A neuron's excitability is how easily and how fast it fires in response to input. It is set largely by which ion channels the cell carries in its membrane. The best existing attempt to predict excitability from a neuron's RNA, built on 955 mouse motor-cortex cells recorded with Patch-seq (a method that records electrical activity and sequences RNA from the same cell), explains only 17% of the variance in fitted channel strengths (Bernaerts et al.). That number is now often read as evidence that RNA is a poor guide to channel function, with the gap blamed on "transcript-protein discordance": mRNA levels that do not track protein levels. Protein-level evidence makes the story plausible. Across the human proteome, mRNA tracks protein only 40–60% of the time (Adhikari et al.), and the channels that matter most are among the hardest proteins to measure.
We argue that the 17% figure cannot carry that conclusion, because nobody knows what the maximum achievable number is. Three things suppress it that have nothing to do with biology. Patch-seq transcripts are noisy. The prediction targets are estimates from a model fit, and that fit is degenerate: many different channel combinations produce the same electrical behavior. And a score computed per parameter penalizes predictions that are wrong only in directions the recording cannot detect. This proposal estimates a noise ceiling for transcript-to-excitability prediction: the best score any predictor could reach given measurement noise and degeneracy. It then asks how much of the gap between 17% and that ceiling a flexible, uncertainty-aware model can close. Only the gap that remains after this is a real candidate for discordance. Last, we test whether that remaining gap varies systematically with a cell's position in the cortex, using a spatial atlas of the same brain area.
Core Claim
The question "does RNA predict channel function?" has never been asked against a calibrated ceiling, so the field cannot tell three explanations apart: noisy measurements, a misleading score, or real post-transcriptional control. We will compute that ceiling on the motor-cortex Patch-seq data and score predictions in electrophysiology space, by simulating the predicted channel parameters and comparing the result to the recorded traces. If a reasonable model reaches most of the ceiling, then transcripts are not the bottleneck at this level of measurement, and the discordance explanation is unsupported for these cells. If a large gap remains, the size of that gap is the first quantitative upper bound on how much excitability variance post-transcriptional mechanisms could account for. Whether the gap is laminarly structured within cell types then tells us whether it reflects something about the tissue environment.
Relation to Prior Work
Two papers motivate and supply the setting. Bernaerts et al. fit single-compartment Hodgkin-Huxley models (simulations of a neuron's voltage built from equations for each ion channel type) to the 955 Patch-seq cells. They use neural posterior estimation with added noise (NPE-N), a neural network that returns a distribution over channel parameters consistent with each recording. They then regress point estimates of those parameters on gene counts with sparse reduced-rank regression, a linear method that predicts many outputs at once through a few shared gene combinations, and obtain $R^2 = 0.17 \pm 0.03$. They also find that predictability collapses at fine cell-type resolution. We reuse their simulator and posteriors. We differ in scoring against a ceiling, predicting distributions instead of points, and evaluating in electrophysiology space.
Gouwens et al. fit over 130 million multi-compartment models to 230 visual-cortex neurons. They show that channel-strength profiles classify cell subclasses with 69% accuracy, against 81% for raw electrophysiology features. They also document that different channel combinations can produce indistinguishable firing. That degeneracy is exactly why per-parameter $R^2$ is a biased score. Their transcriptomes come from different cells than their recordings, so they cannot test within-cell prediction. We can.
On the method side, scChord shows on immune-cell data that predicting the average protein level from RNA over-smooths: it erases bimodal and heavy-tailed protein distributions, which flow matching (a generator trained to transport noise into the target distribution) recovers. We adapt that distributional idea to channel parameters. We also note what it constrains: scChord was trained on about 8,600 to 161,000 cells. With 955 cells, a heavy generative model is a risk rather than an obvious win.
Bush & Jin model how activity drives transcription and local translation with lags. Tang et al. show that cholinergic signaling changes a single potassium current, the afterhyperpolarization current that slows firing after spikes, without any change in transcription. Both papers motivate the hypothesis that part of the residual is real post-transcriptional control. Neither gives a way to measure how large that part is.
On the spatial side, the BARseq motor-cortex atlas maps 11 sublayers and about 24 projection types but contains no electrophysiology. SpaOTsc supplies optimal-transport mapping of dissociated cells into spatial coordinates. Spot-deconvolution methods such as stereoscope and the leaders of the 50-dataset benchmark (CARD, Cell2location, Tangram) recover the mix of cell types within a spot, not per-cell physiology. Our spatial step uses mapping only as a diagnostic for structure in the residual, not as the headline product.
Research Program
The first step builds the ceiling. It combines three measurable quantities.
Transcript reliability comes from count thinning. Each cell's reads are randomly split into two halves, and the agreement of the channel-gene profiles between halves is measured, with the correction for halving applied.
Target reliability comes from repeated current-injection sweeps within a cell, applied to the electrophysiology features themselves: spike width, rheobase (the minimum current that makes the cell fire), and adaptation.
Degeneracy is quantified from the NPE-N posteriors. If $\theta$ is a channel parameter and $x$ a cell's recording, the largest per-parameter $R^2$ that any predictor of the fitted point estimate can earn is bounded by
$$R^2_{\max} \approx r_{gg}\, r_{yy}\left(1 - \frac{\mathbb{E}_x[\mathrm{Var}(\theta \mid x)]}{\mathrm{Var}(\theta)}\right),$$
where $r_{gg}$ and $r_{yy}$ are the transcript and target reliabilities, and the bracketed term is the fraction of parameter variance that the recording actually pins down. The reliability factor is the classical attenuation correction for correlations between noisy variables. The bracket follows from the law of total variance. The bound is approximate because it assumes roughly additive, independent noise, so we will check it on simulated cells where the true parameters are known.
