Inferring neuron-neuron communications from single-cell transcriptomics through NeuronChat

Wei Zhao, Kevin G. Johnston, Honglei Ren, Xiangmin Xu, Qing Nie

Abstract

Neural communication networks form the fundamental basis for brain function. These communication networks are enabled by emitted ligands such as neurotransmitters, which activate receptor complexes to facilitate communication. Thus, neural communication is fundamentally dependent on the transcriptome. Here we develop NeuronChat, a method and package for the inference, visualization and analysis of neural-specific communication networks among pre-defined cell groups using single-cell expression data. We incorporate a manually curated molecular interaction database of neural signaling for both human and mouse, and benchmark NeuronChat on several published datasets to validate its ability in predicting neural connectivity. Then, we apply NeuronChat to three different neural tissue datasets to illustrate its functionalities in identifying interneural communication networks, revealing conserved or context-specific interactions across different biological contexts, and predicting communication pattern changes in diseased brains with autism spectrum disorder. Finally, we demonstrate NeuronChat can utilize spatial transcriptomics data to infer and visualize neural-specific cell-cell communication.

Introduction

Cell-cell communication (CCC) is central to biological function, and several computational tools already exist to infer these networks from single-cell RNA sequencing (scRNA-seq). However, existing tools like CellChat or CellPhoneDB were largely designed for paracrine or autocrine signaling, where molecules diffuse over short distances or require physical contact.

Neurons present a unique challenge: they communicate over long distances via axons and dendrites using specialized small molecules like glutamate and GABA. Traditional CCC models often fail in neural contexts because standard databases exclude non-peptide neurotransmitters, and the abundance of these molecules depends on the coordinated expression of multiple genes (synthesis enzymes and transporters) rather than a single transcript. NeuronChat was developed to bridge this gap by providing a neural-specific framework for inferring connectivity.

Methods

NeuronChat consists of two primary components: a curated database and a logic-based computational model.

NeuronChatDB is a manually curated database of 373 neural-specific molecular interactions for both humans and mice. It categorizes interactions into five groups: small-molecule neurotransmitters, neuropeptides, gap junctions, gasotransmitters, and synaptic adhesion molecules.

The Computational Model addresses the complexity of neurotransmitter release. To estimate the abundance of small-molecule neurotransmitters, the model uses AND/OR logic. For example, to "count" glutamate, the model requires the presence of synthesis enzymes AND vesicular transporters. If multiple genes can perform the same task, it applies OR logic (arithmetic mean).

The computational framework of NeuronChat, illustrating the AND/OR logic used to estimate neurotransmitter abundance and the subsequent permutation testing for significance.

Following abundance estimation, the tool uses a permutation test to determine the statistical significance of communication links, producing a strength matrix that identifies significant neural connections between cell types.

Key Results and Insights

NeuronChat was validated using high-quality datasets, including scRNA-seq paired with monosynaptic retrograde labeling—the "ground truth" of neural connectivity. In these benchmarks, NeuronChat achieved high accuracy (AUROC of 0.832 in the primary visual cortex), significantly outperforming general-purpose tools that struggle to detect neurotransmitter-mediated links.

Disease Application: Autism Spectrum Disorder (ASD)

Applying NeuronChat to postmortem cortical tissue from patients with ASD revealed specific signaling disruptions. The analysis supported the "local overconnectivity" hypothesis and pinpointed a specific deficit: a significant downregulation of NLGN3 signaling from neurons to oligodendrocyte precursor cells (OPCs). This provides a potential molecular explanation for the myelination issues often observed in ASD.

Comparison of communication links between disease and control contexts and a multi-layered visualization approach for spatial transcriptomics data.

Spatial Transcriptomics

The framework also integrates spatial data (e.g., seqFISH+, MERFISH). While functional proximity in neurons often transcends physical distance, spatial constraints help refine predictions for local interneuron connections. The tool offers multi-layered visualization to map these communications onto raw tissue images.

Discussion and Limitations

NeuronChat provides a specialized lens for viewing neural networks through transcriptomic data. However, like all such tools, it assumes mRNA levels correlate with protein and metabolite levels. While its "abundance surrogates" perform well, they are not a perfect substitute for direct proteomic measurements. Future work will focus on expanding the database to include more niche neuromodulators and refining the handling of heteromeric receptor stoichiometry to better capture receptor sensitivity nuances.

Summary updated

Abstract

Neural communication networks form the fundamental basis for brain function. These communication networks are enabled by emitted ligands such as neurotransmitters, which activate receptor complexes to facilitate communication. Thus, neural communication is fundamentally dependent on the transcriptome. Here we develop NeuronChat, a method and package for the inference, visualization and analysis of neural-specific communication networks among pre-defined cell groups using single-cell expression data. We incorporate a manually curated molecular interaction database of neural signaling for both human and mouse, and benchmark NeuronChat on several published datasets to validate its ability in predicting neural connectivity. Then, we apply NeuronChat to three different neural tissue datasets to illustrate its functionalities in identifying interneural communication networks, revealing conserved or context-specific interactions across different biological contexts, and predicting communication pattern changes in diseased brains with autism spectrum disorder. Finally, we demonstrate NeuronChat can utilize spatial transcriptomics data to infer and visualize neural-specific cell-cell communication.
BibTeX
@inproceedings{zhao2023inferring,
  author = {Wei Zhao and Kevin G. Johnston and Honglei Ren and Xiangmin Xu and Qing Nie},
  title = {{Inferring neuron-neuron communications from single-cell transcriptomics through NeuronChat}},
  booktitle = {Nature Communications},
  year = {2023},
  doi = {10.1038/s41467-023-36800-w},
}
Figure