1. Neuroscience
Download icon

Common cell type nomenclature for the mammalian brain

Tools and Resources
  • Cited 3
  • Views 2,669
  • Annotations
Cite this article as: eLife 2020;9:e59928 doi: 10.7554/eLife.59928

Abstract

The advancement of single-cell RNA-sequencing technologies has led to an explosion of cell type definitions across multiple organs and organisms. While standards for data and metadata intake are arising, organization of cell types has largely been left to individual investigators, resulting in widely varying nomenclature and limited alignment between taxonomies. To facilitate cross-dataset comparison, the Allen Institute created the common cell type nomenclature (CCN) for matching and tracking cell types across studies that is qualitatively similar to gene transcript management across different genome builds. The CCN can be readily applied to new or established taxonomies and was applied herein to diverse cell type datasets derived from multiple quantifiable modalities. The CCN facilitates assigning accurate yet flexible cell type names in the mammalian cortex as a step toward community-wide efforts to organize multi-source, data-driven information related to cell type taxonomies from any organism.

Introduction

Cell type classification has been central to understanding biological systems for many tissues (e.g., immune system) (Lees et al., 2015) and organisms (e.g., Caenorhabditis elegans) (Packer et al., 2019). Identifying and naming cellular components of the brain has been an integral part of neuroscience since the seminal work of Cajal, 1899. Many neuronal cell types, such as neurogliaform, chandelier, Martinotti, and pyramidal cells, have been identified based on highly distinct shape, location, or electrical properties, providing robust and consistent classifications of neuronal cell types and a common vocabulary (Greig et al., 2013; Markram et al., 2004). However, the recent application of high-throughput, quantitative methods such as single-cell or -nucleus transcriptomics (scRNA-seq) (Hodge et al., 2019; Macosko et al., 2015; Saunders et al., 2018; Tasic et al., 2018; Tasic et al., 2016; Zeisel et al., 2018, Zeisel et al., 2015), electron microscopy (Zheng et al., 2018), and whole brain morphology (Winnubst et al., 2019) to cell type classification is enabling more quantitative measurements of similarities among cells and construction of taxonomies (Zeng and Sanes, 2017). The use of scRNA-seq, in particular, for cell type classification has increased exponentially since its introduction a decade ago (Tang et al., 2009), with nearly 2000 published studies and several hundred tools for data analysis (Zappia et al., 2018). These methodological advances are ushering a new era of data-driven classification, by simultaneously expanding the number of measurable features per cell, the number of cells per study, the number of classification studies, and the computational resources required for storing and analyzing this information.

This data explosion has enriched our collective understanding of biological cell types, while simultaneously introducing challenges in cell type classification within individual studies. In the retina, neurons with shared morphology also have consistentconnectivity (Jonas and Kording, 2015), spacing, arbor density, arbor stratification (Seung and Sümbül, 2014), and gene expression signatures (Macosko et al., 2015), often with one-to-one correspondences between phenotype and function (Zeng and Sanes, 2017). However, studies combining scRNA-seq with traditional morphological and electrophysiological characterizations in the brain have found a more complicated relationship in the brain than in retina, with cell types defined by morphology and electrophysiology sometimes containing cells from several cell types defined using gene expression (Gouwens et al., 2020; Kozareva et al., 2020), and some transcriptomically defined types containing cells with multiple morphologies (Hodge et al., 2020; Hodge et al., 2019). Further complicating classification is the overlay of discrete cell type distinctions with graded/continuous properties such as cortical depth (Berg et al., 2020), anterior/posterior and other trajectories across neocortex (Hawrylycz et al., 2012), activity-dependent cell state (Wu et al., 2017), or all simultaneously (Yao et al., 2020b). Furthermore, functional properties observed in matched cell types may diverge across species (Bakken et al., 2020a; Berg et al., 2020; Boldog et al., 2018; Hodge et al., 2019) and as cells advance along trajectories of development (Nowakowski et al., 2017), aging (Tabula Muris Consortium, 2020), and disease (Mathys et al., 2019).

Given this complex landscape, determining fundamental criteria for cell type definition in a given study, and then establishing correspondence to a cell type defined in another study, is often nontrivial and sometimes impossible. Substantial progress has been made toward solving this challenge of ‘alignment’, whereby datasets collected with genomics assays such as scRNA-seq or snATAC-seq can be used to anchor diverse cell types in a common analysis space (Barkas et al., 2018; Butler et al., 2018; Johansen and Quon, 2019). Alignment has proven effective for matching cell type sequence data collected on different platforms, across multiple data modalities, and even between species where few homologous marker genes show conserved patterns (Bakken et al., 2020a; Bakken et al., 2020b; Hodge et al., 2020; Hodge et al., 2019; Yao et al., 2020a). When combined with experimental methods such as Patch-seq (Cadwell et al., 2016; Fuzik et al., 2016; Gouwens et al., 2020; Scala et al., 2020), which involves application of electrophysiological recording and morphological analysis of single patch-clamped neurons followed by scRNA-seq of cell contents, autoencoder-based dimensionality reduction (Gala et al., 2019) can extend these alignments to bridge distinct modalities. Such analysis strategies provide a mechanism for classifying cell types using data from disparate data sources, allow for annotation transfer between experiments, and are a critical step toward unifying data-driven cell type definitions. However, as new cell type classifications are continually emerging, it is unrealistic to expect complete alignment of all published datasets, but creation of standardized systems for alignment becomes even more important.

