AI for science
What does it take for AI to make a scientific discovery?
Hodgkin and Huxley’s 1952 model of the nerve impulse includes a fourth power that simpler equations could not replace. An AI agent that writes and fits equations like theirs would also need to report what failed and which other models fit just as well.
- A fit means more when the alternatives fail. Hodgkin and Huxley’s fourth power pointed to four moving parts in each potassium channel. In a refit of their data, a first-power equation misses the delayed rise about three times as badly.
- Measurements can leave a model’s parameters loose. One fluorescence reading cannot separate the number of molecules from their brightness. Simulation-based inference reports every combination that fits, which shows what to measure next.
- AI agents now propose the equations as well. A method called ModelSMC, presented at a major machine-learning conference in 2026, keeps a weighted population of candidate models. Features that every good model shares then stand apart from features only some of them have.
- The best next experiment is where good models disagree. In Figure 5, a single measurement at high stimulus settles a question that six low-stimulus measurements left open.
- A variant can predict risk long before anyone knows how it works. Obesity variants in the gene FTO were found in 2007. Studies in 2014 and 2015 pointed to two other genes, IRX3 and IRX5.
In 1951 Andrew Huxley had a set of equations and no computer to solve them. He and Alan Hodgkin, physiologists at Cambridge, had recorded electrical currents from the giant nerve fibre of the squid. Their equations were meant to explain how a nerve impulse travels along it, and the only way to test them was to solve them and compare the answers with the recordings. The university’s computer was out of action for six months, according to an essay by Jakob Macke, Daniel Gedon and Stefan Wahl, who build AI tools for scientific modeling at the University of Tübingen. Huxley therefore worked through the numbers on a hand-cranked mechanical calculator. In his Nobel lecture he recalled that one “propagated action potential took a matter of weeks.”
The model they published in 1952 earned them a share of the 1963 Nobel Prize in Physiology or Medicine. Macke and colleagues call it the standard description of electrical activity in neurons to this day. Most of it followed from known physics. The cell membrane behaves like an electrical circuit. Charged atoms called ions, such as potassium and sodium, carry current across it through ion channels, pores in the membrane that open and close. One part was closer to a guess. To describe how easily potassium ions cross the membrane, a quantity called the potassium conductance, Hodgkin and Huxley raised one of their variables to the fourth power.
Why the fourth power
The inside of a resting nerve fibre sits at a negative voltage compared with the outside. When Hodgkin and Huxley suddenly made it less negative, the potassium conductance rose with a delay, slowly at first and then fast, along an S-shaped curve. When they returned the voltage to its resting level, the conductance fell at once. The simplest equation for a quantity that relaxes toward a new value, called a first-order equation, changes fastest at the very start and then slows down. It described the fall well and could not produce the delayed rise. In their words, “the end of the record can be fitted by a first-order equation but a third- or fourth-order equation is needed to describe the beginning.” Raising a first-order variable to the fourth power produces a curve of that higher-order kind, with a slow start and a steep climb.
Which power fits the delayed rise?
Hodgkin and Huxley’s choice. Typical miss on the rise 0.25, on the fall 0.15.
In the refit shown in Figure 1, the fit improves quickly with the power. With the first power, the best curve misses a typical measured point on the rise by 0.72 units of conductance, and the fourth power misses by 0.25. On the fall the powers differ much less, with typical misses between 0.13 and 0.22. Higher powers do slightly better still on these points. Hodgkin and Huxley saw a hint of the same thing in other recordings, where the measured rise had more delay than their curves, and wrote that “better agreement might have been obtained with a fifth or sixth power, but the improvement was not considered to be worth the additional complication.”
Four parts that all have to be ready
They also offered a physical picture of the fourth power. Potassium ions, they suggested, “can only cross the membrane when four similar particles occupy a certain region of the membrane.” If each particle is in place with some probability, independently of the others, then all four are in place with that probability multiplied by itself four times. At 80% per particle, the chance that all four are ready is 0.84, about 41%, and at 50% per particle it drops to 6.25%. The same picture explains the shape of the recordings. After the voltage jumps, each particle arrives at its own random moment. A channel conducts only when the last of its four particles arrives, so the average conductance starts late and then climbs steeply. When the voltage drops back, the first particle to leave closes the channel again, so the fall starts at once.
At 1.3 ms each particle is in place with probability 81%, so all four are in place with probability 42%. Open channels: 42 of 100.
Hodgkin and Huxley had cautioned that their interpretation was “unlikely to provide a correct picture of the membrane.” Part of it held up. In 1998 Roderick MacKinnon’s laboratory worked out the atomic structure of a potassium channel from a bacterium and found that “four identical subunits create an inverted teepee” around the pore. Macke and colleagues judge that the four-particle idea “turned out to be largely right.” Huxley himself warned against reading the equations as final. In his Nobel lecture he noted that Cole and Moore had found the potassium rise “can in some conditions be much more delayed than is accounted for by our fourth-power formulation.”
