Single-neuron models linking electrophysiology, morphology and transcriptomics across cortical cell types

Anirban Nandi, Thomas Chartrand, Werner Van Geit, Anatoly Buchin, Zizhen Yao, Soo Yeun Lee, Yina Wei, Brian Kalmbach, Brian Lee, Ed S. Lein, Jim Berg, Uygar Sümbül, Christof Koch, Bosiljka Tasic, Costas A. AnastassioubioRxiv 2020 · DOI

TL;DR

This paper presents a novel computational workflow to generate over 130 million biophysically detailed single-neuron models that integrate electrophysiology, morphology, and transcriptomics. These "all-active" models not only accurately replicate experimental data but also reveal how specific ion channel distributions explain cell-type differences and link molecular profiles to functional properties.

What Problem Does This Paper Address?

Understanding the diversity of cell types in the brain is a fundamental challenge in neuroscience. While we can characterize neurons by their electrical activity (electrophysiology), physical structure (morphology), and molecular makeup (transcriptomics), linking these modalities and understanding the causal relationships between them has been difficult. Specifically, how do genetic differences manifest as specific ion channel distributions, and how do these distributions, in turn, shape a neuron's electrical behavior? Traditional modeling approaches often simplify neurons or focus on limited datasets, hindering a comprehensive understanding of cell-type diversity.

Methods

The authors developed an automated, multi-stage optimization workflow to generate "all-active" single-neuron models. This workflow uses evolutionary algorithms to fit model parameters (ion channel conductances and passive properties) to experimental electrophysiology data while incorporating reconstructed neuron morphologies. The key innovation is the "all-active" approach, meaning ion channels are distributed across the entire neuron (soma, dendrites, axon), not just the soma, leading to more biophysically realistic models. The optimization process involves three stages: fitting passive properties, then h-channel conductances, and finally all active conductances. They also employed sensitivity analysis (Sobol indices) to understand the contribution of individual conductances to specific electrophysiological features. Classification was performed using Support Vector Machines and Random Forests.

Datasets

The study utilized two primary datasets:

  • Single-cell RNA sequencing (scRNA-seq) data: From the Allen Brain Atlas, providing transcriptomic profiles of over 1600 individual neurons from the mouse visual cortex.
  • Electrophysiology and morphology data: Also from the Allen Cell-Types Database, comprising whole-cell patch-clamp recordings and reconstructed morphologies of 230 individual neurons from the mouse visual cortex.

These datasets were used to train and validate the generated single-neuron models.

Key Insights and Ideas

  • "All-active" models are crucial for realism: Distributing ion channels across the entire neuron, not just the soma, significantly improves biophysical realism and is essential for capturing nuanced electrophysiological behaviors, even if perisomatic models can achieve similar overall training performance.
  • Conductance vectors are cell-type fingerprints: Distinct patterns of ion channel conductances across the neuron's morphology serve as reliable identifiers for different cortical cell types, even distinguishing between subclasses.
  • Molecular differences translate to functional differences: Differences in ion channel conductances predicted by the models correlate strongly with experimentally measured gene expression differences, providing a direct link from genotype to phenotype.
  • Causal links between modalities: The models offer a systematic way to establish causal relationships, showing how specific gene expression differences (e.g., in Kv3.1 or HCN channels) lead to predictable changes in ion channel conductances, which in turn explain observed electrophysiological differences between cell types.
  • Robustness to variability: Despite biological variability and the inherent ill-posed nature of model optimization, the generated models robustly capture cell-type identity, with within-class variability being smaller than between-class variability.

