The Complete Parts List: What Biology Knows and What It Can Predict

Biology · Prediction · Evidence

The Complete Parts List

Biology can enumerate almost everything and predict almost nothing. On the questions that are genuinely open — and on the harder question of how you tell a demonstrated result from an announced one.

October 2026 · Luis Matos Ferreira

Imagine being handed a complete inventory of an aircraft. Every rivet, every wire, every alloy, catalogued to the gram, with a photograph of each part from six angles. Nothing is missing. Now predict how it flies.

You cannot, and the reason is not that the inventory is incomplete. It is that an inventory is a different kind of object from a theory. The list tells you what exists; flight is a claim about what happens, and the two are separated by a body of rules that no amount of additional cataloguing will produce.

That is the position modern biology is in, and it is worth being precise about how strange it is. We have the parts list. For an increasing number of organisms we have something close to a complete parts list — every gene, every cell type, every protein fold, every synapse in a fly brain. The catalogue is one of the great achievements of the century. And handed a genome nobody can tell you what the organism will look like, how it will behave, or when it will die.

Almost every serious open problem in biology is a version of that gap. What follows is a map of them, and then a harder question that the map raises but does not answer: given how much gets announced, how would you actually know which parts of the gap have been closed?

An inventory is not a theory

Start by being fair about the size of the achievement, because the argument only works if the catalogue is genuinely as good as claimed.

Sequencing has gone from a fifteen-year international programme to an overnight service. Protein structure prediction, which was a standing embarrassment for fifty years, was substantially solved for well-behaved single-domain folds within about eighteen months.[1] Single-cell atlases have enumerated cell types in tissues where the textbook count was off by a factor of three. The complete synaptic wiring diagram of an adult fruit fly brain — roughly 140,000 neurons and 50 million synapses — was published in 2024.[2] These are not incremental results. They are the kind of thing that will still be in the textbooks in a century.

~200MProtein structures predicted and released in the AlphaFold databaseDEEPMIND / EMBL-EBI

~140kNeurons in the complete adult Drosophila connectomeFLYWIRE 2024

~90Genes essential to the minimal synthetic cell whose function is unknownJCVI-SYN3.0

That last number is the one to sit with. JCVI-syn3.0 is a bacterium stripped to 473 genes — the smallest genome that will sustain a self-replicating cell.[3] Every one of those genes is essential: remove it and the cell dies. And roughly a fifth of them have no known function. This is the most reduced, most controlled, most deliberately understood organism ever constructed, and a fifth of its irreducible core is a black box.

If you want a single image for the state of biology, that is it. Not ignorance at the frontier — ignorance at the centre of the simplest thing we have built.

The claim, stated carefully

This is not the tired complaint that biology is “merely descriptive”. Description at this scale is hard and valuable, and molecular biology has produced genuine mechanistic theory — the genetic code, the operon, the mechanism of DNA replication, population genetics. The claim is narrower and more awkward: the mechanistic wins are local, and they do not compose. We understand transcription initiation and we understand receptor binding, and we cannot put five hundred such facts together and predict a cell, let alone an organism.

The map nobody has

The central unsolved problem is the genotype-to-phenotype map: given a genome, predict the organism. It sounds like one problem. It is at least four, and each is open in its own way.

Protein function, as opposed to structure

Structure prediction was solved. Function prediction was not, and the difference is routinely elided. Knowing the fold of a protein frequently does not tell you what it does — and the failure cases are systematic rather than random. Roughly a third of the human proteome is intrinsically disordered, with no single structure to predict. Many proteins do their work by switching between conformations, and a static prediction shows you one. Allostery, the mechanism by which binding at one site changes behaviour at a distant one, is the basis of most biological regulation and is still not reliably predictable. And none of it addresses how a protein finds its correct partner in a cytoplasm at 300–400 mg/ml of macromolecule, which is closer to a gel than to a solution.

The dark proteome

Something like 15–20% of human protein-coding genes have no meaningful functional characterisation, and the neglect is not random either. Research attention correlates strongly with when a gene was first described and how easy it was to assay in the 1990s, not with its likely importance.[4] Whole protein families — large parts of the kinome, most of the “understudied” GPCRs and ion channels — remain dark for reasons that are sociological as much as technical.[5] A field that studies what it already has reagents for will keep finding that its own map is complete.

Regulatory grammar

We can locate enhancers. We cannot read them. Predicting expression from non-coding sequence, or predicting what a non-coding variant will do, remains unreliable — which is why the overwhelming majority of genome-wide association hits still have no mechanism attached to them twenty years in.[6] The uncomfortable synthesis of that literature is the omnigenic model: for a typical complex trait, essentially every gene expressed in the relevant tissue contributes a little, through networks, and the handful of genes with obvious biological relevance account for a minority of heritability.[7] If that is right, then the reason we cannot predict the trait is not a missing piece. It is that the causal structure is diffuse in a way that resists the kind of explanation biology is set up to produce.

