Engineered Endosymbiosis

Speculative If a host cell can be made to keep a working endosymbiont and hand it to its daughters without anyone applying selection every generation — which no published experiment has shown — the prize is a metabolic capability installed as a compartment instead of rebuilt as a pathway. A heterotroph acquires photosynthesis, or nitrogen fixation, by taking in one cell that already does the chemistry, rather than by importing and balancing the dozens of genes that encode it. That is the shortcut evolution used twice; the question this brief asks is whether it can be used on purpose.

Frontier State the magnitude at the numbers the laboratory record actually contains. The longest-lived engineered endosymbiosis on record is an Escherichia coli strain carrying a bacterial ADP/ATP translocase inside a respiration-deficient yeast, reported stable for more than 40 host doublings. The most recent photosynthetic version grew yeast under light with 0.38 mM bicarbonate as the only carbon source in the medium. The baseline those displace is an industrial yeast fermentation that buys its carbon as sugar, and a metabolic-engineering project that installs carbon fixation gene by gene.

Established The organelle-transplant route has a smaller and more precisely bounded prize, and the bound is the honest headline. In the 2024 red-algal experiment, about 20% of hamster CHO-K1 cells took up one to three isolated chloroplasts and photosystem II electron transport persisted for at least two days, with thylakoid stacks visibly degrading by day four. No carbon fixed by those chloroplasts was traced into host metabolism. So the upside on that branch is bounded by a quantity no group has reported: grams of fixed carbon delivered per gram of host per day. The only published arithmetic on the point, from 2011, estimated roughly 25 cyanobacterial cells inside a CHO cell, and about 14,000 inside a macrophage, before the light-driven supply would cover the host’s demand.

Speculative The gains, if they arrive, are unusually cheap to copy and unusually narrow in who first collects them. The reagents here are strains and plasmids rather than a beamline, so a working chimera diffuses at the cost of a shipping tube; that is a different distribution from a technology gated on capital equipment. But the medical framing that draws the press — photosynthetic tissue, oxygen delivered by light — accrues only to whoever solves cytosolic sensing of bacteria in a mammalian cell, a problem the industrial yeast version does not have to touch. Everything above is conditional on one result: retention through a host lineage with selection switched off, and a measured contribution to host growth rate. Section 7 specifies that experiment, and section 5 says why nobody has run it yet.

Established Engineered endosymbiosis means one organism living and dividing inside the cytoplasm of another and contributing metabolism to it. That is a narrower object than it sounds. It is not a microbial community in a gut, a lumen or on a root surface — those are covered by Microbiome Engineering, which is about populations sharing a compartment with a host, not about a cell inside a cell. It is not a co-culture, and it is not an engineered consortium exchanging metabolites across a membrane in shared medium.

Established The distinction is mechanical rather than semantic: an endosymbiont is inherited with the host cell, and a community member is reacquired. A gut community is re-established at each generation from the environment. An endosymbiont partitions at cytokinesis. Everything hard about this subject — and everything the field has not yet done — sits in that second sentence.

Established Nature has completed the transition at least twice in ways we can still read. Mitochondria and plastids are the ancient cases. The recent ones are the Paulinella chromatophore and the nitroplast of the haptophyte alga Braarudosphaera bigelowii, whose cyanobacterial ancestor was taken up on the order of 100 million years ago — recent enough that the intermediate state is still visible.

Frontier Reviewers of the field converge on five molecular hallmarks that separate an organelle from a lodger. The 2025 review by Cournoyer, De and Mehta lists genome minimisation with gene transfer to the host, protein import and export machinery, metabolic crosstalk, alterations to the peptidoglycan wall, and host-controlled replication. An engineered system that has none of these is a bacterium surviving in a vacuole; a system that has all five is an organelle.

Established Engineered work to date has achieved exactly one of the five in a durable form. Metabolic crosstalk is solved: a bacterial ADP/ATP translocase moves ATP across the endosymbiont membrane and the host uses it. Genome minimisation has been demonstrated piecemeal, with excisions of up to 46 kb from a candidate endosymbiont genome. Protein import into an engineered bacterial endosymbiont has not been demonstrated at all, and neither has host control of the endosymbiont’s division.

Established There are three distinct experimental routes and they should not be pooled. Route one installs a whole living bacterium — E. coli or a cyanobacterium — inside a eukaryote. Route two transplants an isolated organelle, usually a chloroplast, into a cell that never had one. Route three is comparative: read a natural endosymbiosis caught mid-transition and copy what it did. The three routes have different failure modes and very different track records, and press coverage routinely merges them.

Frontier Route two is the one that produced the 2024 headlines and it is the weakest of the three on inheritance. An isolated chloroplast is not a cell. It cannot divide on its own without the nucleus-encoded division machinery of its former host, so a transplanted chloroplast cannot in principle be inherited indefinitely by a mammalian cell line — it can only be diluted. That is a structural limit, not a technical one, and it is the reason the two-day persistence figure is not on a trajectory toward a two-hundred-day one.

Speculative The interesting target is therefore route one with route three’s parts list. A live, dividing, genetically tractable endosymbiont that can be minimised, made dependent, and eventually handed a protein-import channel is the only architecture in which the five hallmarks could all be assembled. Whether that assembly is reachable in a laboratory timescale rather than an evolutionary one is the open question this brief is about.

Established The anchor result the field cites for organelle transplantation is a 2024 paper putting isolated red-algal chloroplasts inside hamster cells, and its measured content is narrower than its title suggests. Aoki and colleagues isolated chloroplasts from the thermoacidophilic red alga Cyanidioschyzon merolae and co-cultivated them with CHO-K1 cells at a 1:100 cell-to-chloroplast ratio. Approximately 20% of the CHO-K1 cells contained one to three chloroplasts immediately after co-cultivation, and about 1% were chloroplast-rich, carrying seven to 45 each.

