Synthetic Biology

Synthetic biology is not a synonym for genetic engineering. It is a specific programme with a citable origin: Drew Endy’s Foundations for engineering biology (Nature 438:449–453, 2005) proposed importing three abstractions from other engineering disciplines — standardisation of parts, decoupling of design from fabrication, and hierarchical abstraction so that a designer need not think in base pairs. The claim under test on this page is the slogan that grew out of that paper: that biology is becoming an engineering discipline.

The honest way to test it is to ask whether Endy’s three foundations were delivered, and the answer splits cleanly down the middle. Fabrication was delivered. Writing megabases of DNA is a solved manufacturing problem with a twenty-year cost curve behind it. Design was not. The failure rates are now published, by the groups whose programmes they indict, and they are the numbers the slogan omits. Endy said so himself in the founding paper: our ability to engineer biological systems that behave as expected “remains quite limited”. Twenty-one years later that sentence is still the most defensible summary of the field.

Established Start with the half that worked, because it is genuinely impressive and it is what the rhetoric is built on. Gene synthesis is a commodity. The 1centbp cost tracker, assembled from Carlson-curve data, Twist Bioscience SEC filings and GenScript pricing, puts synthesis at $0.07 per base pair in 2025 against $25.00 per base in 1999 — a 357-fold fall averaging about 21% a year. Frontier The series mixes promotional list pricing with filings, so it is evidence about what a buyer pays rather than about marginal cost.

Established What that buys is a real and growing catalogue of written genomes. JCVI-syn1.0 (2010) booted a chemically synthesised Mycoplasma mycoides genome in a recipient cell. JCVI-syn3.0 (2016) minimised it to 531 kb and 473 genes, smaller than any autonomously replicating cell in nature. The Sc2.0 consortium has now synthesised all sixteen Saccharomyces cerevisiae chromosomes plus a designed tRNA neochromosome; the last of them, synXVI, is 902,994 bp redesigned from a 948,066 bp wild-type original, published in Nature Communications on 20 January 2025. Recoded Escherichia coli exists as Syn61 and its virus-resistant derivative Syn61Δ3. Established Nobody serious disputes any of this. The dispute is about what it implies.

Established The design failure rate is now measured, and the cleanest measurement is the 57-codon E. coli project. Ostrov and colleagues began in 2016 with a design removing 62,029 instances of seven codons through 162,521 base-pair changes. A decade later Nyerges, Chiappino-Pepe, Ostrov, Church and 27 co-authors published the outturn in Nature Communications 17:5311 (22 June 2026). 19 of 88 synthetic chromosomal segments — 22% — did not support growth once the parental copy was deleted; defects were not isolated but densely distributed, with roughly one lethal and several fitness-decreasing problems per segment; and recoding, in the abstract’s own words, “frequently proves lethal, with fitness impacts remaining poorly characterized”. Established After ten years the best strains carry up to 45.8% synthetic genome. The whole-genome 57-codon organism does not exist.

Established The rescue method matters as much as the failure rate. They did not fix the design by designing better. They fixed it by multi-omics-guided evolution — profile the broken cell, then let selection find the repair. Frontier That is this brief’s reading rather than the authors’ claim about their field, but it is hard to read the method any other way. Established And the interest runs hard against the finding: this is the recoding programme’s own laboratory reporting that its flagship design was pervasively wrong.

Established Yeast tells the same story in a different organism. Erpf, Meier, Walker, Goold, Boeke, Paulsen and Pretorius published a Nature Biotechnology comment in December 2025 whose entire subject is the design- and construction-related defects revealed across sixteen synthetic chromosomes and the debugging methods invented to chase them. In synXVI specifically, three separate defects — at CTR1, GIP3 and MRPL51 — traced to one systematic design error: loxPsym recombination sites placed at the termini of “dubious” open reading frames that turned out to overlap verified genes, disrupting their 5-prime untranslated regions. The designers did not know what the sequence did.

Frontier Protein design is the sub-field with the strongest engineering claim and a Nobel Prize, and its first-pass hit rate is about one in nine. A meta-analysis of 3,766 experimentally characterised de novo binders across 15 targets (bioRxiv 2025.08.14.670059v2) found 436 — 11.6% — bound in vitro, with per-target success ranging from 0.1 to 1.0 and the authors stating that “design success remains very inconsistent, and there are no standard criteria to prioritise binders for experimental testing”. Frontier It is a preprint, not peer-reviewed, and is carried at that weight. Frontier An eleven-per-cent first-pass hit rate is a screening technology, not a design discipline — that characterisation is this brief’s, and it is the sentence the rest of the page rests on.

Established A third of the minimal cell remains a black box, and closing it has not been attempted. Of JCVI-syn3.0’s 473 genes, 149 — 31.5% — have unknown biological function. The follow-up is more damaging to the engineering framing than the original result: syn3.0 divided abnormally, and restoring normal division required adding back seven genes, five of them with completely unknown roles, found by systematic add-back rather than by design. Established The figure of nineteen genes that circulates for this step, including on this Institute’s own earlier page, is wrong; JCVI’s account of the Cell paper says seven, and the correction is recorded here rather than quietly made.