We can enumerate the causes. We cannot compose them.

Condensates, and a field correcting itself

Worth flagging because it is the clearest live example of the cycle. Liquid–liquid phase separation went from a curiosity to a universal explanation in under a decade: nucleoli, stress granules, transcriptional hubs, synapses, everything. Some of that is real. A considerable amount rests on in-vitro concentrations and overexpression conditions that do not obtain in a cell, and the field's own methodologists have said so in print.[8] Distinguishing a functional compartment from an artefact of the assay is now an active research programme in its own right — which is healthy, and also a reminder of how long a fashionable mechanism can run before anyone checks.

From one cell to a body

A fertilised egg becomes an organism with the right number of fingers, in the right places, at the right scale, almost every time. This is so familiar that it takes an effort to notice it is unexplained.

The classical framework is positional information: cells learn where they are from concentration gradients of signalling molecules, then interpret that position to choose a fate. This is not hand-waving — it is one of the few places in biology where a quantitative theory genuinely predicts. Work on the early fly embryo showed that the gap gene network transmits enough information about position for a nucleus to locate itself to within about one percent of the embryo's length, which is roughly a single cell width, and that an optimal decoder of those four gene expression levels reproduces the observed pattern boundaries.[9] That is a real result of the kind physics would recognise: a measured information capacity, a decoding model, and a prediction that matches.

The rare case where developmental biology has predictive theory. No single morphogen gradient carries enough information to place a nucleus to within one cell. Read jointly, the gap gene expression levels do — and an optimal decoder built from them reproduces the observed boundary positions. The diagram is schematic; the result it illustrates is quantitative and was measured.

The trouble is how little of development looks like this. Three questions immediately outrun the framework.

Scaling. Embryos of different sizes — the same species, or a surgically halved embryo — produce correctly proportioned bodies. A gradient with a fixed decay length cannot do that on its own; something has to measure the whole and adjust. Mechanisms have been proposed for particular systems. A general account does not exist.

Termination. Your liver is the size it is. Transplant a small dog's liver into a large dog and it grows to the recipient's proportions, then stops. The Hippo pathway is a genuine mechanistic advance here — activating its effector YAP produces a reversible several-fold liver overgrowth, which is about as clean a demonstration of a size-control switch as biology offers.[10] But a switch is not a ruler. What the system is actually measuring — mechanical tension, cell density, morphogen dilution across a growing field, competition between cells — remains unsettled, and without that you have identified the brake without finding the speedometer.

Regeneration. An axolotl regrows a limb, complete, with the correct pattern, from the correct level. A planarian regrows a head. A mouse regrows a fingertip and nothing else. The genome is sequenced,[11] the blastema has been atlased at single-cell resolution, and connective tissue cells have been shown to carry the positional memory that tells the regenerate what level it is replacing.[12] Nobody has restored limb regeneration in a mammal, and whether mammals lost the capacity or never had it in a recoverable form is open.

The connectome does not contain the animal

The wiring diagram of C. elegans — 302 neurons, every synapse — has existed since 1986, and was substantially revised and extended for both sexes in 2019.[13] Forty years on, nobody can predict the worm's behaviour from it.

This is the cleanest available demonstration that structure is not sufficient, and it should be more chastening than it is. The connectome omits neuromodulation, which reconfigures the same circuit into functionally different circuits depending on internal state; it omits synaptic strength and its history; it omits extrasynaptic signalling, where a peptide released into the neighbourhood addresses whichever cells carry the receptor, ignoring the wiring entirely. The diagram is necessary. It is very far from the animal.

Two further problems sit alongside it. The memory engram can now be tagged, silenced, reactivated, and in rodents artificially implanted — a genuinely remarkable body of work — and yet the physical substrate of a particular memory remains undecided between synaptic weight changes, intrinsic excitability, and molecular storage inside the cell.[14] And sleep has no agreed function after more than a century of investigation, despite being universal, expensive, and lethal to lose.

Consciousness deserves a sentence, mostly for what has changed. It is now partly an empirical subject: adversarial collaborations, in which proponents of competing theories agree in advance on the experiment and what would falsify each position, have tested integrated information theory against global neuronal workspace and returned results that fit neither cleanly.[15] That is real progress in method, whatever one thinks of the theories. Whether the underlying question is tractable at all remains contested, and honest people disagree.

The questions that only happened once

Three of the largest open problems share an awkward property: they are historical, singular, and therefore only indirectly experimental.

