The Unreasonable Effectiveness of Cell Types in Describing Neuronal Physiological Features

Cluster-level cell type representations generally outperform highly variable gene selection, ion channel-coding gene selection, and context-enriched scGPT embeddings in predicting neuronal electrophysiological features. Modeling the relation between transcriptomic and electrophysiological modalities remained a challenge; this work systematically compares traditional discrete cell classification and foundation-model representations on the same paired dataset. Based on paired transcriptomic and electrophysiological patch-sequencing data from 495 human neurons, providing a common data basis for comparing multiple approaches.

Prediction performance varies across model architectures and initializations. This indicates that modeling choices lead to performance differences rather than one approach dominating under all settings. Derived from comparisons across multiple model architectures and initializations.

The best results are obtained by combining the outputs of separate cell type and scGPT-based models. This suggests traditional discrete cell classification can complement pretrained transformer models rather than being replaced by them. Derived from comparing combined model outputs against individual approaches.

The results apply to human neurons from neurosurgical tissue in a paired transcriptomic and electrophysiological patch-sequencing setting, providing a basis for choosing cell type representations or combined models on similar paired data; they are directly relevant to researchers seeking to combine traditional classification with foundation models.

This is a summary-level reading without figures or full methodological detail; the specific behavior of model architectures and initializations and the reproducibility of the combination approach require confirmation from the original text, and applicability to other species, brain regions, or recording conditions remains an open question.