Frontier The counter-move is simulation, and it produced a genuine result in 2026. JCVI announced on 9 March 2026 that Luthey-Schulten, Thornburg, Maytin, Mehta, Ha, Gilbert and Glass had simulated a complete 105-minute cell cycle of JCVI-syn3A — replication, translation, metabolism and division — in about six days on two GPUs, landing within two minutes of the real cell cycle on average. Frontier The source is an institutional press release about a Cell paper, the model averages molecular dynamics rather than simulating atoms, and it models the one organism whose parts list we most nearly have. It is nonetheless the most substantive answer yet to the interpretation gap.

Established The commercial record is the least-cited evidence here and among the most probative, because business failure is a hard readout that does not respond to framing. Zymergen raised about $530 million at IPO in April 2021; on 13 September 2024 the SEC charged it with misleading investors and imposed a $30 million civil penalty, finding that its claimed $1 billion market for the Hyaline film rested on “flawed and unreasonable assumptions” and that the company “misled investors with what amounted to unsupported hype”. It went bankrupt in 2023. Amyris entered Chapter 11 in 2023 with a $190 million debtor-in-possession facility. Established Ginkgo Bioworks, whose foundry is the physical embodiment of the biology-as-engineering thesis, reported FY2025 revenue of $170.2 million against $227.0 million the prior year, a net loss of $312.8 million, and R&D cut from $424.1 million to $243.8 million. The foundry is shrinking, not scaling.

Established The purest clinical test of “design a cell to a specification” failed with the mechanism intact. Synlogic engineered E. coli Nissle to consume phenylalanine in the gut of phenylketonuria patients. On 9 February 2024 the phase 3 Synpheny-3 trial was stopped when its data monitoring committee judged it unlikely to meet the primary endpoint; the company ceased operations and cut over 90% of staff while holding $47.7 million. Safety and tolerability were acceptable. The organism did what it was designed to do and the patients did not get better.

Established And the field’s showcase success cost more than the thing it replaced. Semi-synthetic artemisinin — engineered yeast from the Keasling laboratory, Gates-funded, manufactured by Sanofi — was priced at about $400/kg on a no-profit-no-loss basis while plant-derived artemisinin ran cheaper. Production peaked at roughly 8% of global supply in 2013, fell to zero by 2015, and recovered only to about 20 metric tons in 2017 after Huvepharma bought the plant; rival artemisinin-combination-therapy manufacturers were reluctant to buy from a competitor. In 2018 the Gates Foundation went looking for a route at $100/kg or less. Frontier The ETC Group, a campaigning organisation opposed to synthetic biology, puts the same figures at $370–400/kg against natural artemisinin at $250–270/kg and argues the real costs were hidden by philanthropic support; its agreement on the price arithmetic carries weight precisely because its politics run the other way. Its further claim — that the launch coincided with a decade-low artemisinin price and a two-thirds fall in plantings, threatening about 100,000 smallholders — is a causal attribution and is flagged as such.

Established Two plant-based platforms now hold regulatory approvals, and both are this brief's economic thesis in its purest form: the biology worked and the business did not. Plant molecular farming — expressing recombinant proteins in plants or plant cell culture rather than in mammalian cells or microbial fermenters — has cleared regulators more than once. A plant-cell-derived enzyme replacement therapy for Gaucher disease was approved in the United States in 2012. A plant-produced virus-like-particle COVID-19 vaccine was authorised by Health Canada in February 2022, the first plant-made vaccine licensed for human use anywhere. Established Its manufacturer was wound up by its parent roughly a year later and the product never reached distribution at scale. Frontier The proximate reason was not efficacy or safety: the World Health Organization declined to prequalify the product because of the manufacturer's tobacco-industry shareholding, which closed the international procurement channel the volume case depended on. Speculative A platform that satisfied a regulator and was then defeated by an ownership structure is not a technology failure, and reading it as one is how this field keeps mislocating its own constraint.

Established The technical case for plants is real and it is about capital rather than about biology. Transient expression by agroinfiltration of Nicotiana benthamiana can run from sequence to gram quantities of protein in weeks, with no bioreactor train, and scale-up is planting more plants rather than commissioning a larger vessel. Established The counterweight is downstream: recovering a protein from green tissue means separating it from RuBisCO, phenolics and a full complement of plant proteases, and in published accounts of platform economics purification dominates cost. Frontier The advantage plants hold is in the half of the process that is not the expensive half — which is the same structural error the synthesis price curve makes earlier in this section, a cheap input quoted as though it were a cheap product. Established Plants also modify what they express: plant N-glycans differ from human ones, and correcting that required engineering the host's glycosylation machinery rather than the gene. Frontier That is the context-dependence problem of section 4 arriving at protein level, in a platform where the sequence specified is not the molecule delivered.

