Conductance-Aware Signature Reversal: Ranking Drugs by Predicted Rescue of Ion Currents

Many drug-discovery pipelines now work like this: measure which genes are abnormally high or low in a diseased cell, then search a reference database of drug-induced gene changes for compounds that flip the disease pattern back toward healthy — a strategy called transcriptomic signature reversal. The trouble is that the endpoint being optimized is mRNA abundance, not function. For ion channels — the membrane proteins that carry the tiny electrical currents controlling heartbeat, nerve firing, and secretion — this gap is not a technicality. In the best paired dataset available, where the same individual neurons gave both a full transcriptome and a patch-clamp recording (a technique in which a glass pipette seals onto the cell and measures picoampere ionic currents while holding voltage constant), ion-channel mRNA levels explained only a modest fraction of the variance in actual membrane conductances: cross-validated $R^2 = 0.17 \pm 0.03$ Combined statistical-biophysical modeling links ion channel genes to physiology of cortical neuron types. Worse, functional deficits can arise entirely outside the transcriptome: a missense mutation in titin — a structural protein, not a channel — cut peak sodium current by 51.64% and L-type calcium current by 71.56% in patient cardiomyocytes through failed channel trafficking, a defect invisible to expression profiling A pathogenic titin missense mutation in hiPSC-derived cardiomyocytes predisposes to ventricular fibrillation in acute ST-segment elevation myocardial infarction.

This proposal asks a single, sharp question: when signature reversal nominally "reverses" a disease state, does it actually restore the electrical currents? The answer today is that nobody knows, because no reversal pipeline has ever been checked against an electrode. We will build the first end-to-end test: a learned map from transcriptome to biophysical conductance, trained on paired Patch-seq data, used to re-rank reversal candidates by their predicted restoration of ion currents rather than their transcript inversion, and then validated by real patch-clamp recordings scored with ROC analysis (receiver operating characteristic curves, which sweep every possible decision threshold and summarize discrimination as the area under the curve, AUC — equivalently, the probability that a randomly chosen positive case scores above a randomly chosen negative one).

Core Claim

The claim is two-sided and falsifiable. First, ranking drugs by predicted conductance rescue — pushing the predicted conductance vector back toward the healthy configuration — should discriminate electrophysiologically rescuing compounds from non-rescuing ones better than ranking by transcript reversal alone, because the transcript–conductance gap ($R^2 \approx 0.17$) decouples the two objectives exactly where the therapy must act. Second, ROC analysis gives this comparison an honest, threshold-free yardstick where the electrophysiology field currently has none: neither the functional-clustering work on 2,378 ion-channel models (211 clusters, with no reported discrimination metric — no reported result found on Lacuna for a classification AUC in that study) Mapping the function of neuronal ion channels in model and experiment nor the causal multi-channel variational autoencoder on simulated cardiotoxicity data (which reports only reconstruction error; no reported numerical MSE or AUC value found on Lacuna) Assessing ion channel blockade and electromechanical biomarkers' interrelations through a novel Multi-Channel Causal Variational Autoencoder has ever quantified how well any score separates rescuing from non-rescuing drugs across all thresholds.

Research Program

The program has three stages. Stage A builds the map. Starting from the 955-neuron adult mouse motor cortex Patch-seq dataset — the only resource in this space pairing single-cell transcriptomes with whole-cell recordings from the same cells — we reproduce the sparse reduced-rank regression baseline of Bernaerts et al., which predicts conductance parameters from channel-gene expression, and then try to beat it, using their simulation-based-inference conductance posteriors as richer prediction targets and nonlinear regressors on the same cross-validation splits Combined statistical-biophysical modeling links ion channel genes to physiology of cortical neuron types. We deliberately rejected a brute-force design that simulates every compound's effect forward through a generative model: on the perturbation-prediction benchmarks, inverse formulations trained up to 25 times faster than iterative simulators like scGen while retrieving up to 13.37% more ground-truth targets Combinatorial prediction of therapeutic perturbations using causally inspired neural networks, and our map is exactly such an inverse object. Critically, we do not need the map to be accurate point-by-point; we need it to be directionally accurate, so Stage A reports both cross-validated $R^2$ and the cosine agreement between predicted and measured conductance-change vectors on held-out perturbations. A map that cannot beat the sRRR baseline on either metric is a published negative result, and the project stops there honestly.