Abiogenesis. Prebiotic chemistry has improved considerably — plausible routes to activated nucleotides under conditions a young planet could supply are a serious achievement.[16] What does not exist is a continuous path from chemistry to a self-sustaining Darwinian system. The old dichotomy between replication-first and metabolism-first has softened into a recognition that both are hard, and the protocell question — how a boundary, a metabolism, and a replicator came to be the same object — is where the difficulty concentrates.

Eukaryogenesis. The discovery and eventual isolation of Asgard archaea narrowed the host lineage considerably.[17][18] The order of events — mitochondrial acquisition before or after the nucleus, before or after phagocytosis-like capability — remains genuinely unsettled, and the event appears to have happened once in four billion years, which is exactly the sample size that makes inference hardest.

The major transitions. Multicellularity arose independently dozens of times; the eukaryotic cell, once. Sex is maintained across most of the tree despite a twofold cost that ought to be decisive. Whether evolvability itself evolves — whether lineages are selected for their capacity to generate useful variation — is an old argument that keeps returning because the evidence is suggestive and never conclusive.

The planetary version

Two open problems here are unusual in that the cost of not solving them is being paid now.

Predictive ecology. Ecosystems undergo abrupt regime shifts — a clear lake turns turbid, a fishery collapses, a forest becomes savanna — and they do not always come back when the pressure is removed. Theory says such transitions should be preceded by generic warning signals: critical slowing down, rising variance and autocorrelation as the system's recovery from small perturbations gets sluggish.[19] The theory is elegant, works in lakes and in models, and has an uneven record in the field, where the time series are short and noisy and the shifts are often only obvious afterwards. Forecasting a tipping point in advance, in a real ecosystem, at policy-relevant lead times, is not a solved problem.

Dark taxa. Most species on Earth have never been described — overwhelmingly fungi, bacteria, nematodes and insects. Estimates of the total remain uncertain to within a factor of several,[20] and metabarcoding now generates sequence clusters far faster than taxonomy can turn them into named, described organisms. We are documenting a biosphere while it changes, with an instrument that produces identifiers rather than knowledge.

Interlude: biology has no equations

Step back and the pattern across all of this is one absence.

Physics has theory that constrains data: a measurement that violates conservation of energy is presumed wrong before it is presumed revolutionary. Biology, outside population genetics and a few pockets of biophysics, has no comparable constraint. It has data, and it has narrative fitted to data. This is not a moral failing — living systems are historical, contingent, and layered in a way that resists compression — but it has a consequence that is rarely stated plainly: in biology, almost nothing is ruled out in advance. There is no equation to check a result against. Any given finding is plausible, because the field's theoretical structure is too loose to make many things implausible.

Which is precisely what makes the next question hard.

So how would you know?

Everything above is a map of what is open. The harder problem for anyone reading about biology from outside it is a different one: given the volume of announcement, how do you tell a demonstrated result from a promising one?

The two areas where the gap between the two is widest right now are aging and developmental biology — not because they are worse fields, but because they are the two where the results are most striking, the commercial pressure is highest, and the distance between a mouse and a claim is longest. They also happen to be the two areas where the field has produced an unusually honest instrument for measuring itself. Start there.

The attrition ladder. Nearly all longevity claims that reach the public are still on the top rung. The third rung exists, is run by the US National Institute on Aging, and has eliminated most of the compounds sold on the strength of the first. The fourth has, so far, eliminated everything.

Aging: the one honest scoreboard

The Interventions Testing Program is the most useful object in the field and the least discussed outside it. Run by the NIA at three independent sites since 2004, it tests compounds for lifespan extension in genetically heterogeneous mice — deliberately not an inbred strain, so a result is not an artefact of one genetic background — with adequate power, both sexes, and a protocol that does not let the testing lab talk itself into a positive.

Its record is brutal, and it is the single best calibration available for anyone trying to judge this literature.

Compound	ITP result	Public profile before the test
Rapamycin	Extends lifespan, both sexes, even when started at ~20 months of age[21]	Modest — an immunosuppressant
Acarbose	Extends lifespan, strongly in males, weakly in females[22]	Low — a diabetes drug
17-α-estradiol	Extends lifespan in males only[23]	Low
Canagliflozin	Extends lifespan in males only[24]	Low
Resveratrol	No effect[25]	Enormous — the “red wine molecule”
Nicotinamide riboside	No effect[26]	Enormous — a major supplement category
Fisetin	No effect[26]	High — marketed as a senolytic
Metformin (alone)	No effect[23]	Very high — the subject of a proposed human trial
Curcumin, green tea extract	No effect[25]	High

Two patterns are worth naming. The first is the inverse correlation between public profile and result: the compounds with the largest consumer markets are, almost without exception, the ones that failed. The second is the sex asymmetry. Three of the four hits work mainly or only in males, which is a real and unexplained biological finding, and also a warning about how much of the pre-ITP literature was single-sex.