Established Sprayable RNA is the second platform, and it is the cleanest test this field has ever had of the design thesis. Double-stranded RNA applied to a crop can silence a chosen gene in a chosen pest by sequence complementarity; the design object is a sequence, and its specificity is in principle computable against a genome database before anything is built. Established A dsRNA targeting a proteasome subunit of the Colorado potato beetle was registered by the United States Environmental Protection Agency in late 2023 as the first sprayable RNA pesticide, following plant-expressed RNAi against western corn rootworm registered several years earlier. Frontier If rational design from first principles works anywhere in applied biology, it should work here.

Established It half works, and the half that fails is the informative half. Target selection from sequence is genuinely rational. Delivery is not: dsRNA uptake efficiency varies by orders of magnitude across insect orders — beetles are highly susceptible, while many moths and sap-feeding bugs degrade the molecule in the gut or fail to internalise it — and none of that is predictable from the target sequence. Established Environmental persistence runs the wrong way for efficacy: dsRNA degrades quickly on leaves and in soil, which is excellent for non-target exposure and means the product must be reapplied. Established And resistance, when it was selected in the laboratory, was not sequence-specific. A western corn rootworm population selected against one dsRNA showed impaired luminal uptake and cross-resistance to unrelated dsRNAs. Frontier That single result breaks the platform's central promise. The pitch for RNA pesticides is that resistance is answered by redesigning the sequence, which is cheap; a resistance mechanism that blocks the delivery route rather than the message makes the entire chemistry one mode of action rather than an inexhaustible library of them.

Handwave The strong engineering thesis — specify, design, build, works — is stated in company prospectuses and foundry marketing and is not supported by the evidence offered for it. The evidence cited is fabrication evidence: megabase synthesis, sixteen yeast chromosomes, a 473-gene cell, the price curve. Every one of those is a claim about writing, deployed to support a claim about designing. What would settle it is a published first-pass success rate above about 50% for non-trivial designed constructs, in more than one laboratory. No such figure exists.

Frontier The weak thesis — fabrication is engineered, design is high-throughput search — is the best-supported position on the table, and it has a serious defence. Castle, Stock and Gorochowski (Nature Communications, 29 April 2024) argue for an “evolutionary design spectrum” in which rational design and directed evolution are points on one continuum, because both generate variants, test physically, select and iterate. They note that about 10% of random 100-nucleotide sequences show some promoter activity — functional solutions are dense enough in sequence space that search beats derivation. On that reading an 11.6% binder hit rate is not a scandal; it is throughput multiplied by generations. Handwave But it costs the slogan its content. If “engineering discipline” means high-throughput search plus debugging, then bridge-building is not the analogy, and phrases such as “designed to specification” are doing their work by assertion.

Established The interpretation-gap thesis holds that the binding constraint is the inability to predict phenotype from sequence, and that no amount of synthesis capacity relaxes it. As a description of the present constraint this is established: 149 unknown-function essential genes in the smallest genome ever built, and synXVI’s defects traced to misannotated reading frames. Frontier As a claim about permanence it is contested, and the test is specific: close the unknown-function fraction of one minimal genome. Nobody has.

Frontier Whether whole-cell simulation closes that gap is the live question and it now has a real result attached. The syn3A cycle was reproduced to within two minutes over 105 minutes for six GPU-days. Frontier What would settle it is different and harder: a simulation that correctly predicts the phenotype of a genome design before it is built, for an organism carrying unknown-function genes. Simulating a cell you have measured is not the same as predicting one you have not.

Speculative The reproducibility thesis says the field’s apparent unpredictability is substantially measurement error rather than biology, and it is a hypothesis with almost no direct measurement behind it — which is itself the point. Lux, Strychalski and Vora (Synthetic Biology 8(1):ysad014, 2023) revisited Kwok’s five hard truths of 2010 — undefined parts, unpredictable circuitry, unwieldy complexity, incompatible parts, variability that crashes systems — found all five live, and reported that “only a handful of studies directly measuring interlaboratory reproducibility can be found in the synthetic biology literature”. Established Their anchor is the Reproducibility Project: Cancer Biology, where 0 of 193 experiments from 53 papers carried enough detail to attempt without contacting the authors, and after eight years and $1.5 million only 26% could be attempted, of which 46% of binary effects reproduced. Frontier Transferring that base rate to synthetic biology is an inference, not a measurement.

Frontier Whether recoded organisms become the standard industrial chassis is open and has a clear settlement condition: a recoded chassis in commercial production. Syn61Δ3’s virus resistance and freed codons are the argument for it; nothing is yet manufacturing at scale on one.

Speculative Writing a human chromosome is the largest new commitment in the field. SynHG was announced on 27 June 2025 with £10 million from Wellcome over five years across Cambridge, Kent, Manchester, Oxford and Imperial, led by Jason Chin with Patrick Yizhi Cai as Chair of Synthetic Genomics, targeting foundational tools plus a fully synthetic human chromosome — roughly 2% of the genome. It runs a parallel empirical social-science programme, Care-full Synthesis, under Joy Zhang. Established The funding and the project exist. Speculative That a synthetic human chromosome will exist within five years does not follow from either.

