The smallest genome that still counts as alive

Writing life as a fasta file sounded like a stunt. It was an experiment with a twenty-year methods section. In March 2016, Clyde Hutchison, Ray-Yuan Chuang, Hamilton Smith, Daniel Gibson, John Glass, Craig Venter and colleagues at the J. Craig Venter Institute published JCVI-syn3.0 in Science: a synthetic Mycoplasma mycoides genome, 531,560 base pairs, 473 genes, transplanted into a recipient cytoplasm that then grew and divided. Doubling time about 180 minutes. Colonies that looked roughly like those of the parent synthetic cell, JCVI-syn1.0. Under the microscope the cells were polymorphic — blobs, branches, pearled filaments — which is already a confession that alive and tidy are different achievements. They had started from syn1.0, a 1.08-megabase, 901-gene synthetic genome from 2010, and stripped 428 genes. What remained was smaller than the genome of any autonomously replicating cell found in nature, smaller even than Mycoplasma genitalium, the natural record-holder the same group had sequenced in 1995. The paper's most honest figure isn't the pretty cell. It's the list of genes they couldn't name but also couldn't remove.

In short. In 2016 a team wrote a bacterium with only 473 genes and it still grew. Many of those genes had no known job. That's the result, not the photograph.

They didn't create life from scratch. They synthesised a genome and transplanted it into a recipient cell whose cytoplasm already had a history. The lipids, the ribosomes, the residual proteins of Mycoplasma capricolum were there at the moment of transplantation; over subsequent generations those molecules were replaced as the new chromosome directed synthesis. Both halves of that sentence matter, and skipping the cytoplasm half is how this story gets oversold. Lartigue, Glass, Venter, Science 2007, is the genome-transplantation paper the 2010 and 2016 results sit on: a whole bacterial chromosome, moved, taking over a related cytoplasm. Gibson, Glass, Lartigue, Science 2010, is the first cell controlled by a chemically synthesised genome, JCVI-syn1.0, 1.08 megabases, watermarks and all. Syn3.0 is that method pointed at a reduction. Sequence designed, oligonucleotides made, assembled in yeast, transplanted, a dividing culture. The cytoplasm had a mother. The sequence didn't. Hold that distinction for the rest of the essay, because every honest paragraph below is a commentary on it.

In short. They wrote a genome and put it into a cell that already existed. The cell's insides had a past. The DNA sequence didn't.

The design rule was knock-out until it dies, then put the gene back. That isn't a catchphrase we coined for this page. It's the experimental logic of the 2016 paper, stated without the adjectives. Bombard a genome with transposons, classify every locus as essential, quasi-essential or dispensable, delete the dispensable, resynthesise, transplant, and see whether a colony appears. If it doesn't, you put genes back until it does. If it does, but doubling time collapses or the cells become monsters, you put more back. Engineering, not a theory of life. A theory of life would have predicted the 149 unknowns, or at least predicted that a knockout screen and an annotation were different objects. The screen told them which genes the cell couldn't lose. It didn't tell them what those genes were for. If you write syn3.0 as a solved parts list has skipped the unknown column. If you write it as a failure of annotation has skipped the colony. Both columns are the result.

In short. The method was simple and rude: remove a gene, and if the cell dies, put it back. That tells you what is needed. It doesn't tell you what the gene does.

Mycoplasma was the right chassis because evolution had already done most of the deleting. Mollicutes are wall-less bacteria, parasites of animals, with genomes that had been stripped by a long residence in rich host fluids. They don't make peptidoglycan. Many of them lack a tricarboxylic-acid cycle and a respiratory chain. They take sterols from the medium and write them into a membrane that would otherwise collapse. They use UGA as tryptophan rather than as stop, a codon-table eccentricity that has wrecked more heterologous expressions than the field likes to admit. Fraser, Gocayne, Venter, Science 1995, the complete genome of Mycoplasma genitalium, 580 kilobases, already looked like a floor. Syn3.0 went under that floor, but only in laboratory media stuffed with nucleotides, amino acids, lipids and vitamins. Autonomous replication, in this literature, means a colony on a plate, not a cell that can live in a pond on sunlight and minerals. That's the honesty test the word minimal has to pass before anyone is allowed to talk about the origin of life.

In short. They started from a parasite that had lost many genes. It only grows in a rich broth. Minimal here means smallest that still divides in the lab, not a cell that can live on air.

The 1995 genitalium paper is the origin of this particular obsession, and it's worth reading as a census rather than as a celebrity. Claire Fraser, Jeannine Gocayne, Owen White, Venter and the then-TIGR group reported 470-odd protein-coding genes in a 580-kilobase circle, the smallest cellular genome known at the time, and asked out loud which of them a cell actually needed. Comparative tables against Haemophilus influenzae, sequenced the same year, suggested a core of a few hundred. Comparative tables aren't experiments. Hutchison, Peterson, Glass, Smith, Venter, Science 1999, then Glass, Assad-Garcia, Alperovich, PNAS 2006, made them experiments: global transposon mutagenesis of M. genitalium, insertions recovered or not, an estimate of about 382 essential genes. That number haunted a decade of reviews. It was a lower bound from a parasite already near the floor, in rich medium, scored by whether an insertion clone could be grown. It wasn't a genome anyone had written. Writing is a different verb from disrupting, and the group spent the next years learning to conjugate it.

In short. In 1995 they read the smallest known bacterial genome. Later they broke genes one by one to guess which were required. Guessing isn't the same as writing a genome.

Chemical synthesis of a whole bacterial chromosome was the missing verb. Gibson, Benders, Andrews-Pfannkoch, Science 2008: complete chemical synthesis, assembly and cloning of a Mycoplasma genitalium genome, 582,970 base pairs, built from oligonucleotides, hierarchical assembly, and yeast as a living glue. The genome sat in Saccharomyces as a yeast artificial chromosome. It didn't, in that paper, boot a bacterium, because M. genitalium grows so slowly that transplantation-and-debug would have been a career. The 2008 result is still the one to cite for the foundry: overlapping cassettes, recombination in yeast, a megabase-scale DNA molecule that a chemist didn't ligate by hand. If you start the syn3.0 story at the 2016 cover, you have skipped the year they learned to assemble a chromosome they couldn't yet transplant. The foundry is why a later reduction was a design cycle rather than a prayer. That's the size of the claim, and we'll keep it there.

