Shotgun Genetic Engineering Builds Amino Acid Pathways Into Mammalian Cells

A team at NYU Langone Health has described a method that sidesteps one of the oldest bottlenecks in mammalian metabolic engineering: instead of designing a pathway, building it and testing it, they deliver a library of individual gene cassettes at high enough copy number that every cell in a population assembles a different random pathway, then let growth selection identify the combinations that work. The approach, which the authors call shotgun genetic engineering (SGE), was published in Nature Biotechnology in October 2026, with an accompanying News & Views commentary. A preprint version has been available on bioRxiv since July 2025, and the quantitative detail below comes from that version.

Using it, the group engineered Chinese hamster ovary (CHO) cells and Jurkat T cells to grow without valine in the medium, and CHO cells to grow without isoleucine — which the authors report as the first isoleucine prototrophy achieved in mammalian cells.

Why mammalian cells cannot make their own amino acids

Mammals lost the biosynthetic routes to nine amino acids somewhere in evolutionary history and have obtained them from diet ever since. For a whole animal this is unremarkable. For cultured cells it is a persistent operational cost: every essential amino acid has to be supplied in the medium, which makes formulations complex, expensive and a significant share of the cost of goods in anything grown at scale. That applies to therapeutic protein manufacturing in CHO cells, to viral vector production, and to cultivated meat, where media cost dominates the economics.

Restoring those pathways is an obvious target and a hard one. The enzymes exist in bacteria, but transplanting a multi-step pathway into a mammalian cell means getting several genes expressed at workable levels, in the right subcellular compartment, with the right relative stoichiometry. The number of plausible configurations is combinatorially large, and mammalian cells are poor hosts for iterating through them: doubling times are measured in days rather than hours, and delivering long synthetic DNA constructs is itself difficult. A previous rationally designed valine-prototrophic CHO line, pMTIV, did work — but at a doubling time of 3.77 days, which is far too slow to be useful.

How the method works

SGE inverts the usual order of operations. Rather than assembling candidate pathways in advance, the researchers built a library of barcoded individual transcription units — each one a single coding sequence with its own promoter and localisation tag — and delivered them by lentivirus at high multiplicity of infection, with random integration. Because each cell takes up a different number and assortment of cassettes, the population becomes a library of pathway combinations rather than a library of single genes.

Selection is then purely phenotypic: the cells are grown in medium lacking the amino acid of interest, and only those that happen to have assembled a functional pathway survive. The surviving clones are sequenced by amplicon-seq to read their barcodes, which reveals which cassettes each one carries. The design question is answered by reading out the winners rather than by predicting them.

The library itself was deliberately small: ten coding sequences — four valine-pathway genes, three variants of ilvA, plus ilvG, ilvM and a GFP control — in four promoter and localisation combinations, for 40 transcription units in total. The enzymes are bacterial, with the reference list pointing to Escherichia coli ilv genes as the source. Scale came from the delivery rather than the library: roughly 2.42 million CHO cells were screened at a multiplicity of infection of 8.8, and about 3 million Jurkat cells at 3.4. Simulation suggested that corresponded to on the order of 1.56 million distinct pathway combinations in the CHO screen.

What the screens produced

• Valine in CHO cells. Two clones grew in valine-free medium with doubling times of about 1.1 and 1.05 days, against 3.77 days for the earlier rationally designed pMTIV line — roughly a threefold improvement, and close to a workable growth rate.
• Isoleucine in CHO cells. One clone grew in isoleucine-free medium at 1.69 days per doubling, carrying a six-gene solution (ilvM, ilvG, ilvB, ilvC, ilvD and the ilvA-L481F variant). The authors report this as the first isoleucine prototrophy in mammalian cells.
• Valine in Jurkat cells. Clones grew at about 1.9 to 2.0 days per doubling in valine-free medium, establishing that the approach is not confined to one cell background — CHO cells are adherent, Jurkat cells grow in suspension.
• Construct size. Functional clones carried roughly 23 to 52 kilobases of integrated synthetic DNA, a quantity that would be laborious to assemble and deliver as a single designed construct.