Speculative The bottom-up cell — self-replicating, metabolising, dividing, assembled from defined molecules with no living ancestor — does not exist. Vesicles that grow and divide exist; encapsulated expression exists; minimal metabolic modules exist. Build-a-Cell and MaxSynBio are the organised attempts to integrate them. What would settle it is one vesicle that grows, replicates its genome and divides through multiple generations. Established And note the boundary that “from scratch” quietly crosses: every synthetic genome to date has been booted inside a pre-existing cell. The DNA is synthetic; the ribosomes, membranes and cytoplasm that read it are inherited.

Handwave De novo genome design from first principles — writing a genome for an organism nobody has sequenced — appears in field rhetoric and is held by nobody as a near-term claim. Every synthetic genome in the record is a copy-and-modify of a natural one. There is no evidence for it and no programme pursuing it.

Frontier The economic thesis is that the problem is unit cost rather than biology: engineered routes lose to agriculture and to chemistry. Artemisinin at $400/kg against a plant at $250–270/kg, market share to zero within two years of launch, and Ginkgo’s revenue decline are the evidence. What would settle it is a synthetic-biology product that wins on price against a non-biological incumbent, at scale, without subsidy.

Frontier Delivery is where the money and the papers are, and it is a materials problem rather than a genetics one. Layered clay nanosheets, chitosan and peptide carriers and assorted nanoparticle formulations have each been shown to extend leaf retention and improve uptake in demonstration experiments. Speculative None has been shown to move a refractory insect order into the susceptible category at field rates and field prices, which is the result that would open the market beyond beetles. Frontier The protein side has the same shape: the open question for plant molecular farming is not whether a plant can express a biologic but whether downstream processing can be industrialised to a cost per dose that beats a depreciated mammalian-cell facility. Speculative Both platforms are waiting on a manufacturing result rather than a biological one, which is the position this brief argues the whole field occupies. Frontier The nearer commercial opening for plant expression is probably not human therapeutics but animal-free growth factors and feed proteins, where the incumbent price is high and the regulatory burden far lower — and where a success would say nothing about whether the therapeutic case closes.

Established The binding bottleneck is sequence-to-phenotype prediction, and every other bottleneck on this list is downstream of it. A field that can write a megabase for a few thousand dollars and cannot say what 22% of its designed segments will do is not constrained by fabrication. Established The two hardest numbers in the subject — 19 of 88 segments lethal and 11.6% of binders binding — are both measurements of prediction failure, in different organisms, at different scales, by different communities.

Frontier Standardised parts did not deliver what the name promised. The Registry of Standard Biological Parts and the BioBrick assembly standard were the concrete instantiation of Endy’s standardisation foundation; what followed was a stream of replacement standards, BglBricks among them, because the original prefix-and-suffix scheme scarred junctions and constrained protein fusions, and because parts were deposited without characterisation data. Frontier The proliferation of successor standards is documented; reading it as failure is this brief’s inference. Established The load-bearing critique is narrower and harder to answer: a part is not standard if its behaviour changes with host and context, and the field has never established the baseline reproducibility that would let anyone separate measurement noise from biological noise.

Established Booting remains a hosted operation. Synthetic DNA has never started a cell on its own. Installation into a recipient cell supplies ribosomes, membranes, polymerases and cytoplasm, so the boundary between “synthesised” and “created” sits exactly where the popular coverage erases it.

Established Assembly and error correction still gate scale even though per-base synthesis is cheap. Oligonucleotides are short; genomes are long; the cost that matters at chromosome scale is hierarchical assembly with verification at each step, and the price series that falls 21% a year is a price for the input, not for the assembled and verified construct. Frontier No published series tracks the assembled-and-verified cost the way 1centbp tracks the raw one.

Frontier The field cannot measure its own failure rate, which makes every improvement claim unfalsifiable. With near-zero interlaboratory reproducibility studies there is no denominator: a construct that fails in a second laboratory could be a bad design, a bad part, a different host, or a different plate reader, and nothing in the literature separates those.

Established Organism-scale synthetic biology waits on genome writing as a capability, and that is a result another brief on this map produces. Genetic Engineering owns editing and genome writing generally; everything in section 2 that involved a chromosome — syn1.0, syn3.0, Sc2.0, Syn61, the 57-codon project, SynHG — consumed that capability rather than producing it. The typed edge below records it.

Frontier What it waits on from research is narrower and is not being produced by anyone on this map. Three things: a method that predicts phenotype from a designed sequence well enough to cut build-test cycles by an order of magnitude; a characterised parts set whose behaviour is stable across hosts; and a measured interlaboratory failure rate to compare any of it against. Speculative The first is what whole-cell simulation is aimed at and has not yet demonstrated on an unbuilt design. The third is the cheapest of the three and nobody has run it.

Established Two dependencies are not research results at all and are recorded as typed non-brief constraints in section 13: synthesis priced and provisioned for genome-scale, error-corrected orders, and a screening regime with an enforcement mandate. Both are things the world could choose to supply. Frontier One of them was supplied and then withdrawn, which is a fact about institutions rather than about biology.