In short. In 2008 they chemically built a whole bacterial chromosome and grew it in yeast. That factory is why a later, smaller genome could be designed rather than wished for.

Transplantation had its own paper, and it's ruder than synthesis. Lartigue, Glass, Alperovich, Science 2007: a native Mycoplasma mycoides chromosome, moved into Mycoplasma capricolum, taking over, converting the recipient into the donor species as the new genome was expressed and the old one was lost. Restriction systems had to be disarmed or the incoming DNA was chopped on arrival; Lartigue, Vashee, 2009, is the methylation-and-restriction patch. Mycoplasmas were tractable for this trick because they have no cell wall to strip and because related species will tolerate a foreign chromosome long enough for selection to finish the job. The method isn't a pipette step you run on Escherichia coli after lunch. It's a species-pair, a restriction map, a selection, and a sequence confirmation that the only DNA left is the donor's. Syn1.0 and syn3.0 both inherit that sentence. A synthetic chromosome that can't be transplanted is a very expensive plasmid in a yeast.

In short. They learned to move a whole chromosome from one bacterium into another closely related one. Without that move, a written genome stays stuck in yeast.

Gibson, Glass, Lartigue, Noskov, Chuang, Smith, Hutchison, Venter, Science 2010, is the boot. A 1.08-megabase Mycoplasma mycoides genome, designed from a finished native sequence, synthesised as overlapping cassettes, assembled in yeast, transplanted into restriction-deficient M. capricolum, yielding JCVI-syn1.0: a cell whose only chromosome is the designed one, watermarks included, deletions of a glycerol-transporter virulence cassette included, a handful of building mutations included. The cells grew. They looked like mycoides. The cytoplasm of the recipient was overwritten, generation by generation, by the synthetic genome's instructions. That's the first bacterial cell controlled by a chemically synthesised genome, and it's the parent of syn3.0, not a side exhibit. Nine hundred and one genes. Already a designed reduction from wild-type GM12. Already a fasta that divides. The 2016 paper is what happened when they asked how many of those 901 they could throw away and still have a colony. The answer wasn't a philosophy. It was 473, plus the ones they had to put back after the first designs died.

In short. In 2010 a fully synthetic chromosome took over a cell and the cell grew. That parent strain had 901 genes. The 2016 work asked how many of those could be thrown away.

Yeast is the quiet factory under every one of those sentences, and it deserves its own paragraph because the public story always jumps from oligonucleotide to colony. Saccharomyces cerevisiae will take up overlapping DNA fragments and stitch them by homologous recombination as if that were a reasonable request. Transformation-associated recombination cloning, plus Gibson's isothermal assembly at the smaller scales, plus yeast centromeric plasmids to hold a bacterial chromosome as if it were just another large construct: that's how a 531-kilobase designed genome exists as a physical object you can extract and transplant. Boeke's later synthetic-yeast project, Sc2.0, sits on the same shelf from the other direction — rewriting the host, not using the host as glue — and we'll come back to it. For syn3.0 the yeast is a foundry, not a chassis. Pull the chromosome out, transplant it into a mycoplasma cytoplasm, select, sequence. If the design is lethal, the yeast still has the previous version. Design, build, test is only a cycle if the build step doesn't kill your only copy.

In short. Baker's yeast is the factory that stitches the designed DNA together. The bacterium is where that DNA has to work. Two different jobs, two different cells.

Transposon bombardment is how you classify a genome when you aren't yet ready to synthesise every hypothesis. A Tn5-derived insertion, mapped by sequencing, either is recovered in a growing population or it isn't. Hutchison's group, working on syn1.0, sorted 901 annotated genes into three heaps. Essential, e-genes, about 240: insertions absent or vanishingly rare, the cell apparently can't lose them. Non-essential, n-genes, about 432: insertions common, growth intact. Quasi-essential, i-genes in their notation, about 229: insertions recovered, but the clones grow poorly, or the population later sheds them, or a deletion that looks tolerable in a mixed culture is lethal once it's fixed. That third heap is the one a comparative-genomics paper never sees. It's also the one that killed their first designed minimal genome. The classification is a growth-assay, not a function. An essential gene of unknown function is still essential. A non-essential gene with a beautiful annotation is still dispensable in this medium, in this species, at this doubling-time demand. Medium, species, demand: three variables people drop when they quote 473 as if it were Avogadro's number.

In short. They fired jumping DNA into the genome and asked which genes could still be broken in a living cell. Some breaks were fine, some were fatal, and some only made the cells weak.

The first designed minimal genome failed, and that failure is the methods result the 2016 paper is actually about. They took the annotation-plus-transposon table, kept what knowledge said a cell must have, deleted the rest, synthesised the reduced circle, transplanted, and got nothing that grew. Collective knowledge of molecular biology, plus a limited insertion map, wasn't enough to specify a living cell. That sentence is more important than the later successful number. It means a parts list compiled from textbooks and from other genomes under-counts the genes this organism needs to run at a workable growth rate. Hutchison said the goal was a cell for which the precise biological function of every gene is known. The failed design showed they didn't yet know which genes belonged on the list, let alone what each did. Quasi-essential genes, the ones a mixed culture can limp without and a pure deletion can't, were the missing class. They improved the transposon map, put those genes back into the design, and started again.

In short. The first version they built, using textbooks plus a rough gene map, didn't grow at all. They had thrown away genes the cell quietly needed.

Quasi-essential is a growth-rate word, not a metaphysics. A gene can be dispensable if you're willing to wait a day for a colony, and lethal if you want the culture to keep up with a chemostat. During minimisation there's a trade-off between genome size and doubling time, and the 2016 authors said so. JCVI-syn3.0 is a working approximation of a minimal cellular genome, a compromise between small size and a workable growth rate for an experimental organism. Three hours is workable. Eighteen hours, the M. genitalium pace, is how you lose a grant cycle to a contamination. Keep a quasi-essential gene and the cell is larger than a philosopher's minimum. Delete it and you may still have a living cell, once, that you can't study because it won't give you biomass. Engineering, again. The 229 i-genes on the syn1.0 map are why the successful design is 473 and not the 300-and-something a comparative table had been promising since the 1990s. The comparative table was a core. The organism was a rate.

In short. Some genes aren't strictly fatal to lose, but losing them makes the cell grow too slowly to study. The final design kept those. Smallest, and still usable, are different goals.

