A grid cell model ties the reach of prediction to ion channel kinetics

What the work claims

The central claim is that predictive coding in the medial entorhinal cortex does not require a single dedicated mechanism: an intrinsic, cell-autonomous mechanism and a circuit-level mechanism together account for the different magnitudes of predictive bias observed in layer II and layer III grid cells. Layer II cells show a modest prospective bias of about 5 percent of the grid scale; layer III cells show a bias reaching nearly 25 percent, the range reported for predictive grid cells recorded in vivo.² In the model, the first comes from HCN channel rebound, the second from asymmetry in interneuron-to-stellate connectivity, and reversing the asymmetry flips the bias from prospective to retrospective.¹

Two supporting claims strengthen the case. First, the same model reproduces a cluster of independent phenomena: grid-like firing from interneuron-mediated connectivity, the depolarizing membrane potential ramp observed as a rat enters a grid field (which the analysis traces to gradual disinhibition rather than tuned excitation), and the expansion of grid scales and field sizes seen experimentally when HCN channels are knocked out, which the model explains as a reduction in the velocity gain of the attractor rather than a change in bump geometry.¹ Second, a corollary prediction: because HCN time constants are known to vary along the dorsoventral axis of the entorhinal cortex, the intrinsic component of the predictive bias should change monotonically along that axis, giving a direct link between the biophysics of one channel type and the spatial reach of a population code.¹

This is a computational modeling paper constrained by published biophysics and anatomy. The conductance parameters come from experimental characterizations of stellate cells and fast-spiking interneurons, and the connectivity motif reflects the layered organization of entorhinal layer II. It is not a data paper: every result is produced by simulation, and the key anatomical asymmetry it proposes has not yet been demonstrated in paired recordings or connectomics.¹

How it works

The network is a continuous attractor model built from five neuronal sheets of 71 by 71 neurons each, with opposite edges coupled to form a torus: four sheets of stellate cells and one sheet of inhibitory interneurons. Stellate cells do not excite each other directly; all interaction runs through disynaptic, interneuron-mediated inhibition with a center-surround profile. Speed-dependent depolarizing input, delivered uniformly and inverted through a measured input-velocity curve to linearize the response, translates the activity bump across the sheet so the network path-integrates a simulated trajectory.¹

The stellate cell model is the biophysical load-bearing element. Stellate cells carry a hyperpolarization-activated depolarizing current, HCN, together with a persistent sodium current, and they show post-inhibitory rebound, subthreshold membrane oscillations in the theta range of 4 to 12 Hz, resonance, and a depolarizing sag. As the activity bump approaches a stellate cell's field, inhibition from interneurons wanes; because HCN conductance peaks 100 to 200 milliseconds before field onset, the cell rebounds and spikes slightly early, so the population vector decoder, which infers position from collective activity, places the animal ahead of its reference position. This intrinsic mechanism shifts the inferred position by about 5 percent of a grid field diameter.¹

The circuit mechanism is geometrical. Interneuron-to-stellate connectivity is modeled as a Gaussian profile; skewing that profile makes inhibition on the trailing edge decay more steeply, so the same uniform depolarizing drive elicits spikes earlier. In simulations without HCN conductance, to isolate the effect, increasing the skew increases the prospective bias up to roughly 25 percent of the grid scale, and flipping the skew direction produces a retrospective bias. The authors map the small effect onto layer II stellate cells and the large effect onto layer III, noting that layer III pyramidal cells lack the HCN-dependent rebound properties and are plausibly a direction-biased readout of the layer II attractor.¹

The knockout result ties the biophysics to the code. Removing HCN conductance hyperpolarizes stellate cells, shrinks activity bumps, and reduces interneuron firing; restoring inhibition rescues the expansion of grid scales and field sizes. The model reads this as HCN normally lowering the velocity gain of the attractor, which matches the experimentally observed expansion of grid scale after HCN knockout in mice.¹