Established The photosynthetic measurement is a pulse-amplitude-modulation fluorescence signal, not a carbon balance. The incorporated chloroplasts maintained photosystem II electron transport activity in the mammalian cells for at least two days. The maximum quantum yield of photosystem II in the isolated chloroplasts, before transfer, averaged 0.32 against 0.39 for whole algal cells — a modest loss on isolation. Electron microscopy showed thylakoid membranes spreading apart at two days and the layered structure degraded by four.

Frontier The paper reports that co-cultivated CHO-K1 cells grew faster than controls, and does not establish that the chloroplasts caused it. This is the single most over-read sentence in the coverage. A growth difference in a co-culture has many available explanations, including material released by ruptured chloroplasts acting as nutrient. No isotope tracing was reported, so no fixed carbon has been followed from the chloroplast into a host metabolite.

Established The commission’s anchor is correct in substance with one wording correction worth making. The paper claims activity for at least two days rather than up to two days; the upper bound is set by where the authors stopped measuring, and the structural evidence at four days is of degradation rather than of function. The brief uses the paper’s own phrasing.

Established None of this is new in kind. Isolated chloroplasts were first taken up by mammalian cells in 1969. Nass reported uptake of isolated chloroplasts by mammalian cells in Science that year. What 2024 adds is a demonstration that the transplanted organelle is still doing photochemistry inside the animal cell rather than merely sitting there, which is a real advance and a small one.

Frontier The second anchor, engineered bacterial endosymbionts in yeast, is the stronger result and its headline number is 40 host doublings. Mehta and colleagues at Scripps engineered an E. coli expressing a bacterial ADP/ATP translocase and introduced it into a respiration-deficient Saccharomyces cerevisiae cox2 mutant. The bacterium exported ATP to the host; the host supplied thiamin to a thiamin-auxotrophic bacterium. Two independent secondary accounts of that work give the same figure: the chimera was stable for more than 40 doublings.

Established The 40 doublings were obtained under continuous reciprocal selection, and that qualification is the whole subject. Growth was on glycerol without thiamin, so a yeast that lost its bacterium could not respire and a bacterium that escaped its yeast could not make thiamin. The experiment measures how long a mutual auxotrophy holds under selection. It does not measure inheritance, because nothing was asked to survive the removal of the selection.

Established Avoiding degradation required borrowing virulence factors from an intracellular pathogen. The engineered bacteria expressed SNARE-like proteins derived from Chlamydia to avoid lysosomal destruction. This is the only trick in the field that reliably works, and it is worth stating plainly that the enabling technology for artificial endosymbiosis is, at present, a set of proteins evolved by an obligate intracellular pathogen for the same purpose.

Frontier The photosynthetic version of the same architecture reached 15 to 20 host generations. Cournoyer and colleagues fused engineered Synechococcus elongatus PCC 7942 into yeast spheroplasts by polyethylene-glycol fusion — the protocol previously used to install mitochondria — carrying an ADP/ATP translocase from Protochlamydia amoebophila and SNARE-like proteins from Chlamydia. Chimeras propagated for at least 15 to 20 generations under defined photosynthetic conditions.

Established The internal controls in that experiment are what make it credible. On the most stringent selection medium the best chimera managed about 14 doublings against roughly two for the host without endosymbionts, a sevenfold difference in permitted growth. Growth failed completely in the dark, and the photosystem II inhibitor DCMU abolished it, so the ATP was demonstrably photosynthetic. Doubling time was about 10 hours against roughly seven for a respiring wild-type yeast, so the rescue was partial.

Established The same group reported the chimeras losing viability after about 20 doublings, probably by loss of the cyanobacteria. That is the field’s own statement of its ceiling, published by the authors, and it is a cleaner piece of evidence about the state of the art than any of the press framing around it.

Frontier Two 2024 papers from the Mehta laboratory at Illinois extended the platform in directions that matter. In the first, engineered cyanobacteria were made to secrete both glucose and ATP, and the resulting chimeras propagated under light with 0.38 mM bicarbonate as the only carbon source, through four consecutive rounds of regrowth on progressively stringent media. Limonene was detected by gas chromatography–mass spectrometry from photosynthetically grown chimeras, as a proof that the architecture can carry a product pathway. No titre was reported in the main text.

Frontier The second 2024 paper is a translocase comparison and it produced the most useful negative result in the literature. De and colleagues put ATP/ADP translocases from red algae, glaucophytes and land plants into the cyanobacterial endosymbiont. Red-algal translocases exported about 2.4 times more ATP than wild-type controls and about 1.5 times more than the bacterial reference, and their chimeras survived six rounds of regrowth, at least 24 doublings. Land-plant translocases preferentially imported ATP instead of exporting it, and those chimeras managed roughly six doublings.

Established That result says the direction of the transporter, not its presence, sets the ceiling. It also explains something about evolution: the plastid translocases of land plants run the other way because a mature chloroplast is a consumer of host ATP in the dark, not a supplier. An engineer copying the most familiar plastid transporter would have built the wrong machine.

Established The structure of the transporter the whole field depends on was solved in 2025. Lin and colleagues reported cryo-electron-microscopy structures of the plastid/parasite ATP/ADP translocator, identified the substrate-binding pocket at the interface between domains, showed that a single conserved asparagine controls substrate specificity, and described a rocker-switch alternating-access mechanism. This is the point at which rational redesign of the exchange step becomes possible rather than combinatorial.

Established The mammalian branch of route one has a 2011 record that is clearer than most of what followed. Agapakis and colleagues microinjected up to ten million S. elongatus cells into single-cell zebrafish embryos and saw autofluorescence persist for up to twelve days, declining throughout; engineered bacteria carrying invasin and listeriolysin O entered CHO cells, with 4.8% of cells scoring positive by flow cytometry; and in macrophages the bacteria grew for two to three days and then declined.