Three cycles of design, synthesis and testing, with retention of the quasi-essential set, produced JCVI-syn3.0. Five hundred and thirty-one kilobase pairs. Four hundred and seventy-three genes. Smaller than M. genitalium. Doubling time about 180 minutes, against about 60 minutes for the mycoides parent and about 18 hours for genitalium. Colonies morphologically similar to syn1.0. Cells polymorphic when examined microscopically, a first hint that division genes sat near the edge of what they had kept. Syn2.0, in the same paper, is the less-stripped intermediary, a viable reduced genome that wasn't yet the floor. The successful cycle is the one that treated the insertion map as a rate map rather than as a binary. Build eight overlapping segments, swap them against syn1.0 pieces, find the combinations that transplant, reduce further, stop when the next deletion kills the culture or the morphology. That isn't how a theorist derives a living system from first principles. It's how an engineering group finds a local minimum in a high-dimensional deletion space. Local is doing work in that sentence. A different medium, a different starting species, a different doubling-time requirement, and the floor moves.

In short. After several rounds of building, testing and putting genes back, they arrived at 473 genes and a three-hour doubling time. That number is a useful minimum, not a unique one.

Watermarks, antibiotic cassettes and designed deletions travelled from syn1.0 into the reduced descendants, because a synthetic chromosome that can't be distinguished from a contaminant is a bioinformatics headache and a biosafety one. The 2010 genome carried encoded names, a quotation, a URL; the 2016 reduction kept the logic of a designed object you can authenticate by sequence. Authentication isn't vanity. When the claim is that the only DNA in the cell is the DNA you wrote, you need a string that nature didn't. You also need to know which mutations the build introduced — eight new single-nucleotide polymorphisms in the 2010 sequenced transplant, a reminder that synthesis-plus-yeast-plus-transplantation is a living process, not a photocopier. Syn3.0's 531,560 base pairs are a designed sequence with a laboratory history. Quote the accession if you're going to argue about a gene. CP014940.1 is the syn3.0 chromosome. A cartoon of 473 coloured arrows isn't.

In short. The synthetic genomes carry designed DNA stamps so you can prove the cell is running the written chromosome. Building also introduces a few accidental spelling changes.

Four hundred and seventy-three genes is 438 proteins and 35 RNAs. The RNAs are the ones a first-year course already knows how to name: ribosomal RNAs, transfer RNAs, the small stable RNAs a bacterium uses as tools rather than as messages. The proteins are the argument. Hutchison's table sorted what they could name into four working groups, and a remainder they couldn't. About 195 genes, on their count, sit in expression of genome information: transcription, ribosome biogenesis, aminoacyl-tRNA synthetases, initiation, elongation and termination, chaperones that fold what the ribosome emits. About 34 sit in preservation of genome information: replication, repair, chromosome handling. About 84 sit in membrane structure and function: lipid handling, transporters, the proteins that keep a wall-less cell from bursting in a hypotonic moment. About 81 sit in cytosolic metabolism: the stripped fuel and building-block chemistry a parasite still has to run even when the medium is a pantry. Seventy-nine genes, on that same partition, couldn't be dropped into those four bins. Different classifiers give 149 as the unknown-function count, because generic hydrolases and domains of unknown function aren't the same thing as a named pathway. Both numbers are in the paper.

In short. Most of the 473 genes build proteins and RNAs, copy DNA, keep the membrane intact, or run a stripped metabolism. A large remainder still had no named job.

Translation dominates, which is the least surprising sentence in microbiology and still the one a minimal-cell paper exists to make visible. Ribosomal proteins, rRNA operons, tRNAs, twenty aminoacyl-tRNA synthetases, initiation factors IF1–IF3, elongation factors EF-Tu, EF-Ts, EF-G, release factors, ribosome recycling, a handful of RNA-modifying enzymes: that neighbourhood is most of what a cell can't negotiate away. Breuer, Earnest, Luthey-Schulten, Glass, eLife 2019, looking at the slightly expanded syn3A, put translation at about 40 percent of the protein-coding set. Forty percent of a genome spent on the machine that reads the rest of the genome isn't a luxury. It's the overhead of being a cell rather than a virus. Viruses steal that machine. Syn3.0 has to own it. Fancy traits — a tardigrade radiation shield, a mammoth coat, a CRISPR array — are optional equipment on top of that overhead. We already knew that. Seeing it as a fasta file, with the unknowns sitting in the same circle as EF-Tu, is different. The unknowns aren't a footnote to the ribosome. They're why the ribosome's neighbours aren't a closed book.

In short. A huge share of the remaining genes exist just to read the genome and make proteins. That machinery isn't optional. Viruses steal it. A cell has to own it.

Preservation of the genome is a smaller bin than people expect, and that's information. Thirty-four genes to replicate a 531-kilobase circle, to repair what replication gets wrong, and to hand copies to daughters, sounds under-staffed if your reference is E. coli. E. coli has mismatch repair, nucleotide-excision repair, SOS, several polymerases, a sophisticated recombination shop. A mycoplasma living in rich medium, with a small genome and no outer world of ultraviolet to speak of inside a culture tube, has already discarded most of that shop. What remains is a core replication fork, a ligase, a handful of repair proteins, and a division neighbourhood that, as the morphology showed, was sitting on the edge of sufficiency. Error rate is therefore a design parameter. A minimal cell isn't a high-fidelity cell unless you kept the fidelity genes. Syn3.0 kept what the insertion map said it couldn't lose. It didn't keep a genome-integrity programme you would want in a chassis you planned to take outside a dish. That's a warning to anyone whose next sentence is industrial production.

In short. Few genes remain for copying and repairing DNA. A tiny genome in rich broth doesn't keep the full repair shop that a hardy outdoor bacterium runs.

The membrane is the other non-negotiable, and in a mollicute it's doing extra work because there's no wall. Eighty-four genes for structure and function: lipid synthesis and salvage, sterol handling, transporters that pull the pantry inward, proteins that set curvature and that, poorly, help the cell divide. A wall-less sphere 400 nanometres across has a surface-to-volume problem that E. coli doesn't. It also has a mechanical problem: without peptidoglycan, osmotic shock is a membrane problem, and division is a membrane-and-cytoskeleton problem with a short parts list. Pelletier's later work would show just how short. For the 2016 census the point is simpler. You can delete a metabolic luxury. You can't delete the bag. Transporters without named substrates sit in this bin and in the unknown bin at once — an ABC cassette whose cargo nobody has identified is still, operationally, a membrane gene. Annotation lags necessity. The insertion map doesn't care whether you know the cargo. The cell does.

