Machine learning strain design, explained: what ML4SD does, and what has actually been proven

Research 41 primary sources read, 35 cited: the ML4SD preprint in full, the algorithm papers it builds on, and every wet-lab campaign we could find 4 Nov 2026

0 strains

In September 2026 a group at the Spanish National Centre for Biotechnology (CNB-CSIC, Madrid) posted a preprint describing ML4SD, a machine-learning loop for designing production strains. It reports needing 2.5 to 7.1 times fewer designs than the same search without machine learning ^S01. We covered it as a news note at the time. Several readers asked the obvious question, which is what the number actually means, so we read the full 42-page preprint and 40 other primary sources to answer it properly. This is a satellite of our pillar on AI in bioprocessing: what actually runs in a plant, and it is written for someone who runs a process, not for a computational biologist. Every term is explained the first time it appears.

The five findings in short

• The saving is real and the unit is a simulation. ML4SD reached its best carbon yield after searching 4,316 designs, against 10,869 to 30,615 for the plain search, which is where 2.5 to 7.1 times fewer comes from ^S01. Every one of those designs is a calculation on a model. No strain was built, no fermentation was run, and no titre, meaning grams of product per litre of broth, was measured ^S01. The work is a preprint and has not been through peer review. Confidence: high. We read the full text.
• The interesting result is not the headline, it is that you have to keep the failures. The loop only works if the starting library of designs is large and varied and includes the bad and the unbuildable ones. Libraries made only of the best designs produced models that looked excellent on their own test data and then rated non-viable designs as high performers ^S01. Confidence: high.
• There is a real distinction between two kinds of machine learning here, and the authors draw it themselves. ML4SD trains on the output of a metabolic model, not on laboratory measurements ^S01. The active-learning loops that have actually raised titres tested their recommendations in living cells in every cycle ^S01. Confidence: high.
• With real laboratory data the approach has a track record, and the gains spread wide. Published wet-lab campaigns run from 21% more product ^S31 through five times more ^S22 to 46 times for the best designs of one biofoundry round ^S24, and one industrial report is a win on one product and a loss on the other ^S37. The comparator decides the headline: the same yeast campaign is 74% or 106% better depending on what it is measured against ^{S21 S20}. Confidence: high.
• A model-computed design failing in a living cell is not hypothetical. In the same organism and the same class of feedstock as the ML4SD case study, a complete model-computed set of deletions was built and the strain would not grow on the target substrate, so it made nothing. Three months of serial passaging did not rescue it ^S14. Confidence: high.

How we built this

We read 41 primary sources, 31 of them in full and 10 as abstracts only, and cite 35. The ML4SD preprint itself was read in full, including Materials and Methods. Every number in this article is quoted from the source it came from, and for each source we say who is speaking: the group that wrote the algorithm, an academic laboratory, a public biofoundry, or a company reporting on its own process. Where we did arithmetic on published figures we label it as our own calculation. Our verifier checked every sentence with a number against the stored source text. What we could not read is listed at the end.

What a growth-coupled strain is, in plain terms

A production strain is a microbe that has been modified to make something you want. The modification is often a knockout, which means switching a gene off so its reaction no longer runs, closing a route the cell would otherwise use.

Growth-coupled design is a particular trick: wire the cell so that it cannot grow without making the product ^S01. The point is not chemistry, it is evolution. If a mutant that stops producing no longer has any advantage, production survives repeated passaging and continuous culture instead of drifting away ^S01. That is an argument the field makes, not a measured result. There are two strengths of it: weak coupling, where the cell only has to produce when it is growing as fast as it can, and strong or obligatory coupling, where it has to produce at any growth rate ^S01.

Finding which genes to switch off is a search problem, and the map it searches is a genome-scale metabolic model, usually shortened to GEM: a list of every chemical reaction a cell is known to be able to run, with the genes responsible for each one. The reference model of Escherichia coli covers 1,515 genes, 2,719 reactions and 1,192 metabolites ^S04. The model of Pseudomonas putida that ML4SD runs on covers 1,462 gene products, 2,929 reactions and 2,155 metabolites, and those 1,462 genes are only 38% of the proteins annotated in the genome ^S05.