Standardized cell type classification needs to include (1) standard nomenclature and (2) centralized and standardized infrastructure associated with cell type classification. Such standards provide a mechanism for storing key information about cell types and assigning explicit links between common cell types identified in different studies. Currently, no standard convention of naming brain cell types is widely followed. Cell types have historically been named by their shape, location, electrical properties, selective neurochemical markers, or even the scientist who discovered them (Betz, 1874; Szentágothai and Arbib, 1974). Now, quantitative clusters that cannot obviously be matched with these types are named on an ad hoc basis, either by assigning generic names like ‘interneuron 1’ or ‘Ex1’ and then linking these names to associated figures, tables, or text (Gouwens et al., 2019; Lake et al., 2016; Zeisel et al., 2015), or by chaining critical cell type features in the name itself, resulting in names like ‘Neocortex M1 L6 CT pyramidal, Zfpm2 non-adapt GLU’ (Shepherd et al., 2019). All of these proposals are reasonable for stand-alone projects but make direct comparisons between studies daunting. While several public databases for data storage have been developed (e.g., dbGaP, NeMO, NeuroElectro, Neuromorpho, HuBMAP, etc.), a community-recognized repository for storing and tracking cell type assignments and associated taxonomies does not currently exist. This challenge has been recognized by many (Armañanzas and Ascoli, 2015; DeFelipe et al., 2013; Shepherd et al., 2019) and has been a focus of recent conferences seeking community participation toward a solution (Yuste et al., 2020). Any solution devised to tackle this question should ideally be effective and user-friendly and should directly address some of the ongoing challenges of ontology, data matching, and cell type naming described above in its implementation, providing some amount of immediate standardization of any cell type classifications included therein. This challenge was also addressed at a Cell Type Ontology Workshop (Seattle, June 17–18, 2019; hosted by the Allen Institute, Chan Zuckerberg Initiative [CZI] and the National Institutes of Health [NIH]), where input from representatives from the fields of ontology, taxonomy, and neuroscience made recommendations, highlighted best practices, and proposed conventions for naming cell types.

To begin to address these challenges and driven by a practical need to organize vast amounts of multimodal data generated by the Allen Institute and collaborators, we have developed a nomenclature convention aimed at tracking cell type information across multiple datasets. Here we present a generalizable nomenclature convention, the common cell type nomenclature (CCN), for matching and tracking cell types across studies. This convention was motivated by methodologies used for management of gene transcript identity tracked across different versions of GENCODE genome builds, allowing comparison of matched types with a common reference or any other taxonomy (Frankish et al., 2019; Harrow et al., 2012). Motivated by gene nomenclature conventions from HGNC (Bruford et al., 2020), the CCN also facilitates assigning accurate yet flexible cell type names in the mammalian cortex as a step toward community-wide efforts to organize multi-source, data-driven information related to cell type taxonomies from any organism. An initial version of the CCN was introduced at https://portal.brain-map.org/explore/classes (October 2019), with the intent to encourage discussion and gather feedback for improving subsequent versions, to facilitate collaboration, and to improve shared understanding of the many cell types in the brain.

Results

Overview of proposed nomenclature convention

The problem of defining and naming cell types has many similarities to those of genes in genomics, where there is a practical need to track individual sequencing and assembly results as distinct and self-contained entities, while simultaneously recognizing the goal for a singular reference that the community can use to map sequencing data into a common context (Frankish et al., 2019; Harrow et al., 2012; Kitts et al., 2016). Here, a similar strategy is proposed for cell type nomenclature: Use of a standardized series of identifiers for tracking cell types referenced to individual studies, in addition to providing a mechanism for defining common identifiers (Figure 1A). At the core of the schema are two key concepts: (1) a taxonomy, defined as the output of a computational algorithm applied to a specific dataset, which must be generated prior to implementation of this schema, and (2) a cell set, which can represent any collection of cells within a specific taxonomy (see Table 1 for definitions of key terms). These components are generated through the input of data and information generated from analysis that identifies provisional cell types (sometimes called cell types for convenience). These are analytically relevant cell sets that represent quantitatively derived data clusters defined by whatever classification algorithm generated the taxonomy. Provisional cell types can be organized as the terminal leaves of a hierarchical taxonomy using a dendrogram, as a non-hierarchical community structure, or both. Taxonomies and cell sets are assigned unique identifier tags, as described below, and additional metadata can be stored alongside these tags for use with future databasing and ontology tools. These properties can be tracked using a relational graph or other database service, in a qualitatively similar manner to how transcripts are tracked across different versions of GENCODE genome builds (Frankish et al., 2019).

Overview of common cell type nomenclature (CCN) and application to human middle temporal gyrus (MTG).

(A) Schematic of CCN components and process. (B–D) Example outputs from the CCN. (B) Annotated dendrogram of cell types in human MTG, along with associated cell type names, reproduced from Hodge et al., 2019. Internal nodes with a term (teal circles) represent cell sets with preferred alias tags. (C) CCN annotations for a putative cell type (outlined in blue) and an internal node (outlined in orange) of this dendrogram. (D) Snippet of an output file from the CCN showing cell to cell set mappings as applied to human MTG.

Table 1
Glossary of terms.

