Subtype-Selective Ligands as Actuators of the Heterogeneity Knob

Neurons of the same type are not interchangeable. Even within a nominally uniform population of cortical pyramidal cells, membrane time constants and spike thresholds vary widely from cell to cell, and a growing body of work shows this variation is not noise but a computational resource: spreading out spiking thresholds desynchronizes populations and smooths their collective response, and a spectrum of time constants lets a network hold a fading memory of input at many timescales at once. In recurrent spiking networks trained by gradient descent, heterogeneous time constants improve accuracy on temporally complex audio benchmarks by 15–20% over homogeneous networks with the same neuron count Neural heterogeneity promotes robust learning. Yet this entire line of work treats the shape of these parameter distributions as exogenous: someone picks a Lorentzian or a learned histogram, and the biology that might actually set and adjust it in a living circuit is left as an acknowledged open problem.

This proposal argues that the missing actuator is pharmacological, and that it has a name drug discovery already cares about: receptor subtype selectivity. Neuromodulators such as acetylcholine and serotonin act on families of receptor subtypes that are differentially distributed across excitatory and inhibitory cell classes and couple to different intracellular cascades, ultimately shifting membrane conductances, time constants, and firing thresholds. A ligand's subtype selectivity profile — which subtypes it activates, and how strongly, relative to its sisters — therefore determines which cells' intrinsic parameters move, and by how much they differ after the drug arrives. Selectivity is not just a safety metric for avoiding peripheral side effects; it is a structured control input that reshapes the variance of the heterogeneity distribution within each cell class, and thereby steers the circuit between computational regimes.

Core Claim

The claim is that the circuit-level effect of a neuromodulatory ligand is governed less by its mean effect on excitability than by which receptor subtypes it engages, because subtype engagement — filtered through cell-type-specific receptor expression — determines whose intrinsic parameters shift and how unevenly. Formally, for an intrinsic parameter $\theta$ (a time constant or spike threshold), let cell $i$ in the circuit express subtype $s$ at level $e_{is}$, let the ligand have affinity $a_s$ for subtype $s$, and let $g_s$ be the coupling from activation of subtype $s$ to a change in $\theta$. The induced per-cell shift is

$$\delta\theta_i = \sum_s a_s \, e_{is} \, g_s(\theta_i),$$

and the within-class variance of $\delta\theta$ is set by how the affinity vector $a$ projects onto the cell-to-cell covariance of expression. A subtype-selective ligand picks out one direction of that covariance; a pan-subtype ligand averages over many. This reframes a result from systems biology that currently points the opposite way: Correlated receptor transport processes buffer single-cell heterogeneity documents that receptor surface abundance varies substantially from cell to cell and treats that variation as a fluctuation cells must buffer to preserve signaling fidelity. We instead treat expression heterogeneity as the substrate a drug acts on: it is precisely because receptor levels differ across cells that a ligand converts a uniform dose into a distribution of parameter shifts, i.e., into heterogeneity. The buffering view and the computational-resource view stop contradicting each other once you ask what the drug does rather than what the cell does.

The payoff is a computational endpoint for selectivity itself. Structural and pharmacological work can already achieve exquisite subtype discrimination — constrained catecholamines gain over 100-fold selectivity for $\beta_2$AR over $\beta_1$AR through differences in pocket dynamics even though the two receptors share identical orthosteric residues Constrained catecholamines gain β2AR selectivity through allosteric effects on pocket dynamics — but nothing in that literature predicts what a selective compound does to the circuit downstream. Meanwhile, the computational literature shows the downstream consequences are sharply asymmetric: threshold heterogeneity in inhibitory interneurons enables flexible signal gating, while the same heterogeneity in excitatory cells raises dynamic dimensionality but destroys attractor bistability and degrades spatial working memory in ring networks Neural heterogeneity controls computations in spiking neural networks. A compound selective for a subtype concentrated in interneurons and a compound selective for a subtype concentrated in pyramidal cells should therefore produce opposite computational outcomes at equal target occupancy. No study in the corpus tests this.

Research Program

Three components. First, an expression-informed neuromodulation layer for spiking networks: each simulated cell carries a receptor expression vector, drawn from a measured cell-to-cell distribution, and each subtype carries a coupling rule from activation to parameter change. The subtype-specific couplings are parameterized from the signaling literature — intracellular loop 2 as the switch determining G-protein coupling preference across galanin receptor subtypes Molecular basis for allosteric agonism and G protein subtype selectivity of galanin receptors, the structural templates for M4-selective modulation The structural study of mutation-induced inactivation of human muscarinic receptor M4, and divergent serotonin subtype coupling Structural insights into the ligand binding and Gi coupling of serotonin receptor 5-HT5A. The cell-type specificity of channel-mediated excitability changes is motivated empirically by Kv1.1 channel dysfunction in parvalbumin-positive interneurons contributes to anxiety-like behaviors in young adult presenilin 1/2 conditional double knockout mice, which shows that altering a single channel type only in parvalbumin-positive interneurons changes circuit behavior and the animal's phenotype — exactly the cell-class selectivity our model formalizes, but without the computational framing.