On top of that map runs flux balance analysis (FBA), which calculates how material could flow through the network under a stated assumption about what the cell is trying to do. FBA needs no kinetic parameters, which is exactly why it can be run across a whole genome quickly ^S03. The classical design algorithms sit on top of FBA: OptKnock, published in 2003, searches for knockouts that make growth impossible without product formation, and its authors already noted that this gives you a growth selection for the overproducer ^S02. RobustKnock followed because the earlier method ignored competing pathways and was therefore over-optimistic about what the strain would really make ^S17.

All of this sits inside a loop the field calls design-build-test-learn, or DBTL: pick the modifications, build them into a host, measure what comes out, learn, and go round again ^S01.

Without machine learning: a huge search and a very small number of strains

The search space is the first problem. A ten-knockout search over 150 candidate reactions is more than a quadrillion combinations, and one 2010 study had to cap each run at a week of computing ^S07. Even in a small model, five deletions out of 250 reactions is already more than 7.8 billion combinations ^S12. The exact methods are slow by construction, as the ML4SD authors put it plainly about the whole family ^S01.

What is counter-intuitive is that feasibility is almost never the obstacle. For more than 96% of the metabolites that E. coli and yeast can make from glucose, strong growth coupling is feasible with aeration ^S06. The obstacle is the number of knockouts it takes. In that same study the average set of cuts was 20.6 reactions, and eight extra days of computing only brought the average down to 12.9 and the largest from 58 to 34 ^S06. The authors themselves call strategies needing more than 20 reaction knockouts possibly unrealistic, and 4.4% of theirs did ^S06. The OptCouple group states the objection without hedging: almost everything can be coupled on paper, but often only by deleting an infeasible number of genes ^S11.

Then you have to build it. A standard technique for deleting a gene in E. coli, published in 2000, takes about six days per knockout; a newer one managed seven knockouts in seven days ^S06. That is in the easiest host there is.

The money figure usually quoted for this work is 50 to 300 person-years and 10 to 100 million dollars for an industrially competitive production strain ^S22. Read it as the price of a whole programme, not of one strain: it is a figure the authors cite from three earlier sources, and we found no published cost or schedule for building one specific knockout strain. Do not divide one by the other.

The useful contrast is how few designs ever get measured. In the real laboratory, a yeast campaign explored about 3% of its design space, 250 designs out of 7,776 ^S21. A robotic platform faced 13,824 possible expression combinations and measured under 1% of them ^S39. A cell-free study had more than four million possible buffer compositions to choose from ^S23. The gap between what a computer can propose and what a bench can build is the whole reason anyone wants machine learning here.

What machine learning adds, and the distinction that matters most

Machine learning changes the learn step. Instead of reading results by hand, a model maps design choices to outcomes and says what to build next, so fewer constructs are made per round ^S01. Active learning is the stronger version, where the algorithm chooses the next experiment itself ^S23, and it is chosen in this field precisely because it needs less data than ordinary machine learning ^S01. ML4SD picks its next designs by mixing two signals: predicted performance, which exploits what the model already believes, and predicted error, which explores where it is unsure ^S01.

Here is the distinction that this whole article turns on, and it is not our framing. There are two different things both called machine learning for strain design:

• Models trained on laboratory measurements. Real strains were built, grown and measured, and the model learns from those numbers. The yeast campaign below rested on more than 124,000 measured data points from over 500 real designs ^S21.
• Models trained on the output of a metabolic model. No cells involved. In the words of the ML4SD preprint, in the absence of experimental data it uses the GEM as a generator of synthetic data ^S01.

ML4SD is the second kind. Its test step could in principle be either a simulation on the full model or a real high-throughput measurement ^S01, and in this paper it is the simulation. The authors say why they went that way: there are very few designs ever tested in living cells per bioprocess, and that scarcity is what blocks ordinary machine learning ^S01. The ceiling is well known in the field: datasets in synthetic biology are typically under 100 instances, too small for deep learning ^S20.

And the authors draw the line themselves. The other active-learning DBTL cycles that did raise titres, they write, recommend promoter combinations, promoters being the switches that set how hard a gene is read, or CRISPRi constructs, which dial a gene down instead of deleting it, from sparse data measured in living cells, and test those recommendations in living cells within each cycle ^S01. Theirs does not.