Stage B turns the map into a ranking. We project drug perturbation signatures from LINCS L1000 / Connectivity Map (66,612 profiles covering 12,442 compounds, as used in the RGES study Reversal of cancer gene expression correlates with drug efficacy and reveals therapeutic targets) into the map's input space, obtaining for each compound a predicted conductance shift $\Delta g_c$. Given a disease with a measured conductance deficit $\delta = g_{\text{disease}} - g_{\text{healthy}}$, the conductance-rescue score is simply alignment against the deficit:

$$S_c = -\,\Delta g_c \cdot \hat{\delta}$$

where $\hat{\delta}$ is the unit vector of the deficit, so compounds predicted to push conductances opposite to the disease direction score high. The same candidate pool is scored in parallel by the existing transcript-reversal methods — the Reverse Gene Expression Score of Chen et al. and the module-level latent-variable matching of PhenoPLIER, which projects both disease associations and perturbation profiles into 987 co-expression modules Projecting genetic associations through gene expression patterns highlights disease etiology and drug mechanisms. We also address a known weakness head-on: LINCS measures only ~1,000 landmark genes in immortalized cell lines, so we bridge cell types through the module-level latent space and, in Stage C, calibrate the map in the correct cell type directly.

Stage C is the electrophysiological verdict, in the titin-mutant hiPSC cardiomyocyte system of Ji et al. — chosen because the conductance deficit is already quantified ($I_{\text{Na}}$ down 51.64%, $I_{\text{Ca-L}}$ down 71.56%, shortened APD50/APD90) and because that study's own gaps define our improvements: we add the isogenic CRISPR-corrected controls it lacked, RNA-seq of diseased versus corrected cells under both baseline and hypoxic/ischemic stress, and drug testing under the stress conditions it never used A pathogenic titin missense mutation in hiPSC-derived cardiomyocytes predisposes to ventricular fibrillation in acute ST-segment elevation myocardial infarction. We treat cells with a panel of roughly forty compounds: the top conductance-rescue picks, the top transcript-reversal picks (deliberately including compounds that rank high on one score and low on the other, plus random controls), record standardized multi-step voltage-clamp protocols (activation, inactivation, ramp, action-potential clamp, following the protocol suite used to functionally classify ion-channel behavior Mapping the function of neuronal ion channels in model and experiment), and label each compound a rescuer or non-rescuer by pre-registered criteria on $I_{\text{Na}}$, $I_{\text{Ca-L}}$, and action-potential duration. ROC curves then convert each continuous score into a discrimination measure. Because most compounds will likely fail to rescue, the labels will be class-imbalanced — a setting where ROC curves can flatter a marker — so precision–recall curves serve as the secondary, imbalance-robust readout, and any eventual clinical operating point will be chosen with cost-weighted ROC analysis rather than raw AUC Weighted ROC Curve in Cost Space: Extending AUC to Cost-Sensitive Learning.

Baselines

The naive reference point is a random compound ranking and vehicle-only cells: random ranking gives AUC 0.5 by construction, and vehicle-treated cells define zero rescue, so any useful score must clear both. The strongest existing competitor on the ranking question is module-level transcriptomic reversal: PhenoPLIER's latent-variable matching achieved an AUC of 0.632 against known drug–disease indications, beating raw gene-level matching at 0.566 on the same benchmark Projecting genetic associations through gene expression patterns highlights disease etiology and drug mechanisms, and gene-level reversal itself has experimental credibility through the sRGES–efficacy correlation of Spearman's rho 0.61 ($P = 1.34 \times 10^{-6}$) in liver cancer Reversal of cancer gene expression correlates with drug efficacy and reveals therapeutic targets. These are the right competitors because they are the current best reversal methods with published, threshold-free performance — and 0.632 is the number our conductance-aware score must beat in the patched panel. All ranking methods will run on the identical disease signature, identical candidate pool, and identical compute and tuning budget, with hyperparameters selected on the computational benchmark before any wet-lab compound is chosen, so no method sees the validation labels early. For the sub-task of predicting conductances from expression, the baseline is sparse reduced-rank regression, whose cross-validated $R^2$ of 0.17 ± 0.03 on the 955-neuron Patch-seq dataset Combined statistical-biophysical modeling links ion channel genes to physiology of cortical neuron types is the number our map must materially exceed — in $R^2$ or in directional accuracy. Finally, the key ablation isolates the proposal's central ingredient: the identical pipeline, identical compounds, identical measurements, but with compounds ranked by transcript reversal alone, bypassing the conductance map. If the ablation matches the full method, the conductance map adds nothing and the transcript–function gap is irrelevant to drug ranking — a clean, informative null.

Why not just…?

