Ion channels are membrane proteins that open and close to let specific ions cross the cell boundary. In neurons, it is settled fact that a channel's biological impact depends less on how many copies exist than on where on the cell they sit: the axon initial segment concentrates sodium-channel conductance density roughly 30- to 60-fold above the soma in cortical pyramidal neurons, and 4- to 9-fold in midbrain dopamine neurons, because that spatial gradient is what makes spikes initiate where they should Yang et al.. Strikingly, the same logic has begun to matter in cancer. Leslie et al. showed that breast tumors and their associated fibroblasts carry functional voltage-gated sodium channels (Nav1.5), that persistent sodium leak forces the sodium–potassium pump to burn ATP, that this drives accelerated glycolysis and extracellular acidification at the invasive tumor edge, and that high Nav1.5 protein in a 1,740-patient cohort independently confers a 2.18-fold metastasis hazard (95% CI 1.63–2.92).
But that study, like every prognostic ion-channel or metabolic signature we could find, measures channel abundance on homogenized scales — bulk immunohistochemistry intensity or bulk RNA-seq hazard coefficients. Nobody has asked the question neuroscience answered decades ago: does the prognostic and metabolic power of Nav1.5 come from where it clusters on the invading cancer cell? We propose to map the nanometer-scale topography of Nav1.5 and its metabolic partners at invasive protrusions, build a spatial model of the resulting ion–metabolic feedback, and test whether position-based features of the channel predict patient outcome better than the abundance-based marker already in the literature.
Core Claim
The core claim is twofold. First, mechanistically: Nav1.5 in breast cancer cells is not homogeneously distributed but concentrated, together with the sodium–potassium pump and proton extruders, in submicron actin-rich invasive protrusions (invadopodia), analogous to the ankyrin-and-spectrin-scaffolded clustering of sodium channels at the neuronal initial segment demonstrated by Abiraman et al. and mapped electrophysiologically by Yang et al.. Second, clinically: the spatial clustering of the channel — not its total abundance — is the variable that best predicts metastatic risk, because local channel density sets local sodium influx, local pump ATP debt, and hence the steepness of the pericellular proton gradient that degrades extracellular matrix at the invasive front.
The quantitative skeleton of the claim is a clustering index $C$, the ratio of invadopodial membrane conductance density to the whole-cell average $\bar{g}_{\text{Nav}}$, with the spatial model predicting that peak pericellular acidification scales with the product of the two:
$$C = \frac{g_{\text{Nav}}^{\text{invadopodia}}}{\bar{g}_{\text{Nav}}}, \qquad \Delta\mathrm{pH}_{\text{peak}} \approx \beta \, C \, \bar{g}_{\text{Nav}}$$
where $\beta$ is calibrated from measured extracellular acidification rates. The point of this relation is that two tumors with identical bulk Nav1.5 expression but different $C$ should produce very different invasive behavior — exactly the case that every existing prognostic marker, bulk or otherwise, is blind to.
Research Program
The program has three linked parts, with breast cancer as the application regime and subcellular channel mapping as the core mechanism.
The first part is direct mapping. We will perform multi-target super-resolution immunofluorescence on breast cancer cell lines, patient-derived primary tumor cells, and fresh surgical slices, co-labeling Nav1.5, the sodium–potassium pump, the proton exchanger NHE1, lactate transporters MCT1/4, ankyrin/spectrin scaffold components, and invadopodia markers such as cortactin and Tks5. This transplants the co-localization methodology Yang et al. used to resolve Nav isoforms at the neuronal initial segment onto the non-neuronal protrusions that Leslie et al. explicitly flagged as unmapped. Our concrete benchmark is neuronal: if the invadopodial-to-somatic Nav1.5 ratio is comparable to the 4- to 9-fold initial-segment gradient of midbrain dopamine neurons (44.9 vs 5.0 pS/µm² in substantia nigra cells), the clustering hypothesis has biophysical legs. We will then patch-clamp outside-out patches pulled from protrusion-rich membranes versus cell bodies — the same technique Zhang et al. used on intact axons — measuring persistent sodium current density at physiological and acidic extracellular pH, since Leslie et al. found acidosis doubles the persistent current and Yang et al. showed compartments sharing one gene can still differ in gating kinetics.
The second part is a spatial computational model: a multi-compartment reaction–diffusion system in which each membrane patch carries its measured local Nav1.5 conductance, pump density, and proton-extrusion capacity, coupled to intracellular ATP dynamics and extracellular proton diffusion. This extends the vicious cycle diagrammed by Leslie et al. from a well-stirred cartoon into a spatially resolved predictor of pericellular pH, in the spirit of the compartmental biophysical modeling of Abiraman et al. and the microdomain calcium–potassium coupling measured by Maureira et al., where channels and their downstream effectors sit within nanometers of each other. The model's decisive ablation compares matched cells that differ only in spatial arrangement: total channel number fixed, clustering index set to one (homogeneous) versus measured (clustered). If the homogeneous ablation reproduces the measured pericellular pH equally well, the topography claim is dead and we will say so.