Terminology used with the common cell type nomenclature (CCN), definitions for use, and examples of how terms are applied. Terms are presented in bold upon first use in the text. This glossary is intended to clarify use for the purposes of the CCN since some terms are open to multiple interpretations, and effective classification requires disambiguation. Asterisks denote terms that represent specific components of the CCN.

TermDefinitionExample
TaxonomySet of quantitatively derived data clusters defined by a specific computational algorithm on a specific dataset(s). Taxonomies are given a unique label and can be annotated with metadata about the taxonomy, including details of the algorithms and relevant cell and cell set IDs.Any clustering result in a cell type classification manuscript
DatasetFeature information (e.g., gene expression) and associated metadata from a set of cells collected as part of a single project.Gene expression from 6000 human MOp nuclei
OntologyA structured controlled vocabulary for cell types.Cell Ontology
Marker gene(s)A gene (gene set) which, when expressed in a cell, can be used to accurately assign that cell to a specific cell set.GAD2; PVALB;
CHODL
Taxonomy ID*An identifier uniquely tagging a taxonomy of the format CCN[YYYYMMDD][#].CCN201910120
CellA single entry in a taxonomy representing data from a single cell (or cell compartment, such as the nucleus). Cells have metadata including a unique ID.N/A
Cell setAny tagged group of cells in a taxonomy. This includes cell types, groups of cell types, and potentially other informative groupings (e.g., all cells from one donor, organ, cortical layer, or transgenic line). Cell sets have several IDs and descriptors (as discussed below) and can also have other metadata.A cell type; a group of cell types; all cells from layer two in MTG; all cells from donor X
Provisional cell typeQuantitatively derived data cluster defined within a taxonomy. This is a specific example of a cell set that is of high importance, as most other cell sets are groupings of one or more provisional cell types. Here, the term ‘cell type’ is synonymous with ‘provisional cell type.’ .A cell type defined in a specific study
DendrogramA hierarchical organization of provisional cell types defined for a specific taxonomy. Dendrograms have a specific semantic and visualizable structure and include nodes (representing multiple provisional cell types) and leaves (representing exactly one). Not all taxonomies include a dendrogram (e.g., if the structure of cell sets is non-hierarchical).N/A
Community structureNon-hierarchical relationships between cell types defined as groups of cell types in a graph.N/A
Cell set accession ID*A unique ID across all tracked datasets and taxonomies. This tag labels the taxonomy and numbers each cell type. CS[taxonomy id]_[unique # within taxonomy]CS201910120_1
Cell set label*An ID unique within a single taxonomy that is used for assigning cells to cell sets defined as a combination of multiple ‘provisional cell types’.MTG 12
MTG 01–08
Cell set alias*Any cell set descriptor. It can be defined computationally from the data, or manually based on new experiments, prior knowledge, or a combination of both. Cell aliases beyond the ‘preferred’ or ‘aligned’ are defined as ‘cell set additional aliases’.(Any ‘cell set aligned alias’); Interneuron 1; Rosehip
Cell set preferred alias*The primary cell set alias (e.g., what cell types might be called in a publication). This can sometimes match the aligned alias, but not always, and can be left unassigned.Inh L1-2 PAX6 CDH12; ADARB2 (CGE); Chandelier; [blank]
Cell set aligned alias*Analogous to ‘gene symbol’. At most one biologically driven name for linking matching cell sets across taxonomies and with a reference taxonomy.L2/3 IT 4; Pvalb 3; Microglia 2
Cell set structure*The location in the brain (or body) from where cells in the associated set were primarily collected.Neocortex
Cell set ontology tag*A tag from a standard ontology (e.g., UBERON) corresponding to the listed cell set structure.UBERON:0001950
Cell set alias assignee*Person responsible for assigning a specific cell set alias in a specific taxonomy (e.g., the person who built the taxonomy or uploaded the data, or a field expert).(First author of manuscript)
Cell set alias citation*The citation or permanent data identifier corresponding to the taxonomy where the cell set was originally reported.(Manuscript DOI); [blank]
Reference taxonomyA taxonomy based on one or a combination of high-confidence datasets, to be used as a baseline of comparison for datasets collected from the same organ system.Cross-species cortical cell type classification
Morpho-electric(ME) typeA provisional cell type defined using a combination of morphological and electrophysiological features.ME_Exc_7
Governing bodyA forum of subject-matter experts to guide policy and manage change of the CCN and associated ontologies and databasing efforts.N/A

