Why this is not simply 'infants have fewer mutations'
The trivial version predicts proportionally fewer of every mutation. The substitution version predicts something stronger and separable: that the share of resistance attributable to genetic escape falls faster than mutational burden does, because the non-genetic route absorbs the demand. Those two predictions come apart quantitatively and can be distinguished if anyone measures resistance mechanism and mutational burden in the same children.
Where the analogy overreaches
Portfolio variance is a coherent objective for an investor holding many positions; a single patient is one draw, not a portfolio, and minimising variance is not obviously what you want for one child — you may rationally prefer a high-variance strategy with a chance of cure over a low-variance strategy with none. The covariance framing belongs at the level of trial design and population strategy, where many patients really are a portfolio, and should not be pushed down to the bedside.
The analogy is also silent on timing. Portfolio correlations are estimated over a period; resistance routes unfold sequentially, and a route that is unavailable in month one may open in month six. A static covariance matrix cannot represent that, and the honest version of this analysis reports correlations conditional on line of therapy.
| domain | finance |
|---|---|
| source concept | portfolio variance, the covariance matrix, and the requirement that hedges be uncorrelated with what they hedge |
| relationship to existing | Extends hyp-menin-plus-engager-orthogonal-pre-emption from a qualitative claim about the sign of mechanism correlation to a quantitative claim about magnitude, and adds the possibility of negative correlation, which layer-based reasoning does not generate. |
mapping
[
{
"bio": "each named resistance route (MEN1 mutation, adaptive rewiring, PRC1.1 loss, signalling co-mutation, lineage switch)",
"analog": "an asset in a portfolio, each with its own risk"
},
{
"bio": "adding a combination partner",
"analog": "adding a holding intended as a hedge"
},
{
"bio": "two partners that both act in the transcriptional layer",
"analog": "two positively correlated holdings — one position with extra fees and extra toxicity"
},
{
"bio": "the rate at which each route is used under selection",
"analog": "the weight the market puts on each asset"
},
{
"bio": "a cell choosing the cheapest available escape, so that plasticity pre-empts mutation",
"analog": "negatively correlated holdings, which are substitutes rather than complements"
},
{
"bio": "infant leukemia with a silent mutational landscape",
"analog": "a portfolio in which one asset class is simply unavailable, forcing weight onto the others"
},
{
"bio": "MEN1 sequencing as a resistance biomarker",
"analog": "monitoring one asset class and inferring the health of the whole portfolio from it"
},
{
"bio": "the correlation matrix over all routes",
"analog": "the covariance matrix, which is the object that actually determines the optimal combination and which nobody has estimated"
}
]generated hypotheses
[ "hyp-resistance-routes-negatively-correlated-genetic-vs-plastic" ]
- If MEN1 mutation is rare in children, then the field's main resistance biomarker, and the main rationale for switching to a differently-binding inhibitor, both fail in paediatric patients while appearing to work in adults.
- Negative correlation between routes means closing one route pushes traffic onto another, so a combination that successfully suppresses mutation may simply increase the rate of non-genetic escape without changing the total.
- Optimising a portfolio requires estimating a covariance matrix from limited data, which is notoriously unstable. A badly estimated matrix would confidently recommend the wrong combination.
- Has anyone measured resistance mechanism and mutational burden in the same menin-treated children? If not, the paediatric menin trials now open could answer it by banking serial samples with both DNA and transcriptomic readouts.
- Are the four named routes mutually exclusive in the CRISPR resistance screens that discovered them, and has anyone looked at the co-occurrence structure rather than reporting each hit separately?
- Does the correlation structure change with line of therapy, so that a static matrix is the wrong object entirely?
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.review
Across the menin-inhibitor class, acquired MEN1 mutations were described in 39% of post-revumenib relapses and have not been identified following ziftomenib or bleximenib; composite complete remission rates are 20-35% with differentiation syndrome in 10-20%.
The promise of menin inhibitors: from approval to triplet regimens · Thakur RK, Wang ES · Hematology, American Society of Hematology Education Program 2025What this evidence cannot show
Review-level and adult-dominated. Differential follow-up and sequencing intensity are strong alternative explanations for the compound differences.
Evidence that argues against this (1)
- The 39% comparator is regimen-dependent, not a population constant: emergent menin-binding-site mutations occurred in 13% of 42 relapsed/refractory patients on a revumenib triplet, 40% of whom were KMT2A-rearranged. Escape-route frequencies move with regimen, response depth and follow-up. A covariance matrix whose entries are not stable across the regimens you intend to choose between is not a decision object. Issa GC et al. 2026
- E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.paper
In serial patient-derived xenografts, heavily pretreated KMT2A::AFF1 ALL samples were far less menin-inhibitor-sensitive than earlier samples from the same patients despite no prior menin-inhibitor exposure, with sustained on-target transcriptional effect, emergent RAS-pathway and TP53 co-mutations that were individually insufficient in vitro, and possible KMT2C/KMT2D involvement.