In short. Without a cell wall, the membrane has to hold the cell together, feed it, and help it split. Those jobs take a surprising number of genes, including ones whose cargo is still unnamed.

Cytosolic metabolism, 81 genes, is a stripped pantry-user, not an autotroph. Glycolysis is there. Substrate-level phosphorylation is how ATP is minted; there's no oxidative phosphorylation to speak of, no Complex I, no Mitchell gradient to run a turbine. Nucleotide salvage, amino-acid handling, a limited lipid chemistry, cofactors the medium doesn't supply in the right form: those are the remaining reactions. Breuer's 2019 metabolic reconstruction of syn3A is the map to read after Hutchison, because it names which pathways actually carry flux in the model and which annotated arrows are traces. Central carbon metabolism and membrane transport dominate that later table. Biosynthesis is thin, as it should be in a parasite grown on a soup. Energy metabolism, in the oxidative sense, is almost gone. Anyone picturing a tiny E. coli has the wrong picture. Anyone picturing a mitochondrion has a more interesting wrong picture, and the mitochondria heading is where we'll take it. For now: 81 genes is what you keep when the medium is a pantry and the job is to turn pantry into polymer and phosphate.

In short. The remaining chemistry is mostly sugar-burning and scavenging. The broth supplies the rest. This isn't a cell that builds its world from minerals and air.

Name what is missing, because headlines skip the absences. No cell wall. No flagellum. No two-component signalling forest. No SOS system worth the name. No TCA cycle as a full wheel. No endotoxin, no outer membrane, no periplasm. No CRISPR array. No secondary metabolism. No developmental programme. Almost no transcriptional regulation: a handful of factors, not a network. Mycoplasma had already discarded most of those before the first oligonucleotide was ordered. Syn3.0 discarded more. The industrial reading of a chassis is that you can now add modules onto a known floor. The scientific reading is that the floor is a mycoplasma floor, a rich-medium floor, a three-hour-doubling floor, and that adding a module doesn't make it E. coli. Comparative genomics had guessed a core of around 250 to 300 genes. Syn3.0 is 473 because this organism, in this medium, at this rate, needed the extras, including the ones nobody could name. A different starting bacterium would give a different floor. That isn't a defect in the experiment. It's the experiment.

In short. Gone are the cell wall, most of the wiring that turns genes on and off, and the extra pathways a free-living microbe would keep. What remains is a parasite's floor, not a universal one.

One hundred and forty-nine genes of unknown function, at publication, in a genome that had been reduced until the next deletions died. Venter, in the accompanying remarks, put it at about a third of essential life, and said bioinformatics over twenty years had underestimated the essential set by staring only at the known world. That isn't a press flourish. It's the experimental content of the unknown column. Homology had been run. Operon context had been stared at. The remaining genes included domains of unknown function conserved across a surprising phylogenetic range, including, in some cases, hits that reach out of bacteria. Conservation without a name is a particular kind of humiliation: the sequence isn't new, the biology is. Fussenegger's reaction at the time — we've sequenced everything on this planet and still don't know 149 genes that are most essential for life — is the correct emotional register. The 2016 paper didn't hide the column. Coverage that treated syn3.0 as a solved cell hid it for them.

In short. About one gene in three had no known job, even though the cell died without it. Having the sequence isn't the same as knowing the biology.

A knockout screen isn't an annotation. That sentence is the claim of syn3.0, and it's why this page refuses to treat 473 as a parts catalogue with a few blank SKUs. Essential means: insertions weren't recovered, or a designed deletion didn't transplant. Function means: a reaction, a substrate, a pathway, a structure, a phenotype more specific than death. You can have the first without the second. Biochemistry has always known this — lethal alleles of unnamed loci are older than genomes — but a minimal cell makes the gap unignorable, because there's nowhere to hide an unknown in a 4.6-megabase neighbourhood. In E. coli you can ignore a few dozen y-genes. In syn3.0 they're a third of the object. Some of the 149 will turn out to be transporters whose cargo was unmarked, hydrolases whose substrate was unmarked, membrane proteins whose job is curvature or a division assist. Some may be genuinely new processes. The paper can't tell you which. A later blot might. Until then, the honest line is Hutchison's: the goal is a cell for which the precise biological function of every gene is known. Goal, not result.

In short. Knowing a gene is required isn't knowing what it does. In a genome this small, those blanks are a third of the cell, not a rounding error.

Later work chipped the unknown column and didn't empty it. Zhang, Freddolino, Zhang, Journal of Proteome Research 2021, took the 438 proteins and, with homology, gene ontology and contact-map structure models, proposed molecular-function labels for a large majority; nutrient acquisition, host interaction and nucleotide metabolism were enriched among the previously blank set, which is what you would guess for a parasite's remaining mysteries. Structure models, including the post-AlphaFold generation, turn a sequence into a fold you can stare at. A fold is a hypothesis with a colour bar, as this journal has said elsewhere, not a substrate. Breuer's metabolic reconstruction assigned reactions where a metabolite and an enzyme family lined up. Pelletier's division study, below, took unnamed membrane proteins and gave them a phenotype: without them the cell doesn't make a clean sphere. That's function in the genetic sense. It's still not a mechanism. The 149 is no longer 149. It's also not zero. Anyone quoting only the 2016 number is out of date. Anyone claiming the cell is now fully understood hasn't read the 2021 discussion sections.

In short. Later papers named likely jobs for many of the mystery genes, especially transporters and membrane proteins. A predicted shape isn't a finished explanation.

AlphaFold belongs in this heading because the neighbouring essay in this journal is about the fold becoming cheap, and because a third of a minimal cell is exactly the sort of set people now paste into a browser. Paste the 149. Get ribbons. Get confidence maps. Some will match a known fold family and the annotation will move one notch, from unknown to putative hydrolase, putative transporter, putative nucleic-acid binder. Some will remain stubborn ribbons with no family and no ligand. The fold doesn't tell you why the insertion map said the gene was essential. It tells you what shape the essential mystery takes. Baker's generative tools can, in principle, design a binder to that shape, which is a different experiment again. The 2024 chemistry Nobel made the fold a public object. Syn3.0 makes the remaining gap a public object of another kind: even with a fold, even with a knockout, you can still lack a mechanism. We like having both objects on the same desk. They keep each other honest.