Established The cheapest high-value experiment in this subject is an interlaboratory ring trial on a standard construct, and it has never been run. The field’s own reproducibility analysts report only a handful of interlaboratory studies in the whole literature. A properly powered ring trial — one plasmid, one host, one readout, ten laboratories, pre-registered — would settle whether the unpredictability everyone reports is biology or measurement, and it costs a rounding error against a single foundry’s annual budget.

Established Second: publish first-pass design success rates as a standing metric. The two figures that exist — 19 of 88 segments and 436 of 3,766 binders — came from a genome project reporting its own lethality and from a meta-analysis of other people’s papers. Neither is a routine disclosure. Frontier A field that reported hit rates the way clinical trials report endpoints would make the strong engineering thesis testable for the first time.

Frontier Third: make the whole-cell simulation predict something unbuilt. The syn3A result reproduced a cycle that had been measured. The discriminating experiment is to hand the model a designed genome variant, record the prediction, then build it — and to do it on an organism with unknown-function genes, because that is the case the interpretation gap is about.

Established Fourth: close the unknown-function fraction of one minimal genome. 149 genes in syn3.0. It is a bounded, finite, unglamorous target, it is the single most direct test of the interpretation-gap thesis, and a decade after the genome was published it has not been done.

Frontier Fifth, and uncomfortable: the Paraphrase Project’s redesigned toxin sequences were never expressed. The demonstration that screening can be evaded is in silico. Whether the evading designs would have retained function is unknown, and it is the difference between a screening gap and a weapons capability. Handwave It is also an experiment that should probably not be run as stated, which is a real methodological problem rather than a rhetorical one: the biosecurity literature is arguing about the severity of a vulnerability whose severity cannot responsibly be measured.

Established The engineering that works is the foundry, and its economics are now public. A biofoundry industrialises the build-test-learn loop: automated assembly, high-throughput transformation, plate-scale screening. It is a search engine with liquid handlers. Established Ginkgo built the largest one and its FY2025 filing shows revenue falling from $227.0 million to $170.2 million, R&D cut from $424.1 million to $243.8 million and an adjusted EBITDA loss of $167.0 million. The throughput thesis is not being disproved by biology; it is being disproved by the income statement.

Established The synthesis supply chain is the input that actually got engineered. Column and array-based oligonucleotide synthesis, enzymatic synthesis, hierarchical assembly and sequence verification are mature industrial processes with named vendors and published prices. Frontier They are also concentrated: the cost series is built on two companies’ pricing plus one company’s filings, which is thin for a curve carrying this much argumentative weight.

Established Recoding is the most demanding engineering the field attempts, and its cost is now legible. 162,521 base-pair changes across a genome, ten years, 88 segments designed, 19 dead, six partially recoded strains, 45.8% maximum synthetic content. Established The per-segment debugging workflow — multi-omics profiling followed by selection — is the actual deliverable of that decade, and it is a debugging method rather than a design method.

Frontier Biocontainment is the one place where synthetic biology engineers a safety property into the substrate rather than around it. Recoded chassis can be made dependent on synthetic amino acids unavailable in nature and unable to exchange genes with wild organisms. Speculative No genetic firewall has been tested at environmental scale, and escape frequencies quoted from laboratory work are not field measurements.

Established Containment for a plant platform is an agronomic problem rather than a genetic-firewall one, and it has a documented precedent. A recoded chassis in a fermenter can be made dependent on a synthetic amino acid; a field of engineered plants cannot be, and its containment is physical distance, dedicated equipment and control of volunteer plants that germinate the following season. Established The reference case is a 2002 United States enforcement action in which volunteer maize engineered to express a veterinary vaccine protein appeared in a following soybean crop and the harvest was impounded. Frontier The regulatory response shaped the field: plant-made pharmaceuticals in food crops became administratively expensive, which is why the platform converged on a non-food tobacco relative and on contained indoor growing rather than on open acreage. Speculative That is a genuine case of a containment failure selecting a technology, and it sits awkwardly beside the observation above that no genetic firewall has been tested at environmental scale: the one platform that did go outdoors was disciplined by an incident rather than by a firewall.

Within this map, three boundaries need stating precisely because the subjects overlap in ordinary speech. Genetic Engineering owns editing and genome writing as a general capability — the tools, the therapies, the regulatory record for editing existing organisms. This brief is the engineering-discipline claim instead: standardised parts, decoupled design and fabrication, hierarchical abstraction, and whether Endy’s three foundations were delivered. Xenobiology owns alternative chemistries — expanded genetic alphabets, xeno-nucleic acids, and the mirror-life governance question that produced the December 2024 call to withdraw funding. Synthetic biology as treated here rearranges the natural alphabet; xenobiology changes it. And Designer Organisms is the specification claim: whether an organism can be built to spec, tested against de-extinction, gene drives and organisms actually released.

Also adjacent within the map: Biological Computing, which inherits this brief’s circuit-predictability problem in a harder form; Microbiome Engineering, where Synlogic’s failure is directly load-bearing; Synthetic Ecosystems, which asks the design question one level up; Future Agriculture, where the artemisinin economics recur as a general pattern; and Biological Energy Systems, whose engineered-route-versus-incumbent arithmetic is the same arithmetic.