What this table does and does not say

It does not say these compounds do nothing. Several have real metabolic effects; resveratrol's biochemistry is not fictional. It says they do not extend lifespan in mice under controlled multi-site conditions — which is the claim that was made for them, and the reason people buy them. It also does not say rapamycin is safe for healthy humans: it is an immunosuppressant with a real side-effect profile, tested for lifespan in mice, not in people.

Senolytics: strong in mice, unproven in humans

The senescence story is genuinely good mouse biology. Genetically clearing p16-positive cells delayed age-associated pathology in a progeroid model,[27] and in normal mice extended median lifespan by roughly a quarter while improving function in several tissues.[28] Drug combinations that preferentially kill senescent cells reproduce parts of this.

Human translation has gone considerably less well. The most advanced programme, a senolytic for osteoarthritis, missed its endpoints in a phase 2 trial; the human studies that exist elsewhere are small, open-label, or biomarker-based.[29] There is also an unresolved problem underneath: p16 is a poor marker, senescent cells are heterogeneous, and some of them are doing useful work — in wound healing and in limiting fibrosis. “Clear the senescent cells” may turn out to be the wrong verb for a population that needs sorting rather than deleting.

The result that is genuinely new

Partial reprogramming is the one development in aging biology that deserves the attention it gets, so it is worth being exact about what has and has not been shown.

Cyclic expression of the Yamanaka factors — enough to move cells toward a younger epigenetic state, not enough to send them back to pluripotency — extended lifespan in a progeroid mouse model.[30] The strongest single result restored vision in aged mice and after optic nerve crush injury, using three factors and showing the effect depends on active DNA demethylation, which is what makes it mechanistically informative rather than merely striking.[31] Long-term partial reprogramming has since been shown to be tolerated in normal mice with improvements in skin and kidney.[32] A companion line of work argued from mice engineered to accumulate repair-driven epigenetic disruption that information loss alone, without mutation, drives aging phenotypes — an important claim which has drawn substantive methodological criticism and should be read as contested.[33]

What has not been shown: lifespan extension from systemic partial reprogramming in normal mice, a delivery method that works in a whole organism without viral vectors and inducible transgenes, and any control of dose and duration that keeps the intervention on the useful side of teratoma formation. There is no human data. Several billion dollars of well-staffed private research has been directed at this for some years now without a published result that changes the summary above.

The biology is real. The product is not close.

Clocks, and the hazard of measuring what you want

Epigenetic clocks — methylation-based predictors of chronological and then biological age — are excellent correlates,[34] and second-generation versions trained on health outcomes rather than birthdays perform considerably better.[35] The CALERIE trial, two years of roughly 12% caloric restriction in healthy humans, reported a slowed pace of aging by one such measure.[36]

The hazard is structural. A clock is a model fitted to predict age; intervening on the features it uses is not the same as intervening on aging, and the field has not established that resetting a clock resets biology. That is the whole question, and it is being routinely assumed rather than tested — most visibly in the consumer market for biological-age tests, where test–retest reliability is often poorer than the effect sizes being reported back to customers.

What has been tested in humans

Very little, and the results are sobering. An mTOR inhibitor improved immune response in older adults in phase 2a and then failed its phase 3 endpoint.[37] The metformin trial designed to establish aging as a regulatory indication has been proposed and discussed for a decade without being fully funded and completed.[38] Plasma-based interventions collapsed commercially after a regulatory warning, and the most interesting recent work suggests the effect in parabiosis may owe more to diluting old factors than to adding young ones[39] — alongside a report that extended heterochronic parabiosis produced a measurable lifespan increase in mice.[40] The largest well-designed mammalian trial currently running outside mice is in companion dogs.[41]

The honest summary: no intervention has been shown to extend human lifespan. One class of drug has repeatedly extended mammalian lifespan across independent sites in both sexes with late-life dosing. Everything else is mouse-only, sex-specific, biomarker-only, or has failed.

Development in a dish: what the pictures do not show

Organoids are the other place where the image and the object have diverged. The achievement is real: pluripotent stem cells, given the right medium and left alone, self-organise into structures containing the correct cell types of a brain, gut, kidney or retina, in roughly the correct developmental sequence.[42] Nobody instructs them where to put anything. That is a genuine demonstration that a great deal of developmental patterning is intrinsic to the cells.

Then there are the constraints, which are rarely in the caption.

