There is a bacterium in your gut right now, possibly several thousand species, that we have never been able to grow in a lab.
We know they exist. We have sequenced their DNA from your stool, from soil, from seawater, from hot springs. They sit in databases as ghost organisms: full genomes with no organism behind them. We can read every gene they have. We just cannot culture and study them.
this is true for 99% of microbial life on earth
It is not for lack of trying. Microbiologists have been at this for over a century. The standard procedure is depressing: take a sample, dilute it, plate it on agar, try every growth medium you can think of, wait days to weeks, see what grows. Most labs run this with perhaps five media. The most ambitious labs run two hundred. The result is always the same: a handful of fast-growing generalists colonise the plates, and everything interesting stays invisible.
Some of what we are missing. Christensenella, the genus most strongly associated with leanness in human gut microbiome studies, of which we have grown less than one percent.1 Patescibacteria, an entire bacterial superphylum that lives by attaching to the surface of other bacteria, discovered through metagenomic sequencing in 2015, almost none cultured, their lifestyle so unusual that nobody is confident standard methods will ever work on them.2 Asgard archaea, the closest living relatives of the cell type that became all complex life. The first one, Prometheoarchaeum syntrophicum, was finally cultured in 2019 after twelve years of trying with a single sample.3 And the deep biosphere, bacteria living kilometres below the Earth's surface, in rock pores, with doubling times measured in thousands of years: sequenced, but not a single one grown.4
These are not edge cases. These are the rule. The cultivable microbes (the E. coli, the Bacillus, the lactobacilli upon which we have built a hundred years of microbiology) are weird, fast-growing outliers that happened to tolerate the conditions a human in a lab coat decided were "standard."
Why does this matter?
Because every working antibiotic we have was discovered from cultivable microbes. Penicillin came from Penicillium, vancomycin from Streptomyces, a hundred others from a handful of soil genera. The drug discovery pipeline is running dry. We have squeezed the cultivable one percent for everything they have, and resistance is winning. The next generation of antibiotics, almost by definition, has to come from the uncultured 99%.
Here is the surprise: we know these uncultured organisms make novel compounds. Metagenomic mining (reading the DNA without growing the organism) has shown they encode biosynthetic pathways for hundreds of thousands of small molecules we have never characterised.5 We can see the genes for the antibiotics. We just cannot grow the microbes to make them.
The same is true for probiotics. The next generation of microbiome therapeutics (drugs based on actual gut bacteria, not yogurt) is blocked by cultivability. Seres Therapeutics brought an FDA-approved microbiome drug to market;6 their pipeline is constrained by what they can grow. The next breakthrough probiotic, the next Akkermansia (itself only cultured in 2004, and now a billion-dollar market opportunity7), almost certainly already exists as a sequence in someone's database. We just cannot get it into a flask.
The same is true for environmental remediation. Microbes that eat plastic, microbes that fix nitrogen without fertiliser, microbes that capture carbon. The actual best ones, the ones evolution spent four billion years optimising, are mostly uncultured.
And it is not only industry that depends on this. Unknown microbes remain a mystery in both the human body and the environment. Recent work has linked oral bacteria to Alzheimer's risk, gut bacteria to cognitive function, and extremophiles to unusually efficient metabolism, yet for most species we still do not know what they do. Microbial communities are central to health, but the majority remain functionally unmapped.
The unseen is not the unknowable.
So here is the question: what would it take to grow them?
The current approach is brute force. Try more media. Screen more conditions. Hope something works. Some labs run two hundred conditions per sample. Pacific Northwest National Lab runs more than a thousand. None of this scales. Each condition takes weeks, requires expert microbiologists, and most attempts fail.
Exhaustive search over biological permutations is intractable. But evolution has already run the experiment: billions of years of microbial biology encode hints about function in the sequences that survived.
What has changed in the last five years is that we now have foundation models that can read DNA and predict function. ESM-2 reads proteins and predicts structure. Evo 2 reads genomes and predicts mutational effects. These models contain a vast amount of implicit biological knowledge, learned from hundreds of millions of sequences.
