Le calcul naturel : quand 10¹⁷ variants battent toute simulation

The old picture of R&D is comforting. You form a hypothesis, build a model, simulate the system, and test a few promising candidates. It feels orderly. It also assumes the space of possibilities is small enough to think through.

Biology no longer plays by that rule.

George Church has been arguing for a very different method: manufacture absurd numbers of biological variants, expose them to reality, and let the world do the filtering. In electronics, nobody fabricates a billion different materials in one afternoon just to see which ones work. In modern molecular biology, that kind of scale is becoming ordinary enough to discuss without lowering your voice.

That shift matters because some problems are hostile to simulation. Protein folding, gene regulation, cell behavior, molecular binding, materials discovery—these domains contain too many coupled variables, too many context effects, and too many hidden assumptions. We can model pieces. We can even model them well. But there are places where the cleanest route is to stop asking a computer to approximate nature and instead arrange nature so it computes the answer directly.

Church’s claim is not that software stops mattering. It is almost the opposite. The most powerful labs may be the ones that combine machine learning with massive biological search, using models to decide what to build and experiments to decide what is true. That creates a form of discovery that looks less like classical engineering and more like guided evolution at industrial speed.

Biology learned to search at industrial scale

Church likes to frame the scale bluntly. In vitro systems can reach around 10^14 to 10^17 variants. Cell-based systems usually run in the billions. Those numbers do not feel real until you compare them with older experimental habits, where researchers would synthesize a handful of candidates, maybe a few dozen if they had patience and grant money.

A library at that scale is not just a bigger spreadsheet of options. It changes the kind of question you can ask. If you can generate and test billions of sequence variants, you no longer need a neat mechanistic theory before you begin. You can search a landscape first and understand it more clearly afterward. The experiment stops being a final check on the model. It becomes the main engine of discovery.

This did not arrive overnight. Church points back to work from 2004 on synthesizing genes on chips, published in Nature. The basic capability existed early, but the field took years to absorb what it meant. That pattern is familiar in technology. A tool appears, looks exotic, then sits around half-understood until the surrounding ecosystem catches up. Gene synthesis, high-throughput sequencing, cheaper DNA assembly, better assay design, and machine learning each mattered on their own. Together, they turned combinatorial exploration from a stunt into a platform.

The phrase “library” can sound abstract, so it helps to be concrete. Imagine an enzyme with a few positions that might change its activity. Traditional engineering might test a short list of educated guesses. A library approach encodes huge numbers of versions at once, introducing controlled variation across those positions, then exposing the pool to a selection pressure: heat, acidity, a substrate, a pathogen, a cell state. Variants that survive or perform well become visible through sequencing or other readouts. The experiment is not trying to explain every molecular interaction in advance. It is trying to let the interaction happen and count the winners.

That is a different philosophy of science. It trusts search, not because theory failed, but because reality contains more useful information than the theory can currently express.

Evolution became a design tool

Church’s most provocative line on this is about time. Evolution might incorporate a few base-pair changes over enormous spans. A synthetic biology lab can make billions of changes in an afternoon. The comparison is intentionally dramatic, but the underlying point is simple: mutation and selection used to be nature’s slow, blind process. Now they can be made fast, intentional, and instrumented.

The “intentional” part matters as much as the speed. Natural evolution wastes plenty of motion on dead ends, neutral drift, and local adaptations that do not help with our chosen objective. Directed libraries do not need to be random in that sense. They can be biased toward regions of sequence space that are likely to preserve function while still creating novelty. Church describes focusing on variants that are close enough to remain viable yet different enough to become consequential.

That sounds technical, but it is really a search strategy. If you randomize everything, most variants are broken. If you never stray far from the known sequence, you mostly rediscover what you already had. The sweet spot lies in changes that are “quasi-neutral,” as Church puts it: tolerated by the system, but still capable of producing a meaningful shift in behavior. That is where useful enzymes, improved binding affinities, new regulatory properties, and surprising emergent effects often live.

