The core problem SGE addresses is combinatorial explosion. A synthetic metabolic pathway is not simply a collection of genes; it is a carefully tuned arrangement of coding sequences, promoters, expression levels and subcellular addresses. Each design variable multiplies the number of possible configurations, and in mammalian cells the search space quickly becomes unmanageable. Microbial engineers sidestep this with rapid design-build-test cycles: Escherichia coli doubles every twenty minutes and budding yeast every ninety, so iterative refinement is cheap. Mammalian cells, with doubling times exceeding twenty hours, make such iteration painfully slow. Worse, delivering large DNA constructs into mammalian cells is exponentially less efficient as construct size grows, so brute-force screening of complete multigene pathways has been impractical beyond roughly ten kilobases.

SGE flips the logic of construct design. Instead of building and delivering one large pathway at a time, the researchers assembled a pooled library of small, individually barcoded transcription units, each consisting of a coding sequence paired with a promoter and, in some cases, an organellar localization signal. Because small DNA fragments are vastly easier to synthesize, assemble and package into lentivirus, the library could be delivered at high multiplicity of infection, so each cell received a random assortment of components. Every infected cell thereby became a unique pathway variant, exploring different gene content, stoichiometry and localization simultaneously. After selection for a desired phenotype, sequencing the barcodes from surviving cells reveals which combinations of parts conferred the function.

As a test case, the team targeted branched-chain amino acid biosynthesis. Mammals lost the ability to synthesize nine essential amino acids deep in evolutionary history and must obtain them from food, making amino acid auxotrophy a stringent and convenient selection scheme: cells simply die unless they acquire the pathway. In previous work, the group had shown that four E. coli genes, ilvN, ilvB, ilvC and ilvD, could confer valine prototrophy on CHO cells, but growth remained slow at 3.8 days per doubling, and the same genes failed to enable isoleucine biosynthesis despite overlapping enzymatic steps. For the SGE library, they expanded the toolkit to ten coding sequences, adding ilvA to supply the isoleucine-specific substrate 2-oxobutanoate, two feedback-resistant ilvA variants, and the isozymes ilvG and ilvM, which form an acetohydroxy acid synthase II complex with a strong preference for the isoleucine substrate and insensitivity to end-product inhibition.

Each coding sequence was paired with either the strong EF1a or the medium-strength PGK promoter, with or without a mitochondrial targeting sequence, yielding a forty-member transcription unit library. The researchers infected 2.42 million CHO cells at a multiplicity of infection of 8.8 and used Monte Carlo simulations to confirm that the resulting diversity was sufficient to capture all plausible four- and five-component pathway combinations. After fifteen days of selection in low-valine medium, clonal outgrowths appeared only in the experimental population. All sixteen isolated clones proved valine prototrophic, and the best performers doubled in 1.1 days in valine-free medium, roughly 83 percent of their growth rate in complete medium and a dramatic improvement over the rationally designed pathway of the earlier study.

Barcode sequencing revealed a consistent architectural principle. In strongly prototrophic clones, the winning configuration placed mitochondrially targeted ilvN, ilvB, ilvC and ilvD under PGK promoter control, a solution rational design had never identified. Mitochondrial colocalization of all four genes was enriched forty-eight-fold relative to random expectation, and no clone showed all four enzymes in the cytoplasm. The researchers reconstructed this SGE-informed pathway on a single DNA construct using yeast homologous recombination and Flp-In integration, and it drove valine-free growth at 1.25 days per doubling, substantially outperforming the original design. Isotope tracing with carbon-13-labeled glucose and pyruvate confirmed that the engineered cells were genuinely synthesizing valine endogenously at rates comparable to the native mammalian synthesis of the nonessential amino acid alanine.

The isoleucine results were even more striking. Rational design had failed to produce isoleucine-prototrophic CHO cells, but SGE succeeded. Among forty-six clones selected on low-isoleucine medium, nineteen showed improved prototrophy, and the strongest clone grew in isoleucine-free medium at 1.69 days per doubling. Every strongly prototrophic clone carried barcodes for six genes: ilvB, ilvC, ilvD, ilvG, ilvM and the partially feedback-resistant ilvA(L481F). Again, mitochondrial localization dominated, with the six-gene mitochondrial configuration enriched 217-fold over expectation and the fully cytoplasmic alternative never observed. Mass spectrometry detected the expected carbon-13 isotopologs of isoleucine, confirming endogenous synthesis of a metabolic function last present in the mammalian lineage over 500 million years ago. The integrated constructs spanned roughly 23 to 52 kilobases of synthetic DNA, far beyond what conventional screening approaches can deliver and test.

To demonstrate generality, the team deployed the same lentiviral library in Jurkat cells, a human T lymphocyte line that grows in suspension. Three million infected cells at a lower multiplicity of 3.4 yielded valine-prototrophic clones after selection, with 71 percent of 119 isolated clones passing the prototrophy threshold and the best performers doubling in under about two days without valine. The pathway solution mirrored the CHO result: mitochondrially localized ilvN, ilvB, ilvC and ilvD under PGK control, enriched 283-fold over random expectation. Isoleucine prototrophy, predicted by simulation to be vanishingly rare at the achieved library coverage, did not emerge, illustrating how Monte Carlo modeling of transcription unit distributions and multiplicity can guide experimental design and set realistic expectations before a screen begins.

The datasets generated by SGE also proved rich enough for machine learning. The researchers trained a random forest classifier on the barcode-derived transcription unit composition of each Jurkat clone, using it to predict whether a clone would be prototrophic or auxotrophic. With stratified fourfold cross-validation, the model achieved a mean accuracy of 0.942 and an area under the ROC curve of 0.993, and its top predictive features were precisely the four mitochondrially localized, PGK-driven genes that manual analysis had identified as the optimal pathway. This demonstration points toward a future in which large-scale synthetic perturbation data, rather than observational datasets shaped by natural evolution, train predictive models capable of generative metabolic design in mammalian systems.

The authors acknowledge limitations, including the current reliance on selectable phenotypes, biases introduced during lentiviral packaging that underrepresented certain promoters and coding sequences, and the fact that barcode sequencing reveals integration but not expression, which RNA sequencing must supplement. Even so, the implications are broad. Cells engineered for amino acid prototrophy could reduce the cost and complexity of media for biomanufacturing, cultivated meat and viral vector production, while immune cells metabolically augmented to withstand the nutrient-poor, acidic and inflammatory tumor microenvironment could improve cell therapies for solid cancers. The team anticipates extending SGE to larger libraries, additional essential nutrients, vitamins and growth factors, stress tolerance traits, and integration with CRISPR-based host factor modulation, establishing a scalable framework for pathway-scale engineering of increasingly complex biosynthetic traits in mammalian cells.

Subject of Research: Highly multiplexed mammalian metabolic engineering using barcoded shotgun genetic engineering to screen millions of synthetic amino acid biosynthesis pathways

Article Title: Highly multiplexed mammalian metabolic engineering with a shotgun approach

Article References: Trolle, J., Sessa, S., Wudzinska, A., Grivainis, M., Rincones, K., Rodrick, T., Jones, D. R., Fenyö, D., Pinglay, S., & Boeke, J. D. (2026). Highly multiplexed mammalian metabolic engineering with a shotgun approach. Nature Biotechnology. https://doi.org/10.1038/s41587-026-03318-7

Image Credits: AI Generated

DOI: 10.1038/s41587-026-03318-7

Keywords: shotgun genetic engineering, metabolic engineering, synthetic biology, mammalian cells, CHO cells, Jurkat cells, amino acid biosynthesis, valine, isoleucine, mitochondrial localization, machine learning, high-throughput screening