The third part is prognostic translation. We will re-score the Nottingham tissue microarray of 1,740 breast cancer cases scored by Leslie et al. with an automated image-analysis topography score — punctate membrane clustering at the invasive edge versus diffuse staining — and ask, in nested Cox models, whether topography adds prognostic power beyond Nav1.5 intensity and the Nottingham Prognostic Index. We will complement this on TCGA-BRCA with a coupling-module score (co-expression of SCN5A with pump and proton-transporter genes), which the spatial model predicts should carry signal that SCN5A abundance alone does not, and we will validate the model's perturbation predictions against GSE228621, the Nav1.5-knockdown xenograft RNA-seq dataset deposited by Leslie et al., by testing whether model-predicted flux reduction tracks the observed downregulation of invasion genes. A natural extension, if the static results hold, is the dynamic question raised by Li et al. — whose longitudinal serological metabolomics tracked radiotherapy response with an AUC of 0.818 — namely whether therapy-induced acidosis adaptively upregulates persistent current and clustering; we treat this as follow-on, not core.
Baselines
The naive reference point is standard clinical staging alone: the Nottingham Prognostic Index and TNM-type variables in the same 1,740-patient cohort, which is the "do nothing" comparator any clinically readable result must clear. Beating it means the topography score remains an independent predictor of cancer-specific survival in a multivariate Cox model and improves the concordance index over a clinical-covariates-only model with a bootstrap confidence interval excluding zero.
The strongest existing competitor is the Nav1.5 bulk immunohistochemistry intensity score of Leslie et al., which is precisely the abundance-based version of our marker: in this cohort it carried a metastasis hazard ratio of 2.18 (95% CI 1.63–2.92), independent of tumor size, grade, and node status. This is the fairest possible comparison because it uses the same patients, the same slides, the same event definitions, and the same staining; we add only the image-derived spatial score. We beat it if adding topography to intensity yields a significant likelihood-ratio test and a concordance-index improvement of at least 0.02 with a bootstrap interval excluding zero; if the intensity-only model absorbs the topography signal entirely, our clinical claim fails even if the mechanistic mapping succeeds.
The strongest method-level competitor is bulk-transcriptomic deep prediction: BulkRNABert, a transformer pretrained by masked modeling on discretized expression values, which reports a pan-cancer survival concordance index of 0.765 Gélard et al., with public checkpoints and fine-tuning pipelines in the official codebase instadeepai/multiomics-open-research. We will fine-tune it on TCGA-BRCA ourselves under a tuning budget matched to our own Cox and coupling-module models, with identical cross-validation folds; no reported result found on Lacuna for BulkRNABert on the breast-cancer subset specifically, so we establish that number ourselves and require our spatial-coupling features to add to it, not merely match it. Finally, the key ablation is internal: the same spatial model with channel distribution forced homogeneous at matched total expression. This isolates spatial arrangement — the proposal's single new ingredient — from abundance, and it is the ablation whose failure would most cleanly refute the core claim.
Why not just…?
Why not just measure Nav1.5 intensity, as Leslie et al. already did? This is the real threat, because that measurement is already validated at n = 1,740 with a strong hazard ratio, and if topography adds nothing beyond intensity the proposal's clinical half collapses. Our answer is partly prior evidence and partly our own decisive test. In neurons, Yang et al. showed that somatic and axonal channels of the same isoform have different inactivation and recovery kinetics purely because of location, and Zhang et al. showed functional differences between neuronal classes track conductance density gradients, not expression identity — so abundance is known, in the best-characterized system available, to be an incomplete summary of channel function. But we do not lean on the analogy: the nested-model comparison on the Nottingham array is designed exactly to settle whether position adds predictive value, and we commit in advance to reporting a negative result if the likelihood-ratio test is null.
Why not just use bulk RNA-seq models — LASSO signatures or BulkRNABert — and skip biophysics? Because the bulk representation homogenizes exactly the compartment the mechanism lives in. Leslie et al. measured tumor periphery pH markedly below the core, meaning a bulk average mixes the acidic invasive front with the quiescent core and with cancer-associated fibroblasts that also carry currents; and BulkRNABert's pan-cancer concordance index of 0.765 Gélard et al. leaves most survival variance unexplained. There is precedent for channel genes slipping through such pipelines as unexplained statistics — the 9-gene colon signature of Wu et al. includes KCNJ11 (the ATP-gated potassium channel Kir6.2) as a bare Cox coefficient with no electrophysiology or localization, achieving 1-, 3-, and 5-year AUCs of 0.740, 0.723, and 0.763 — which is precisely the gap we fill rather than repeat. Our TCGA-BRCA experiment, adding coupling-module features to a BulkRNABert fine-tune under a matched budget, is the concrete test of whether the biophysical layer carries bulk-invisible signal.
