Predicting cellular electrophysiology with generative modeling

Neuronal cells exhibit distinct molecular, electrophysiological and morphological features, yet current technologies offer limited capacity to capture these jointly at high throughput. Here, we present PREPS (PRedicting ElectroPhysiology in Single-cell data), a computational approach that learns mappings from transcriptomes to electrophysiological features using coupled measurements of electrophysiology, transcriptome, and morphology by Patch-Sequencing (Patch-Seq). PREPS fine-tunes a transformer on >8 million brain cells to obtain contextual embeddings and attention-ranked gene lists, trains elastic-net models to predict electrophysiological features (e.g., resting potential, input resistance, action-potential metrics) from these embeddings, and assigns electrophysiology- and morphology-based cell-type labels (inhibitory, pyramidal, non-spiking, hybrid). We also introduce DEEPS (Deep ElEectrophysiology Pathway Scores), a weighted, attention-informed electrophysiology module-scoring scheme that quantifies synaptic and excitability programs. Applied to glioma single-cell datasets, PREPS identifies neuronal-like tumor subpopulations with concordant predicted electrophysiology, consistent with neuronal program adoption by glioma cells. Applied to the Allen Brain non-neuronal atlases, PREPS identifies rare neuronal-like cells within committed oligodendrocyte precursors (COP), OPCs, and astrocytes, with elevated ionotropic receptor, ion channel, and action-potential machinery scores. In summary, PREPS enables atlas-scale inference of electrophysiology from RNA alone, generating testable hypotheses on neuron-glia signaling in cancer and health. It also helps prioritize targets for validation and therapeutic repurposing.