Established The same paper did the stoichiometry that later work mostly skipped. It estimated that roughly 25 cyanobacterial cells inside a CHO cell, or approximately 14,000 inside a macrophage, would be required for the photosynthetic supply to meet the host’s demand. Those are the numbers that decide whether an engineered endosymbiont in an animal cell is a power source or a passenger, and no subsequent mammalian experiment has reported achieving them.

Frontier The longest co-existence in the literature is 101 generations, and it is not endosymbiosis. Azuma and colleagues maintained a coculture of Synechocystis sp. PCC 6803 with the ciliate Tetrahymena thermophila for at least 101 generations across 143.5 days and 41 transfers. The ciliate engulfed the cyanobacteria, but the relationship was predation: the ciliate used the cyanobacteria for growth by eating them, no metabolic benefit from a retained intracellular symbiont was measured, and the authors state the system has no stable equilibrium. A serial-transfer experiment that evolved pairs did improve coexistence, with both populations maintained in seven of twelve tested replicates against zero of four for the original pairing.

Established The ledger, stated plainly: the engineered record is more than 40 host doublings under mutual auxotrophy, at least 24 under photosynthetic selection, 15 to 20 in the founding photosynthetic system, and about two days for a transplanted organelle. No published experiment has removed the selection and asked whether the endosymbiont is still there. That is not a gap in the reading; it is the state of the field.

Established Nature’s most recent completed case gives the target specification. The nitroplast of Braarudosphaera bigelowii derives from the cyanobacterium UCYN-A, whose lineage traces back to nitrogenase-gene amplification from oligotrophic ocean water in 1998. Its genome is 1.44 Mb and it lacks photosystem II, the tricarboxylic acid cycle and carbon fixation, along with biosynthetic pathways for several amino acids and purines — it cannot live alone.

Frontier The 2024 argument for calling it an organelle rests on protein import, synchronised division and size scaling. Proteomic work reported by Coale and colleagues found that over 350 proteins encoded in the algal genome were more abundant in UCYN-A than in the alga itself; the structure divides in sequence with the host cell cycle and is split evenly between daughters; and cell sizes scale linearly, larger algal cells carrying larger nitroplasts. The origin is dated to roughly 100 million years ago.

Frontier The organelle label is a judgement about where a line sits, and the researchers themselves phrase it carefully. The published framing is that UCYN-A has evolved beyond endosymbiosis and functions as an early evolutionary stage nitrogen-fixing organelle; the quoted formulation from the lead author is that it is behaving as an organelle. The criterion doing the work is that endosymbionts encode their own functional proteins while organelles have most of theirs encoded by the host nucleus.

Established The natural comparison set contains several systems caught at different points on the same road, and they are informative about what is missing from the engineered ones. The citrus mealybug carries Moranella endobia, roughly 406 to 410 kb, inside Tremblaya princeps, roughly 139 kb, inside its own cells, maternally transmitted. A rickettsial endosymbiont of the placozoan Trichoplax has a 1.47 Mb genome with about 1,500 genes and mixed mutualistic and parasitic traits. A marine choanoflagellate hosts an obligate associate that carries its own ATP/ADP translocase to import host energy.

Established In the natural systems the host controls the symbiont’s life cycle, and this is the capability no engineered system has. Dual RNA-sequencing across the cereal weevil Sitophilus oryzae and Sodalis pierantonius shows coordinated host and symbiont gene expression governing the symbiont’s proliferation and its subsequent elimination in emerging adults, with the bacteria confined to bacteriocytes and female germ cells. The host decides how many, where, and when they stop.

Established The same acquisition also happens on the parasitic side of the ledger, which bounds how special it is. A single horizontally transferred nucleotide transporter into the ancestor of Microsporidia gave that lineage the ability to steal host ATP, after which it lost mitochondrial and cytosolic energy generation. Acquiring an ATP transporter is evidently easy; becoming an organelle is what is hard.

Frontier The first open question is whether any reported persistence figure measures inheritance at all. Every number in the engineered literature — 40 doublings, 24 doublings, 15 to 20 generations — was collected while a selection that kills the host on symbiont loss was running. A population under strong selection can show high retention with a high per-division loss rate, because the losers die and are never counted. Retention and loss rate are different quantities and the literature reports only the first.

Frontier Nobody knows the per-division loss probability for an engineered endosymbiont. This single number determines everything downstream. If it is 10−2 per division the system is a laboratory curiosity requiring permanent selection; if it is 10−5 it is a heritable trait needing only a weak fitness benefit to be maintained. It is measurable with existing instruments and has not been published.

Frontier Whether host-controlled division is required, or whether high copy number substitutes for it, is genuinely contested. Mitochondria and plastids are partitioned stochastically at high copy number in many lineages, which is tolerant of imprecision; the nitroplast appears to divide in synchrony with the host cell cycle and split evenly. If stochastic partitioning at copy numbers of tens is sufficient, engineered systems do not need a cell-cycle coupling they have no idea how to build.

Frontier It is unresolved whether a transplanted chloroplast contributes any fixed carbon to a mammalian host. Photosystem II electron transport can proceed while the Calvin cycle is starved of the nucleus-encoded enzymes and stromal conditions it needs, and isolated chloroplasts lose stromal contents readily. A fluorescence signal reports that photochemistry is happening; it does not report that sugar is being made, still less that the host is getting it.

Frontier The reason mammalian cells refuse endosymbionts is not agreed. The candidates on the table include lysosomal destruction, cytosolic pattern recognition and the interferon response, autophagic capture, antigen presentation on MHC, and apoptosis triggered by recognition. A 2021 review names mammalian-cell inflexibility toward even benign endosymbionts as an unexplained barrier rather than a solved engineering problem, which is an accurate statement of where the question stands.