Where a skeptic should push

The most load-bearing assumption is that the proposed asymmetric interneuron-to-stellate connectivity exists in layer III. The authors state plainly that this asymmetry is a constraint of the model, not a measured wiring diagram, and that it has not been demonstrated directly. The entire layer III story - the five-fold larger predictive bias, the dissociation from the intrinsic mechanism - rests on anatomy that paired recordings or connectomic reconstruction could either confirm or eliminate. Until then, the skew parameter is a fitted explanation, and reversing it to produce retrospective bias shows the model can produce either sign of bias on demand.¹

Second, this is an in silico network with a single attractor module. Real entorhinal cortex contains multiple grid modules with different scales along the dorsoventral axis, and the model simulates one. Predictions about monotonic variation of predictive reach with HCN time constants are therefore about the mechanism as modeled, and real populations could confound the gradient with module-scale effects the model does not represent.

Third, the path-integration input is idealized. The population speed signal enters as a single instantaneous depolarizing drive, uniform across stellate cells, with the input-velocity curve doing the linearization work. Real speed signals are heterogeneous and temporally offset across cells; the authors themselves note that offset speed drive could change the magnitude or even the sign of the bias they attribute to network asymmetry. That is an honest caveat, but it means the two-mechanism decomposition is not unique.

Fourth, the ramp and rebound analyses lean on filtering choices: spikes are removed and the membrane potential is low-pass filtered at 3 Hz to compare with experimental ramp measurements. Different processing could shift the apparent timing relationships on which the 5 percent estimate depends.

Channel kinetics set a living code's lookahead

The non-obvious implication for organoid intelligence is that what a living neural network computes is set, in part, by which ion channels its cells express, and that this is a dial rather than a destiny. In this model, changing one channel type changes the temporal reach of the population code: HCN rebound turns a present-position code into a few-percent-ahead code, and the channel's time constant gradient across the tissue predicts a spatial gradient of predictive reach. A cortical organoid is a tissue whose channel expression, and therefore its intrinsic timescales, drifts with maturation state and batch conditions. That implies the computational mode of an organoid - how far its dynamics anticipate, how slowly it integrates, which timescales dominate - is a moving property that must be characterized and controlled before any benchmark result can be trusted or reproduced.¹

The opportunity is that intrinsic excitability is an underused control input. Organoid computing today mostly treats the tissue as a fixed recurrent dynamical system and works through stimulation and readout. This paper suggests a second axis: modulating expressed channels or their kinetics (pharmacologically, genetically, or through maturation protocols) changes the family of computations the circuit can natively perform, before a single electrode fires. Rebound-based anticipation is also a property a living substrate might offer that is awkward to engineer in silicon: memory of recent inhibition carried by slow channel kinetics, converting history into prospective spike timing for free at the single-cell level.

The threat is subtler. If a self-organized circuit's computational behavior depends on channel biophysics that vary dorsoventrally, across cell types, and across developmental time, then two organoids with identical wiring diagrams could compute differently, and the difference would be invisible to standard electrophysiological summaries like firing rate or synchrony. The field already struggles with batch variability; this work says some of that variability is computational, not noise. Benchmarks that do not measure intrinsic timescales (sag, resonance, rebound latency) alongside task performance will keep conflating the two.

There is also a design lesson. The model's predictive code emerges from a marriage of an attractor (the population structure) and cell-autonomous dynamics (the channels). Organoid systems built to path-integrate or anticipate may need both: structure imposed by connectivity or patterned input, and intrinsic timescales tuned by maturation. Tissue engineered with only the first half may plateau at the computation its architecture alone supports.

The bottom line

Established in simulation: a continuous attractor network of conductance-based stellate cells and interneurons path-integrates accurately, reproduces the depolarizing ramp in grid fields through disinhibition, accounts for the modest about 5 percent prospective bias of layer II cells through HCN rebound, generates up to about 25 percent prospective bias through skewed inhibitory connectivity, and explains HCN-knockout grid expansion as reduced velocity gain. Established as prediction, not fact: the inhibitory asymmetry itself, and the monotonic dorsoventral gradient of predictive reach.

For organoid intelligence, the durable message is that intrinsic channel kinetics are part of the computational specification of living tissue. What would confirm the transfer is experimental: organoids characterized for HCN-dependent properties and their dynamics' anticipation horizons, then asked whether manipulating those properties shifts the family of computations the tissue can perform.