A major goal of the CCN is to track taxonomies and their associated cell sets by providing an easy-to-understand schema that is widely applicable to new and published taxonomies and that can be implemented through a user-friendly code base. The CCN is compatible with taxonomies generated from either single or multiple modalities, taxonomies applied to cells from overlapping datasets, and reference taxonomies (discussed in detail below). Each taxonomy is assigned a unique taxonomy id of the format CCN[YYYYMMDD][#], where ‘CCN’ denotes this nomenclature convention; Y, M, and D represent year, month, and day, respectively; and # is an index for compiling multiple taxonomies on a single day. Each taxonomy can also be assigned metadata, such as species, but such details are outside the scope of the CCN. Within each taxonomy, cell sets (and therefore also provisional cell types) are assigned multiple identifier tags, which are used for different purposes. Cell set accession IDs track unique cell sets across the entire universe of taxonomies and are defined as CS[YYYYMMDD][#]_[unique # within taxonomy], where CS stands for ‘cell set’ and the date and number match the taxonomy id. Cell set labels are useful for constructing cell sets from groups of provisional cell types, but can otherwise be ignored. Cell set aliases represent descriptors intended for public use and communication, including data-driven terms, historical names, or more generic cell type nomenclature. For convenience these are split into at most one preferred alias, which represents the primary tag for public consumption (e.g., the cell type names used in a manuscript), and any other additional aliases. Additionally, each cell set can have at most one aligned alias, which is a biologically driven term that is selected from a controlled vocabulary. Aligned aliases generally are assigned to only a subset of cell sets by alignment to a reference taxonomy, but in principle can be assigned in any taxonomy or taxonomies (e.g., if a rare type is identified that is missing from the reference). The CCN includes a specific system for assigning such aliases in the mammalian cortex using properties that are predicted to be largely preserved across development, anatomical area, and species, which will be discussed in detail. Furthermore, the CCN includes a series of metadata tags tracking the provenance and anatomy of cell sets. The cell set alias assignee and cell set alias citation indicate the person and permanent data identifier associated with each cell set alias. The cell set structure indicates the location in the brain (or body) from where associated cells were primarily collected. Ideally, this will be paired to an established ontology using the cell set ontology tag; in this case, we use UBERON since it is designed as an integrated cross-species anatomy ontology (Haendel et al., 2014). Finally, the CCN is compatible with incorporation of additional taxonomy-specific or future global cell set metadata or descriptors. This could include donor metadata (e.g., age or sex), summarized cell metadata (e.g., cortical layer or average reads), or additional cell set tags. In particular, the concept of a cell set level is often useful for distinguishing highly specific but statistically less confident provisional cell types from the more general and more statistically robust cell sets.

The CCN is currently in use by the Allen Cell Types Database for transcriptomic taxonomies (http://celltypes.brain-map.org/rnaseq/) and is being applied to taxonomies generated by the BRAIN Initiative Cell Census Network (BICCN; https://biccn.org/) (Bakken et al., 2020a; Adkins et al., 2020; Yao et al., 2020a), a consortia of centers and laboratories working collaboratively to generate, analyze, and share data about brain cell types in human, mouse, macaque, and other non-human primates.

Application of the CCN to cell types in human middle temporal gyrus

A detailed walk-through of how to apply the CCN to a published study on cell types in human middle temporal gyrus (MTG) (Hodge et al., 2019) is presented in Materials and methods. In short, Figure 1B recapitulates the cell types and associated hierarchy previously published for MTG (Hodge et al., 2019). After applying the CCN, each leaf (provisional cell type) and internal node of the dendrogram is assigned the series of cell set tags described above (Figure 1C), and every cell is mapped to every cell set (Figure 1D). This was all done using a user-friendly set of scripts (https://github.com/AllenInstitute/nomenclature). These output files are intended to be directly included as supplemental materials in manuscripts performing cell type classification in any species, and such output for human MTG (and for 17 additional taxonomies) is presented in Supplementary file 1.

Naming cell types in mammalian cortex

Mammalian brain cell types inhabit a complex landscape with fuzzy boundaries and complicated correspondences between species and modalities, leading to a variety of disparate solutions for naming cell types. Thus, a challenging and potentially contentious question in cell type classification is how these newly identified cell types should be named, or in the context of the CCN, what should be put in the ‘cell set aligned alias’ identifier. The CCN utilizes a strategy for naming cell types in the mammalian cortex that includes properties that are cell intrinsic and potentially well conserved between species (Table 2). This convention is used as the cell set aligned alias tag in the CCN and ideally should directly map to cell types defined in a relevant ontology (i.e., Cell Ontology [Diehl et al., 2016] or Neuron Phenotype Ontology [Gillespie et al., 2020]). While admittedly underdeveloped, this convention has been applied to multiple studies of the primary motor cortex (M1; as discussed below) and represents only a starting point for discussion.

Table 2
Proposed strategy for naming cortical cell types.
ClassFormatExample
Glutamatergic[Layer] [Projection] #L2/3 IT 4
GABAergic[Canonical gene(s)] #Pvalb 3
Non-neuronal[Cell class] #Microglia 2
Any class[Historical name] #Chandelier 1

For glutamatergic neurons, cell types are named based on predominant layer(s) of localization of cell body (soma) and their predicted projection patterns. The relatively robust laminarity of glutamatergic cell types has been described based on cytoarchitecture in multiple mammalian species for many years (e.g., Rakic, 1984), and has been confirmed using RNA in situ hybridization (Hodge et al., 2019; Tasic et al., 2018; Zeng et al., 2012), and a combination of layer dissections and scRNA-seq (Hodge et al., 2019; Tasic et al., 2018). While in humans many cell types do not follow the layer boundaries defined by cytoarchitecture entirely, laminar patterning is still generally well conserved between human donors and mice (Hodge et al., 2019). In adult mouse visual cortex, projection targets for cell types have been explicitly measured using a combination of retrograde labeling and scRNA-seq (Tasic et al., 2018; Tasic et al., 2016). By aligning cell types across species, the projection targets in mice can be hypothetically extrapolated to putative projection targets in human or other mammalian species. For example, von Economo neurons are likely to project subcortically (Hodge et al., 2020). For GABAergic interneurons, developmental origin may define cell types by their canonical marker gene profile established early in development, with Pvalb and Sst labeling cell types derived from the medial ganglionic eminence and Vip, Sncg, and Lamp5 labeling cell types derived from the caudal ganglionic eminence (DeFelipe et al., 2013). Non-neuronal cell types have not been a focus of the studies cited and hence they are labeled at a broad cell type level only. However, knowledge from other single-cell transcriptomics studies on microglia (Hammond et al., 2019; Li et al., 2019), astrocytes (Batiuk et al., 2020), and oligodendrocytes (Marques et al., 2016) could be included in subsequent versions of this convention. In all cases, multiple cell types are present within a given class. While it may not be possible to directly translate every feature to other brain structure or other organs, most of the concepts proposed here could still be followed.

