The 156-Year Wall: Why Fixing Aging Still Leaves Your Neurons on the Clock
Russian researchers built a mathematical model of the human body as a system of critical organs, then asked a single question: if we cured every reversible cause of aging and left only the one process we currently cannot reverse, the slow accumulation of random DNA mutations in our cells, how long would we live? Their answer is roughly 146 to 194 years of median lifespan, about twice today’s figure. The model reveals a sharp split between organs. Tissues that constantly renew themselves, like the liver, shrug off mutation damage almost indefinitely. But irreplaceable cells, specifically neurons in the brain and cardiomyocytes in the heart, become the hard bottleneck. Because these cells are never replaced, mutations pile up until the organ fails. The headline implication is sobering for the optimists: somatic mutations alone cannot explain why we die at 80, meaning other aging processes contribute at least as much, and even a near-total cure of aging would not deliver immortality.
For decades the longevity field has ranked its enemies into a tidy list of aging hallmarks, from worn-out telomeres to exhausted stem cells and tired mitochondria. Most of these, in principle, look reversible. Drugs and cell therapies can imagine undoing them. But one hallmark stands apart. Somatic mutations, the random typos that accumulate in the DNA of ordinary body cells across a lifetime, represent pure information loss. Once written, they are extraordinarily hard to erase. So a team at the Skolkovo Institute of Science and Technology and the AI Research Institute in Moscow asked the question the whole field has been circling: if everything else were fixed, what would this one irreversible process leave on the table?
To answer it they borrowed a tool from engineering rather than biology. Reliability theory treats a machine as a chain of critical components, and the machine fails when any essential part fails. The body, in their hands, becomes a series of organs, and death arrives when the first indispensable organ crosses a failure threshold. They fed the model real mutation rates measured from single-cell sequencing of human neurons, heart muscle, liver, and airway tissue, then let simulated populations of a million virtual people age under mutation pressure alone.
The results split the body cleanly in two. Post-mitotic organs, made of cells that never divide again once mature, are the losers. Neurons drive median lifespan down from a theoretical non-aging ceiling of 1,759 years to under 200. Heart muscle behaves similarly. These cells cannot be replaced, so every lethal mutation is permanent subtraction. Proliferating tissues are the winners. The liver, continuously restocked by dividing cells and stem cells, never failed within a hundred thousand simulated years. Its damaged cells are simply swept out and replaced by fresh copies.
Stitch the organs back together and the model predicts a median human lifespan of about 156 years under its central assumptions, with a plausible range of 146 to 194. That is roughly double the current 79 years seen in the longest-lived populations. It is a stunning number and a deflating one at the same time. Stunning because it doubles the human span from a single mechanism. Deflating because our real median is 79, not 156, which means somatic mutations account for only about half the gap. The rest must come from the other hallmarks. The dream of longevity escape velocity, where medicine outruns death indefinitely, finds no support here. Even a near-total victory over aging runs into a wall built by the brain and the heart.
Actionable Insights
Honesty first: this is a modeling paper with no direct intervention, so it offers strategic direction rather than a protocol you can start Monday. The take-home messages are about where the field, and your attention, should point.
The dominant effect size is the predicted lifespan multiplier. Curing all reversible hallmarks but not somatic mutations yields a modeled median of 146 to 194 years, which is 197 percent of the observed Swiss median of 79 years under the central estimate, and a maximum of 470 years, which is 385 percent of the observed human maximum of 122 years. Read the other direction, the magnitude is humbling: even this near-total cure closes only about half the distance between a normal life and a non-aging body, whose theoretical median is 1,759 years.
The strategic message for anyone allocating effort or money is that not all organs age equally. Interventions that renew proliferating tissues (liver, gut, blood, skin) buy little additional lifespan once other hallmarks are handled, because those tissues already self-repair. The real bottleneck is irreplaceable tissue: brain and heart. The practical implication the authors draw is that neuronal and cardiac protection, and eventually neuronal replacement therapy, is the only plausible route past roughly 150 years. For a present-day biohacker the defensible read is narrow: protect post-mitotic cells from excess mutagenic load now (avoid known mutagens, manage oxidative and inflammatory burden), because damage to non-dividing cells is the one thing you genuinely cannot undo later.
Paper:
Somatic mutations impose an entropic upper bound on human lifespan
Anyone want to argue with this? Maybe John Hemming.
Worked for 13m 42s
Bottom line
The basic calculus is mostly competent, and the headline numbers can be reproduced from the authors’ chosen parameters. But the paper does not establish a credible mathematical upper bound on human lifespan.
One tissue model—the lung model—is internally inconsistent and implemented differently from the published equations. More importantly, the 146–194-year estimate is dominated by a highly speculative “probability that a mutation kills a cell.” Under the paper’s own stated plausible range for that parameter, predicted median lifespan spans roughly 68 to 1,738 years.
My overall assessment:
Aspect Finding Core survival mathematics Good Arithmetic from chosen parameters Good Statistical uncertainty treatment Poor Lung model Substantively broken Reproducibility from released code Poor Claim of an “entropic upper bound” Not justified Quantitative credibility of 146–194 years Low I reviewed the s41514-026-00421-6_reference.pdf and checked the authors’ released code.