Frontier Whether the peptidoglycan wall must go is open, and the evidence cuts both ways. Wall remodelling is listed among the molecular hallmarks of the transition, and plastids retain only vestigial peptidoglycan. But a wall-less endosymbiont is osmotically fragile in a cytoplasm it did not evolve in, and no engineered system has tested removal against retention.

Speculative The direction-of-travel question raised by the 2024 translocase comparison has not been followed up. If land-plant plastid translocases import rather than export ATP, the transporter complement of a mature organelle is not a guide to the transporter complement of a nascent one. That implies the engineering target should be the transporters of recently acquired endosymbionts, which is a different and much smaller catalogue.

Frontier How genome reduction is initiated is unknown in any engineered system. In nature it proceeds by drift under relaxed selection inside a small, asexual, bottlenecked population, over timescales no laboratory runs. Whether reduction can be driven deliberately, by deleting and then supplying the missing function from the host, is untested past excisions of tens of kilobases.

Speculative The most consequential unknown is whether protein import into a bacterial endosymbiont is engineerable at all. Every organelle has one and no engineered endosymbiont has any. The natural machinery involved — the translocons of the outer and inner membranes — is elaborate and coevolved. Whether a minimal synthetic import channel can be bolted onto a gamma-proteobacterium is a question nobody has published an attempt at.

Established Delivery is solved separately for every host and generalises to none of them. Yeast takes endosymbionts by polyethylene-glycol fusion of spheroplasts. Zebrafish embryos take them by microinjection at the single-cell stage. Macrophages take them by phagocytosis. Other mammalian cells need engineered invasion, using invasin and listeriolysin O, and even then the 2011 record is 4.8% of cells positive. There is no method that works across hosts and no method that works at scale.

Established Escaping degradation currently depends on borrowed pathogen proteins with unknown mechanism. The SNARE-like proteins from Chlamydia that keep an engineered endosymbiont out of the lysosome improve stability by a route the authors of the cyanobacterial work explicitly state is still unclear. An unexplained enabling step is a bottleneck even when it works, because it cannot be optimised or ported deliberately.

Frontier There is no way to set or hold the endosymbiont copy number. The chloroplast-transplant data show the problem in miniature: 20% of cells with one to three organelles, 1% with seven to 45, and the number falling with time. A population in which copy number is uncontrolled has no defined dose and cannot be partitioned reliably at division.

Established Nothing couples the endosymbiont’s division to the host cell cycle. In the natural comparison the weevil symbiont’s proliferation and elimination are governed by coordinated host and symbiont transcription, and the nitroplast divides in sequence with host cell division. Engineered endosymbionts divide on their own schedule, which is why they are either diluted out or overgrow the host.

Speculative The absence of a protein-import pathway is the bottleneck that separates this field from an organelle, and it has no partial solution. Organelle proteins are largely encoded in the host nucleus, imported after translation. That is the criterion the nitroplast argument turns on, with over 350 host-encoded proteins found enriched in UCYN-A. Until a host-encoded protein can be put inside an engineered endosymbiont, gene transfer to the host nucleus is a one-way loss of function rather than a step in a transition.

Frontier Photon supply bounds every photosynthetic application in an animal. Tissue attenuates light steeply; a chloroplast or cyanobacterium more than a few millimetres deep receives almost nothing. The 2011 stoichiometry — roughly 25 cyanobacteria per CHO cell — already assumes culture-dish illumination. Any tissue-scale claim needs both the copy number and a light budget, and published claims supply neither.

Frontier The metabolic cost of hosting is real and is usually left out. The best photosynthetic chimera doubled in about 10 hours against roughly seven for a respiring wild-type yeast. A 40% slower host is a substantial tax; under relaxed selection that tax is precisely the force that removes the endosymbiont. The engineering target is not merely retention but retention with a net positive contribution to growth rate.

Established Transplanted organelles cannot self-replicate, which caps route two at dilution. Chloroplast division depends on nucleus-encoded machinery from the original host. A CHO cell has none of it. However long an isolated chloroplast lasts, its copy number per host cell halves at each division, so persistence and inheritance are different problems and only one of them is being worked on.

Frontier Genome minimisation tooling stops well short of what the transition requires. Excisions of up to 46 kb have been achieved in a candidate endosymbiont genome. The natural endpoint is 139 kb total for Tremblaya princeps, and 1.44 Mb for a nitroplast that is still described as reduced. Getting from a 4.6 Mb E. coli to anything in that range requires serial deletion with host complementation at every step, which needs the import pathway that does not exist.

Frontier Measurement is a bottleneck in its own right. The field’s central quantity — endosymbionts retained per host cell per division under no selection — requires single-cell tracking over dozens of generations with a marker that does not itself impose a burden. Flow cytometry on a fluorescent marker, the usual approach, silently counts dead endosymbionts and loses dim ones.

Established One upstream dependency was cleared in 2025 and it matters more than its citation count suggests. The cryo-electron-microscopy structure of the plastid/parasite ATP/ADP translocator, with its substrate-binding pocket and the conserved asparagine controlling specificity, converts transporter selection from a screening problem into a design problem. The 2024 finding that transport direction differs between red-algal and land-plant translocases is exactly the kind of result that structure can now explain.

Frontier The field waits on protein design for a synthetic targeting-and-import route, and that dependency runs outside it. Designing a channel that accepts a host-encoded, targeted protein across two bacterial membranes is a de-novo design problem of a kind that Generative Biomolecular Design is the brief for. Nothing in the endosymbiosis literature suggests the field intends to solve it internally.