In short. Predicted structures can hint at what a mystery protein looks like. They don't, on their own, explain why the cell dies without it.

Even a minimal cell isn't a fully understood cell. Keep that sentence at this length; it's the one the 2016 paper earned and the one a chassis brochure keeps trying to shorten. Understood, in the Hutchison sense, would mean every gene has a molecular function, every function has an assay, and a whole-cell model built from those assays predicts growth, division and a perturbation. Thornburg, Luthey-Schulten, Cell 2022, built a remarkable three-dimensional simulation of syn3A in which fundamental behaviours emerge from the parts they could name. Emergent is a success. It's also an admission that the unnamed parts were handled by approximation. We'll take the model. We won't take it as a closed theory of life. The design rule that produced the cell — knock-out until it dies, then put the gene back — never claimed to be that theory. It claimed to be a way to get a genome small enough that understanding might finish. Ten years later the finishing is underway and not done. That's a healthy pace for a third of essential life.

In short. A cell with 473 genes still isn't a cell we fully understand. The small genome makes the remaining blanks visible. It doesn't fill them.

Bacterial RNA polymerase is a core of five subunits — two α, β, β′, ω — plus a σ factor that points it at a promoter. Mycoplasmas run a reduced version of that machine and almost none of the transcription-factor menagerie that E. coli uses to decide which operons fire. Syn3.0 keeps the polymerase, a minimal σ, and a handful of associated proteins. It doesn't keep a regulatory network worth drawing as a circuit diagram. Most genes, most of the time, are on. That's what a genome looks like when you have deleted the luxuries of choice. Regulation, in a rich-medium parasite with 473 genes, is largely the slow logic of growth rate, ribosome number, and whether the membrane can keep up, not a set of dedicated repressors. Breuer's later census put transcription at a few percent of the protein-coding genes, against translation's 40 percent. Polymerase is cheap, in gene count, relative to the ribosome it feeds. The messages are short, the operons are dense, the intergenic spaces aren't a mammalian enhancer landscape. Reading isn't the scarce step. Having something worth reading is.

In short. This tiny genome keeps the basic machine that copies DNA into RNA, and almost none of the switches that turn genes on and off. Most genes run most of the time.

Constitutive expression is a phenotype, and it's one a chassis-builder has to love and fear in equal measure. Love: if you add a module, you may not need to invent a regulatory cascade to keep it visible. Fear: you also can't easily turn it off, and you can't easily make it tissue-specific, because there's no tissue, and you can't easily make it conditional, because the conditionals were among the genes they deleted. Promoters in mycoplasmas are compact. There's no chromatin in the eukaryotic sense, no nucleosomes to open, no Mediator complex, no promoter-proximal pausing. There's supercoiling, polymerase crowding, and a nucleoid that's mostly the chromosome folded around itself in a 400-nanometre bag. An added gene still has to have a promoter the remaining polymerase will recognise, a ribosome-binding site the remaining 16S will find, and a codon table that treats UGA as tryptophan. Heterologous modules fail that last test more often than slide decks admit. The transcription diagram on this page is here because a designed genome is still a read genome. Writing 473 genes is the engineering. Getting polymerase to treat them as messages is the cell.

In short. With so few switches, extra genes you add will tend to stay on. They still need a matching on-switch, a ribosome landing site, and the right genetic code.

Most of the RNA a growing cell makes, by mass, is ribosomal RNA, not messenger RNA. That's as true in syn3.0 as it's in a hepatocyte, and it's the sentence a journal that also writes about nucleoli keeps repeating because Pol II chauvinism is a bad habit. Two rRNA operons in a 531-kilobase circle aren't a decoration; they're how you get enough ribosomes to double the protein inventory in three hours. tRNAs are the other non-negotiable RNA species, 20-odd flavours for a reduced codon usage, charging the synthetases that dominate the protein list. Messenger RNAs are the minority product, short-lived, encoding the 438 proteins, many of them themselves ribosomal. A transcription essay that only talks about mRNA hasn't visited the factory floor. Syn3.0 makes that floor visible because there's almost nothing else. No long non-coding menagerie. No spliceosome — bacteria don't splice in that way. No nucleus, so transcription and translation couple in the same compartment, ribosomes grabbing messages as they emerge, which is how bacteria have always done it and how a 400-nanometre cell has no choice.

In short. Most of the RNA the cell makes is ribosome material, not messages for proteins. In a genome this small that factory is obvious, because there's almost no other kind of RNA.

A module added to this chassis is a transcription-and-translation problem before it's a product. A metabolic pathway, a sensor, a DARPA-shaped party trick: polymerase has to fire, the ribosome has to finish, the proteins have to fold in a cytoplasm that has 438 native proteins and a lot of ribosomes and not much else. Chaperones are in the essential set for a reason. Codon optimisation has to respect UGA-as-Trp. Toxic products have nowhere to be compartmentalised; there's no periplasm, no mitochondrion, no specialised organelle. The industrial reading of syn3.0 as a chassis isn't wrong. It's incomplete until you have said those sentences. Colossal's mammoth alleles have to be read in an elephant follicle; a syn3.0 module has to be read in a mycoplasma nucleoid. Different readers. Same rule. Sequence isn't a phenotype. Phenotype is processed RNA, folded protein, and a membrane that noticed. The 149 unknowns include, almost certainly, proteins that help that noticing happen. Adding a module without understanding those is how a chassis becomes a mysterious death.

In short. Anything you add still has to be read and built by this cell's own machinery, in a tiny bag with no extra rooms. Sequence on a disk isn't a working trait.

The codon table is a methods landmine and a philosophical souvenir. Mycoplasmas reassigned UGA from stop to tryptophan. Any gene you import from E. coli, from a human, from a tardigrade, will, if it contains TGA in the DNA, insert Trp instead of stopping, or will fail to stop where you wanted, and the protein will be garbage or worse. Recoding the module, or recoding the cell, is the actual first step of chassis work, and it's the step that connects this page to Church's 57-codon E. coli and to Chin's Syn61, a 61-codon Escherichia that has had TAG freed as a blank codon. Syn3.0 didn't recode the table; it inherited the mollicute one. A designed cell with a designed table is a later object. The souvenir is that the genetic code isn't quite universal, and that a minimal cell is allowed to keep an eccentricity if the insertion map says it can't lose the tRNA that makes the eccentricity work. Engineering, again. The code is part of the parts list.