The fourth power meant something because the first power failed. If a first-power equation had fitted the rise just as well, nobody would have had a reason to picture several moving parts. A fitted model tells us something about nature when we also know which other models the data rule out. Macke, Gedon and Wahl build their essay on AI and scientific models around this point.
Three jobs in every model
A scientific model is a simplified mathematical description of how something works. The essay splits the work of making one into three jobs. Building means choosing the parts and the shape of the equations. Constraining means setting the adjustable numbers in those equations, called parameters, so that the model matches measurements. Interpreting means asking what the finished model says about the real system. In practice scientists loop through the three jobs many times, going back to rebuild when a fit fails or an interpretation makes no sense.
1. Build
Treat the membrane as an electrical circuit, and make the potassium conductance depend on the fourth power of one variable.
Assume the drug leaves the blood at a rate proportional to how much is left, so its concentration falls exponentially, C(t) = C0e−kt.
2. Constrain
Set the rates at which the particles move at each voltage so the curves match the recorded conductance.
Estimate the elimination rate k from a few blood samples taken over a day.
3. Interpret
Read the fourth power as four particles that must all be in place before a channel opens.
Ask whether an unusually slow k points to kidneys that clear the drug poorly.
All three jobs were once done by people, and AI systems are now involved in each. Machines have arguably made the most progress on the second, the essay says. A close fit there can still leave the parameters undecided.
When the data cannot pin down the numbers
Suppose a detector reads 12 units of light from a tube of fluorescent molecules, and the reading equals the number of molecules times the brightness of each one. Twelve molecules at brightness 1 explain the reading, and so do three molecules at brightness 4, along with every other pair whose product is 12. The measurement is correct and so is the equation, yet the number of molecules remains unknown.
A parameter is identifiable when the data, together with the model, allow only one value for it. In the fluorescence example only the product of count and brightness is identifiable. Neither factor is identifiable on its own until a different kind of measurement arrives.
Real models have the same problem on a larger scale. In models of nerve cells, Macke and colleagues write, “different combinations of gate parameters can produce near-identical electrical activity,” where the gates are Hodgkin and Huxley’s particles. A standard fit returns one combination and says nothing about the others. Those others may predict different things when the cell is pushed into conditions nobody has tested.
Bayesian inference keeps all of those combinations in view. It starts from a prior, the range of values considered possible before the experiment, and returns a posterior, the values that remain plausible after the data, each with a probability. For a model as small as Hodgkin and Huxley’s the posterior can be calculated directly. For a network of thousands of ion channels that open and close at random, the calculation becomes impossible, because it needs the probability of the observed data under every combination of settings.
Simulation-based inference avoids that calculation. It picks parameter values at random and runs the model to produce simulated data, and it repeats this perhaps 100,000 times for a model like Hodgkin and Huxley’s. A neural network, a flexible function that learns from examples, is then trained on the pairs of parameters and simulated data. Given a real measurement, it returns the range of parameters that could have produced it. The simulations and training can take weeks, the essay notes, and after that each answer comes in seconds. In 2017 a sky map made from gravitational waves helped astronomers find the glow of two colliding neutron stars 11 hours after the waves arrived. A network trained this way, published in 2025, does the full analysis of such a signal in about one second.
Kept 51 of 1,000 simulations. Counts that fit: 2 to 12 molecules. One reading pins down the product and leaves the count open.
The cloud of kept dots is the useful answer. It shows that the reading fixes the product and leaves the count open, and it hints at what kind of measurement would help. Reading the same tube a second time keeps the same curve. Measuring the brightness of a single molecule cuts across it and leaves only counts near four.
AI now writes the equations as well
Until recently, people chose which equations to try and software found the parameters. Large language models, or LLMs, the systems behind chatbots such as Claude, have changed that division of labour. An LLM agent is a language model set up to write and run code by itself. Given data and a description of a system, it writes code for a candidate model and runs it. Then it compares the output with the data and revises the equations, going around the loop until the fit is good enough. Macke and colleagues cite six such systems from the past two years.
A loop like this tends to stop at the first model that fits. Run an agent twice on the same data, the Tübingen group writes, “and it might return two different, and equally good, models.” An agent that reports one of them hides the other, in the same way that a single best-fitting set of parameters hides the rest of the posterior.
Their own method, ModelSMC, presented at the ICML machine-learning conference in 2026, treats the search for a model as inference too. It keeps a population of candidate models, and the LLM proposes and refines them. Each candidate gets a weight for how well it explains the data, and the LLM writes variations on the strongest. The team tested it on problems with known answers and on real measurements, including recordings from neurons and data for a model of how the kidney regulates potassium in the blood.
If every surviving model shares a feature, the data probably favour that feature. If the survivors disagree, the evidence is thin at that point, and the disagreement shows where to measure next.
At stimulus 8.0 the three models predict 13.0, 6.9 and 9.9, a spread of 6.1. A measurement here would separate them.
Every model in the figure fits the low-stimulus data, so more measurements there barely move the weights. At high stimulus the models split, and one measurement there can settle the question. The essay gives a version from neuroscience. Sodium channels let sodium ions rush in at the start of each nerve impulse. Models of a neuron that agree on everything measured so far should also agree on what happens when every sodium channel is blocked. They can disagree about blocking only some of them, and that disagreement marks the next experiment worth running.
