Industries / Frontier science / Synthetic biology
Engineering strains
This work alters the genes of a microbe, or rewrites its genome outright, so that it makes a chosen substance, and then tests many versions to find one that holds up in a tank.
7
Milestones achieved, from 1991 to 2025.
2
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1991
Metabolic engineering gets a name
James Bailey set out in Science how recombinant DNA methods, which move genes between organisms, could be used to redirect a cell's chemical pathways to raise output or to make products the cell never made before. He also wrote that cells respond to genetic changes in complex ways, which complicates predictive design.
Lessonthe field's founding paper said in its own abstract that the tools for rational design existed but that cell behaviour would not follow the plan.
Credit
James E. Bailey, in Science
2006
A malaria precursor in yeast
A team rebuilt the mevalonate pathway in brewer's yeast and added three genes from sweet wormwood, including a cytochrome P450 enzyme that carries out a three-step oxidation. The yeast made up to 100 mg per litre of artemisinic acid, the immediate precursor of the antimalarial drug artemisinin, and pushed it outside the cell, where it could be collected.
Credit
Dae-Kyun Ro, Jay Keasling and colleagues, in Nature
2009
Many edits in one cycle
Multiplex automated genome engineering, or MAGE, changes many places on a chromosome at once using a pool of synthetic DNA, then repeats the cycle. Its authors modified 24 genetic components of one pathway in Escherichia coli at the same time, creating over 4.3 billion combinations, and selected strains that made more of the red pigment lycopene. They also built prototype machines to run the cycles.
Credit
Harris Wang, Farren Isaacs, George Church and colleagues, in Nature
2013
Twenty-five grams, then a rethink
Further engineering of the same yeast raised artemisinic acid to 25 grams per litre in fermenters, and commercial production of semi-synthetic artemisinin began that year. The chemical step after the fermentation did not hold up: a later review by the chemists involved says the scaled route reached its performance limits and would not be cost-effective, so Sanofi went back to a photochemical method, which now makes up to 60 tonnes a year.
Lessonthe fermentation was not the deciding step, and the chemistry on either side of it decided whether the product could compete.
Credit
Chris Paddon and colleagues at Amyris; Sanofi for the manufacturing route
2016
The smallest working genome
Starting from JCVI-syn1.0, a bacterium whose 1,079,000-letter genome had been chemically synthesised, a team cut the genome down to 531,000 letters and 473 genes, smaller than that of any self-replicating cell found in nature. Their first design, based on what was then known about which genes are essential, produced no living cell.
Lessonthe missing class was not essential genes but quasi-essential ones, needed to grow well rather than to survive, and the design only worked when those were kept.
Credit
Clyde Hutchison, Daniel Gibson, John Glass, Craig Venter and colleagues at the J. Craig Venter Institute, in Science
2023
Half a yeast genome rewritten
The Sc2.0 project, which is building a synthetic yeast genome, reported that all sixteen individual synthetic chromosomes had been assembled, and combined them step by step into single strains, reaching one strain with 6.5 synthetic chromosomes and then, by substituting the largest chromosome, over half the genome.
Lessonsome faults appeared only in combination, so the team built a mapping method, CRISPR D-BUGS, to trace one defect to synthetic chromosome II and then found a defect that showed up only when synthetic chromosomes III and X were in the same cell.
Credit
Yu Zhao, Jef Boeke and the Sc2.0 consortium, in Cell
2025
A 57-codon genetic code
The genetic code uses 64 three-letter codons. A group designed and built an E. coli with a four-million-letter synthetic genome in which every known instance of six sense codons and one stop codon was swapped for a synonym. The result, named Syn57, uses 55 codons to encode the 20 amino acids.
Lessonan independent group then measured the cost of such rewrites across six partly recoded strains, found widespread fitness defects and codons left unassigned in both Syn61 and Syn57, and had to revise the design and evolve the strains to recover growth.
Credit
Wesley Robertson, Jason Chin and colleagues, in Science
Watch for
A plug-in that lifts titres
Vivici, which ferments the whey protein beta-lactoglobulin, says that adding Enduro Genetics' Enduro Sense to its production strain raised titres and yields by 30% with no change to the fermentation equipment or the amount of feedstock. Enduro Genetics says the plug-in links a cell's survival to how much product it makes, so that high-producing cells come to dominate the tank.
Achieved whenan outside party publishes the before-and-after figures from the same process, or a second customer reports a gain of the same size.
Claimed by
Vivici and Enduro Genetics, April 2026
A hundred grams of emulsifier
Ruby Bio says its yeast-based process passed titres of 100 grams per litre for food emulsifiers, which it says brings fermented emulsifiers close to the price of synthetic ones. Its chief executive, Charlie Silver, said a fermentation titre "is a milestone, not a finish line" and that the test is repeating it at scale.
Achieved whenthe run data is published, or a customer or independent laboratory confirms the titre at production scale.