In short. In these bacteria one of the usual stop signs means tryptophan instead. Genes copied from other species will be misread unless you rewrite them to match.

Syn3.0 mints ATP by substrate-level phosphorylation. Glycolysis takes glucose, or whatever sugar the remaining transporters admit, through the familiar ten steps, and the net phosphate is made at the kinase steps, not at an ATP synthase driven by a proton-motive force. There's no Complex I waiting for NADH, no Q-cycle, no cytochrome oxidase, no Mitchell turbine. NADH still has to be reoxidised, because GAPDH will stall if the pool sits reduced; lactate dehydrogenase, or a related dump, is how a wall-less fermenter keeps the hydride moving. The NAD+ essay in this journal is about a dinucleotide that mitochondria spend at Complex I and that sirtuins spend on chromatin. Syn3.0 has the dinucleotide as a redox coin and, as far as the census shows, doesn't have the eukaryotic spending programmes. No PARP. No sirtuin family. No CD38. A millimolar cofactor on a glycolytic path is still not a lifestyle accessory. It's how sugar becomes phosphate when you have thrown the respiratory chain away.

In short. This cell makes its energy by burning sugar in the cytoplasm, not by running a membrane turbine. It has no mitochondria, and no breathing chain.

Rich medium is the other half of the energy story, and it's the half a origin-of-life headline always drops. Nucleotides, amino acids, fatty acids, sterols, vitamins: the plate supplies what the 81 metabolic genes don't make. Glass's group has always been clear that autonomous means axenic growth in the laboratory, not autotrophy. A truly free-living minimal genome, a cell that builds itself from minerals and a carbon source, would be a different and much larger object — Pelagibacter, at about 1.3 megabases, is a better pointer to that floor than syn3.0 is, and even Pelagibacter is an auxotroph for several cofactors. Endosymbionts such as Nasuia and Tremblaya went smaller than syn3.0, but they aren't autonomous; they live inside insect cells that finish their metabolism for them. Syn3.0 sits in the narrow band of organisms you can plate. That band is the experimental prize. It isn't the primordial soup. Anyone moving from 473 genes to a sentence about how life began has changed the medium without saying so.

In short. The growth broth is doing a lot of the chemistry. A cell that had to make everything from scratch would need many more genes. Smaller genomes exist, but they live inside other cells.

Mitochondria went the other way, and that's why their diagram belongs on this page rather than only on a cell-desk energy essay. An α-proteobacterial endosymbiont, a billion-plus years ago, handed most of its genes to the host nucleus and kept a remnant chromosome: in humans, 16,569 base pairs, 13 proteins, 22 tRNAs, 2 rRNAs. The 13 proteins are still written on-site, hydrophobic respiratory subunits that are awkward to import. The other thousand-odd mitochondrial proteins are nuclear-encoded and imported. That's reductive evolution as outsourcing. Syn3.0 is reductive evolution as deletion without a host to outsource to, plus, in the last step, reductive design. A mitochondrion isn't a minimal cell. It's a former cell that became an organelle, and it can't live on a plate. Syn3.0 is a cell that can live on a plate, and it can't run a nervous system. Comparing their genome sizes without comparing their bargains is how a scale bar becomes a muddle. The organelle kept a respiratory chain and discarded autonomy. The synthetic mycoplasma kept autonomy and discarded the respiratory chain.

In short. Mitochondria also shrunk their genomes, but they did it by handing jobs to the cell around them. Syn3.0 had no host to hand jobs to. Different bargains, both called minimal.

Two jobs, one campus, if you're a eukaryotic cell; one invoice, no campus, if you're syn3.0. A human cell runs glycolysis in the cytosol and cashes most of the reducing power in the matrix, at Complex I, on a NAD+ pool that isn't in free exchange with the cytosol. SIRT3 deacylates the matrix neighbourhood if that pool is there to spend. MOTS-c, a 16-mer from mitochondrial 12S rRNA, talks to AMPK, which is a different invoice again. None of those objects exist in syn3.0. The mycoplasma has a cytoplasm, a membrane, a chromosome, and a glycolytic path. It's a useful reminder, on a desk that also stocks lyophilised β-NAD+ and a mitochondrial peptide, that the eukaryotic energy story is a late elaboration. The elaboration is real. It isn't the floor. Peter Rich's whole-body ATP turnover, 40 to 60 kilograms a day in a human, is a mitochondrial number. Syn3.0's ATP turnover is a tiny glycolytic number in a 400-nanometre sphere. Scale is why both belong in the same journal and why they must not share a catchphrase.

In short. Our cells split energy work between the liquid inside and the mitochondria. The minimal bacterium does it all in one room, on a much smaller budget.

Syn3.0 divided, and it divided as a mess. Polymorphic, the 2016 paper said; under later microfluidic chemostats, Pelletier, Strychalski, Glass and colleagues watched pearled filaments, branches, irregular blobs, a geometry that said the division machine had been stripped past the point of a clean septum. Population growth can hide that. A culture can double every three hours while individual cells fail to make two daughters of similar size, because some compartments get a chromosome and some get a neck of membrane and a prayer. Fitness at the colony level isn't shape at the single-cell level. That distinction is why a minimal genome selected on plating isn't automatically a model of bacterial physiology. It's a model of whatever you selected for. They selected for a colony. Shape was optional until someone looked. A plating assay scores biomass. A camera scores whether two daughters exist. The camera came later, and it wasn't kind.

In short. The 2016 cell grew as a population, but under a microscope many cells were misshapen. Growing isn't the same as dividing neatly in two.

JCVI-syn3A is the 19-gene apology, and it's the version most later papers actually use. Restore 19 genes that syn3.0 hadn't retained, and the cells become spheres of about 400 nanometres, doubling in something like 105 minutes, morphology closer to syn1.0. The genome is 543 kilobases, 493 genes, 452 of them proteins. Still smaller than any natural autonomous cell. Still a designed object. Pelletier, Sun, Wise, Assad-Garcia, Karas, Deerinck, Ellisman, Gibson, Glass, Strychalski, Cell 2021, then asked which of those 19 did the shape work. Seven, together, restored a phenotype similar to the parent: two known division genes, ftsZ and sepF, a hydrolase of unknown substrate, and four membrane-associated proteins of unknown function. Cluster 1, the ftsZ neighbourhood, wasn't enough on its own. Unknowns again, this time with a microscope readout. A named tubulin homologue and a named partner, plus five genes the 2016 table couldn't annotate, are what it takes this cell to look like a bacterium rather than a mistake. The unknowns weren't waiting for a better BLAST. They were waiting for a chemostat and a camera.