What ML4SD actually is

ML4SD is an active-learning DBTL cycle whose model ensembles, meaning groups of models whose predictions are pooled, are trained on scores that a genome-scale metabolic model gives to knockout designs, and which samples the next designs from predicted performance and predicted error ^S01.

Its real methodological finding is about the starting library. The cycle generalises only if that library is large and varied and includes suboptimal and non-viable designs; libraries restricted to the minimal or best designs produced models that overfitted ^S01. That failure is specific, not rhetorical. Trained on a library of only the best designs, the ensembles had a median error of 58% of the target range on fresh designs ^S01, and rated designs as high as 0.89 carbon yield, which is the fraction of the carbon fed that ends up in the product, when validation on the full model showed those designs were not viable at all ^S01. In the authors’ summary, the lack of negative examples, the designs that produce nothing, is what caused the overfitting ^S01.

So they wrote a new search algorithm, gcSwarms, to supply those negatives: a particle swarm search that keeps its entire search history rather than only the winner ^S01. Its libraries averaged 59,000 unique qualifying designs, which the authors say is more than any strain design algorithm they know of ^S01, and the same ensembles trained on them dropped from 58% to a median error of 4.4% of the target range ^S01. Note what that comparison is: model against model. Nothing here is measured against a cell.

The case study is 4-hydroxybenzoate, an aromatic compound obtainable from lignin, converted to 6-caprolactam, the monomer of nylon-6 ^S01. The campaign started from a deliberately small library of 1,200 designs, which the authors chose partly because it is what a wet laboratory could afford ^S01. The interpretable output is a motif of three deletions that redirects flux out of the TCA cycle, the central loop the cell burns carbon in, toward the product through acetyl-CoA ^S01. It was recovered from the model and never built.

One resource claim deserves reading twice. ML4SD ran on a laptop with 6 cores in about 74 minutes, against 32 cores of a computing cluster for 44 minutes for the plain search ^S01. That is cheaper hardware, but it is more wall-clock time, not less.

The authors’ own label for the work is a computational proof of concept ^S01, and their own warning is blunt: without measured production, the recommendations stay ranked lists on a digital twin, a computer copy of the cell, rather than designs shown to yield a product in the host ^S01.

The 164 per cent, and the sibling campaign that gained 1.6 per cent

The preprint’s other headline number is a carbon yield reaching 164.4%, and it is worth seeing what the denominator is. It is 164.4% of the best design in the starting library of 1,200 ^S01. Not of the wild type, meaning the unmodified organism, not of the best published result, and not of any measured titre.

The reason to trust the authors on this is that they published the number that undercuts it. The second campaign, on the same bioprocess but scoring a different objective, gained 101.6% of the best starting design, which is to say 1.6% ^S01. Their explanation is the honest one: for that objective the plain search had already found a design at 92.7% of the theoretical maximum, so there was nothing left to win ^S01. For carbon yield it had only reached between 52.5% and 77% of the maximum, and that gap is what the learner exploited ^S01. Their own conclusion scopes the claim accordingly: the method is more advantageous when the search algorithm hands you designs far enough from the global maximum ^S01.

That is a useful thing to know before anyone sells you a loop like this. The size of the improvement is mostly a statement about how bad the baseline was.

What machine learning has actually done with real laboratory data

Outside this preprint there is a genuine track record, and it is worth separating by how it was measured.

In yeast, models trained on real strains found promoter combinations giving up to 74% higher tryptophan titres than the best design they were trained on ^S21, resting on more than 124,000 measured time points from over 500 designs ^S21. In Pseudomonas putida, active learning plus laboratory automation raised isoprenol titres five-fold across six DBTL cycles of two to three weeks, narrowing the design space to 347 arrays out of more than 800,000, and building and measuring 472 unique strains in triplicate ^S22. At the Agile BioFoundry, two design-build-test-learn cycles on 60 E. coli strains, with machine learning trained on the first cycle to pick the second, bought a 21% titre increase in that second cycle ^S31, which is the most sobering published number and probably the most representative. A biofoundry run on pinocembrin had 19 of 25 constructs beat the previous best, ten of them by up to 46 times ^S24. A Bayesian-optimisation robot found a lycopene strain 1.77 times better than the best of a random library it had also built and measured ^S39. In cell-free protein production, which is real measurement but not a growing cell, active learning over a four-million-point space raised yield 34 times against the starting buffer, measuring 1,017 compositions in ten rounds of 102 to get there ^S23.