Frontier Serial genome minimisation depends on synthesis and assembly capacity this subject does not own. Deleting a megabase from an endosymbiont genome in controlled increments, with complementation, is a genome-writing task. Genetic Engineering and Synthetic Biology carry that capability; the endosymbiosis groups are consumers of it.

Frontier Comparative genomics of recent endosymbioses is the only source of a parts list, and it is sample-limited. The informative systems are the ones caught mid-transition: the nitroplast at roughly 100 million years, the Paulinella chromatophore, the mealybug nesting, the placozoan rickettsia, the choanoflagellate associate. Each new one is a discovery rather than a product, and their discovery rate is not something a programme can schedule.

Established Mammalian work depends on immunology that is being developed for other reasons. Selectively blunting cytosolic bacterial sensing without producing a generally immunodeficient cell is the same problem Immune Engineering works on from the therapeutic side. The endosymbiosis field has no independent route to it.

Speculative Long-run experiments depend on automation that exists and is not being pointed here. A relaxed-selection lineage experiment over hundreds of generations with per-generation sampling is a scheduling problem more than a scientific one; Cloud Laboratories describes infrastructure suited to exactly this shape of work. The binding constraint is that nobody has commissioned it.

Speculative The decisive experiment is vertical inheritance under relaxed selection: a host lineage that retains a functional endosymbiont through 100 or more divisions with no selection applied in any generation, with a measured contribution to host growth rate. Every claim in this brief that engineered endosymbiosis is on a path to a new organelle depends on that result, and no published experiment has attempted it. It is the one result that would most change this brief’s assessment.

Frontier The design is straightforward and uses only what the field already has. Take an established chimera — the E. coli/yeast system that ran more than 40 doublings, or the cyanobacterial system that ran 24 — and split the population. One arm keeps the selection. The other is moved to complete medium in which the host can grow perfectly well without its endosymbiont. Both arms are passaged in parallel for at least 100 host divisions.

Frontier The readout is a loss rate, not a survival time. Score the fraction of host cells carrying at least one endosymbiont at every 10 divisions, by a method that distinguishes live from dead symbionts rather than by bulk fluorescence. Fit the decline to get a per-division loss probability with a confidence interval. That number, not a headline duration, is what the field is missing.

Frontier The growth-rate arm is what turns retention into inheritance. Measure the doubling time of endosymbiont-carrying and endosymbiont-free hosts in the same permissive medium. If carriage costs growth — and the 10-hour against seven-hour comparison in the photosynthetic chimera says it does — then retention under relaxed selection is decaying and the only question is how fast. A positive contribution to growth rate in a medium where the endosymbiont is not required is the result that would make this a heritable trait.

Speculative A clear threshold should be set in advance, because this is a subject where a partial result will be reported as a success. Retention above 90% of host cells after 100 divisions without selection, with a growth-rate difference against endosymbiont-free controls that is not negative, would be a qualitative change in what this field has shown. Retention below 10% would say that everything published so far is a measurement of selection pressure.

Frontier The second experiment is an isotope trace through a transplanted chloroplast. Supply labelled bicarbonate to CHO-K1 cells carrying isolated Cyanidioschyzon chloroplasts under light, and look for label in host central-carbon metabolites against a dark control and an inhibitor control. This settles whether the reported growth-rate difference in the 2024 paper reflects photosynthate transfer or the nutritional value of a degrading organelle.

Frontier The third is a partitioning measurement at cytokinesis. Image dividing chimeras and count endosymbionts into each daughter. Whether the distribution is binomial, as stochastic partitioning predicts, or tighter, as host control would produce, decides whether the field needs to engineer a cell-cycle coupling at all. This experiment needs a microscope and a marker, and has not been reported for any engineered system.

Speculative The fourth is the first attempt at protein import, and it is the highest-value negative result available. Express a host-encoded reporter carrying a candidate targeting sequence and ask whether any of it reaches the endosymbiont interior. A clean failure across a panel of candidates would be the strongest evidence yet that the organelle transition needs evolutionary time rather than engineering, and would be worth publishing as such.

Frontier The fifth is a directed reduction ladder with complementation. Delete an essential biosynthetic operon from the endosymbiont and ask whether the host can supply the product. Each successful step is a genuine instance of the reductive process nature used; each failure marks a function that cannot be outsourced without an import pathway. The current record is excision of up to 46 kb, which is a start on a megabase-scale problem.

Frontier The sixth is a translocase-direction screen informed by the 2025 structure. With the substrate pocket and the specificity-determining asparagine now known, transporters can be designed for export rather than screened for it. The 2024 comparison — roughly 2.4-fold higher ATP export for red-algal translocases and about six doublings for land-plant ones — provides the assay and the benchmark.

Speculative The natural experiment worth watching costs nothing and is already running. The nitroplast lineage is a completed transition roughly 100 million years old, and finding more systems at intermediate stages is the cheapest way to learn which of the five hallmarks comes first. That is a survey rather than a scheduled experiment, and its yield cannot be planned.

Handwave The claim that the decisive experiment is merely a matter of time is the weakest step in the field’s own case. It assumes the relaxed-selection loss rate is small enough that a 100-division run will show retention rather than a clean exponential decay to zero, and there is no measurement supporting that assumption in either direction. The experiment is cheap, uses existing strains, needs no new hardware, and nobody has announced one.

Established The engineering requirement list is short and almost entirely unbuilt. A working artificial organelle needs a delivery method, a degradation-escape mechanism, a bidirectional metabolite exchange, copy-number control, division coupling, a protein-import channel, and a reduced genome. The field has one of these in a form that works reliably, part of a second, and nothing on the rest.

Frontier The exchange step is the mature component and it is now designable. Rickettsial and chlamydial ADP/ATP translocases work in a heterologous host, red-algal variants export about 2.4 times more ATP than wild type, and the transporter’s structure and specificity determinant are known. A second-generation endosymbiont should carry a designed exporter rather than a borrowed one.