These systems “have mostly been tested on rediscovering models that scientists already know,” Macke and colleagues write, and evidence that they work on problems with no known answer “is sparse.” Rediscovering Hodgkin and Huxley’s equations shows that a method can find an answer people already have. A discovery would be a model that predicts something new, followed by an experiment that confirms the prediction.
One association, three explanations
Human genetics has its own version of several explanations fitting one result. A variant is a spot in the genome where people’s DNA differs, and a genome-wide association study scans many people for variants that are more common in those with a trait. In 2007 such a search for diabetes genes turned up a common variant in the gene FTO. Its link to body mass index, a measure of weight relative to height, then held up in 13 groups totalling 38,759 people. The 16% of adults with two copies of the risk version weighed about 3 kilograms more than those with none. The FTO region became the strongest known genetic association with obesity, and the obvious reading was that the variant changes what FTO does.
Two other explanations fitted the same association. The variant could control a gene further along the chromosome, or it could simply be inherited together with another variant that does the real work. In 2014 a team led by Marcelo Nóbrega found that the obesity-linked stretch of FTO reaches the gene IRX3 across “megabase distances,” meaning stretches on the scale of a million DNA letters. The variants tracked the activity of IRX3 “but not FTO” in human brains, and mice without Irx3 were 25 to 30% lighter. A year later Melina Claussnitzer, Manolis Kellis and their colleagues described a mechanism in fat-cell precursors. A single-letter change in the DNA, called rs1421085, doubles the activity of IRX3 and IRX5 in developing fat cells and shifts them from burning energy to storing it. Correcting that letter with CRISPR gene editing in fat cells from a patient who carried it raised their heat production sevenfold.
In 2015, a single-letter change, rs1421085, was found to break a binding site for a protein that keeps genes switched down. A stretch of DNA that boosts IRX3 and IRX5 then doubles their activity in fat cell precursors.
The 2007 association could already flag carriers as being at higher risk. It could not say which gene to study or aim a drug at, because the competing explanations point to different places in the genome. The studies that narrowed it down were the ones where the explanations made different predictions, such as editing the single letter and measuring IRX3 and IRX5.
An AI agent asked to explain a genetic association can produce a detailed pathway, a chain of molecular steps from variant to disease, with supporting citations and a match to every measurement on hand. If two or three other pathways match the same measurements, the detail makes one of them look better supported than the others, though the data favour none of them. In 2007 the plausible story pointed at FTO, and later experiments pointed at IRX3 and IRX5.
Models too large to read
Hodgkin and Huxley’s model has a handful of variables, small enough to reason about on paper. Philosophers of science call a theory intelligible when you can work out its consequences without running it, and the essay points out that this gets harder as models grow. A fruit fly’s brain has more than 100,000 neurons. Researchers now have a complete wiring diagram of it and have started building models on top of that diagram. At this scale, the essay argues, no single correct model exists, and the useful questions change. They become which properties all the models compatible with the data share, and which experiment would rule out the most of them. Climate science already works this way, reporting its projections as the spread across many large simulations built on slightly different assumptions.
What a useful AI scientist would report
Macke and colleagues write that an AI system has to deliver “not just the model that fits, but an account of which parts are pinned down by the data, which are still open, which alternatives failed along the way, and what new puzzles were uncovered in the process.” Hodgkin and Huxley’s paper, written with a hand-cranked calculator, already gives part of such an account.
The essay also describes a proposal for testing understanding directly, by checking how well a system predicts what happens when the real system is disturbed. The sodium-channel case is that kind of test, with the system committing to a prediction for a partial block before anyone runs the experiment. Experiments cost money and sometimes use samples that cannot be replaced, so a system that picks the informative experiment would be useful before it discovers anything new.
The 1952 paper runs to 45 pages. On page 506 Hodgkin and Huxley report that a first-order equation fits the end of the potassium record and not its beginning. On page 509 they note that a fifth or sixth power might have fitted better still.
Further reading
- Jakob Macke, Daniel Gedon and Stefan Wahl, Teaching AI to build scientific models (6 November 2026). The three jobs of modeling and how AI changes each, with interactive figures of a Hodgkin and Huxley fit.
- A. L. Hodgkin and A. F. Huxley, A quantitative description of membrane current and its application to conduction and excitation in nerve (Journal of Physiology, 1952). The fourth power is on pages 506 and 507.
- A. F. Huxley, Nobel lecture (1963). How the calculations were done by hand.
- Kyle Cranmer, Johann Brehmer and Gilles Louppe, The frontier of simulation-based inference (PNAS, 2020). An introduction to fitting models that can only be simulated.
- Stefan Wahl and colleagues, A Probabilistic Framework for LLM-Based Model Discovery (International Conference on Machine Learning, 2026). The ModelSMC method.
- Melina Claussnitzer and colleagues, FTO obesity variant circuitry and adipocyte browning in humans (New England Journal of Medicine, 2015). From an association to a mechanism.