Key Results

  • Massive model generation: The workflow generated and evaluated over 130 million single-neuron models.
  • High accuracy: The best models accurately recapitulated experimental electrophysiology data, with errors typically within 2 standard deviations of experimental values.
  • Classification performance: Conductance vectors alone could classify broad cell subclasses with 69% accuracy, approaching the 81% accuracy achieved with experimental electrophysiology features. Combining conductance vectors with morphology improved classification to 75%.
  • Transcriptomic validation: Model predictions of differential ion channel conductances between inhibitory subclasses (e.g., Kv3.1 in Pvalb, Sst, Htr3a) were experimentally validated by scRNA-seq data. Similarly, differences in h-channel conductance between excitatory subclasses (Rbp4 vs. Nr5a1) correlated with HCN gene expression.
  • Electrophysiological explanation: The identified conductance differences explained key electrophysiological distinctions between cell types, such as AP width, AHP depth, and firing rate.

Limitations and Future Work

  • Model Recipe Dependency: The models are constrained by the initial set of ion channels included in the optimization recipe. Future work could incorporate class-specific channel recipes informed by transcriptomics to improve model accuracy and reduce computational demand.
  • Computational Cost: While the workflow was optimized, generating millions of models remains computationally intensive.
  • Patch-seq Integration: The authors anticipate that future datasets combining morphology, electrophysiology, and transcriptomics from the same cell (e.g., Patch-seq) will enable even more precise mapping between model parameters and cell types.
  • Gene Encoding Complexity: Reconciling differences in ion channel conductances with the expression of multiple genes encoding subunits of a single channel remains a future challenge.

Summary updated

Abstract

Abstract Identifying the cell types constituting brain circuits is a fundamental question in neuroscience and motivates the generation of taxonomies based on electrophysiological, morphological and molecular single cell properties. Establishing the correspondence across data modalities and understanding the underlying principles has proven challenging. Bio-realistic computational models offer the ability to probe cause-and-effect and have historically been used to explore phenomena at the single-neuron level. Here we introduce a computational optimization workflow used for the generation and evaluation of more than 130 million single neuron models with active conductances. These models were based on 230 in vitro electrophysiological experiments followed by morphological reconstruction from the mouse visual cortex. We show that distinct ion channel conductance vectors exist that distinguish between major cortical classes with passive and h-channel conductances emerging as particularly important for classification. Next, using models of genetically defined classes, we show that differences in specific conductances predicted from the models reflect differences in gene expression in excitatory and inhibitory cell types as experimentally validated by single-cell RNA-sequencing. The differences in these conductances, in turn, explain many of the electrophysiological differences observed between cell types. Finally, we show the robustness of the herein generated single-cell models as representations and realizations of specific cell types in face of biological variability and optimization complexity. Our computational effort generated models that reconcile major single-cell data modalities that define cell types allowing for causal relationships to be examined. Highlights Generation and evaluation of more than 130 million single-cell models with active conductances along the reconstructed morphology faithfully recapitulate the electrophysiology of 230 in vitro experiments. Optimized ion channel conductances along the cellular morphology (‘all-active’) are characteristic of model complexity and offer enhanced biophysical realism. Ion channel conductance vectors of all-active models classify transcriptomically defined cell-types. Cell type differences in ion channel conductances predicted by the models correlate with experimentally measured single-cell gene expression differences in inhibitory (Pvalb, Sst, Htr3a) and excitatory (Nr5a1, Rbp4) classes. A set of ion channel conductances identified by comparing between cell type model populations explain electrophysiology differences between these types in simulations and brain slice experiments. All-active models recapitulate multimodal properties of excitatory and inhibitory cell types offering a systematic and causal way of linking differences between them.
BibTeX
@misc{nandi2020single,
  author = {Anirban Nandi and Thomas Chartrand and Werner Van Geit and Anatoly Buchin and Zizhen Yao and Soo Yeun Lee and Yina Wei and Brian Kalmbach and Brian Lee and Ed S. Lein and Jim Berg and Uygar Sümbül and Christof Koch and Bosiljka Tasic and Costas A. Anastassiou},
  title = {{Single-neuron models linking electrophysiology, morphology and transcriptomics across cortical cell types}},
  year = {2020},
  doi = {10.1101/2020.04.09.030239},
  howpublished = {bioRxiv},
}
Figure