The closest wet-lab precedent to the ML4SD case study is worth naming, because it shows what the finished article looks like. In the same organism and from the same lignin-derived class of feedstock, a group explored 4,114 growth-coupling solutions and ended with a strain producing 7.3 g/L of indigoidine at 77% of maximum theoretical yield ^S34. That is a number with a unit of mass in it.

Two cautions on reading any of this. First, the comparator decides the headline. The yeast campaign above is 74% better than the best training design ^S21 and 106% better than the base strain ^S20: same experiment, two denominators. Second, some results are less prospective than they look. A frequently cited result, active learning finding a good E. coli valine strain in about eight experiments rather than the 89 originally run ^S19, is largely a replay over data the same group had already generated; they did go back to the bench, and all four newly suggested experiments came out as predicted, with one at 53.4% of maximum theoretical yield against a best measured 54.7% ^S19. That is a real but small prospective validation.

Adjacent work is more mature than strain design itself. Bayesian optimisation of culture media cut the experimental burden up to three-fold against classical design of experiments, and 10 to 30-fold with nine factors in play, reaching its result with 24 real experiments ^S28. An autonomous software agent drove 20 days of a 27-day perfusion cultivation, as reported by the company and university team that built it ^S36. In cell line development, a company ranking model picked a clone with 68.5% higher titre than the legacy choice for one antibody, and one with 6.6% lower titre for the other ^S37. We quote both halves of that one on principle: it is the only industrial head-to-head we found, and it is one win and one loss.

Where the models break

The sharpest counter-example available is in the same organism as the ML4SD case study. A group completed a model-computed set of cuts in P. putida KT2440, intended to couple growth to a product on a lignin-derived substrate. The finished strain grew slowly on rich medium and failed to grow at all on the target substrate, so it produced nothing ^S14. Three months of serial passaging recovered no improved clone, and the experiment was stopped ^S14. Their conclusion names what the model could not see: conditional synthetic lethality and gene dosage requirements among the deletions the design demanded ^S14. Even the partial version of the design produced less than predicted ^S14.

That is not an unlucky one-off, it is the known limit of the method. A metabolic model accounts for mass balance and thermodynamics and cannot account for kinetics, allosteric control, or regulation outside metabolism ^S13. The primer written by the authors of flux balance analysis says it cannot predict concentrations and, without modification, ignores regulation, so its predictions may not always be accurate ^S03. Benchmarked against intracellular fluxes measured with carbon-13 tracers under six conditions, no single FBA objective reproduced wild-type E. coli everywhere, across 99 combinations of objective and constraint tried ^S15.

And the specific model underneath ML4SD gets about one growth call in five wrong: its published accuracy on carbon and nitrogen growth screens is 79% and 84% ^S05. The ML4SD authors concede the general point about their own tooling, naming annotation gaps and promiscuous enzymes as reasons model predictions are sloppy ^S01. On top of that, metabolic burden, the cost to the cell of carrying the modification, is a known cause of designs underperforming, and reviewers of the field call it a black box, with reduced growth, impaired protein synthesis and genetic instability among the symptoms ^S29.

Finally, a model trained on a model can be confident and wrong. The ML4SD ensembles trained on the narrow library could not tell non-viable designs from high performers ^S01, and that failure mode is invisible without a validation step. Neither search algorithm in the paper found a library for every target, and the authors note that some apparent infeasibilities are false negatives from the computing time budget ^S01.

What would have to be published to believe the 2.5 to 7.1 times

• Strains. Nothing was built. The authors list constructing the high-importance knockouts, and especially the minimal design, as the immediate next step ^S01, and plan to fold real measurements in by transfer learning onto the existing ensemble, also as future work ^S01.
• Titres. There is not a single g/L in the preprint, because nothing was fermented. The comparator exists and is published: 7.3 g/L at 77% of maximum theoretical yield in the same organism and the same class of feedstock ^S34.
• Scale. Nothing above simulation, so nothing about what happens in a bioreactor.
• A comparison against the state of the art. The 2.5 to 7.1 times is against the authors’ own new search algorithm, not against OptKnock, gcOpt, minimal cut sets or gcFront on equal terms ^S01.
• Independence. The metabolic model, the base search algorithm and the framework are all from the same group, and the last author of the preprint is also an author of the model ^{S05 S01}. We found no external reproduction.
• Settled parameters. The authors concede that the cycle’s own settings are unsettled and that answering those questions needs extensive proofs ^S01, and that the design landscape is rugged, with the best designs not where the good ones are densest, which is itself a reason to expect surprises in living cells ^S01.
• Peer review. This is version 1 of a bioRxiv preprint, with no journal version indexed at the time of writing.

