Which single-neuron model can you trust with living tissue?

What the work claims

This is a classification paper, not an experimental result. The authors review mathematical models of individual neurons and sort them into two families: single-entity (single-compartment) models, where one computational unit stands in for the whole cell, and composite multi-compartment models, where axon, dendrites and synapses each get their own equations. Only the single-entity family is treated here; the multi-compartment survey is deferred to future work¹.

The survey's useful move is to grade each model on three axes at once: whether it is built from biophysics (ion-channel kinetics) or from curve fitting, how many state variables, parameters and floating-point operations it consumes, and whether it carries experimental support for the firing patterns it claims to reproduce. The headline judgments are blunt. Integrate-and-fire variants are described as biophysically meaningless and explicitly not recommended for biological simulation. The Hodgkin-Huxley model is the biophysical reference point but is analytically complex and computationally demanding. The Izhikevich model sits in the pragmatic middle: two differential equations, parameters fitted to cortical recordings, cheap enough for large networks, but unfit for ion-channel-level questions.

How it works

The integrate-and-fire family integrates input current until the membrane voltage crosses a threshold, emits a binary spike, and resets. Leaky versions add a constant leak conductance; adaptive versions raise the threshold after each spike; quadratic versions gain bistability. What all of them omit is the action potential itself: the emitted spike has no duration, amplitude or shape, and there are no ion channels. The survey treats this as disqualifying for biological simulation, while acknowledging that the same simplicity is exactly what makes these models effortless to implement in neuromorphic silicon¹.

The Hodgkin-Huxley model, in contrast, is built from Markov-style kinetics of sodium and potassium channels: four coupled differential equations whose parameters were determined empirically from squid giant axon experiments in 1952². The survey flags two structural compromises: activation and inactivation are treated as kinetically independent, which real channels violate, and the model is expensive enough to choke large-scale simulations. Simplifications such as FitzHugh-Nagumo, Morris-Lecar and Izhikevich recover tractability by fitting reduced dynamics to observed firing patterns. Izhikevich's two-equation model, with its quadratic voltage term and recovery variable, was derived from bifurcation analysis and fitted to cortical neuron dynamics, and it can be discretized at one-millisecond resolution to run tens of thousands of neurons on modest hardware¹³.

Two less conventional entries matter for our purposes. The dendritic neuron model replaces the weighted-sum synapse with a multiplicative structure: sigmoidal synapses feed branches whose outputs are multiplied, then summed, letting the model discard unnecessary synapses through pruning. The survey notes it does not describe dynamics and suffers from the curse of dimensionality, but points out it can be built from comparators and AND, OR and NOT gates, and names its most promising application as analyzing how real synapses and dendrites work¹. The second is the use of trained artificial networks as neuron emulators: a convolutional-LSTM architecture reported in earlier work replicates membrane potential, sodium and potassium conductances, AMPA and NMDA receptor non-linearities and multi-compartment dendritic signals, at about 1.95 million parameters per emulated neuron class, with a custom Izhikevich layer fixing its tendency to ignore long-duration firing patterns at a cost of five extra trainable parameters¹⁴.

The comparison table puts hard numbers on the trade. In the survey's own counting convention, a Hodgkin-Huxley update costs roughly 1200 floating-point operations, Morris-Lecar about 600, FitzHugh-Nagumo about 72, Izhikevich about 13, and a leaky integrate-and-fire update about 5. The table also carries an experimental-evidence column: Hodgkin-Huxley, Izhikevich and Morris-Lecar match the full range of observed firing classes, while the cheap integrate-and-fire variants match none¹.

Where a skeptic should push

The most load-bearing assumption is that matching firing-pattern catalogs is a reasonable proxy for biological plausibility. It is not nothing, but it is a static, input-driven test: a model can reproduce the spike shapes it was fitted to and still compute the wrong thing under closed-loop perturbation, which is precisely the regime organoid intelligence lives in. The survey does not address transfer under feedback, drift, or pharmacological perturbation at all.

Second, the cost column is convention-bound, as the authors themselves warn: membrane capacitance and external current are excluded, and every elementary mathematical operation is counted as one floating-point operation. The 1200-to-5 ratio between Hodgkin-Huxley and leaky integrate-and-fire is therefore a rough ordering, not a benchmark. Treat it as two orders of magnitude of spread, not as a precise figure.

Third, the neural-network emulator entry deserves particular suspicion. The survey reports its out-of-domain failure as a known limitation, and this is the crux: a trained emulator is only guaranteed to behave like the neuron on the distribution it was trained on. And the training cost is paid before simulation starts, which can erase the speed advantage for short studies. A four-page classification of models it calls biologically meaningless also sits oddly next to the fact that those same models dominate neuromorphic engineering; the survey's standards are biologist's standards, and engineers reading it should keep that lens in mind.

What model fidelity means for organoid computing

The non-obvious implication is that organoid intelligence currently runs on an unexamined epistemic loan. Closed-loop experiments with living cultures are almost always designed, and their controllers trained, against simulated networks of cheap neurons, typically leaky integrate-and-fire or Izhikevich units, because whole-dish simulation in Hodgkin-Huxley detail is unaffordable. The survey's table states plainly that the affordable end of that ladder is the end its authors would not use for biological simulation. That does not make the practice wrong; organoids are not squid axons either, and no single-compartment model captures dendritic computation. But it means transfer claims, this protocol worked in simulation and therefore should work on tissue, inherit a fidelity gap that is rarely disclosed and never quantified. A protocol validated on integrate-and-fire dynamics has been validated against a computer that fires binary spikes with no action potential, which is not what the electrode sees.

The opportunity is concrete: the table is a rational procurement chart for a two-tier simulation stack. Fast, biologically meaningless units are legitimate for prototyping the controller and the stimulation pipeline, where only timing logic matters. Biophysical or fitted-reduced models belong in the qualification stage, where the question is whether the mechanism, not the code, will survive contact with the dish. The dendritic neuron model hints at what the qualification stage should be looking for: real dendrites multiply rather than merely sum, and a substrate whose computation lives in multiplicative branch interactions will look unimpressive to any readout built on the weighted-sum assumption.

The threat is the digital twin. The CNN-LSTM emulator approach, about 1.95 million parameters to stand in for one neuron's response class, is the only credible route to real-time twins of organoid tissue for closed-loop training, and its documented failure mode is out-of-domain behavior. A twin goes out of domain exactly when the living substrate does something new, which is also when a closed-loop trainer pushes the tissue hardest. A controller certified against a silently broken twin is not a hypothetical hazard: it is the default failure mode of a stack that never states which neuron model stands in for the tissue, and this survey is the reminder that the choice is consequential, made by default, and currently nobody's job to defend.

The bottom line

Established: the taxonomy, the cost ordering and the plausibility judgments are a fair summary of mainstream modeling practice, and the survey's own numbers document a real cost-versus-plausibility trade of roughly two orders of magnitude between Hodgkin-Huxley detail and integrate-and-fire speed. Hypothesis, not established: that matching firing-pattern catalogs predicts behavior under closed-loop perturbation, which is the only claim that matters for organoid computing and the one nobody has tested directly. What would confirm it: a benchmark that runs identical closed-loop protocols against simulated networks at two or three fidelity tiers and against real cultures, and reports where the transfers break. What would break the field's current habit: evidence that controllers tuned on cheap models destabilize cultures that biophysical models predicted would be pushed out of regime.