Why organoids stop. With no blood vessels, oxygen and nutrients arrive only by diffusion, which runs out at a couple of hundred micrometres. Past a few millimetres the interior becomes hypoxic and stressed — and cell stress in cortical organoids has been shown to distort the very subtype identities the model is used to study. The size ceiling is not an engineering detail; it is a limit on what the model can be evidence for.

Three limits, in order of how often they are omitted. No vasculature, hence a size ceiling of a few millimetres and a stressed or dying interior beyond it — and that stress is not inert, it demonstrably impairs the specification of cortical cell subtypes, which is precisely what such organoids are used to model.[43] Variability: batch-to-batch reproducibility has improved markedly with directed protocols[44] but remains the field's central methodological problem. And developmental stage: cerebral organoids correspond transcriptionally to fetal tissue, do not reproduce proper cortical layering, and have no reproducible circuit-level function. The most credible structural advance has been assembloids — fusing separately patterned regions so that, for instance, interneurons migrate between them as they do in a real forebrain.[45]

None of which is a criticism of the work. It is a criticism of the word “mini-brain”, which has done real damage to public understanding, and of the recurring suggestion that these structures might be approaching sentience. Nothing in the data supports that, and the people closest to the work say so.

Embryos that are not embryos

Stem-cell-derived embryo models are the most striking recent result in developmental biology and the one most in need of careful reading. Mouse models built from stem cells alone have completed gastrulation and reached neurulation and early organogenesis, with a beating heart and a neural tube.[46][47] Human models have reached structures corresponding to roughly day 14 post-implantation.[48] Separately, mouse embryos have been cultured entirely outside the uterus from pre-gastrulation to late organogenesis, which is what made much of this tractable.[49]

The qualifiers matter. Formation efficiency is typically low single-digit percentages — most aggregates do not form anything usable, and the published images are the successes. Fidelity to real embryos is actively disputed, with legitimate disagreement about how much of the resemblance is morphological rather than molecular. And the field moved fast enough that governance followed rather than led: the ISSCR revised the fourteen-day rule to case-by-case review in 2021,[50] which tells you the science outran the framework built to anticipate it.

Bioelectricity: real experiments, premature framing

Manipulating the resting voltage of cells produces some of the most arresting results in modern biology. Blocking gap junction communication in planaria yields worms that regenerate two heads — and keep doing so through subsequent rounds of cutting, with an unaltered genome.[51] Changing membrane potential in Xenopus induces well-formed eyes in tissue that has no business making eyes.[52] Dissociated frog cells reassemble into motile constructs that do not resemble any frog.[53]

These experiments are real and they are not adequately explained by the standard morphogen account. What runs ahead of the evidence is the interpretive layer: the proposal that bioelectric state constitutes a stored, rewritable “target morphology” — a pattern memory the tissue consults. That is a strong claim, the mechanism connecting voltage to the transcriptional programmes that would have to implement it is thin, and independent replication outside the originating group is limited. This is the highest hype-to-evidence ratio in developmental biology, which is not the same as saying it is wrong. It is saying that a striking result and an established mechanism are different things, and the gap between them is where most of the discussion currently sits.

Against which, quietly: engineered cell–cell signalling circuits have been used to program cells to self-organise into multi-layered structures from scratch.[54] Small scale, hundreds of cells, no press cycle — and a genuine proof that patterning rules can be written rather than only observed.

The ledger

Claim in circulation	Where it actually stands
NAD boosters (NR / NMN) slow aging	Failed the multi-site lifespan test; human trials show biomarker movement without demonstrated functional benefit
Consumer “biological age” tests measure how fast you are aging	Clocks are good correlates; causality unestablished, and test–retest reliability is frequently worse than the reported effect
Young plasma rejuvenates	Commercially shut down after regulatory warning; dilution of old factors is a better-supported explanation than addition of young ones
Senolytics are ready for people	Strong in mice, one outright phase 2 failure, no positive functional outcome trial
Reprogramming will reverse human aging soon	Real biology, unsolved delivery and safety, no wild-type lifespan result, no human data
Brain organoids are “mini-brains”	Fetal-stage, unvascularised, stressed cores, no reproducible circuit function
Synthetic embryos are embryos	Low formation efficiency, contested molecular fidelity, published images are the successes
Bioelectricity encodes a target morphology	Experiments real and unexplained; the stored-pattern interpretation is a hypothesis with thin mechanism
Rapamycin extends mammalian lifespan	Holds up. Repeated, multi-site, both sexes, effective from late life. Human safety is a separate and unsettled question

Why the gap keeps opening

It would be comfortable to blame journalists. The causes are structural and mostly sit inside the field.