What is good
- Equations 6–22, covering survival functions, constant hazards, exponential cell loss, failure time and population integration, are essentially correct.
- The central post-mitotic result is mathematically transparent:tfailure≈ln(K/Xcrit)μ.t_{\text{failure}}\approx \frac{\ln(K/X_{\text{crit}})}{\mu}.tfailure≈μln(K/Xcrit).With their parameters:
- Brain: ln(1/0.60)/0.00244≈209\ln(1/0.60)/0.00244\approx209ln(1/0.60)/0.00244≈209 years.
- Heart: ln(1/0.55)/0.00261≈229\ln(1/0.55)/0.00261\approx229ln(1/0.55)/0.00261≈229 years.Population variability and background mortality then produce their approximately 194- and 208-year medians.
- The derivation of average proliferative potential in Models IIIA and IIIB is algebraically reasonable as a mean-field approximation.
- The Fréchet–Hoeffding formulas in Equations 41–43 are correct. The authors also correctly acknowledge that the lower bound generally is not attainable for four marginals.
- I independently recomputed the multi-organ results using only brain, heart and background mortality:
Dependence assumption My median Paper My “maximum” Paper Lower Fréchet bound 146.93 146 209.90 210 Independence 156.43 156 470.72 470 Upper Fréchet bound 194.28 194 557.34 557 So the arithmetic behind the headline is real. It is almost entirely a brain-and-heart result; the elaborate liver and lung models contribute essentially nothing to those numbers.
Major problems
1. The published lung model cannot generate the published lung result
Equation 39 gives:
dBdt=−μbB+replication term (fbb−fxx).\frac{dB}{dt}=-\mu_b B+\text{replication term},(f_{bb}-f_{xx}).dtdB=−μbB+replication term(fbb−fxx).
But on PDF page 30 the paper explicitly sets fbb=fxxf_{bb}=f_{xx}fbb=fxx at every basal-cell abundance. Therefore:
dBdt=−μbB.\frac{dB}{dt}=-\mu_b B.dtdB=−μbB.
All replication effects cancel. The delayed, replication-supported plateau followed by collapse at 4,423 years in Figure 2 cannot follow from the published equations.
Using the paper’s mean parameters, the equation actually gives approximately:
tfailure≈ln(1/0.23)0.00208≈706 years,t_{\text{failure}}\approx\frac{\ln(1/0.23)}{0.00208}\approx706\text{ years},tfailure≈0.00208ln(1/0.23)≈706 years,
not 4,423 years.
The released implementation obtains the plotted behavior by using undocumented and erroneous alternatives:
- fbb/fxx=1.1f_{bb}/f_{xx}=1.1fbb/fxx=1.1, rather than equality.
- The softmax intercept sign is reversed.
- Its denominator assumes fbb=fxxf_{bb}=f_{xx}fbb=fxx even while computing them differently, so the three “probabilities” do not sum to one.
- At homeostasis, the intended probabilities are about 3%, 3%, and 94%; the code produces approximately 49%, 45%, and 1.6%.
These discrepancies are visible in the authors’ Model III implementation and configuration.
This invalidates the lung-specific conclusions. It barely affects the 146–194-year headline because lung survival is still nearly 100% during the first 200 years.
2. The decisive lethality parameter is not empirically identified
The model requires plethalp_{\text{lethal}}plethal, the probability that one mutation immediately kills a cell. No such measurement exists. The paper constructs it by multiplying:
- an essential-gene indicator,
- an ad hoc transformed pathogenicity score,
- a haploinsufficiency score.
That product is not a validated probability model. Independence and calibration are assumed rather than demonstrated. Worse, mutations are measured in living cells: mutations that actually killed their cells generally cannot appear in the sequencing dataset. That introduces direct survivorship bias.
Failure time scales approximately as 1/plethal1/p_{\text{lethal}}1/plethal, so this is not a peripheral uncertainty—it controls the answer.
Putting SNV and indel lethality at the paper’s own lower and upper bounds gives, under the same model:
Lethality setting Median lifespan range Stated lower bounds approximately 1,719–1,738 years Authors’ defaults 147–194 years Stated upper bounds approximately 68–88 years Thus 146–194 is one selected parameterization, not a robust bound. The paper itself concedes that sensitivity bands span one to two orders of magnitude in time.
3. The reported confidence intervals do not follow from the described propagation
For example, Table 4’s brain slope intervals and Table 1’s fixed lethality probabilities imply a lethal-rate interval around 2.20–2.65×10−32.20–2.65\times10^{-3}2.20–2.65×10−3 using the endpoints, or approximately 2.25–2.60×10−32.25–2.60\times10^{-3}2.25–2.60×10−3 using ordinary independent-error propagation.
Table 1 instead reports 1.53–3.95×10−31.53–3.95\times10^{-3}1.53–3.95×10−3.