Alignment of established cell sets using reference taxonomies

The CCN presents a flexible data structure for storing key information about taxonomies and cell sets, implemented through reproducible code with standard output files, along with a specific convention for naming mammalian neocortical cell types. It is applicable to taxonomies defined on any data type using any classification algorithm, including hierarchical cell type classification using scRNA-seq. While useful for these reasons alone, a primary utility of the CCN is to facilitate cross-study integration of cell type classifications, in particular when applied in the framework of a reference taxonomy. A reference taxonomy (or reference cell type classification) is any taxonomy based on one or a combination of high-confidence datasets, which can be used as a baseline of comparison for other datasets collected from the same organ system. For example, many researchers favor building a gene expression-based reference taxonomy based on high-throughput, high-resolution single-cell transcriptomics assays and then layering on additional phenotypic data as they become available (Yuste et al., 2020). Molecular, physiological, and morphological characteristics of cortical neurons are highly correlated based on simultaneous measurement in individual cells using Patch-seq (Berg et al., 2020; Gouwens et al., 2020; Scala et al., 2020), making such a strategy feasible. Many groups are currently performing scRNA-seq analysis in different areas of the brain, from all organs in the human body (Rozenblatt-Rosen et al., 2017), from multiple mammalian species (Geirsdottir et al., 2019), and across trajectories of development (Nowakowski et al., 2017), aging (Tabula Muris Consortium, 2020), and disease (Mathys et al., 2019). Application of the CCN to these datasets will allow future reference taxonomies to evolve to accommodate these additional complexities by overlaying a common data structure and associated nomenclature.

Reference taxonomies and the CCN are two components of a multi-staged analysis workflow for aligning cell type classifications using datasets collected across multiple labs, from multiple experimental platforms, and from multiple data modalities (Figure 2). This workflow accommodates methodological differences in cell type definitions across studies and accommodates changes in reference taxonomies over time. The proposed workflow can be broken down into four broad stages:

Workflow for assigning types to a given dataset with taxonomy.

(1) Cell type classification will initially be performed separately on all taxonomies. (2) One, some, or all of these datasets will be combined into a high-confidence reference taxonomy which can be used as a comparator for any related datasets, by (3) mapping existing and new datasets to the reference taxonomy. (4) The reference will periodically be updated as new datasets and taxonomies are generated.

  1. First, many research teams will independently define cell types, identify their discriminating features, and name them using one of many available experimental and computational strategies. This represents the current state of the field. The CCN may be applied to each dataset independently at this stage.

  2. Second, an initial reference cell type classification will be defined by taking the results from one or more (ideally validated) datasets and integrating these data together in a single analysis, if needed. Being high dimensional, high throughput, and relatively low cost, transcriptomics strategies are immediately applicable to many organs and species, and the goal is for reference cell types to be defined using this modality (Yuste et al., 2020). The CCN will then be applied to the reference taxonomy as described above – the CCN treats reference taxonomies identically to any other taxonomy. Importantly, aligned aliases should be defined in the reference taxonomy at this stage using a standard naming convention such as the one proposed above.

  3. This reference cell type classification can now be used as a comparator for any related datasets, providing a mechanism for transferring prior knowledge about cell types across datasets. Cell sets from existing taxonomies can be renamed using one of the many validated alignment algorithms (e.g., Barkas et al., 2018; Butler et al., 2018; Gala et al., 2019; Johansen and Quon, 2019) by integrating data from this taxonomy with the reference, and then updating the cell set aligned alias to match terms defined in the reference. For new datasets, taxonomies can be generated using any clustering or alignment strategy followed by the same mapping and annotation transfer steps.

  4. Finally, new versions of the reference cell type classification should be periodically generated using additional data and/or computational methods, and this new classification will now be used as comparator for related datasets. Steps 3 and 4 can iterate at some to-be-defined cadence.

This workflow provides two complementary strategies to compare between taxonomies without needing to look at gene expression or other quantitative features. First, each taxonomy draws upon a common set of aligned alias terms, which allows for immediate linking of common cell sets between taxonomies (in cases where such information can be reliably assigned). A second strategy is through inclusion of common datasets across multiple taxonomies (reference or otherwise); if cells are assigned to the same cell sets in more than one taxonomy, then the cell sets can be directly linked. As a whole, this workflow provides a general outline for versioned cell type classification that could be specialized as needed for communities studying different organ systems and that provides a starting point for design of future cell taxonomy and nomenclature databases.