This is one place where many outsiders misunderstand the field. They imagine biologists replacing thought with brute force. The actual craft lies in how the diversity is shaped. Which residues are varied? Which codons are allowed? Which structural motifs are preserved? Which cellular contexts are used? What counts as success, and what hidden failure mode will the assay miss? The scale is immense, but it is not indiscriminate. A good library is not a pile of noise. It is a compressed hypothesis about where reality might be hiding something valuable.

The result feels like evolution, but accelerated and constrained by design goals. You are not waiting for the environment to randomly reward a lucky organism. You are specifying the search space, compressing the cycle time, and reading the output with high-throughput instruments. Darwin, rewritten for liquid handlers and sequencers.

Reality computes differently

Church’s sharpest claim is that this approach can beat simulation because it is not simulating at all. In his words, “It’s a hundred percent precision, because you’re not simulating. You’re not making assumptions.” He then contrasts this with the stack of approximations we typically rely on, moving from more fundamental physics toward practical molecular models, each layer losing something along the way.

Taken literally, “hundred percent precision” is too clean. Real experiments have noise. Assays have bias. Cells are fickle. Batch effects exist to humble the confident. A flawed selection setup can send you racing toward the wrong answer with impressive statistical significance. Anyone who has spent time around wet labs knows that reality is generous with surprises and stingy with explanations.

But Church is pointing to something deeper than laboratory neatness. A simulation is always a translation. It encodes a view of the system, and that view leaves things out. Sometimes the omissions are harmless. Often they are the only reason the simulation finishes this century. Molecular dynamics, quantum chemistry, docking models, force fields, coarse-grained approximations—these are useful because they simplify. Yet the simplifications become dangerous when the neglected interactions are exactly where the interesting behavior lives.

A real selection experiment does not need to approximate solvent effects, folding intermediates, crowding, metabolic interference, membrane composition, or the thousand other details that make biological systems annoying. Those details are simply present. Physics takes care of the bookkeeping without asking your permission. That does not mean the experiment answers every question. It means the answer it gives was produced by the actual system, under actual constraints, rather than by a model’s best effort to stand in for it.

This difference becomes decisive when the landscape is rugged and poorly understood. Simulations are strong when the objective is clear and the abstraction holds. They struggle when hidden variables dominate. Biological search shines in exactly those messy zones. You can expose 10^9 or 10^14 candidates to a real environment and ask a narrow question: which ones bind, survive, catalyze, grow, suppress, fluoresce, resist, or assemble? If the assay captures the property that matters, reality does the hard part for you.

There is a weird elegance to this. We tend to think computation means pulling the world into silicon. Church’s argument flips it around. Sometimes the more efficient move is to push the problem back into matter and let matter resolve it.

Models and molecules form a tighter loop

This does not make AI secondary. It gives AI a more grounded role.

The easiest version of large-scale biological search is brute mutagenesis followed by selection. That works, but it wastes bandwidth. If you have the ability to build enormous libraries, the valuable question becomes which enormous library to build. That is where machine learning earns its keep.

Models can predict which regions of sequence space are worth exploring, which changes are likely to preserve structure, and which combinations might produce useful non-linear effects. They can rank candidates, design assays, infer hidden variables from sparse measurements, and recommend the next round of exploration. After the experiment, they get something more precious than synthetic benchmark data: high-volume observations from the real system.

Church describes this as a cycle. AI helps generate better libraries. Biology executes those libraries in the world. The resulting data gets harvested efficiently and fed back into the models for another round. That loop matters because each side corrects the other’s weakness. Models narrow the search and increase sample efficiency. Experiments rescue the process from hallucinated certainty.

This is already visible in protein engineering. A model may suggest mutations that preserve folding while shifting binding preference or catalytic activity. A high-throughput assay then tests huge numbers of those variants in actual biochemical conditions. Sequencing reveals which variants enriched under selection. The next model learns not from toy labels but from the real fitness landscape uncovered by the experiment. The method is less like proving a theorem and more like building an increasingly sharp instrument for interrogating a complicated world.