The second step changes the score. Predicted parameter distributions are pushed through the simulator. Performance is measured as the error between simulated and recorded electrophysiology features, scaled by within-type variance. This score does not penalize errors along degenerate directions.
The third step fits predictors of increasing flexibility against this ceiling. These range from linear models on ion-channel genes, through kernel and tree models, to a conditional density model over $\theta$ that is regularized toward cell-type priors. We report each model's fraction of the ceiling, both overall and within transcriptomic types.
The fourth step maps the Patch-seq cells and their residuals into the BARseq atlas by transcriptomic optimal transport. It then asks whether the within-type residuals vary with inferred sublayer more than expected under permutation. One constraint matters here. BARseq uses a targeted gene panel that may contain few channel genes, so inferred position depends on the panel genes the cell shares with the atlas. We treat any spatial signal as a hypothesis to follow up, not as a measurement.
Baselines
The naive reference point assigns every cell the mean channel parameters, or the mean electrophysiology features, of its transcriptomic type. This is the "cluster equals biophysical archetype" view. It is the strongest cheap competitor, because Gouwens et al. show that much of the conductance signal is between-type. We found no reported result on Lacuna for type-mean $R^2$ on the motor-cortex Patch-seq parameters, so we will compute it. Any proposed model must beat it on within-type explained variance, since overall $R^2$ can be inflated by type identity alone.
The strongest existing method is sparse reduced-rank regression from Bernaerts et al., with a reported cross-validated $R^2 = 0.17 \pm 0.03$ on predicting biophysical parameters across 955 motor-cortex neurons. We will rerun it on identical cross-validation folds, gene sets, and posterior targets, and confirm that we reproduce roughly that value before comparing. All models get the same hyperparameter-search budget and the same nested cross-validation. Beating it means a higher fraction of the ceiling, with non-overlapping fold-level intervals, on both the parameter-space and electrophysiology-space scores.
A direct regression from genes to electrophysiology features, with ridge and random-forest variants, skips the simulator entirely and serves as the second strong baseline.
Two ablations isolate the key ingredients. The first scores the identical best model in parameter space versus electrophysiology space. The second compares point-estimate targets with full-posterior targets. If neither change moves conclusions, the degeneracy argument is wrong.
Why not just…?
Why not just regress genes directly onto electrophysiology features? This is a real threat, and we include it as a baseline rather than argue against it. Direct regression avoids degeneracy by construction. The contribution here is the ceiling, not a cleverer predictor. If ridge on channel genes reaches most of the ceiling, that is a clean and useful answer: transcripts suffice, and the 17% reflected the choice of target. The simulator-based route matters only if it explains different variance, for example by sharing structure across features through biophysics. The side-by-side fraction-of-ceiling comparison settles this.
Why not just assign each cell its type's average excitability? For between-type questions this may be enough, and Gouwens et al. suggest types carry much of the signal. But the discordance question is a within-type question: do two cells of the same type with different channel mRNA levels differ in function? Type means cannot answer it by definition. Reporting within-type ceiling fractions shows directly whether anything beyond type identity is predictable.
Why not just run a top spot-deconvolution method on Visium cortex and attach type-level physiology? Methods ranked highly in the benchmark estimate cell-type proportions per multi-cell spot. Attaching physiology then reduces to the type-mean baseline painted onto tissue. It inherits the "platform effect" bias the benchmark documents and adds no per-cell information. We use single-cell-resolution BARseq instead and reserve spatial analysis for testing residual structure.
Why not just assume channel mRNA equals channel function? That assumption is exactly the null hypothesis the ceiling tests. It is not something to grant up front, given the 40–60% concordance reported by Adhikari et al. and the transcription-independent current modulation shown by Tang et al.. The experiment quantifies how much room is left for it to fail.
Proposed Experiments
The primary dataset is the 955-neuron motor-cortex Patch-seq set with the NPE-N model of Bernaerts et al.. Spatial residual analysis uses the BICCN BARseq motor-cortex atlas. A secondary consistency check uses the Allen visual-cortex data from Gouwens et al.. Because its transcriptomes and recordings come from different cells, that check can only compare subclass-level conductance predictions.
The first decisive check validates the ceiling on synthetic cells. We draw parameters, simulate recordings, generate transcripts from a known linear-plus-noise map, and confirm that the estimated ceiling matches the achievable $R^2$. If the bound is badly off, it is revised before any real-data claim is made.
The second check is the headline number: the best model's fraction of the ceiling, overall and within type. A within-type fraction near 1 argues against discordance as the main limit for somatic excitability in these cells. A fraction well below 1 with a high ceiling bounds the post-transcriptional share.
The third check is a permutation test of within-type residuals against atlas-inferred sublayer. Because inferred position is itself computed from transcripts, a positive result is interpreted as atlas-borrowed structure that 955 cells alone could not learn. We confirm or refute that reading by testing whether adding sublayer improves held-out prediction over the transcript-only model.
All code, splits, and ceiling estimates will be released, so that later work on channel-protein measurement can be scored against the same bound.
References
- Combined statistical-biophysical modeling links ion channel genes to physiology of cortical neuron types
- A high-stringency blueprint of the human proteome
- Single-neuron models linking electrophysiology, morphology and transcriptomics across cortical cell types
- scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation
- A unified computational model of the genetic regulatory networks underlying synaptic, intrinsic and homeostatic plasticity
- Cholinergic Modulation Preserves Spike Timing Under Physiologically Realistic Fluctuating Input
- Cellular anatomy of the mouse primary motor cortex
- Inferring spatial and signaling relationships between cells from single cell transcriptomic data
- Single-cell and spatial transcriptomics enables probabilistic inference of cell type topography
- A comprehensive benchmarking with practical guidelines for cellular deconvolution of spatial transcriptomics