Upfront menin-inhibitor resistance in multiply pretreated leukemias · Mahdavi L, Bernt KM, et al. · Experimental Hematology 2025What this evidence cannot show
Two patients. The observation that named co-mutations were individually insufficient is itself evidence that the routes are not simply additive.
Evidence that argues against this (1)
- Searched for any published co-occurrence analysis of menin-inhibitor resistance routes — MEN1 interface mutation versus PRC1.1 loss versus cell-state escape versus signalling co-mutation — scored jointly in the same samples or screens. None was found. The off-diagonal terms this analogy is built around have never been estimated even once, which means the analogy currently supplies a framework with no data in it. Recorded so the gap is known to have been checked rather than skipped. Q12-analogy-counter (search record) 2026
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
Infant B-ALL, including MLL-rearranged cases, has an exceptionally silent DNA mutational landscape consistent with its prenatal origin, and genome-wide methylome and transcriptome analysis of 69 uniformly treated infants shows the discriminating signal lies in epigenetic rather than mutational features.
Integrative methylome-transcriptome analysis unravels cancer cell vulnerabilities in infant MLL-rearranged B cell acute lymphoblastic leukemia · Tejedor JR, Bueno C, Cobo I, et al. · Journal of Clinical Investigation 2021What this evidence cannot show
Establishes low mutational burden and epigenetic dominance at diagnosis; says nothing directly about resistance mechanism under menin inhibition, which is the inferential leap this analogy makes.
Evidence that argues against this (1)
- The inference from a mutationally silent diagnostic genome to a low rate of acquired genetic escape is weaker than the entry allows. In the founding report of clinical menin-inhibitor resistance, an unbiased base-editor screen generated interface mutants de novo and recovered the same escape positions seen in patients, and the same mutants conferred resistance in xenografts. Mutation supply under sustained drug pressure in a large leukaemic population is governed by per-division mutation rate at a small number of permissive codons, not by diagnostic tumour mutational burden. The prediction that infants will show markedly less genetic escape may still hold, but not for the reason given here. Perner F et al. 2023
- E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
Loss of PRC1.1 subunits PCGF1 or BCOR causes menin-inhibitor resistance by reactivating non-canonical menin targets including MYC, and these cells show reduced monocytic signatures, become BCL-2-dependent and are resensitised by venetoclax — a route whose countermeasure differs entirely from that of MEN1 mutation.
Epigenetic regulation of noncanonical menin targets modulates menin inhibitor response in acute myeloid leukemia · Zhou X, Zhang L, Aryal S, et al. · Blood 2024What this evidence cannot show
Model-derived; PRC1.1 loss has never been sought in a menin-treated patient, so its real-world frequency and its correlation with other routes are both unmeasured.
Evidence that argues against this (1)
- WHERE THE ANALOGY BREAKS, ON ITS OWN TERMS. The operational proposal is to estimate a resistance-route covariance matrix and use it to rank combinations. Portfolio theory's own empirical literature is that this does not work even where the data are vastly better. DeMiguel, Garlappi and Uppal evaluated fourteen mean-variance and extended optimisation models against the naive 1/N rule on seven empirical datasets and found none consistently superior on Sharpe ratio, certainty-equivalent return or turnover, concluding that out of sample the gain from optimal diversification is more than offset by estimation error in the estimated moments. That is with decades of daily returns. The proposed 4x4 leukaemia covariance matrix would be estimated from a handful of differently designed, underpowered CRISPR screens with no common outcome definition; the expected estimation error dwarfs any signal, and the predicted output — a ranking that differs from the layer-based heuristic — is exactly what pure noise also produces. The defensible residue of this analogy is qualitative: asking whether two routes are substitutes or complements, and refusing to assume independence. The optimiser should be dropped. DeMiguel V et al. 2009
estimation-error objection: across seven datasets no optimisation model beat naive 1/N because covariance estimation error dominates (DeMiguel 2009), so the planned combination ranking should be dropped in favour of a qualitative substitutes/complements test
- analogous to →MEN1 drug-interface mutations — the first proof that a chromatin drug can drive Darwinian escape in patients
genetic escape as one asset in the resistance portfolio
- analogous to →Adaptive, non-genetic resistance — the leukemia keeps the fusion and rewires how it is maintained
non-genetic escape as the substitute asset
- analogous to →MYC reactivation through PRC1.1 loss — a named, druggable bypass out of menin dependence
PRC1.1 bypass as a third route with its own countermeasure
puts a magnitude and a possible negative sign on the orthogonality argument
generated by treating resistance routes as substitutes in a portfolio