Nothing is ruled out in advance. This is the interlude's point cashed in. Without theory that constrains, a surprising result carries no prior improbability, so it is published rather than doubted. Physics would demand extraordinary evidence for a claim that violated a conservation law. Biology has almost no claims of that shape to violate.

Effect sizes shrink on replication, systematically. Attempts to reproduce landmark preclinical findings have recovered a minority of them, and where effects survive they are typically much smaller than first reported.[55][56] This is not fraud; it is the predictable output of small samples, flexible analysis, and a publication filter that selects for large effects. It means a first report should be read as an upper bound, not an estimate.

The mouse-to-human step is where things die, and it is skipped rhetorically. A great deal of the aging literature is written as though the mouse result were a preview. The ITP exists precisely because even the mouse-to-mouse step was failing.

Biomarkers substitute for outcomes. Measuring a clock is fast; measuring lifespan takes a lifetime. The temptation to accept the proxy is enormous, and it is strongest exactly where the commercial pressure is highest.

Money arrived before the science was ready. Longevity biotech is now funded at a scale that requires narrative. That does not corrupt the underlying work — much of it is excellent — but it changes what gets said about it, and by whom, and how early.

What would count as progress

Two things, and neither is a discovery.

The first is theory that constrains. The most interesting current bet is that large learned models — the “virtual cell” programmes — can supply prediction without mechanism: train on enough perturbation data, and the model tells you what a cell will do without anyone understanding why. That may well work as engineering. Whether it constitutes understanding, and whether a model that cannot say what would falsify it can play the role that theory plays in physics, is the field's live methodological argument, and it is not settled by pointing at benchmark scores.

The second is more instruments like the ITP. The single most valuable thing in aging research is not a compound. It is a multi-site, blinded, adequately powered testing programme that publishes negative results, run by people with no stake in the answer. It cost relatively little, it has saved the field years of chasing failures, and there is no equivalent in most of biology. Developmental biology has nothing comparable. Neither does microbiome research, which needs it more.

Which returns to the parts list. The reason the inventory does not fly is that we lack the rules; the reason it is hard to tell how close we are to the rules is that we lack the instruments that would tell us. Both gaps are closable, and the second is much cheaper than the first. It is also less interesting to fund, which is roughly why we have two hundred million protein structures and one honest scoreboard.

If you want to go further

The ITP's published cohorts, which are readable and unsparing. Boyle, Li and Pritchard on the omnigenic model, for the clearest statement of why the genotype–phenotype map may not have the shape we wanted. Petkova and colleagues on optimal decoding in the fly embryo, for what a quantitative developmental result actually looks like. And the reproducibility project's cancer biology report, for a picture of what a field looks like when it checks itself.