Claimed by
Ruby Bio, June 2026
Open
01
A table of synthetic-genome bugs
One public table of every published whole synthetic chromosome and recoded genome, with each design feature set beside the defect it caused. One row per defect, giving the project, the organism, the kind of feature (a codon swap, a removed transfer RNA gene, an inserted recombination site, a deleted region), the measured effect and the conditions it appeared under, how it was found and how it was repaired, each with a citation to a figure or table.
The briefCredit · unclaimed02
Score agents on strain design
A public test set that asks an AI agent to propose the next change to a strain, built backwards from published work. Each task gives the host organism, the target molecule, the starting titre and the tools the agent may call, and holds back what the authors did. The agent's proposal is scored against the published change and the gain that was measured.
The briefCredit · unclaimed03
The minimal cell's unknown genes
A public table of the genes in the minimal bacterium JCVI-syn3.0 whose job was unknown, with every later claim about each one: what function was proposed, whether it was predicted by software or measured in an experiment, the evidence given, and the citation. It is refreshed as new papers appear.
The briefCredit · unclaimed
Open 01 · the brief
A table of synthetic-genome bugs
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
One public table of every published whole synthetic chromosome and recoded genome, with each design feature set beside the defect it caused. One row per defect, giving the project, the organism, the kind of feature (a codon swap, a removed transfer RNA gene, an inserted recombination site, a deleted region), the measured effect and the conditions it appeared under, how it was found and how it was repaired, each with a citation to a figure or table.
Why it is still open
The reports are project by project and in each project's own terms. The Sc2.0 team had to build its own mapping method to trace a defect in synthetic chromosome II, and then found one that appeared only when two synthetic chromosomes were combined. A separate group reported widespread defects across six partly recoded E. coli strains, including codons left unassigned in Syn61 and Syn57. Neither paper gives a list that spans projects. The consortium has the pieces for a list: its chromosome designs were version-controlled as they were edited, and the consolidation paper named a method for mapping a defect to the design feature that caused it. Nobody has used them to publish one table across projects.
Built with what exists
The published papers and their supplementary tables; a long-context AI model to pull candidate rows from supplements, with a person checking each against the source; Python and pandas; a public Git repository for the schema and code; Zenodo to archive each version.
First step
Code every defect named in the 2023 Sc2.0 consolidation paper and the 2026 recoding paper into a 30-row draft, and publish the column definitions for comment.
Done looks like
A released CSV covering at least 20 published synthetic chromosomes or genomes and at least 150 defect rows, with written coding rules, every row carrying a paper and a figure or table number, and two people independently coding 50 random rows agreeing on at least 90% of fields.
Open 02 · the brief
Score agents on strain design
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test set that asks an AI agent to propose the next change to a strain, built backwards from published work. Each task gives the host organism, the target molecule, the starting titre and the tools the agent may call, and holds back what the authors did. The agent's proposal is scored against the published change and the gain that was measured.
Why it is still open
The nearest benchmark, BioDesignBench, sets 76 expert-curated tasks across five molecular families and reports that agents call evaluation tools at about 14% of the intensity an expert uses. It covers protein design, not the choice of which gene to add, delete or tune in a production strain.
Built with what exists
Metabolic engineering papers published after the data cutoff of the newest model tested; BioDesignBench's contamination checks, which paraphrase task prompts and measure word overlap with the source paper, as the pattern to copy; the open COBRApy toolkit and published genome-scale metabolic models as the tools an agent may call; Python; open model weights or paid model access. Agents that can call tools over many steps are the advance that makes this testable.
First step
Write ten tasks from 2026 metabolic engineering papers, with the authors' change and the measured gain held back, and check that two people grade the same ten answers the same way.
Done looks like
At least 100 tasks drawn from papers published after that cutoff, a written scoring rule on which two graders agree for at least 90% of a 30-task sample, at least four models scored and ranked, and the tasks, grading rules and code released under an open licence.
Open 03 · the brief
The minimal cell's unknown genes
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public table of the genes in the minimal bacterium JCVI-syn3.0 whose job was unknown, with every later claim about each one: what function was proposed, whether it was predicted by software or measured in an experiment, the evidence given, and the citation. It is refreshed as new papers appear.
Why it is still open
The 2016 paper reported 473 genes, of which 149 had no known biological function. A 2021 study predicted functions and interaction partners for many of the poorly characterised proteins of syn3.0 and its derivative syn3A from predicted protein structures, but those are predictions, not measurements. No public record tracks which of the genes now have experimental support.
Built with what exists
The gene lists in the 2016 paper's tables; Europe PMC's citation data to find papers that cite either paper; Python; an AI model to draft one-line summaries of each claim, with a person checking each against the cited paper before it is published.
First step
Enter the 149 gene names from the 2016 paper's supplementary tables and attach the 2021 predictions to the genes they cover.
Done looks like
A published table with one row per gene of unknown function, each row carrying its dated claims, each claim tagged predicted or measured with its paper and method, the page stating how many genes now have experimental support, and two refreshes at least three months apart.