In short. Putting 19 genes back made the cells round and faster-growing. Seven of those, including two known division genes and five poorly named ones, were needed for a normal shape.

FtsZ is the bacterial tubulin that assembles a ring at the future septum. SepF helps that ring find and hold the membrane. In walled bacteria a much larger dcw cluster builds peptidoglycan in lockstep with the ring. Mycoplasmas already lacked the wall, so their division logic was always a short list. Syn3.0 had shortened it past the point of a reliable ring-and-pinch. The cell still had a way to make more cells — membrane growth, chromosome replication, occasional fission — but the way was stochastic enough to make monsters. Adding FtsZ and SepF back, without the unnamed membrane proteins, didn't finish the job. That's a useful negative. Textbook division genes aren't the whole division phenotype, even in a cell with 493 genes. If your picture of a minimal cell is a sphere pinching in the middle because FtsZ said so, the 2021 paper is the correction. If your picture is that unknowns are an annotation failure rather than a biology, the four membrane proteins are the correction. We'll keep both corrections.

In short. The famous division-ring protein wasn't enough on its own. Mystery membrane proteins had to come back too, which is how you know the mystery was real.

Thornburg, Looker, Kalimeri, Luthey-Schulten and colleagues, Cell 2022, put syn3A in a computer. A 3D whole-cell model, 493 genes, metabolites, ribosomes, a membrane, stochastic transcription and translation, DNA replication racing a volume doubling: fundamental behaviours emerged from the kinetics they could write down. Simulated doubling times landed near the measured 105 minutes. Chromosome-copy numbers, ribosome counts, the crowding of a 400-nanometre cytoplasm, all became numbers you could plot. The model doesn't contain a soul, and it doesn't contain the remaining unnamed mechanisms except as effective rates. What it contains is the first serious attempt to run a living cell as a simulation in which every named gene has a job and the jobs interact in space. That's the Hutchison goal, approached from the other end: not only a genome you wrote, but a physics you can integrate. Gaps in the physics are now as informative as gaps in the fasta. We'd rather have a model that admits its blanks than a photograph that pretends there are none.

In short. A 2022 computer model of the slightly larger version of the cell reproduced growth from the named parts. The remaining blanks are now holes in a simulation as well as holes in a table.

Once you have a minimal cell you can, in principle, add modules: a metabolic pathway, a sensor, a product a fermenter might care about. That's the industrial reading, and it isn't wrong. JCVI, Synthetic Genomics, and the later syn-cell community have said it out loud for a decade. A chassis with 473 genes is a background you can see. Add a cassette, and the new phenotype has fewer native pathways to hide in. Trouble is the same as the promise. The background is a mycoplasma background: UGA-as-Trp, rich medium, no wall, a membrane that's also the division machine, almost no regulation, a doubling time you won't confuse with E. coli's twenty minutes. Industrial microbiology already has chassis that make insulin and artemisinic acid, and they aren't syn3.0. What syn3.0 is good for is the scientific reading. Four hundred and seventy-three genes is a census of necessity. Ribosomes dominate. Membranes dominate. A surprising number of unknowns dominate. Any peptide, any CRISPR edit, any de-extinct allele is a passenger on a cell that still has to do some version of these jobs. Syn3.0 makes that hierarchy visible. Fancy traits are optional. Translation isn't.

In short. People talk about adding extra skills onto this tiny cell. That may work one day. Its real gift already is a clear list of what a cell can't do without.

Yeast 2.0, recoded E. coli, and cell-free systems sit on the same shelf: life as an editable medium, each with a different bargain. Sc2.0, Boeke and a large consortium, rewrites Saccharomyces chromosome by chromosome, with SCRaMbLE as a recombinase-driven scrambler, aiming at a designed eukaryotic genome you can still brew with. Fredens, Wang, Chin, Nature 2019, Syn61: an E. coli whose serine codons have been compressed so that TAG is blank, a 61-codon cell, a slot for a non-canonical amino acid. Church, Isaacs, Ostrov and the 57-codon project push the same logic further. The PURE system, Shimizu and Ueda, is the other extreme — transcription and translation in a tube, no genome, 30-odd purified factors, a cell-free protein factory. Syn3.0 is the genome-minimisation pole of that shelf. Syn61 is the code-rewriting pole. Sc2.0 is the eukaryotic-rewrite pole. PURE is the no-cell pole. The sci-fi is a designed organism. The true sentence is already stranger. We know the lower bound of a free-living genome well enough to synthesise it, and we still don't know what a third of the essential genes do. Humility and power, in the same paper, which is how we like our synthetic biology.

In short. Other labs are rewriting yeast, recoding E. coli, or running protein synthesis in a tube with no cell at all. Syn3.0 is the pole where the genome itself was shrunk until it almost failed.

Optional equipment is the rest of this journal, and it's why a peptide-map diagram belongs at the end of a mycoplasma essay. Dsup, the tardigrade's disordered DNA-binding protein, is a radiation buffer you can move into a human nucleus in a dish; a 473-gene cell doesn't carry it. AlphaFold will give you a ribbon for a syn3.0 unknown on a Tuesday; the ribbon isn't a function. CRISPR–Cas9, base editors, prime editors — the writing tools of the 2010s and 2020s — are how later synthetic biology edits a living genome without resynthesising 531 kilobases from oligonucleotides. Syn3.0 was built the old, heroic way: total synthesis, yeast assembly, transplantation. A Cas9 would now delete a candidate unknown in an afternoon, which is why the post-2016 functional papers could move faster than the design cycles that produced the cell. The enzyme doesn't care whether the genome is synthetic. The enzyme also doesn't annotate. Neighbourhood on this reading list isn't identity of method. Total synthesis is one way to get a minimal cell. Programmable nucleases are a way to interrogate it. A fold-predictor is a way to stare at what you still can't name.

In short. Radiation shields, structure prediction and gene-editing scissors are extras on top of the 473-gene floor. They help us study the cell. They aren't what made it alive.