References

1. Jumper, J. et al. (2021). “Highly accurate protein structure prediction with AlphaFold.” Nature 596, 583–589. See also Abramson, J. et al. (2024), “Accurate structure prediction of biomolecular interactions with AlphaFold 3”, Nature 630, 493–500.
2. Dorkenwald, S. et al. (2024). “Neuronal wiring diagram of an adult brain.” Nature 634, 124–138. The FlyWire whole-brain connectome.
3. Hutchison, C. A. et al. (2016). “Design and synthesis of a minimal bacterial genome.” Science 351, aad6253. JCVI-syn3.0: 473 genes, of which roughly 149 are of unknown function.
4. Stoeger, T., Gerlach, M., Morimoto, R. I. & Nunes Amaral, L. A. (2018). “Large-scale investigation of the reasons why potentially important genes are ignored.” PLOS Biology 16(9), e2006643.
5. Oprea, T. I. et al. (2018). “Unexplored therapeutic opportunities in the human genome.” Nature Reviews Drug Discovery 17, 317–332.
6. Manolio, T. A. et al. (2009). “Finding the missing heritability of complex diseases.” Nature 461, 747–753.
7. Boyle, E. A., Li, Y. I. & Pritchard, J. K. (2017). “An Expanded View of Complex Traits: From Polygenic to Omnigenic.” Cell 169(7), 1177–1186.
8. Alberti, S., Gladfelter, A. & Mittag, T. (2019). “Considerations and Challenges in Studying Liquid–Liquid Phase Separation and Biomolecular Condensates.” Cell 176(3), 419–434.
9. Petkova, M. D., Tkačik, G., Bialek, W., Wieschaus, E. F. & Gregor, T. (2019). “Optimal Decoding of Cellular Identities in a Genetic Network.” Cell 176(4), 844–855.
10. Dong, J. et al. (2007). “Elucidation of a universal size-control mechanism in Drosophila and mammals.” Cell 130(6), 1120–1133. See also Camargo, F. D. et al. (2007), Current Biology 17(23), 2054–2060.
11. Nowoshilow, S. et al. (2018). “The axolotl genome and the evolution of key tissue formation regulators.” Nature 554, 50–55.
12. Nacu, E. et al. (2013). “Connective tissue cells, but not muscle cells, are involved in establishing the proximo-distal outcome of limb regeneration in the axolotl.” Development 140(3), 513–518. See also Nacu, E. et al. (2016), Nature 528, 407–410.
13. Cook, S. J. et al. (2019). “Whole-animal connectomes of both Caenorhabditis elegans sexes.” Nature 571, 63–71. The original: White, J. G. et al. (1986), Phil. Trans. R. Soc. Lond. B 314, 1–340.
14. Josselyn, S. A. & Tonegawa, S. (2020). “Memory engrams: Recalling the past and imagining the future.” Science 367, eaaw4325.
15. Cogitate Consortium (2025). “Adversarial testing of global neuronal workspace and integrated information theories of consciousness.” Nature 642, 133–142.
16. Powner, M. W., Gerland, B. & Sutherland, J. D. (2009). “Synthesis of activated pyrimidine ribonucleotides in prebiotically plausible conditions.” Nature 459, 239–242.
17. Spang, A. et al. (2015). “Complex archaea that bridge the gap between prokaryotes and eukaryotes.” Nature 521, 173–179.
18. Imachi, H. et al. (2020). “Isolation of an archaeon at the prokaryote–eukaryote interface.” Nature 577, 519–525.
19. Scheffer, M. et al. (2009). “Early-warning signals for critical transitions.” Nature 461, 53–59.
20. Mora, C., Tittensor, D. P., Adl, S., Simpson, A. G. B. & Worm, B. (2011). “How Many Species Are There on Earth and in the Ocean?” PLOS Biology 9(8), e1001127.
21. Harrison, D. E. et al. (2009). “Rapamycin fed late in life extends lifespan in genetically heterogeneous mice.” Nature 460, 392–395. The founding ITP positive result.
22. Harrison, D. E. et al. (2019). “Acarbose improves health and lifespan in aging HET3 mice.” Aging Cell 18(2), e12898.
23. Strong, R. et al. (2016). “Longer lifespan in male mice treated with a weakly estrogenic agonist, an antioxidant, an α-glucosidase inhibitor or a Nrf2-inducer.” Aging Cell 15(5), 872–884. Metformin alone produced no significant lifespan extension in ITP cohort testing; the contrasting positive result was obtained in a single inbred strain — Martin-Montalvo, A. et al. (2013), “Metformin improves healthspan and lifespan in mice”, Nature Communications 4, 2192. The discrepancy is itself the argument for heterogeneous-mouse testing.
24. Miller, R. A. et al. (2020). “Canagliflozin extends life span in genetically heterogeneous male but not female mice.” JCI Insight 5(21), e140019.
25. Strong, R. et al. (2013). “Evaluation of resveratrol, green tea extract, curcumin, oxaloacetic acid, and medium-chain triglyceride oil on life span of genetically heterogeneous mice.” Journals of Gerontology A 68(1), 6–16.
26. Harrison, D. E. et al. (2021). “Astaxanthin and meclizine extend lifespan in UM-HET3 male mice; fisetin, SG1002, dinitrophenol, nicotinamide riboside, MIF098 and 17-α-estradiol do not significantly affect lifespan in either sex at the doses and schedules used.” GeroScience 43, 2311–2347.
27. Baker, D. J. et al. (2011). “Clearance of p16^Ink4a-positive senescent cells delays ageing-associated disorders.” Nature 479, 232–236.
28. Baker, D. J. et al. (2016). “Naturally occurring p16^Ink4a-positive cells shorten healthy lifespan.” Nature 530, 184–189.
29. Justice, J. N. et al. (2019). “Senolytics in idiopathic pulmonary fibrosis: Results from a first-in-human, open-label, pilot study.” EBioMedicine 40, 554–563. The osteoarthritis phase 2 failure (UBX0101) was reported by the sponsor in 2020 and remains unpublished as a full trial report.