Why not just use spatial transcriptomics like stMVC? The multi-view graph framework of stMVC does resolve tumor-boundary niches and metastatic transition states by fusing histology, expression, and spatial coordinates — but at 10x Visium spot resolution of roughly 55 micrometers, it averages dozens of cells per spot and cannot see a submicron invadopodium, nor can any transcript count measure ionic current, pump ATP demand, or proton flux, all of which are the quantities our mechanism actually turns on. That said, stMVC is complementary rather than merely a rival: if spot-level boundary features from stMVC fully explain our topography score's prognostic signal, that would genuinely cap the value of subcellular mapping, and we will run that comparison on the invasive-ductal-carcinoma Visium data it used as one of our checks.
Why not just measure glycolytic enzymes like PFKFB3? Because the canonical glycolytic marker demonstrably fails in the very cancer type where the channel marker works: the pan-cancer analysis of PFKFB3 found PFKFB3 lower in breast tumors than in normal breast tissue, while Nav1.5 carries a 2.18-fold metastasis hazard there. The channel sits upstream of glycolysis in this circuit — it is the physical trigger of the pump's ATP demand — so measuring the downstream enzyme loses both the direction and the localization of the driver.
Proposed Experiments and Decisive Checks
Experiment one tests the mapping claim: super-resolution co-localization in cell lines, patient-derived cells, and fresh tumor slices quantifies the invadopodial enrichment ratio $C$ for Nav1.5 and the nanometer distances to NHE1, MCT4, and the sodium–potassium pump, benchmarked against the neuronal initial-segment gradients of Yang et al. (44.9 pS/µm² axonal versus 5.0 pS/µm² somatic in substantia nigra neurons). The decisive readout is a clustered, scaffold-associated metabolon at protrusions versus diffuse membrane staining; scaffold disruption with actin and ankyrin-binding perturbations, motivated by the cytoskeletal packing constraints of Abiraman et al., should disperse the cluster and, per the model, blunt local acidification.
Experiment two tests the functional claim: outside-out patches from protrusion-rich versus somatic membrane, at extracellular pH 7.4 and 6.5, following the tumor-slice electrophysiology of Leslie et al. and the axonal patch methodology of Zhang et al., with acidosis-doubled persistent current as the positive control already established by Leslie et al. and the paired-compartment kinetic differences of Yang et al. as the precedent for what to look for.
Experiment three tests the model: the spatial reaction–diffusion model, calibrated on measured conductance densities and extracellular acidification rates, predicts the pericellular pH profile around single cells and in slice preparations, validated against pH-sensitive microelectrode measurements of the kind used by Leslie et al., and against the homogeneous-distribution ablation at matched total channel number. Success is reproducing the measured peak pericellular acidification with the clustered distribution while the homogeneous ablation underestimates it.
Experiment four is the clinical test and the one that decides the proposal: on the 1,740-case Nottingham array, nested Cox models compare clinical covariates plus intensity (the Leslie baseline, HR 2.18) against the same plus topography score, with concordance index and metastasis hazard as primary metrics and bootstrap intervals throughout; in parallel, on TCGA-BRCA, the coupling-module score must add to a matched-budget BulkRNABert fine-tune, and the model's perturbation direction must rank-correlate with the invasion-gene changes in GSE228621. If the topography score adds nothing anywhere, the honest conclusion is that in cancer, unlike neurons, abundance is the whole story — a negative result that would itself discipline both fields' next steps.
References
- Biophysical Properties of Somatic and Axonal Voltage-Gated Sodium Channels in Midbrain Dopaminergic Neurons
- A novel Nav1.5-dependent feedback mechanism driving glycolytic acidification in breast cancer metastasis
- Modeling of the axon membrane skeleton structure and implications for its mechanical properties
- Axon initial segment potassium channel density in cortical neurons
- P/Q and N Channels Control Baseline and Spike-Triggered Calcium Levels in Neocortical Axons and Synaptic Boutons
- Serum metabolic profiling enables diagnosis, prognosis, and monitoring for brainstem gliomas
- BulkRNABert: Cancer prognosis from bulk RNA-seq based language models
- Comprehensive analysis of mitochondrial-related gene signature for prognosis, tumor immune microenvironment evaluation, and candidate drug development in colon cancer
- Elucidating tumor heterogeneity from spatially resolved transcriptomics data by multi-view graph collaborative learning
- Glycolytic regulatory enzyme PFKFB3 as a prognostic and tumor microenvironment biomarker in human cancers