Outside the map: the origins programme’s treatment remains the entry point into this subject for readers coming through the programme path, and it is shorter, narrower and deliberately kept as an introduction. This brief supersedes it as the slot’s coverage, corrects it on the syn3A add-back count, and adds the failure-rate, economic and biosecurity material it does not carry. Further out: protein biophysics, industrial fermentation, metabolic engineering, and the biosecurity policy literature.

Established DNA-synthesis screening is the field’s principal safeguard, and 2025 was the year it was shown to be leaky and then partly unbuilt. Providers check orders against sequences of concern. Microsoft’s Paraphrase Project, published in Science on 2 October 2025 with Eric Horvitz as lead author, used generative protein-design models including EvoDiff to redesign toxin sequences so that they evaded screening while retaining function in silico. The work was digital only, disclosed to the US government and to screening-software vendors before publication, and patched. Frontier Horvitz’s own assessment is the one to carry: “the patch is incomplete, and the state of the art is changing”.

Frontier The institutional layer underneath the screen is thin, and the people saying so run biosafety consultancies rather than laboratories. Gillum and Moritz, in Frontiers in Bioengineering and Biotechnology on 28 October 2025, argue US policy moved to sequence-level oversight without the capacity to execute it: university biosafety offices typically have fewer than three full-time professional staff; few possess commercial screening software for sequences of concern and fewer can perform entity-wide sequence audits; legacy constructs numbering tens of thousands per laboratory circulate informally through collaborators and repositories; providers show substantial variation in screening sensitivity; and benchtop synthesisers bypass institutional oversight entirely. Established And the 2024 Framework for Nucleic Acid Synthesis Screening was rescinded by Executive Order 14292 in 2025. The chokepoint everyone points to was weakened by demonstration and by policy in the same year.

Established The most uncomfortable fact in the governance record is that the regime responds to a threat that has never materialised. No confirmed malicious incident involving synthetic nucleic acids has been documented. Frontier That cuts both ways and this brief refuses to round it in either direction: it does not show the risk is negligible, because the regime exists precisely to prevent a first instance, and it does not show the regime is working, because nothing distinguishes deterrence from absence of attempt.

Established Internationally, the treaty body declined to restrict anything. CBD Decision 16/21 on synthetic biology, adopted 1 November 2024, imposed no moratorium. It established a new technical expert group and a thematic action plan on capacity-building and technology transfer, emphasised a precautionary approach to environmental release, and does not mention gene drives at all.

Established The securities regulator turned out to be a more consequential institution here than any biosafety body. The SEC’s $30 million penalty against Zymergen is the only case in this brief where a formal finding was made that the engineering claims were unsupported — and it was made about a market projection, by an agency with no view on biology whatever. Frontier Where scientific claims are enforced at all in this field, they are enforced as investor-protection claims.

Established The regulatory categories these two platforms fall into differ in a way that decides their fates. A plant-made biologic is regulated as a drug or a vaccine, by the agency that regulates drugs, on the strength of clinical data. A sprayable double-stranded RNA is regulated as a pesticide, by the agency that regulates pesticides, on the strength of environmental-fate and non-target data. Established Those are different burdens, different timescales and different expert communities, applied to the same class of molecule. Frontier The pesticide route produced the faster approval and the harder market: an approved biologic has a reimbursed price, while an approved pesticide competes against a synthetic insecticide costing a few dollars a hectare. Speculative What the pair says is that regulatory approval was not the binding constraint for either platform, which is exactly what the commercial record in section 2 says about this field generally.

Established The dual-use argument is real, the chokepoint is the screen, and the screen has now been publicly shown to be evadable. That combination puts a genuine burden on the field, and the honest statement of it includes the part that weakens the alarm: the Paraphrase redesigns were never expressed, so it is not known whether they would have been toxic. Handwave Coverage implying that AI has produced working bioweapon designs is asserting the untested step.

Frontier The artemisinin case is the field’s clearest distributive question and it is usually told as a triumph. A philanthropically subsidised engineered route entered a market supplied by smallholder farmers, sold at a higher price than the crop, took about 8% of supply at peak, and collapsed to zero within two years. The ETC Group’s claim that plantings fell by two-thirds and threatened around 100,000 smallholders is contested causal attribution from an interested party. Established What is not contested is that the engineered product was the more expensive one and that the funder later sought a cheaper route.

Established Investor harm is an ethical outcome of overclaiming, not merely a market outcome. A $530 million IPO, a $30 million penalty and a regulator’s finding of “unsupported hype” are the measurable cost of a field describing search as design. Frontier The scientific and the financial claims were not separable: the market projection the SEC found unreasonable rested on the engineering framing this brief is testing.

Established The “creating life” framing is the one the field mostly did not choose and has not corrected loudly enough. Every synthetic genome was installed in a pre-existing cell. Nothing in the record synthesises a cell from non-living components. Frontier Saying so is not pedantry: the public argument about whether synthetic biology crosses a moral line is partly an argument about a step nobody has taken.