Speculative Copy-number control needs a device the field has not proposed. The natural analogue is host transcriptional control over symbiont proliferation, as in the weevil system where the host governs both growth and elimination. A synthetic version would need a host-expressed signal the endosymbiont senses and a growth-rate response to it, which is a two-organism circuit with no published prototype.

Speculative Division coupling may be avoidable and that would be the cheapest win available. If partitioning at copy numbers in the tens is reliable enough, the requirement collapses from a cell-cycle interface to a copy-number floor. The partitioning measurement in section 7 decides this, and it is a week of microscopy.

Frontier The mammalian branch needs an immune-evasion package rather than a single protein. Listeriolysin O for vacuolar escape and chlamydial SNARE-like proteins for lysosomal avoidance address trafficking, not sensing. Cytosolic nucleic-acid and peptidoglycan sensing remain untouched, and a cell engineered blind to them is a cell that cannot be released or transplanted. Immune Engineering owns that trade-off.

Frontier The yeast branch needs none of that, which is why it is ahead and why it should stay the workhorse. Every quantitative advance in this field — 40 doublings, 24 doublings, bicarbonate-only growth, the translocase comparison — came from yeast. A programme that wants the decisive inheritance result should run it in yeast and treat mammalian work as a separate, slower line.

Frontier Product-carrying chimeras are further along than the organelle question, and that is a distortion worth naming. Limonene has already been detected from a photosynthetically grown chimera. Demonstrating a product is easier to fund and publish than measuring a loss rate, so the applied branch is likely to keep outrunning the foundational one. Biomanufacturing Resilience is where those products would land if they scaled.

Speculative The alternative architecture is a consortium, and it currently beats endosymbiosis on every applied metric. Engineered cross-feeding between yeast strains achieved 19.3 times higher 3-hydroxypropionic acid than monoculture, at 4.67 mM, without anyone having to get a cell inside another cell. Anything endosymbiosis offers over a consortium has to come from inheritance and from the shortened diffusion distance, not from the metabolic division of labour, which consortia already do.

Established The nearest adjacent technology is engineered microbial consortia, and it is a competitor rather than a complement. Cross-feeding communities deliver division of labour with none of the delivery, degradation or inheritance problems. The honest framing is that endosymbiosis must justify itself against consortia on the two things consortia cannot do: co-inheritance, and metabolite transfer without a bulk-phase diffusion step.

Frontier Mitochondrial transfer between cells is the adjacent clinical technology and it shares the copy-number problem. Mitochondrial Medicine covers transplantation of organelles into recipient cells; the measurement difficulty there — how many organelles arrived, how long they lasted, whether they functioned — is the same one the chloroplast-transplant work has, and the two literatures should be read together.

Frontier Abiotic carbon fixation is the substitute technology for the photosynthetic use case. Artificial Photosynthesis and Biological Energy Systems pursue the same output through devices and through redesigned free-living organisms. A funder choosing between them should note that the endosymbiotic route currently reports no carbon-fixation rate at all.

Established The origin-of-organelles literature is the parts catalogue. Eukaryogenesis covers the ancient transitions this field is trying to re-run. The recent ones — the nitroplast, the Paulinella chromatophore — are the useful ones, because the ancient cases have had a billion years to erase their intermediates.

Frontier Genome minimisation connects this subject to chassis design. Designer Organisms and Xenobiology both carry the reduced-genome and altered-chemistry work that a purpose-built endosymbiont would draw on. A minimised endosymbiont with a non-standard genetic code would also be a containment device, which is the one place where those programmes and this one clearly converge.

Speculative Closed-habitat life support is the application most often gestured at and least supported. Synthetic Ecosystems needs specified organisms with known stoichiometry. An endosymbiotic chassis would be attractive there only if inheritance were solved, because a habitat cannot re-inoculate its crop every generation.

Established This is a small field concentrated in a handful of laboratories, and that shapes its evidence. The engineered bacterial-endosymbiont line runs from Schultz’s group at Scripps through Mehta’s laboratory at Illinois, which authored the 2022 cyanobacterial system and both 2024 follow-ups and the 2025 review of the field. The chloroplast-transplant line is separate and largely Japanese. When one group produces the platform, the follow-ups and the review, replication by an independent laboratory becomes the missing institutional function.

Established No large named programme with a published budget is working on this. The outputs are individual investigator papers in general journals. There is no consortium, no roadmap document, and no funding line specific to artificial endosymbiosis that this brief could locate. That is a straightforward statement of scale: the field is a few laboratories and their trainees.

Frontier The absence of a replication requirement is the specific institutional failure to watch. The two numbers this brief leans on hardest — more than 40 host doublings, and at least 24 — each come from a single laboratory lineage. Neither has been reproduced elsewhere. The decisive experiment in section 7 would be more valuable run by a group with no stake in the answer.

Speculative There is no regulatory category for an engineered obligate intracellular symbiont. Containment frameworks classify free-living organisms by pathogenicity. An organism whose defining property is that it lives inside another organism’s cells, built partly from the trafficking proteins of Chlamydia and Listeria, does not have an obvious home in those schedules. The institutional gap is not that rules are too strict; it is that no rule addresses the object.

Frontier Publication incentives currently reward the wrong measurement. A longer persistence figure and a detected product are publishable; a per-division loss rate under relaxed selection is a methods result that would be hard to place in a general journal, and it is the number the field needs. This is a case where a funder specifying the deliverable would change the science.

Established The biosafety conversation is downstream of a capability nobody has, which buys time that should be used. No engineered endosymbiont survives outside continuous selection, so none could presently establish anywhere. Writing the containment category while that remains true is cheaper than writing it afterwards, and AI-Biology Governance covers the broader machinery for doing so.