Second, a mean-field extension. Macroscopic dynamics of neural networks with heterogeneous spiking thresholds links threshold-distribution shape to macroscopic population dynamics, and Neural heterogeneity controls computations in spiking neural networks extends this to cell-class-asymmetric roles, but both rely on analytically convenient Lorentzian distributions — an acknowledged limitation. We extend the mean-field treatment to distributions that are reshaped by the modulation term above, and we validate against non-symmetric, heavy-tailed empirical distributions of time constants and thresholds measured in real neurons Dynamic I-V Curves Are Reliable Predictors of Naturalistic Pyramidal-Neuron Voltage Traces. This directly tests whether the smoothing and gating predictions survive biologically realistic distribution shapes, which is itself an open problem in the heterogeneity literature.

Third, a computational screen. Given a ligand's affinity vector and a circuit's expression map, predict the sign and magnitude of change in task accuracy, gating flexibility, and working-memory retention. The output is a quantity we call the computational therapeutic index: the ratio between the dose range that improves a target computation and the dose range that degrades working memory. The hypothesis is that subtype selectivity enlarges this index even when it does not change maximum efficacy, because it confines the variance reshaping to the cell class where heterogeneity helps.

Proposed Experiments

Experiment 1 uses the Spiking Heidelberg Digits and Spiking Speech Commands benchmarks, the temporally rich audio tasks from Neural heterogeneity promotes robust learning, which reported a 15–20% accuracy increase from heterogeneous over homogeneous time constants (no absolute accuracy value is reported on Lacuna for these benchmarks, so we compare in relative terms). We clamp the network's time-constant spread to simulate a pathology or mistuning that collapses it toward homogeneity, then apply a virtual subtype-selective agonist versus a virtual pan-subtype agonist, each matched for total receptor occupancy. The prediction: the selective agonist, projected onto expression covariance concentrated in one cell class, recovers a larger fraction of the 15–20% heterogeneity gain and restores generalization across input playback speeds, while the non-selective agonist at equal occupancy shifts means without restoring variance and recovers less. DVS128 Gesture serves as an intermediate-temporal-structure case, and Neuromorphic MNIST is the falsifying control: Leung et al. found heterogeneity gains negligible there, so modulation should also do nothing — if it "helps" on a static task, something other than the claimed mechanism is at work.

Experiment 2 tests gating. Following the demonstration that inhibitory threshold heterogeneity permits flexible, synchrony-free signal gating Neural heterogeneity controls computations in spiking neural networks and that heterogeneous intrinsic properties stabilize oscillations against fluctuating inhibition Gamma Oscillations in a Nonlinear Regime: A Minimal Model Approach Using Heterogeneous Integrate-and-Fire Networks, we measure the range of input frequencies over which the network entrains and gates reliably as a function of agonist selectivity profile. No reported numeric value for this dynamic range exists on Lacuna, so we define it operationally and report it. A subtype-selective compound targeting an interneuron-enriched subtype should widen the gating range; a pan-subtype compound should narrow it by pushing excitatory heterogeneity toward the synchrony and attractor-collapse regimes.

Experiment 3 tests the attractor trade-off on the ring-network working-memory paradigm, where the same paper showed excitatory threshold heterogeneity shrinks the bistable regime and degrades spatial memory retention (again, no numeric value is reported on Lacuna; we measure bump stability and bistable-regime width ourselves). The decisive comparison is between two compounds at equal occupancy of their intended targets: one selective for a pyramidal-enriched subtype, one selective for an interneuron-enriched subtype. Opposite signs of memory effect at equal occupancy would be strong evidence that selectivity, not affinity, carries the circuit-level outcome.

Experiment 4 replaces the Lorentzian baseline distributions with empirical parameter distributions from dynamic I-V measurements Dynamic I-V Curves Are Reliable Predictors of Naturalistic Pyramidal-Neuron Voltage Traces and the public recording datasets referenced by Leung et al. (mouse cochlear nucleus, mouse visual cortex, human middle temporal gyrus), checking that the mean-field selectivity predictions from component two survive realistic distribution shapes.

Baselines

The do-nothing reference is the unmodulated heterogeneous recurrent spiking network itself, the strongest documented configuration on these benchmarks: Leung et al. report that heterogeneous time constants beat homogeneous ones by 15–20% on SHD and SSC without adding neurons or synapses Neural heterogeneity promotes robust learning, so this baseline defines the ceiling any pharmacological intervention is trying to recover or exceed. It will be run with the published architecture and surrogate-gradient training, on identical data splits.