Frontier And CBD 16/21’s emphasis on capacity-building and technology transfer names an equity problem the technical literature rarely does. Synthesis capacity, foundry capacity and screening capacity are all concentrated in a handful of countries; a precautionary approach to environmental release written by parties who do not host that capacity is a different instrument from one written by parties who do.

Frontier The terminal position is a split verdict, and stating it as a split is the finding rather than a failure to conclude. Fabrication is an engineering discipline: priced, industrialised, falling in cost for twenty-five years, delivering megabase constructs on demand. Design is not: 22% of designed segments lethal, 11.6% of designed binders binding, a third of the minimal genome unannotated, the flagship recoding project rescued by selection rather than redesign. Both halves are measured, and picking a side requires ignoring one of them.

Frontier If the weak thesis is right, the civilizational implication is smaller than the rhetoric and stranger than the criticism. A biology that improves by search rather than by derivation still gets more capable every year — throughput rises, models narrow the search, debugging methods accumulate. What it does not do is become predictable, which means it does not acquire the property that makes civil engineering safe: the ability to certify a design before it exists.

Speculative Writing a human chromosome is the point at which this stops being an industrial question. SynHG has funding, a five-year horizon and an explicit social-science programme running alongside it, which is itself an institutional innovation worth noting. Whether a synthetic human chromosome is produced in that window is unknown; that the attempt is funded and public is not.

Speculative The long-run question is whether biology ever becomes derivable at all, or whether living matter is a domain where search is the only method that works. Nothing in the evidence settles this. The 10% promoter-activity figure hints that biological function is dense enough in sequence space to be found rather than computed; the whole-cell simulation hints the other way. Two hints, opposite directions, no measurement that discriminates.

Handwave And the version of this subject that appears in strategy documents — an economy in which arbitrary materials, foods and medicines are specified and grown — requires the strong thesis, which has no supporting evidence. The bioeconomy projections rest on the design claim; the design claim rests on the fabrication evidence; the fabrication evidence does not support it.

These horizons track construction milestones and disclosure practices rather than a technology curve, because the curve that exists is a price curve for the input:

• 10 yr: Frontier A consolidated all-synthetic S. cerevisiae strain carrying all sixteen chromosomes is the most likely landmark, and no fetched source states a date or a remaining work list for it. Frontier SynHG either produces a synthetic human chromosome inside its five-year grant or reports foundational tools instead. Speculative A completed 57-codon E. coli is plausible but is running at 45.8% after ten years. Speculative An interlaboratory reproducibility standard is cheap, overdue and unclaimed by anyone. Frontier Expect the plant platforms to be decided by manufacturing cost rather than by approvals in this window: a second sprayable RNA registration against a non-beetle pest, or a plant-made biologic that holds a market for five years, would each be more informative than any new construct.
• 25 yr: Speculative Designer genomes for several microbes — the claim the earlier programme page made for this horizon — is optimistic on the record: one designer genome has consumed a decade and is under half built. Speculative A bottom-up protocell integrating growth, heredity and metabolism in one vesicle is the field’s stated goal and remains unintegrated. Frontier Predictive models cutting build-test cycles by an order of magnitude is the outcome that would decide between the strong and weak theses, and it is the thing to watch rather than any construction milestone.
• 50 yr: Speculative A de novo free-living cell with no living ancestor becomes conceivable if the bottom-up subsystems integrate; nothing in the current record forecasts when. Speculative Recoded chassis in routine industrial production is a commercial question rather than a scientific one, and the commercial record in section 2 is the reason to be cautious about it. Handwave “Synthetic genomics as a mature engineering discipline” at this horizon is an assertion about a transition nobody has demonstrated the first step of.
• 100 / 250+ yr: Handwave Designing genomes from principle rather than by copying nature is the endpoint the field names and the one with no evidence at all: every synthetic genome ever built is a modified copy. Handwave Beyond useful forecasting, and the honest statement is that a 250-year projection about a field whose 20-year-old founding paper still describes its limitation correctly is a statement about ambition rather than trajectory.

