henhaobio
Sign in

bioRxivNeurosciencePreprint

The Unreasonable Effectiveness of Cell Types in Describing Neuronal Physiological Features

Harshil SharmaiD, Xiao-Ping LiuiD, Thomas ChartrandiD, Brian E. KalmbachiD, Ed S. LeiniD, Ştefan MihalaşiD, Zhixin LuiD

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Single-cell RNA sequencing (scRNA-seq) captures detailed gene expression profiles at scale, while patch-clamp recordings measure intrinsic neuronal electrophysiological properties. Modeling the relations between these two modalities remains a challenge. Here, we compare how well electrophysiological features can be predicted by traditional transcriptomic cell type classification, representations derived from a foundational model (scGPT) pretrained on large-scale scRNA-seq datasets, ion channel-coding genes, and highly variable genes. Using paired transcriptomic and electrophysiological patch-sequencing data from 495 human neurons from neurosurgical tissue, we find that cluster-level cell type representations generally outperform highly variable gene selection, ion channel gene selection, and context-enriched scGPT embeddings. Notably, performance varies across model architectures and initializations, and the best results are obtained by combining the outputs of separate cell type and scGPT-based models. Together, these findings suggest that traditional discrete cellular classification is highly effective in predicting physiological features. For maximum performance it can be complemented by pretrained transformer models.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (16 Sept 2026) · CC BY 4.0

On 495 human Patch-seq neurons, cell type means predicted electrophysiology better than neural networks on scGPT embeddings or gene panels; fusing scGPT embeddings with cell type labels predicted best.

Predicting physiology from transcriptomes would let large expression atlases be read in functional terms, but Patch-seq (patch-clamp plus RNA-seq in the same cell) data are scarce for human neurons, and earlier gene-to-physiology models (linear models, GLMs, autoencoders) were built mainly on mouse cells. This work compares, on one human dataset with fixed splits, whether scGPT foundation-model embeddings carry physiological information beyond gene panels or cell type labels.

Key results

  1. ShownPer-cell-type averages of electrophysiological features out-predicted networks given only continuous transcriptomic inputs. Of these inputs, scGPT embeddings did best, reaching or slightly passing the baseline on a few features. Highly variable gene and ion channel gene models trailed.

    Per-feature test R^2 on eight percentile-normalized features, 495 neurons, stratified four-fold cross-validation, with fold standard deviations shown in figures (no R^2 values in the text). Inputs: 512-dimensional scGPT embedding, top 512 highly variable genes (HVGs), 328 ion channel genes. Comparator: training-set means for each of 15 cell types. · Baseline comparison; Data used; Models and training

  2. SuggestedWhen a one-hot cell type vector was concatenated to the input, the scGPT model was the only one to beat the cell type means baseline. The HVG and ion channel gene models still fell short even with cell type available.

    Same multilayer perceptron (MLP) and splits, with a 15-dimensional one-hot cell type vector added to each representation. The scGPT plus cell type model had the highest overall R^2 across features (Fig 3). The models reached similar training performance, but loss curves (S10 Fig) showed the HVG and ion channel models failing to generalize to the test set. The text gives no numeric R^2 and no significance test. · Comparison given cell type information; Comparison given cell type information and…

  3. SuggestedStarting the combined networks from a cell-type-only MLP, with weights for the extra inputs set to zero, removed the differences: every model ended at or very near the cell type baseline. The authors read this as the cell type prior keeping optimization at a solution that ignored the transcriptome.

    The pretrained one-hot network had 15 inputs, two hidden layers of 512 units and 8 outputs. To avoid leakage, a separate one-hot network was trained for each inner fold and for each outer training set, and used to initialize the matching larger network. No model clearly beat the others (Fig 4). · Comparison given cell type information and…; Discussion

  4. SuggestedA dual MLP gave the best predictions in the study when paired with scGPT embeddings. It runs cell type and transcriptomic inputs through separate pretrained subnetworks, joined by a linear layer that starts as their average. The HVG and ion channel versions barely changed from the previous test.

    Three dual MLPs (scGPT, 512 HVGs, 328 ion channel genes, each paired with the one-hot subnetwork) were retuned by nested cross-validation. The scGPT dual MLP beat every other model tested (Fig 5). In test loss curves (S10 Fig), the scGPT dual MLP showed some learning while the other two learned negligibly and failed to generalize. · Dual MLP fusion models; Discussion

  5. SuggestedFor this task, embeddings from the scGPT model pretrained on brain cells did slightly better than those from the whole-body human scGPT model.

    Brain model pretrained on 13.2 million human brain cells vs whole-human model pretrained on 33 million cells from across the body. The comparison is reported only in supplementary S16 Fig, with no values in the text. · Data used

Methods, data and reproducibility

Multilayer perceptrons (MLPs) were trained to predict eight percentile-normalized electrophysiological features from scGPT cell embeddings, 512 highly variable genes, 328 ion channel genes or one-hot cell type, scored by per-feature R^2 under stratified four-fold cross-validation. Cell type was then added by concatenation, cell-type-pretrained initialization, or a dual MLP fusion.

  • Allen Institute Patch-seq of human neocortical layer 1 GABAergic interneurons and layer 2/3 glutamatergic neurons (SMART-Seq v4 snRNA-seq plus patch-clamp electrophysiology) · human · n = 495 neurons, 15 cell types (types with fewer than four cells removed)
  • Data and model files released with this study (Zenodo) · human · 10.5281/zenodo.20500624
Methods & tools
scGPT, PyTorch, Optuna, Scanpy, Seurat v3, HGNC database, SMART-Seq v4, Patch-seq
Reproducibility gaps
No package versions, hardware or run times are stated. The source Patch-seq datasets are credited to the original publications, with no accession given in the text.
Limitations (6)
  • 495 neurons is very few compared with the size of a transcriptome and the number of network parameters. This likely limited what complex models could gain and made results sensitive to architecture and initialization. · stated by the authors
  • Action potential features and latency were dropped, partly because they predicted poorly in exploratory analysis. The eight features kept therefore do not represent the full electrophysiological phenotype. · stated by the authors
  • snRNA-seq may miss within-type transcriptional variation, alternative splicing, spatially localized transcription and time lags between transcription and phenotype. This may have held back the gene-panel models. · stated by the authors
  • Folds were stratified by cell type rather than donor, and donors differed in pathology and demographics, so donor-level effects may have contributed to the reported performance. · stated by the authors
  • Only layer 1 interneurons and layer 2/3 glutamatergic neurons from human cortex were used. The contributions of broad classes versus fine types were not separated, and whether the results hold in deeper layers, other regions or larger mouse datasets is untested. · stated by the authors
  • The text gives no R^2 values or statistical tests, and rankings rest on figure bars over only four folds. So the size and reliability of the scGPT dual MLP's lead over the cell type baseline cannot be judged from the text. · noted in this summary

Read this if you are weighing single-cell foundation model embeddings against cell type labels for predicting neuronal physiology, or combining discrete taxonomies with continuous embeddings on small Patch-seq datasets.

Citations
0

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISBernaerts '25Tripathy '18Kendrick '24Bomkamp '19Nandi '22Michiels '20
70% semantic matchNeuroscience
Yves Bernaerts, Michael Deistler, Pedro J. Gonçalves+7 more·Patterns·Aug 5, 2025· 10 citations