Defining a cross-species reference taxonomy in M1

A recent study profiling nearly half a million nuclei in primate and mouse primary motor cortex (M1) presents a taxonomy suitable for defining as a reference taxonomy (Bakken et al., 2020a). This study included single cell data from three separate ‘omics’ modalities (transcriptomics, epigenetics, and methylation) for mouse, marmoset, and human. Datasets were integrated in two ways. First, epigenetics and methylation datasets were integrated with snRNA-seq data within mouse, marmoset, and human independently (as shown in Figure 3A for human), which demonstrates a consistent genomic profile of cell types within species. Second, snRNA-seq from each species were aligned into a single integrated reference, which identifies cell type homologies across species that were presumably present in the mammalian ancestor to rodents and primates. This evidence-based assumption of cross-species homologies provides a strategy for transferring cell type characteristics from rodent studies (e.g., projection targets) into human, where experiments for making such measurements are not yet possible. A total of 11 taxonomies were generated (Figure 3B), and all were included in the same nomenclature schema, and the CCN was applied to this set of taxonomies as described above (see Supplementary Table 3 in Bakken et al., 2020a and Supplementary file 1). Figure 3C shows an example of these cross-species and cross-modality alignments for L6 CT cells, which are divided into two cell sets in the integrated taxonomy (and assigned the aligned alias tags L6 CT_1 and L6 CT_2) and includebetween one and seven cell sets in the single-modality taxonomies.

Series of multimodal, cross-species taxonomies in primary motor cortex (M1) demonstrates utility of nomenclature schema.

(A) Taxonomies based on transcriptomic (‘1’; top), open chromatin (‘2’; middle), and DNA methylation (‘3’; bottom) in human M1. Epigenomic clusters (‘2’, ‘3’; in rows) aligned to RNAseq clusters (‘1’) as indicated by horizontal black bars and are also assigned matching cell sets in the relevant taxonomies. Adapted from Bakken et al., 2020a. (B) Flow chart showing all 11 taxonomies generated for this project and their connections. The integrated (reference) taxonomy included nuclei collected using snRNA-seq from three species (gray box), with nuclei collected from layer five in macaque mapped to this space post hoc (gray line). Separately, epigenetics taxonomies from human, marmoset, and mouse were aligned to their respective transcriptomics taxonomies (black lines). This entire taxonomic structure is captured by the CCN (see Supplementary file 1). (C) An example mapping of corticothalamic (L6 CT) provisional cell types across the human and transcriptomics taxonomies using the CCN (black box in A). Preferred aliases for each taxonomy are used for clarity.

This integrated taxonomy (Figure 3B, gray box) represents a suitable reference taxonomy for several reasons: first, the data generation, data analysis, and write-up spanned multiple BICCN-funded labs as part of a coordinated consortium project, indicating that this taxonomy was approved by a large subset of the neocortex cell typing community; second, while a number of differences were found between species, 45 core provisional cell types could be aligned across all species with transcriptomics; third, the taxonomies generated using epigenetics and methylation are largely consistent with results of this integrated taxonomy (Figure 3A, bottom panels and Yao et al., 2020a); and finally, this taxonomy can be linked with other quantitative features (such as morphology, electrophysiology, and expected projection targets) through comparison with mouse studies using complementary modalities such as Patch-seq (Gouwens et al., 2020; Scala et al., 2020) and Retro-seq (Tasic et al., 2018; Tasic et al., 2016). Using these linkages, aligned aliases of the format proposed in Table 2 were assigned to cell sets in the integrated taxonomy along with the 10 other species-specific taxonomies using a combination of (1) robust gene markers from the literature, (2) highly discriminating gene markers in these data, (3) projection targets in mouse, (4) historical names based on cell shape, and (5) broad or low-resolution cell type names (that directly map to ontologies), providing a starting point for how brain cell types could be named. A complete list of aligned aliases used is shown in Supplementary file 2.

Applying the CCN to existing and new datasets

For a specific convention to be adopted, both in general or in place of other competing conventions, it needs to be easy to use and immediately useful to the community. For example, many cell type classification studies now use Seurat (Butler et al., 2018) for clustering and alignment because it produces believable biological results, and it is implemented in intuitive R code with extensive user guides for non-specialists. As such, Seurat visualizations appear frequently in manuscripts and its file format is used as input for several analysis pipelines. While the usability of the CCN has been established above, the utility of applying it to a single taxonomy in the absence of a centralized database of taxonomies may be less clear. Here five use cases are presented to illustrate how the CCN can be applied to published datasets using scRNA-seq and electrophysiology and morphology in multiple species. These use cases provide immediate utility and also lay a foundation for future databasing and ontology efforts.

Use case 1: Alignment of human MTG taxonomy to M1 reference

The M1 reference taxonomy includes a validated set of aligned aliases that follows the proposed nomenclature for mammalian cortex (Table 2) and that can be applied to any other taxonomy. As part of the original analysis (Bakken et al., 2020a), nuclei from human MTG (Hodge et al., 2019) were aligned to the human M1 dataset. This analysis provides a perfect use case for transferring cell set aligned alias tags from the reference taxonomy to the MTG data (as was done; see Materials and methods). Figure 4 shows a visualization of glutamatergic types in M1 and MTG, with the color of each square representing the fraction of cells from provisional cell types in each brain region that are assigned to the same alignment cluster, and boxes indicating the aligned alias calls in M1 and their corresponding calls in MTG. While alignment is not perfect for provisional cell types, it is sufficient for matching aligned aliases between cortical areas. These mappings enable biological insights such as the presence of L4-like neurons in M1, where an anatomically defined L4 is not apparent. Likewise, such alignment enables prediction of cell properties such as long-range connectivity (e.g., thalamic inputs), as well as electrophysiology measurements in primary sensorimotor cortices or other brain regions inaccessible to techniques such as Patch-seq. Similar alignments were performed for GABAergic interneurons and non-neuronal cell types (Supplementary file 1). Such tagging allows cell sets in human MTG to be directly compared to cell sets from any other taxonomy with the same aligned alias, for example to infer morphological or electrophysiological properties (see Use case 2) or cell class persistence across multimodal phenotypes and developmental stages (see Use case 3) in mouse. Cell sets can even be matched with more distant species using the CCN (see Use case 4), to the extent that such alignment is possible based on the data.