Why not just find the single most dysregulated channel and block or boost it directly? The single-target route fails in two documented ways in this literature. Statistically, single-gene approaches miss functional drivers: in PhenoPLIER's HepG2 CRISPR screen, the causal lipid regulators DGAT2 and ACACA had no significant individual associations and were recovered only through co-expression modules, and gene-level reversal underperformed module-level reversal (AUC 0.566 versus 0.632) Projecting genetic associations through gene expression patterns highlights disease etiology and drug mechanisms. Biologically, the driver may not be a channel at all: in the very model system we use, the lethal current deficit comes from a structural protein (titin) deranging channel trafficking, so no channel-subunit expression screen would ever point at the mechanism A pathogenic titin missense mutation in hiPSC-derived cardiomyocytes predisposes to ventricular fibrillation in acute ST-segment elevation myocardial infarction. Multi-target, network-level reversal is the natural answer, and our proposal simply holds that answer to a functional standard.

Why not just patch-clamp a compound library directly and skip the computation? This is the most serious objection, and we are honest about it: for a small library and a single cell type, a direct automated screen would work and would be simpler. It fails as a general strategy for three reasons grounded in the file. Manual patch clamp is low-throughput and labor-intensive by reputation Electrophysiological Measurement of Ion Channel Currents; combinations of perturbagens explode combinatorially beyond any physical screening capacity, which is precisely the regime where inverse prediction was shown to dominate forward search Combinatorial prediction of therapeutic perturbations using causally inspired neural networks; and pure electrophysiological screening is mechanism-agnostic, blind to off-target liabilities that structured validation can catch — the Aurora A–TPX2 study achieved 19 nM binding with no detectable activity across 97 kinases precisely by profiling selectivity rather than screening raw Selective Aurora A-TPX2 Interaction Inhibitors Have In Vivo Efficacy as Targeted Antimitotic Agents. Still, the objection predicts a measurable outcome: if the top ten computational picks show the same rescue rate as ten random compounds in our panel, the computational layer earns nothing, and the experiment is designed so that this result would be unambiguous.

Why not just pick an expression cutoff and report sensitivity and accuracy at that threshold? Fixed thresholds are exactly where biomarker evaluation goes wrong in this literature: operating points drift across laboratories and platforms due to batch effects and calibration differences, which is why distributed validation work had to build modified ROC generalized linear models to compare markers across sites without sharing raw data Distributed non-disclosive validation of predictive models by a modified ROC-GLM, and why the ROC primer frames AUC as the threshold-free probability interpretation of a continuous marker Quantifying Diagnostic Accuracy Using ROC Curve Analysis. The field's own practice confirms the point: AUC filtering at >0.75 is what separated strong from weak candidates in the endometriosis biomarker study (MRC1 at 0.8864, SYK at 0.7645) Identification and Exploration of Novel Macrophage M2-Related Biomarkers and Potential Therapeutic Agents in Endometriosis, and serum miR-4636 reached 0.963 as a threshold-free discrimination measure in gastric cancer miR‑4636 inhibits tumor cell proliferation, migration and invasion, and serves as a candidate clinical biomarker for gastric cancer. Our comparison of two ranking scores is meaningless at a single arbitrary cutoff; only the full curve makes the comparison fair to both methods.

Decisive Checks

Three checks decide the proposal. First, the go/no-go gate on the map: cross-validated $R^2$ and directional accuracy against the sRRR baseline of 0.17 ± 0.03 on the 955-neuron Patch-seq dataset Combined statistical-biophysical modeling links ion channel genes to physiology of cortical neuron types; if neither improves, the ranking score is noise and we report that the transcript–function gap blocks conductance-aware repurposing. Second, a fully computational check with real labels: on the CiPA-style dataset of 109 drugs (37 torsadogenic, 72 not) used by the multi-channel causal autoencoder Assessing ion channel blockade and electromechanical biomarkers' interrelations through a novel Multi-Channel Causal Variational Autoencoder, we test whether the conductance-rescue score discriminates arrhythmogenic from safe drugs by ROC-AUC — a number that study never reported, so we would be setting the first benchmark rather than beating one. Third, the wet-lab verdict: in the titin-mutant hiPSC cardiomyocytes with isogenic controls and ischemic stress, the conductance-rescue score must achieve a bootstrap-corrected ROC-AUC for electrophysiological rescue significantly above both 0.5 and the transcript-reversal baseline's AUC in the same panel — with PhenoPLIER's published 0.632 as the external reference point for what good reversal discrimination looks like Projecting genetic associations through gene expression patterns highlights disease etiology and drug mechanisms. A further bonus signal, in the spirit of Chen et al.'s xenograft validation of computationally predicted reversal drugs Reversal of cancer gene expression correlates with drug efficacy and reveals therapeutic targets, would be demonstrating that a compound ranked highly only by the conductance-aware score — and poorly by transcript reversal — measurably restores sodium and calcium currents, proving the two objectives genuinely diverge and that the electrode, not the RNA kit, picked the therapeutic.