For liver, the slope intervals imply approximately 0.87–1.64×10−30.87–1.64\times10^{-3}0.87–1.64×10−3, while Table 1 reports 0.07–2.25×10−30.07–2.25\times10^{-3}0.07–2.25×10−3. The Methods do not explain this much wider calculation.
4. Parameter uncertainty is treated as human-to-human variability
The simulation draws each synthetic person’s mutation rate from distributions based largely on confidence intervals or standard errors of estimated mean slopes.
Those are different quantities:
- Confidence intervals describe uncertainty about an estimated population parameter.
- Inter-individual variation describes genuine biological heterogeneity.
Converting the first into the second creates artificial demographic variation and especially distorts extreme-tail survival.
Indel rates inferred from SNV rates are also treated as statistically independent during error propagation, despite being mathematically derived from those SNV rates.
5. “Maximum lifespan” is incorrectly defined
The paper repeatedly invokes an initial population of eight billion but defines “maximum” at S=10−5S=10^{-5}S=10−5, which means one survivor per 100,000—not one survivor among eight billion.
For its baseline:
- Reported 10−510^{-5}10−5 quantile: approximately 29,216 years.
- Expected-one-survivor quantile for eight billion, S=1/(8×109)S=1/(8\times10^9)S=1/(8×109): approximately 57,866 years.
Therefore 470 years is a 99.999th percentile under the model, not a mathematical or population maximum. Comparing that quantity with Jeanne Calment’s record age of 122 is not an apples-to-apples comparison.
6. Normal tissue turnover is largely omitted
The proliferating-tissue models start at X=KX=KX=K. Their logistic replication gate is then exactly zero. Consequently, replicative capacity is not consumed by normal homeostatic turnover; it is used only after mutation deaths create a deficit.
That assumption essentially builds the 37,000-year liver result into the model. With mutation-only turnover, the exhaustion timescale is approximately H/(2μ)H/(2\mu)H/(2μ), which is tens of thousands of years. Normal apoptosis, replacement and injury are intrinsic physiology, not necessarily removable “aging hallmarks.”
7. The stated POLR2A lower-bound calculation misses a factor of two
The paper uses one POLR2A target length but divides by the diploid genome length. If either allele is a lethal target because of haploinsufficiency, both target copies and genome size double, leaving the haploid ratio unchanged.
Keeping all their other assumptions, the lower bound should be roughly 3.9×10−73.9\times10^{-7}3.9×10−7, not 1.94×10−71.94\times10^{-7}1.94×10−7. This is minor compared with the thousand-fold overall uncertainty, but it is a genuine arithmetic/setup error.
8. Released code does not reproduce the manuscript configuration
The current repository contains several additional discrepancies:
- Background hazard 0.0016130.0016130.001613, whereas the reported 1,759-year baseline requires about 0.0003940.0003940.000394.
- LPC replication rate fixed at 0.112/year, versus approximately 10.0/year in Table 9.
- The Fréchet routine uses 1/(8 billion)1/(8\text{ billion})1/(8 billion), while the paper defines its reported “maximum” using 10−510^{-5}10−5.
- Default Monte Carlo counts differ substantially from the Methods.
The released repository therefore does not provide clean reproduction of the published figures without undocumented alterations.
Assumption-heavy but not strictly wrong
- Brain failure at 40% neuronal loss and heart failure at 45% cardiomyocyte loss are judgments, not measured chronic-failure thresholds. The brain criterion is actually cognitive independence rather than organismal death.
- Acute myocardial-infarction loss is used to infer consequences of gradual centuries-long cardiomyocyte decline. Adaptation could make those situations very different.
- The lower Fréchet endpoint is called “perfect negative dependence,” although no such joint distribution need exist for four organs. The numerical curve remains a valid pointwise lower bound, but not necessarily a realizable scenario.
- Background mortality is assumed independent of all organ failures. The advertised dependence bounds therefore do not cover dependence between background mortality and organ deterioration.
- “Entropy” is never calculated. There is no information-theoretic or thermodynamic entropy variable or theorem in the paper. It is a descriptive label for irreversible mutation accumulation.
- The claim that the remaining gap implies other hallmarks contribute “at least as much” is not mathematically established. Lifespan effects and hazards are not additively decomposable.
Final assessment
This is a worthwhile exploratory framework showing that, if neurons and cardiomyocytes die independently at approximately 0.24–0.26% per year and organs fail after losing roughly 40–45% of those cells, failure occurs on a roughly 200-year timescale.
It is not evidence that human median lifespan has a mutation-imposed upper bound of 146–194 years. That precision is unsupported, the lung result is mathematically defective, and the dominant lethality parameter is too speculative to anchor a quantitative lifespan estimate.
Confidence: high on the algebraic and code inconsistencies; medium-high on the overall conclusion that the numerical lifespan claim is not reliable.
As Aubrey de Grey says this is just the next frontier of diseases of aging. As we live longer we will likely uncover new diseases we didn’t even know existed. Damage accumulation that kills is builds up faster than types of damage that would kill us if given the time, and some of that damage type that builds up over longer periods could be more difficult to clear out.
Kind of disturbing in a way lol, but necessary to achieve longevity escape velocity.