Alignment of glutamatergic cell sets in human middle temporal gyrus (MTG) to a reference primary motor cortex (M1) taxonomy.

Cluster overlap heatmap showing the proportion of nuclei from MTG clusters and the reference (M1) clusters that coalesce with a given aligned cluster. Cell sets corresponding to aligned aliases in the MTG and M1 taxonomies are labeled and indicated by blue boxes. Adapted from Bakken et al., 2020a.

Use case 2: Building a morpho-electric taxonomy

While much effort for cell typing is currently focused on taxonomies based on scRNA-seq datasets, the CCN can equally apply to non-transcriptomic and non-hierarchical taxonomies. For example, a study of mouse visual cortex examined ~1800 cells characterized electrophysiologically by whole-cell patch clamp recordings, and for a subset of these (450 cells), morphological reconstructions were also performed (Gouwens et al., 2019). Using a multimodal unsupervised clustering method, the authors identified 20 excitatory and 26 inhibitory morpho-electric types (or me-types), which are cell types defined using a combination of morphological and electrophysiological features. Figure 5 shows the application of the CCN to a subset of excitatory (glutamatergic) me-types of that study (see Supplementary file 1 for application to remaining me-types). The preferred alias and inferred subclass columns show the organization scheme; me-types were organized by broader cell types inferred from transgenic labels, but not placed into a binary hierarchical taxonomic tree (Gouwens et al., 2019). Through application of the aligned alias tag, these cell types can be directly linked to cell types defined based on transcriptomics.

Application of common cell type nomenclature (CCN) to glutamatergic me-types in the mouse visual cortex.

Excitatory (glutamatergic) me-types from Gouwens et al., 2019 that have been incorporated into the nomenclature schema. Eleven of the original 20 excitatory me-types are shown as examples. Representative morphologies and electrophysiological responses are shown to illustrate the differences between types. The ‘inferred subclass’ calls perfectly map to cell set aligned aliases from the reference M1 taxonomy in Figure 3, except that L5 CF (corticofugal) is an additional alias for L5 ET, and cells sets corresponding to L4, L6 IT, and L6 CT (blue boxes) have been added to the taxonomy.

Use case 3: Exploring an interneuron subclass using multimodal attributes: The ‘Sst Chodl’ class persists across cross-taxonomy matching

Somatostatin-expressing interneurons can be divided into multiple cell types (the specific number differs by taxonomy), some of which include cells that express Chodl in the mouse cerebral cortex (Tasic et al., 2018; Tasic et al., 2016). These ‘Sst Chodl’ neurons are rare and, based on expression of specific marker genes, correspond to the only known cortical interneurons with long-range projections (Tomioka et al., 2005). Recent studies using the multimodal cell phenotyping method Patch-seq (Gouwens et al., 2020) confirmed that ‘Sst Chodl’ cell sets characterized based on morphology and electrophysiology (Gouwens et al., 2019) match those defined by transcriptomic profiles (Tasic et al., 2018; Tasic et al., 2016). The CCN can be applied to readily represent these ‘Sst Chodl’ cells (and other cell types) matched between all relevant taxonomies, regardless of species or modality through the use of aligned alias tags. For example, Table 3 shows all cell sets from Bakken et al., 2020a (Figure 3B) associated with Sst Chodl cells, which all have ‘Sst Chodl’ in the aligned alias (with one exception noted below). In mouse, all three modalities have a single ‘Sst Chodl’ cell type, which can be linked to a matched type in VISp due to its highly distinct gene expression patterning that is conserved across brain regions (Tasic et al., 2018; Yao et al., 2020b). This transcriptomic cell type is similarly linked to the ‘Sst Chodl’ cell type in the integrated transcriptomic (reference) taxonomy, which lists ‘long-range projecting Sst’ as an additional alias to formalize the cross-modal correspondence. In human, the RNA-seq and ATAC-seq have one-to-one correspondences, but for DNA methylation (DNAm), Inh L1-5 SST AHR aligns with several Sst cell types including Sst Chodl (likely due to the rarity of this cell type). Cell sets from the methylation- and epigenetics-based taxonomies include an additional alias that list the cell set labels in transcriptomics taxonomy, directly linking these cell types. Therefore, while Inh L1-5 SST AHR does not have ‘Sst Chodl’ as its aligned alias, the cell set label ‘RNA-seq 040, 046–047, 050–052, 068 in CCN201912131’ indicates the inclusion of ‘Sst Chodl’ cells (RNA-seq 040). In marmoset, where fewer cells were collected, an ‘Sst Chodl’ cell set is only found with transcriptomics. Explicitly linking cell sets in this way provides multiple potential points of comparison with other studies, including studies of disease or development. For example, a study of interneuron development in E14 mice found that the ‘Sst Chodl’ cells were severely affected by Sox6 removal during interneuron migration (Munguba et al., 2019), and cell class definitions observed in the mature brain may have foundational roles in cortical patterning.