For context on how young this corner is: the framework the ML4SD authors name as the only other one published for this task reports accuracy gains on model datasets rather than titres ^{S33 S01}, and the non-machine-learning comparator family it is measured against was also validated only on models ^S10. Nobody in this specific niche, machine learning that picks growth-coupled knockouts, has shown a strain yet. Machine learning that picks gene targets in other ways has: one 2026 database-driven model was checked by overproducing geraniol in yeast ^S41.

The AI in the plant series

This piece sits inside our series on artificial intelligence in bioprocessing. The pillar, AI in bioprocessing, explained: what actually runs in a plant, sorts the uses by the evidence behind them and includes a simulator of a model that does not survive scale-up. PAT in practice: sensors, models that fail and where to start covers the probes and the soft sensors that feed any of this. The news note this research grew out of is our brief on the ML4SD preprint.

Questions

What is machine learning strain design? Using a statistical model to decide which genetic modifications to try next, instead of ranking every candidate by hand or by an optimisation algorithm alone. The model maps design choices to outcomes and recommends the next batch, so fewer constructs are built per round ^S01.

What is ML4SD? A framework described in a September 2026 bioRxiv preprint from CNB-CSIC in Madrid: an active-learning design-build-test-learn cycle whose model ensembles are trained on genome-scale metabolic model scores of knockout designs ^S01. It is in silico throughout, has not been peer reviewed, and no strain was built ^S01.

Does machine learning in metabolic engineering reduce the number of experiments? With real laboratory data, yes, in published campaigns, by amounts that range from 21% more product ^S31 to five times more ^S22, and in one case a satisfactory strain from about eight experiments rather than 89 ^S19. How much depends heavily on the comparator, and one industrial report is a win on one product and a loss on another ^S37.

What does growth-coupled mean? That the cell has been modified so it cannot grow without making the product, which means a mutant that stops producing gains no advantage ^S01. For more than 96% of metabolites that E. coli and yeast make from glucose it is feasible on paper ^S06, but often only with a number of knockouts that is not practical to build ^S11.

Is AI used in cell line development? It is being reported by companies on their own processes. One ranking model picked a clone with 68.5% higher titre than the legacy method for one antibody and 6.6% lower for another ^S37. That is one published comparison, not a settled practice.

What we could not verify

The preprint states that all its code is available in a free-access GitHub repository ^S01. We checked that URL on 4 November 2026 and it returns a not-found error, and the account’s six public repositories, four of its own plus two copies of other people’s, include none called ML4SD. The first author does publish a related strain design pipeline covering knockout analysis and flux balance analysis, but it is not ML4SD ^S42. So the method cannot currently be reproduced from the link the paper gives.

We read the supplementary figures of the preprint only through their captions, so the per-bioprocess numbers behind its medians were not checked one by one. We found no published cost or schedule for building one specific knockout strain in a named organism, only whole-programme figures ^S22 and gene-editing timings ^S06. We found no study comparing a metabolic model’s knockout predictions directly against carbon-13 measured fluxes in the same multi-deletion strains, so our statement that these models do not predict real flux well rests on a benchmark in wild-type cells ^S15 plus one implemented design that failed ^S14, not on a head-to-head test. Measured flux data are not a clean reference either: that field reports large discrepancies in quality and studies that cannot be reproduced from incomplete reporting ^S16. Several key older papers were available to us as abstracts only, including the standard review of the pre-machine-learning algorithms ^S08, the most candid biofoundry result ^S31 and the closest wet-lab precedent ^S34; their headline numbers are in the abstracts, their internal breakdowns are not. A search for few-shot learning applied to strain design returned nothing.