• Depends on Genetic Engineering, which owns genome writing and editing as a general capability. Organism-scale synthetic biology consumes that capability and does not produce it: every construct in section 2, from syn1.0 to the 57-codon project to SynHG, is downstream of it. That is the one typed depends-on edge claimed here, and it follows the convention nine other Category III briefs already record.
• Requires (not on this map) Two constraints here are not research results and no brief on this map produces them. The first is industrial: the cost series everyone quotes — $0.07 per base pair in 2025 against $25.00 per base in 1999, a 357-fold fall — prices raw oligonucleotides, not assembled and error-corrected constructs, and no published series tracks the assembled figure. Genome-scale work is gated by hierarchical assembly with verification at every step, and that is a manufacturing capacity a supplier could choose to build and price. The second is institutional and was supplied and then withdrawn inside twelve months: the 2024 Framework for Nucleic Acid Synthesis Screening was rescinded by Executive Order 14292 in 2025, in the same year that Microsoft’s Paraphrase Project demonstrated in Science that generative protein design can redesign toxin sequences past the screens, and that Gillum and Moritz documented university biosafety offices with fewer than three full-time staff, providers with substantially varying screening sensitivity, and benchtop synthesisers outside institutional oversight altogether. A screening regime with an enforcement mandate is something a legislature could pass; it is not something a laboratory can discover. A third constraint arrives with the plant platforms and is industrial in the same way as the first: manufacture priced at the scale the application actually consumes. Sprayable RNA is dosed by the hectare and a plant-made biologic by the dose, and in both cases the published cost sits in downstream processing — enzymatic dsRNA synthesis and formulation for one, purification out of green tissue for the other. Both are plant capacity a manufacturer can choose to build and price, and neither is a result a laboratory can publish its way past.
• Enables Contained bio-manufacturing on recoded chassis; engineered live therapeutics, on a worse footing after Synlogic than before; whole-cell modelling as a design tool rather than a description; and the payload question in Directed Panspermia, where the most developed seeding proposal carries a gene synthesiser rather than organisms. No typed enabling edge is claimed, because in each case what is enabled is a capability this brief has just shown is not delivered.
• Adjacent Xenobiology, Designer Organisms, Biological Computing, Microbiome Engineering, Synthetic Ecosystems; and outside the map, industrial fermentation, metabolic engineering, protein biophysics and biosecurity policy.

Handwave “Synthetic biology has created life.” It has not. syn1.0 and syn3.0 were chemically synthesised genomes installed into pre-existing recipient cells, which supplied the ribosomes, membranes, polymerases and cytoplasm that made the synthetic instructions mean anything. Established No source in this brief, or anywhere this research pass could find, describes the synthesis of a cell from non-living components. The distinction is not a technicality; it is the whole remaining distance to the field’s stated goal.

Handwave “DNA synthesis costs fell 357-fold, so design capability is on the same curve.” The cost fall is real and it is evidence about fabrication only. Established Over the same period the flagship design project produced 22% lethal segments and stalled at 45.8% completion, and the best-characterised protein design pipeline hits 11.6% on first pass. A price curve for writing is not a capability curve for designing, and the slogan depends on treating them as one.

Frontier “Standardised biological parts work like electronic components.” No source shows context-independent part behaviour, and the evidence points the other way. The BioBrick standard was succeeded by replacement standards because its junction chemistry constrained fusions and its parts arrived without characterisation data. Speculative The reproducibility analysts’ sharper point is that the field cannot currently tell an unreliable part from an unreliable measurement, because the interlaboratory studies that would separate them have not been run.

Established “A fully synthetic yeast cell exists.” All sixteen Sc2.0 chromosomes have been synthesised and published, and a designed tRNA neochromosome with them. No fetched source states that a single strain carrying all sixteen exists, and the Macquarie release announcing the final chromosome in January 2025 does not say what remains. The December 2025 Nature Biotechnology comment from the consortium is a defect catalogue, not a completion notice. Consolidation is undated.

Established “The 57-codon organism exists, or is about to.” It does not. Syn61 is the 61-codon organism and it is real; the 57-codon successor is a different and much harder project, and as of June 2026 its best published result is six partially recoded strains at up to 45.8% synthetic genome. Frontier Descriptions that put Syn61 and Syn57 on one trajectory invite a reader to assume a completion that has not happened.

Established “Nineteen genes were added back to restore syn3.0’s division.” Seven. Five of them of completely unknown function. The nineteen-gene figure appears on this Institute’s earlier origins page and is corrected here; it appears to conflate the division-restoration work with a different reduction step, and it has been carried unsourced.

Frontier “AI has made protein design reliable.” Across 3,766 experimentally characterised de novo binders on 15 targets, 11.6% bound, with per-target success from 0.1 to 1.0 and no standard criteria for prioritising which designs to test. Speculative That is a real and useful capability — a one-in-nine hit rate on a task that used to be hopeless is transformative for screening — and it is not reliability.

Handwave “AI has produced working bioweapon designs.” The Paraphrase Project redesigned toxin sequences that evaded biosecurity screening in silico. No protein was made. Whether the evading designs would have retained toxicity is unknown and was deliberately not tested. Established The vulnerability in the screening layer is established and was patched incompletely; the weapons capability is not established and coverage that implies it is over-reading a digital result.

Frontier “Sequence screening is a durable chokepoint.” It was demonstrated evadable, its governing US framework was rescinded, provider sensitivity varies substantially, and benchtop synthesisers bypass institutional oversight entirely. Established The one piece of evidence on the other side is that no confirmed malicious incident involving synthetic nucleic acids has been documented. Both readings of that fact — that the risk is negligible, and that the regime is working — overreach, and the record supports neither.

Handwave And the framing itself: “biology is becoming an engineering discipline” elides two meanings of engineering. If it means high-throughput search plus systematic debugging, it is true, well-evidenced and much less impressive than it sounds. If it means specify-design-build-works, the measured first-pass success rates are 22% lethal segments and 11.6% binders, and the field’s flagship project recovered by abandoning design for selection. The slogan is doing its work at the point where the two meanings are swapped.