A peptide catalogue is a luxury an organism with 473 genes doesn't have. That sentence belongs on this page once, without a product. The 438 proteins of syn3.0 are ribosomal subunits, polymerases, transporters, glycolytic enzymes, membrane proteins, and a third of mystery. They're ribosomes and bags, not incretins or copper-binding tripeptides or a 16-mer from a mitochondrial rRNA. Those ligands exist because eukaryotic bodies run signalling programmes a mollicute never paid for. Occupancy at a GPCR is a later invention. NAD+ as a sirtuin substrate is a later invention. The dinucleotide as a hydride coin isn't; syn3.0 still needs that coin for GAPDH. Keep them apart. A reading list can sit a minimal cell next to retatrutide and next to a tardigrade protein. They are not one experiment. The map that follows is a courtesy to the reader who arrived from the peptide desk, and a warning to the reader who would like the 473-gene floor to license a vial. It doesn't.

In short. The extra signalling proteins sold as research peptides belong to complicated animal bodies. The minimal cell doesn't make them. Shared reading lists aren't shared mechanisms.

Here's the map we'd like you to take home, rather than a creation myth. Starting species: Mycoplasma mycoides, already a parasite, already near a floor. Method: transposon map, design-build-test, yeast assembly, transplantation into a related cytoplasm. Design rule: knock-out until it dies, then put the gene back. Product: JCVI-syn3.0, 531,560 base pairs, 473 genes, 438 of them proteins, doubling in about three hours, polymorphic until 19 genes were restored as syn3A. Census: expression, genome preservation, membrane, cytosolic metabolism, and a remainder that at publication included 149 unnamed essentials. Energy: glycolysis, no mitochondrion, rich medium. Reading: a reduced polymerase, almost no dedicated regulation, UGA as tryptophan. Later: Pelletier on shape, Breuer on metabolism, Thornburg on a whole-cell model, Zhang on folds and ontology. Neighbouring objects on this desk: a recoded E. coli, a synthetic yeast, a cell-free tube, a Cas9, a tardigrade protein, a predicted ribbon. The cytoplasm had a history. The sequence didn't. Both halves, always.

In short. Leave with the map: a written mycoplasma genome, 473 genes, many still mysterious, grown in rich broth, insides borrowed at the start from another cell.

Engineering, not a theory of life. We've now said that enough times that it can sit as the closing claim rather than as a refrain. A theory of life would say why these 473, why these 149, why a wall-less fermenter and not a walled autotroph, why three hours and not three days. The experiment says: this deletion series, this medium, this species, this growth-rate demand, this is what remained. That's a local minimum in a designed space, and local minima are what engineers find. They're enormously useful. They aren't universal. A different group, starting from a cyanobacterium or from a free-living heterotroph, would print a different fasta and a different unknown column. The power is that the fasta exists, that it divides, that you can add a gene or subtract one and ask a clean question. The humility is that a third of the essential column was blank in 2016 and isn't fully filled now. We'd not trade the colony for a tidier philosophy. We also won't let the colony pose as the philosophy.

In short. This is a working design for one tiny cell, not the hidden formula of all life. Useful because it's real. Incomplete because a third of it was unnamed.

The public papers are the reading list, and they're short enough to actually read. Fraser, Science 1995, the genitalium genome. Hutchison, Science 1999, and Glass, PNAS 2006, the essential-gene estimates. Lartigue, Science 2007, transplantation. Gibson, Science 2008, synthesis of a chromosome in yeast. Gibson, Science 2010, syn1.0, the first cell on a synthetic chromosome. Hutchison, Science 2016, syn3.0, 473 genes, 149 unknowns. Breuer, eLife 2019, the metabolic reconstruction. Zhang, J Proteome Res 2021, the later annotation push. Pelletier, Cell 2021, 19 genes, seven of them for shape. Thornburg, Cell 2022, the 3D model. Fredens, Nature 2019, if you want the recoded-E. coli neighbour. Jinek, 2012, if you want the nuclease that now interrogates these genomes without a total resynthesis. Jumper, Nature 2021, if you want the ribbons for the remaining blanks. That's a fortnight of evenings, not a guru. The created-life headlines will still be there when you come back, and they will look smaller.

In short. A short stack of named papers covers the natural tiny genome, the transplant, the 2010 synthetic cell, the 2016 reduction, and the later shape and computer models.

Scoreboard, kept boring on purpose, because that's the only version that will still be readable in five years. Smallest natural cellular genome, plated: M. genitalium, 1995, about 580 kilobases. First synthetic bacterial genome in a cell: JCVI-syn1.0, 2010, 1.08 megabases, 901 genes. First designed minimal genome that grew: JCVI-syn3.0, 2016, 531 kilobases, 473 genes, 149 unnamed, three-hour doubling, ugly shape. Usable near-minimal physiology: JCVI-syn3A, 19 genes restored, 2021, spheres, two-hour doubling. Whole-cell simulation: 2022. Remaining blanks: fewer than 149, not zero. Recoded E. coli and synthetic yeast: parallel shelves, not sequels. No designed autotroph. No cell from a bag of chemicals without a cytoplasm. No theory of life in the strong sense. A fasta that divides, and a table that still has empty cells. CRISPR, AlphaFold and a tardigrade protein are tools and extras. They aren't the floor. We'll update the unnamed count when someone finishes Hutchison's goal. Not before.

In short. 2010: a synthetic genome that runs a cell. 2016: 473 genes, many unnamed. 2021: a tidier, slightly larger version. Still no cell built from scratch without borrowed insides.

This page isn't a catalogue listing, and research-use-only is the wrong legal class for a bacterium in a La Jolla freezer. The objects here are genomes, papers, and a cell that divides. A neighbouring CRISPR medicine is licensed for two blood diseases. A neighbouring peptide vial is a characterised laboratory solid, not a minimal cell. Confusing those classes is how a reader ends up with a protocol they shouldn't have. The physiology in the paragraphs above is public, cited, and older than this journal. Use it to read the next syn-cell paper with the papers in front of you rather than the press release. Name the 473. Name the 149. Name the cytoplasm that had a history. Name the rich medium. Name the 19 genes that made a sphere. Name the fact that a knockout screen isn't an annotation. Then argue about the remaining blanks and about what autonomous was allowed to mean. We won't sell you a mycoplasma. We'll tell you the parts list, and we'll keep the blanks in the list where Hutchison left them.

In short. This isn't a product page. It's a map of a real, unfinished attempt to name every gene a cell needs. Read the next paper against that map, not against a creation headline.