Established The dual-use concern here is unusually concrete because the enabling parts are literally virulence factors. The working engineered endosymbionts carry SNARE-like proteins from Chlamydia, listeriolysin O from Listeria, and invasin. An organism assembled from a pathogen’s cell-entry and lysosome-evasion toolkit, optimised for persistence inside eukaryotic cells, is adjacent to an engineered intracellular pathogen by construction rather than by analogy.

Frontier The mitigating fact is that the field’s central difficulty is the same as the containment mechanism. Every system to date collapses without continuous selection, which is exactly the property that makes it safe. Success on the decisive experiment in section 7 would remove that property. The safety case and the scientific goal point in opposite directions, and that should be said in advance rather than discovered.

Frontier Coverage of the 2024 chloroplast work has over-claimed in a way that has a cost. Two days of photosystem II activity in 20% of cells, with no measured carbon transfer, has been reported as photosynthetic animal cells. Patients and funders reading that will form expectations about light-powered tissue that the measurement does not support, and the correction, when it comes, is paid for by the field’s credibility rather than by the outlets.

Speculative If inheritance were solved in an animal cell, the germline question arrives immediately and has no owner. A heritable endosymbiont in an animal is a heritable modification not encoded in the animal’s genome. Frameworks written around germline genome editing address sequence changes in nuclear DNA and would not obviously reach a cytoplasmic passenger transmitted maternally, as insect endosymbionts are.

Frontier Environmental release is the case where the technology’s value and its risk are the same property. An endosymbiont useful for agriculture — nitrogen fixation inside a crop cell — is useful precisely because it spreads through a lineage without reapplication. That is also what makes it unrecallable. Designer Organisms carries the general version of this argument.

Established Animal-use questions are live now, at small scale. The 2011 work injected up to ten million bacteria into zebrafish embryos and followed them for twelve days. That is ordinary developmental-biology practice, and it is also the point at which artificial endosymbiosis stops being a cell-culture subject; the oversight applied to it is animal-research oversight, which was not designed with self-replicating intracellular passengers in mind.

Speculative The civilizational case rests on one substitution: acquiring a metabolic capability as a compartment rather than as a pathway. If that worked reliably, the unit of biological engineering would change from the gene cluster to the cell, and capabilities that took evolution hundreds of millions of years to package could be moved between lineages in a transfer step. Nothing in the present record demonstrates this, and the brief flags the whole argument accordingly.

Speculative The agricultural version is the one with real magnitude, and it is furthest from demonstration. A nitrogen-fixing endosymbiont inside a crop cell is the obvious analogue of the nitroplast, which fixes nitrogen inside a eukaryotic alga with a 1.44 Mb genome and over 350 host-encoded proteins supporting it. The gap between that specification and anything engineered is the entire subject of section 5.

Frontier The industrial version is nearer and smaller. A photoautotrophic yeast chassis that grows on bicarbonate and light rather than purchased sugar would change the feedstock economics of fermentation. The demonstrated version grew on 0.38 mM bicarbonate under selection with a doubling time around 10 hours, which is not a production organism, and the competing architecture — engineered consortia — is already delivering 19-fold product improvements without the difficulty.

Speculative The distributional character of this technology is unusual and mostly benign. A chimera is a strain. Strains travel cheaply, and a laboratory in any country with a molecular-biology facility could work with one. If this field produces anything, the capability is unlikely to concentrate the way capital-intensive technologies do — which also means containment cannot be achieved by controlling equipment.

Frontier The negative civilizational consequence to weigh is the release case rather than the laboratory case. A self-inheriting endosymbiont deployed in an agricultural species is a permanent addition to a lineage. That is the property that makes it valuable and the property that makes reversal impossible, and no existing framework governs the combination.

Handwave Claims that this route leads to photosynthetic humans or to autotrophic animals do not survive contact with the numbers. They require copy numbers in the thousands per cell, from the only published stoichiometry, and a light flux that tissue does not transmit. The argument works by assertion at the step where the photon budget should appear, and that step has never been supplied.

These horizons track the engineered side of the subject — a living endosymbiont inside a eukaryotic host — rather than the transplanted-organelle branch, which section 2 argues cannot be inherited in principle.

• 10 yr: Frontier The relaxed-selection inheritance experiment gets run and published, and a per-division loss rate exists. Designed rather than borrowed ATP exporters, informed by the 2025 structure, push the selected-growth record past 40 doublings in the photosynthetic system. Independent replication of the Illinois and Scripps results by a third laboratory. A first published attempt at protein import into an engineered endosymbiont, most likely reporting failure.
• 25 yr: Speculative A yeast chimera maintained without selection for hundreds of generations, if the loss rate turns out to be small. Directed genome reduction past a few hundred kilobases with host complementation. A photoautotrophic production chassis at pilot scale, competing against Artificial Photosynthesis and against engineered consortia rather than against sugar fermentation.
• 50 yr: Speculative A synthetic protein-import channel that delivers a host-encoded protein into an endosymbiont, which is the step that converts this field’s product from a passenger into an organelle. Endosymbiont-bearing plant or insect lines under field evaluation, with the regulatory category written for them after the fact.
• 100 / 250+ yr: Handwave Endosymbiosis as a routine engineering operation, with capabilities transferred between lineages as compartments. Every step from the 50-year line to this one works by assertion; the honest statement is that nature took on the order of 100 million years to complete the two transitions we can date, and no argument in the literature explains why a laboratory should take four orders of magnitude less.