The strongest competitor is non-selective modulation: a virtual pan-subtype agonist that engages every receptor subtype in the circuit and produces the same total mean parameter shift as the selective agonist, under the same dose sweep, the same simulation budget, and the same tuning effort. This is the honest stand-in for most real drugs, and beating it means demonstrating that where the variance is reshaped matters at equal target engagement — the core claim. If non-selective and selective modulation produce statistically indistinguishable outcomes across Experiments 1–3, the claim is wrong and we will say so.

A second competitor is weight retraining: after degrading the heterogeneity distribution, retrain the synaptic weights with surrogate gradient descent under the same compute budget as the pharmacological intervention, testing whether plasticity can simply compensate for a collapsed parameter distribution. Duarte et al. found that varying structural connectivity does not match intrinsic parameter diversity as a driver of computational capacity in cortical microcircuit models Leveraging heterogeneity for neural computation with fading memory in layer 2/3 cortical microcircuits, which predicts retraining falls short on temporal tasks, but the experiment settles it directly. Finally, the key ablation is a variance-preserving modulation: a subtype-selective agonist artificially constrained to shift every targeted cell's parameters by the same amount, so that means move but within-class variance does not. If the full model beats this ablation, the mechanism is variance shaping specifically, not parameter shifting in general.

Why not just…?

Why not just use standard machine learning on chemical descriptors to predict selectivity? Because that solves a different problem. Lombardo et al. showed a one-step Extra Trees model on Mordred descriptors reaches 0.936 accuracy for Sigma-1 receptor activity prediction and ROC-AUC 0.964 for Sigma-2 Prediction of Activity and Selectivity Profiles of Sigma Receptor Ligands Using Machine Learning Approaches, and for predicting binding that may well be enough. But that same model correctly identified only 4 of 8 truly selective compounds — the minority class that matters for discovery — and descriptor-based approaches cannot represent the case where selectivity is driven entirely by pocket dynamics between subtypes with identical binding residues Constrained catecholamines gain β2AR selectivity through allosteric effects on pocket dynamics. More fundamentally, none of these models outputs anything about circuits: they stop at the molecule. Our proposal consumes the selectivity profile as an input and predicts what happens to computation. If a reviewer's goal is binding prediction only, the alternative wins and we do not contest it.

Why not just tune the synaptic weights and skip intrinsic heterogeneity? The literature directly answers this: intrinsic parameter diversity outperformed structural connectivity variation as the driver of computational capacity in microcircuit simulations Leveraging heterogeneity for neural computation with fading memory in layer 2/3 cortical microcircuits, optimally tuned homogeneous networks were still beaten by heterogeneous ones on SHD and SSC Neural heterogeneity promotes robust learning, and mean-field analysis shows homogeneous populations synchronizing pathologically under shared input in a way weight tuning does not easily remove Macroscopic dynamics of neural networks with heterogeneous spiking thresholds. We are also honest about the boundary: heterogeneity gains are negligible on static tasks like Neuromorphic MNIST, so we restrict claims to temporal computation and include the weight-retraining baseline above to settle the question empirically rather than by citation.

Why not just solve this with structural biology — find the allosteric pocket and dock? Structure tells you the affinity vector $a$, and we take that as given; indeed the field keeps improving at it, from extended pockets in HCAR2 Structural insights into ligand recognition and selectivity of the human hydroxycarboxylic acid receptor HCAR2 to computationally reprogrammed allosteric pathways Computational design of allosteric pathways reprograms ligand-selective GPCR signaling. But even at the single-protein level, static docking already fails: β2AR selectivity lives in vestibule dynamics and association kinetics, not in any static structure Constrained catecholamines gain β2AR selectivity through allosteric effects on pocket dynamics. And the deeper gap is that binding endpoints cannot predict circuit outcomes at all: the Kv1.1 study shows the same channel change produces distinct circuit and behavioral effects when confined to one interneuron class Kv1.1 channel dysfunction in parvalbumin-positive interneurons contributes to anxiety-like behaviors in young adult presenilin 1/2 conditional double knockout mice. Structure is a real threat to binding claims, not to ours; our Experiments 1–3 are precisely the measurements docking cannot make.

Decisive Checks

The proposal lives or dies on two comparisons. First, selective versus non-selective modulation at matched occupancy across all three circuit benchmarks: if the outcomes are indistinguishable, or if the variance-preserving ablation matches the full model, then selectivity adds nothing beyond mean excitability changes and the core claim is false. Second, the Neuromorphic MNIST control: any improvement there, where intrinsic heterogeneity is known to be computationally inert, would indicate an artifact rather than the proposed mechanism. A positive result — opposite computational outcomes from opposite selectivity profiles at equal occupancy, surviving realistic parameter distributions — would give receptor subtype selectivity a quantitative, circuit-level justification: selectivity as the biological actuator of a computational knob, and heterogeneity theory its downstream readout.