30. Ocampo, A. et al. (2016). “In Vivo Amelioration of Age-Associated Hallmarks by Partial Reprogramming.” Cell 167(7), 1719–1733.
31. Lu, Y. et al. (2020). “Reprogramming to recover youthful epigenetic information and restore vision.” Nature 588, 124–129.
32. Browder, K. C. et al. (2022). “In vivo partial reprogramming alters age-associated molecular changes during physiological aging in mice.” Nature Aging 2, 243–253.
33. Yang, J.-H. et al. (2023). “Loss of epigenetic information as a cause of mammalian aging.” Cell 186(2), 305–326. Read alongside the published critiques of its design and interpretation.
34. Horvath, S. (2013). “DNA methylation age of human tissues and cell types.” Genome Biology 14, R115.
35. Belsky, D. W. et al. (2022). “DunedinPACE, a DNA methylation biomarker of the pace of aging.” eLife 11, e73420.
36. Waziry, R. et al. (2023). “Effect of long-term caloric restriction on DNA methylation measures of biological aging in healthy adults from the CALERIE trial.” Nature Aging 3, 248–257.
37. Mannick, J. B. et al. (2018). “TORC1 inhibition enhances immune function and reduces infections in the elderly.” Science Translational Medicine 10, eaaq1564. The subsequent phase 3 trial of RTB101 did not meet its primary endpoint (sponsor announcement, 2019).
38. Barzilai, N., Crandall, J. P., Kritchevsky, S. B. & Espeland, M. A. (2016). “Metformin as a Tool to Target Aging.” Cell Metabolism 23(6), 1060–1065. The TAME trial rationale.
39. Mehdipour, M. et al. (2020). “Rejuvenation of three germ layers tissues by exchanging old blood plasma with saline-albumin.” Aging 12(10), 8790–8819.
40. Zhang, B. et al. (2023). “Multi-omic rejuvenation and lifespan extension on exposure to youthful circulation.” Nature Aging 3, 948–964.
41. Creevy, K. E. et al. (2022). “An open science study of ageing in companion dogs.” Nature 602, 51–57. The Dog Aging Project, including its rapamycin trial.
42. Lancaster, M. A. et al. (2013). “Cerebral organoids model human brain development and microcephaly.” Nature 501, 373–379.
43. Bhaduri, A. et al. (2020). “Cell stress in cortical organoids impairs molecular subtype specification.” Nature 578, 142–148.
44. Velasco, S. et al. (2019). “Individual brain organoids reproducibly form cell diversity of the human cerebral cortex.” Nature 570, 523–527.
45. Birey, F. et al. (2017). “Assembly of functionally integrated human forebrain spheroids.” Nature 545, 54–59.
46. Amadei, G. et al. (2022). “Synthetic embryos complete gastrulation to neurulation and organogenesis.” Nature 610, 143–153.
47. Tarazi, S. et al. (2022). “Post-gastrulation synthetic embryos generated ex utero from mouse naive ESCs.” Cell 185(18), 3290–3306.
48. Oldak, B. et al. (2023). “Complete human day 14 post-implantation embryo models from naive ES cells.” Nature 622, 562–573.
49. Aguilera-Castrejón, A. et al. (2021). “Ex utero mouse embryogenesis from pre-gastrulation to late organogenesis.” Nature 593, 119–124.
50. International Society for Stem Cell Research (2021). ISSCR Guidelines for Stem Cell Research and Clinical Translation. The revision that moved the fourteen-day limit to case-by-case specialised review.
51. Oviedo, N. J. et al. (2010). “Long-range neural and gap junction protein-mediated cues control polarity during planarian regeneration.” Developmental Biology 339(1), 188–199.
52. Pai, V. P., Aw, S., Shomrat, T., Lemire, J. M. & Levin, M. (2012). “Transmembrane voltage potential controls embryonic eye patterning in Xenopus laevis.” Development 139(2), 313–323.
53. Kriegman, S., Blackiston, D., Levin, M. & Bongard, J. (2020). “A scalable pipeline for designing reconfigurable organisms.” PNAS 117(4), 1853–1859.
54. Toda, S., Blauch, L. R., Tang, S. K. Y., Morsut, L. & Lim, W. A. (2018). “Programming self-organizing multicellular structures with synthetic cell-cell signaling.” Science 361, 156–162.
55. Begley, C. G. & Ellis, L. M. (2012). “Raise standards for preclinical cancer research.” Nature 483, 531–533.
56. Errington, T. M. et al. (2021). “Investigating the replicability of preclinical cancer biology.” eLife 10, e71601.

Published October 2026 · The Complete Parts List

On method and tools

This piece was written collaboratively with Claude Opus 5 (Anthropic): human specification and critical review, machine synthesis and drafting. The three diagrams are schematic and illustrate mechanisms, not measured data — the gradient figure in particular is a drawing of a result, not the result. Where a claim is a demonstrated finding it is cited; where it is contested (the epigenetic-information account of aging, the bioelectric target-morphology hypothesis, the fidelity of embryo models) that is said in the text rather than left to the reader. Negative results are cited as carefully as positive ones, which in this subject is most of the point.

The hardest sentence to write in a piece like this is the one that says a beautiful result has not yet been shown to mean what everyone hopes it means.

Authored by: Luis Matos Ferreira
Physicist & Developer