• Depends on Genetic Engineering for the serial deletion and complementation that a reduction ladder requires, and Synthetic Biology for the two-organism circuits that copy-number control would need. Eukaryogenesis supplies the only worked examples of the transition this field is trying to re-run, and Generative Biomolecular Design is where a synthetic import channel would have to come from. The mammalian branch additionally waits on Immune Engineering. Nothing else on this map blocks the yeast branch, which is why the yeast branch is the one that has produced every number in this brief.
• Requires (not on this map) a demonstration that a host lineage keeps a functional endosymbiont for 100 or more divisions with the selection switched off, which is the result section 7 specifies and nobody has published; a route for getting a host-encoded protein across the endosymbiont’s membranes, without which gene transfer to the host nucleus is a loss of function rather than a step in a transition; a way for the host to set how many endosymbionts there are and how they are shared between daughters at division; a means of blunting cytosolic bacterial sensing in a mammalian cell without producing a generally immunodeficient one; a containment classification for an organism defined by living inside another organism’s cells, which no schedule currently addresses; and an industrial customer willing to buy a light-and-bicarbonate chassis at the price sugar fermentation already charges.
• Enables Biomanufacturing Resilience, in the narrow sense that a chassis eating carbon dioxide and light is a feedstock-independent one, if the growth rate ever approaches a production organism’s; Future Agriculture, if nitrogen fixation can be put inside a crop cell the way the nitroplast sits inside an alga; and Synthetic Ecosystems, which needs organisms whose capabilities are inherited rather than re-inoculated. Each of these is downstream of the same unpublished result.
• Adjacent Microbiome Engineering, which handles the population-in-a-lumen case this brief deliberately excludes; Mitochondrial Medicine, which shares the organelle-transfer measurement problem; Artificial Photosynthesis and Biological Energy Systems, which pursue the same output by other means; Designer Organisms and Xenobiology for reduced genomes and containment by altered chemistry; and Cloud Laboratories, which is where a several-hundred-generation passage experiment would most cheaply be run.

Established Claim: scientists have made photosynthetic animal cells. They have not. Isolated chloroplasts entered about 20% of CHO-K1 cells and kept photosystem II electron transport running for at least two days, with thylakoid structure degrading by day four. No carbon fixed by those chloroplasts was traced into host metabolism, and the chloroplasts cannot divide in a mammalian cell because the division machinery is nucleus-encoded in the alga they came from. The correct description is a transient organelle transplant with a photochemical readout.

Frontier Claim: engineered endosymbionts have been shown to be heritable, because they lasted 40 generations. The 40 doublings were run under a mutual auxotrophy in which a yeast that lost its bacterium could not respire and a bacterium that escaped could not make thiamin. That measures how long a selection holds, not whether a trait is inherited. The distinction is the entire content of section 7, and collapsing it is the most common error in secondary coverage of this field.

Established Claim: endosymbiosis is just a fancy word for the microbiome. A gut community shares a compartment with the host and is reacquired from the environment each generation. An endosymbiont lives inside the cytoplasm of a host cell and is partitioned at cell division. Microbiome Engineering covers the first; this brief covers the second, and almost nothing transfers between them.

Established Claim: mitochondria prove this is easy, since it already happened. It happened, on the available evidence, twice for photosynthesis in a way we can date — the Paulinella chromatophore and the nitroplast, both on the order of 100 million years old — against an enormous number of engulfment events that went nowhere. A process with two known successes over geological time is evidence that it is rare, not that it is easy.

Frontier Claim: the nitroplast discovery shows endosymbiosis is still happening all around us, so engineering it should be routine. The nitroplast is genuinely a recent case and genuinely important. But the specification it reveals is daunting rather than encouraging: a 1.44 Mb genome stripped of photosystem II, the tricarboxylic acid cycle and carbon fixation, more than 350 host-encoded proteins imported into it, division synchronised with the host cell cycle, and linear size scaling with the host. No engineered system has any of those five properties.

Established Claim: a chloroplast is a self-contained photosynthesis module you can move around. It is not self-contained. Organelle proteins are largely encoded in the host nucleus and imported — that is precisely the criterion used to argue the nitroplast is an organelle. A chloroplast removed from its alga is missing a continuous supply of proteins it does not encode, which is the most likely reason it degrades over days rather than persisting.

Speculative Claim: better gene editing will solve this. The binding constraints are a protein-import pathway that does not exist in any engineered system, host control of symbiont division that nobody has prototyped, and a loss rate nobody has measured. None of those is an editing problem. Editing capacity is a dependency of this field, not its bottleneck, and section 5 says why.

Frontier Claim: the 101-generation cyanobacterium-ciliate result shows long-term artificial endosymbiosis works. The coculture was sustained for at least 101 generations across 143.5 days and 41 transfers, and the ciliate did engulf the cyanobacteria — as food. The authors describe growth through predation and state the system has no stable equilibrium point. It is a valuable experiment about how a symbiosis might start, and it is not an endosymbiosis.

Handwave Claim: this leads to photosynthetic humans, or to animals that do not need to eat. The only published stoichiometry puts the requirement at roughly 25 cyanobacterial cells per CHO cell and about 14,000 per macrophage, under culture-dish illumination. Tissue does not transmit that light more than a few millimetres. Every version of this claim skips the photon budget, and skipping it is where the argument becomes assertion.

Frontier Claim: a second 2024 paper already achieved photosynthetic mammalian cells through artificial endosymbiosis. A paper by Hu and colleagues with that framing exists, and a 2025 commentary discusses it alongside the chloroplast-transplant work. Neither could be opened for this brief, so nothing here rests on either, and readers should treat the claim as unassessed rather than as supported or refuted. Saying which sources a brief could not read is part of the product.

Speculative Claim: engineered endosymbiosis is the efficient way to build a metabolic division of labour. On present evidence it is not. Engineered cross-feeding between yeast strains reached 19.3 times higher product than monoculture, at 4.67 mM, with no cell inside another cell. The case for endosymbiosis has to be made on inheritance and on transfer distance, and neither has been demonstrated to pay.