The more interesting result is not any single clone but a pattern across them. Successful solutions strongly favoured mitochondrial targeting of the pathway enzymes, enriched by roughly 48-fold to 283-fold over chance depending on the pathway and cell line. The authors describe this as unexpected, given that the enzymes are of bacterial origin, and offer several possible explanations: mitochondria's own bacterial ancestry, higher local pyruvate concentrations, better cofactor availability, or simply higher effective enzyme concentration in a smaller compartment. The data do not distinguish between these.

That finding is the clearest argument for the method. Compartment preference of this strength would be difficult to arrive at by reasoning alone, and a rational design programme committed to cytoplasmic expression could have spent a long time failing for a reason it never tested. The team also trained a random forest classifier on the barcode data, which reached a mean accuracy of 0.942 and an ROC AUC of 0.993, correctly classifying 112 of 119 Jurkat clones; its most informative features recapitulated the mitochondrial pathway that manual inspection had identified.

Limitations

The authors are direct about the constraints, and they are substantial.

The method requires a selectable phenotype. Growth in dropout medium is close to an ideal selection, but most metabolic engineering goals — higher titre, better glycosylation, altered flux — do not come with a survival readout, and without one the central mechanism of SGE does not apply.

Lentiviral packaging introduced representation bias, with EF1a-driven constructs underrepresented relative to PGK. A library that is not uniformly delivered is not uniformly screened, so some combinations were effectively never tested.

Amplicon-seq reports which cassettes integrated, not which are expressed. Silencing of an integrated unit would make the barcode read-out an imperfect description of the functioning pathway, and the authors suggest pairing the method with RNA-seq to close that gap.

And the winning genotypes are not finished products. Randomly integrated multi-cassette solutions spread over tens of kilobases would generally need rebuilding into compact, defined constructs for any production use, and the authors note that behaviour may change when they are. A clone that grows well is a lead, not a cell line.

One further caveat is worth drawing out, as context rather than criticism. Valine and isoleucine were well chosen: in bacteria their pathways share most of their enzymes, so a single small library could address both, and the isoleucine solution reuses the valine-pathway genes almost entirely. That shared architecture is not a general feature of the remaining essential amino acids, several of which require longer and more independent routes. Two of nine, from the most tractable corner of the problem, is a genuine result but not a near-complete one.

What it could be used for

The authors set out several directions. Reducing the number of nitrogen sources a cell needs would make stable isotope labelling experiments far more interpretable, which is a basic research application rather than an industrial one but potentially the most immediately useful. Simpler media would lower costs in cultivated meat and viral vector production. And metabolically augmented cell therapies are raised as a longer-term possibility: T cells that can synthesise their own amino acids might function better in the nutrient-depleted tumour microenvironment, where amino acid scarcity is one of the mechanisms by which tumours suppress infiltrating immune cells. That the valine work was done in Jurkat cells, a T-cell line, makes this more than rhetorical, though a leukaemia line is a long way from a primary therapeutic T cell.

For the method itself, the proposed next steps include transposon-based delivery in place of lentivirus, combining SGE with compact-construct assembly methods, raising the multiplicity of infection to screen larger libraries, and adding CRISPR activation or interference machinery to tune expression of host genes alongside the introduced ones. Prototrophy for additional essential amino acids, growth factors and vitamins, and for stress tolerance, are listed as targets.

What to watch

• Whether the solutions survive refactoring. The practical test is whether a 23 to 52 kb randomly integrated genotype can be rebuilt as a defined construct in a production CHO line and still support growth at a useful rate.
• The harder amino acids. Extending beyond the branched-chain pathways, where enzyme sharing made a small library unusually powerful, will show whether the method generalises or whether it found the accessible case.
• Selection-free variants. Coupling SGE to fluorescent or sequencing-based readouts rather than survival would open it to the many engineering goals that do not kill cells when they fail.
• Growth rates against industrial benchmarks. Doubling times near one day are a large improvement on 3.77 days, but CHO lines in production typically grow faster still, and media savings have to outweigh any remaining growth penalty.

This article is for information only and is not medical advice. Treatment decisions should be made with a qualified clinician.

Sources: Nature Biotechnology; Nature Biotechnology News & Views; bioRxiv preprint