Yet raw capability is not the same as fit. We have done extensive research on why current genome models and protein language models underperform on these problems, and we are beginning to form a unified theory that pieces together interactions with equations and formalism. It is clear to us why today's state-of-the-art genome models struggle: next-token prediction, weak key-value matching, and compression-first architectures are a poor fit for biology's data distribution and its interaction structure.
And beyond sequence, there is molecular dynamics. In materials chemistry this is already real: MACE-MP-0 is a foundation neural potential trained on the Materials Project at near-DFT accuracy, running stable molecular dynamics on solids, liquids, and battery cells.8 In biology the equivalent is younger but arriving. AlphaFlow fine-tunes AlphaFold under a flow-matching objective so the output is a conformational ensemble rather than a single structure.9 BioEmu, trained on over 200 milliseconds of aggregated MD, generates thousands of protein equilibrium conformations per hour and predicts free energies to roughly 1 kcal/mol.10 MDGen models full molecular dynamics trajectories directly, up to 100,000 frames at a time.11 The training substrate is ATLAS, an open corpus of 1,390 proteins each simulated in triplicate for 100 nanoseconds.12 Sequence gives you the parts list. Dynamics gives you the runtime.
None of this yet predicts key functions for an uncultured microbe. The pieces sit in different communities: machine-learned interatomic potentials in materials,8 protein ensemble generators in computational biology,9 genome-scale metabolic networks at 247,000-microbe scale,13 and amino-acid-composition models that already predict oxygen tolerance, optimum temperature, and salinity from genome alone.14 None of them individually is enough.
Deep learning fundamentals and the world-model framing map naturally onto biology, because living systems are defined by complex interaction dynamics. We are building a world model for biology, starting with microbial dark matter. We have characterised only a small fraction of Earth's microbes, and even that fraction has already shaped drug discovery, agriculture, food systems, and other industries. If the same pattern holds for what remains uncharacterised, the opportunity is enormous.
If this works, the impact could be large: cultivation-medium prediction, function inference from genome sequence, progress toward virtual cells, and a dent in the data-scarcity problem that runs across biology. The next breakthrough will come from filling those gaps, by making microbial dark matter discoverable at scale.
The loop is how the world model learns. From a genome, predict what the cell needs. Simulate the dynamics: how membranes assemble, how metabolites flow, what conditions the cell was built to live in. Test in the lab. Every outcome trains the world model further. Every cycle, predictions get sharper. Every cycle, cultivation gets faster and cheaper. The artifact at the end is the most accurate model in the world for growing an unknown microbe, and a step toward making microbial dark matter discoverable at scale.
This is hard for several reasons. The labels are bad: existing cultivation databases (BacDive, ATCC, DSMZ) record what worked, not what failed; a working ML pipeline needs both, and we will have to generate the negative data ourselves. The same organism behaves differently in different communities. A bacterium that grows in a defined medium with one neighbour may refuse to grow alone, so cultivation is not really a property of an organism, but of an organism plus its context, and most existing models ignore this. Wet-lab cycles are slow: a single round of cultivation is one to four weeks, so to learn fast every experiment must be maximally informative, which means active learning and Bayesian optimisation choosing what to run next, weighted by information per dollar. The first team to close that loop credibly accumulates a data advantage that compounds monthly.
Most pharmaceuticals fail. Most biology is harder than it looks. This will probably fail too. But the upside is large enough that even partial success matters. Every previously uncultured organism we grow is a potentially useful new tool. New antibiotics. New probiotics. New enzymes. New chemistry.
And if it works at scale, the implication is not subtle. The microbial 99% is the largest unexplored library of useful biology on Earth. Unlocking it is the closest thing biology has to computing the missing dark matter and seeing what is actually out there.
that is the bet
In a few years, hopefully, we will be writing the postmortem on this. Either "here is why it failed and what we learned," or "here is a new genus we just cultured, and the antibiotic it makes." Either way, somebody has to try.