The same pattern can extend beyond proteins. Gene circuits, CRISPR guide design, viral capsids, antibody libraries, regulatory elements, cell therapies, metabolic pathways—any domain where you can generate many variants and read out a meaningful signal becomes a candidate for this hybrid workflow. The bottleneck shifts from raw imagination to library design, assay quality, and interpretation.

That last word matters. A billion measurements are not wisdom by default. They are often an expensive form of confusion. The loop only improves when the readout captures the property you care about and when the model can absorb the result without mistaking correlation for mechanism. The lab of the future may produce torrents of data, but the deeper skill will be knowing which few bits of that torrent deserve trust.

Discovery spills beyond medicine

The obvious applications sit in drug discovery and therapeutics, where biological screening has clear homes. Church pushes further. He imagines using these libraries to discover superior materials and even, in a speculative stretch, perhaps room-temperature superconductors.

That sounds like a category error until you unpack the idea. Biology is not restricted to making things that look biological. It is a vast manufacturing system for polymers, peptides, nucleic acids, nanostructures, binding motifs, and templated assemblies. It can explore chemical and structural spaces that are expensive to search with conventional fabrication. Even when the end material is not biological, biological processes can help discover precursors, catalysts, scaffolds, or assembly pathways.

There are precedents for this style of thinking. Researchers already use evolved enzymes for chemical synthesis, engineered microbes for materials precursors, peptides for mineral binding, and DNA structures as nanoscale templates. The broader bet is that high-throughput biological generation plus real-world selection might uncover material behaviors that theory did not predict cleanly. If the space is too large to simulate and too weird to reason through by hand, direct search becomes attractive.

Superconductivity is an extreme example, and it is fair to keep some skepticism handy. The path from a biological library to a room-temperature superconductor is nowhere near straightforward. The screening problem alone would be formidable. But Church’s larger point survives even if that specific prize remains distant. The toolkit we built for manipulating living systems may turn out to be a general engine for searching complex design spaces, including ones that spill into chemistry and materials science.

That possibility is easy to underestimate because our institutional boundaries are so rigid. Biology departments work on life. Materials scientists work on matter. Software people work on abstractions. Reality has less respect for those job descriptions.

Science starts to look more like search

If this paradigm keeps spreading, the center of gravity in research changes. The hero experiment is no longer the one that isolates a single variable with exquisite elegance. It is the one that constructs a meaningful search space, subjects it to the right pressure, and reads the result at scale. Theory still matters, but often as a way of sculpting the search rather than replacing it.

That has consequences for who holds leverage. Better foundation models will help, but they will not be enough on their own. Advantage will also come from synthesis capacity, automation, assay design, sequencing infrastructure, sample handling, and the ability to close the loop quickly. A lab that can iterate ten times in a week learns differently from a lab that takes a month to run a single cycle.

It also changes the culture of explanation. In classic science, we often prefer understanding before action. In this emerging mode, useful action can arrive earlier. You find variants that work, deploy them, then spend longer deciphering why. Purists will dislike that. Sometimes they should. A field that only screens and never understands can drift into expensive empiricism. Yet it is worth admitting how much progress in technology has always depended on methods that worked before anyone fully grasped the mechanism.

The broader lesson is not that simulation loses. Plenty of domains will remain dominated by models, and for good reason. Simulations are cheap to run, easy to copy, and often the only practical way to narrow a gigantic space. The shift is subtler. For a growing class of problems, the best computation is a partnership between prediction and embodiment. Software proposes. Matter disposes.

That is a strange sentence to write in an era obsessed with disembodied intelligence, but it fits the evidence. When you can manufacture 10^17 possibilities, test them under real conditions, and use the outcome to train the next round, the frontier stops looking like a contest between AI and biology. It looks like a new grammar for invention, one where discovery happens by steering vast populations through reality faster than our theories can catch up.