Table 3
Nomenclature for ‘Sst Chodl’ cell sets cited in Bakken et al., 2020a.

Relevant common cell type nomenclature (CCN) entities and taxonomy metadata, including the cell set additional alias that links to cell set labels from relevant transcriptomics taxonomies. All listed cell sets have a cell set structure of ‘primary motor cortex’ and a cell set ontology tag of ‘UBERON:0001384’.

#Cell set preferred aliasCell set labelCell set accessionCell set aligned aliasCell set additional alias
1Inh L1-6 SST NPYRNA-seq 040CS201912131_40Sst Chodl
2Inh L1-5 SST AHRDNAm 12CS202002272_12RNA-seq 040, 046–047, 050–052, 068 in CCN201912131
3Inh L1-6 SST NPYATAC-seq 08CS202002273_8Sst ChodlRNA-seq 040 in CCN201912131
4Inh SST NPYRNA-seq 01CS201912132_1Sst Chodl
5Sst ChodlRNA-seq 028CS202002013_28Sst Chodl
6Sst ChodlDNAm 09CS202002276_9Sst ChodlRNA-seq 028 in CCN202002013
7Sst ChodlATAC-seq 10CS202002277_10Sst ChodlRNA-seq 028 in CCN202002013
8Sst ChodlIntegrated 14CS202002270_14Sst ChodlLong-range projecting Sst
#Cell set alias assigneeCell set alias citationTaxonomy idSpeciesModality
1Nikolas Jorstad10.1101/2020.03.31.016972CCN201912131HumanRNA-seq
2Wei Tian10.1101/2020.03.31.016972CCN202002272HumanDNAm
3Blue Lake10.1101/2020.03.31.016972CCN202002273HumanATAC-seq
4Fenna Krienen10.1101/2020.03.31.016972CCN201912132MarmosetRNA-seq
5Zizhen Yao10.1101/2020.02.29.970558CCN202002013MouseRNA-seq
6Hanqing Liu10.1101/2020.02.29.970558CCN202002276MouseDNAm
7Yang Li10.1101/2020.02.29.970558CCN202002277MouseATAC-seq
8Nikolas Jorstad10.1101/2020.03.31.016972CCN202002270AllRNA-seq

Use case 4: Alignment of cell types from reptilian and mammalian cortex using the CCN

While the focus of this study is the mammalian cortex, the CCN framework is applicable to other organs and more distant species. As an example use case, a single-cell transcriptomics study of turtle and lizard pallium found GABAergic interneuron and non-neuronal cell types to be homologous with those seen in mouse cortex (Tosches et al., 2018). In many cases, these cell types expressed shared gene markers, suggesting a shared evolutionary origin across 320 million years of evolution in amniote vertebrates. These types include astrocytes (GFAP), oligodendrocytes (MBP), oligodendrocyte precursor cells (OLIG1 and PDGFRA), microglia (C1QC), GABAergic interneurons as a whole (GAD1 and GAD2), and Sst+ interneurons (SST). Reptilian analogs for other CGE- and MGE-derived GABAergic types were also identified, although interestingly neither VIP nor PVALB marker genes are expressed in reptiles. Application of the CCN to the taxonomies presented for the turtle demonstrates the utility of this approach (Supplementary file 1).

Assignment of aligned aliases for non-neuronal cells and GABAergic interneurons is straightforward, with ‘PV-like’ interneurons (cell types i11–i13 from Tosches et al., 2018) assigned ‘PVALB’, and similar alignments for other types. In contrast, the correspondence between reptilian and mammalian glutamatergic cells is more complicated. Reptiles have a three layer pallium and only the anterior dorsal cortex (representing a small fraction of pallium) is comparable with the six-layer mammalian neocortex (Jarvis, 2009; Tosches et al., 2018). RNA-seq in combination with in situ hybridization identified two distinct sublayers of turtle layer 2: a superficial L2a (cell types e07–e08) and a deeper L2b (e13–e16), which seem to correspond with mammalian deep layer and upper layer neurons, respectively, suggesting there was likely an inversion of layers in one clade. However, all of these cell types coexpress genes found in mutually exclusive L2/3, L4, and L5a intra-telencephalic neurons (e.g., SATB2, RORB, and RFX3) along with extra-telencephalic projection neurons (e.g., BCL11B, TBR1, and SOX5), suggesting either a lack of homologous cell types between clades or at least a change in the core transcription factor regulatory programs. Thus, with the level of resolution presented in this study, no aligned aliases (beyond the broadest) are assigned for glutamatergic types. This points to the importance of having measurements in other modalities, for example, local and long-range connectivity, that may help establish homologies or bolster claims of clade-specific cellular innovations. Overall, the CCN provides a mechanism for assigning a standard nomenclature for cell types found in the reptilian cortex and linking these types with a mammalian neocortical reference at the level of resolution resolved in the taxonomy.

Use case 5: Comparison of novel to existing taxonomies

The first four use cases represent specific examples of how taxonomies from different brain regions, modalities, and species can be presented in the framework of the CCN to make published inferences more easily accessible to a naive reader. These represent specific examples of a more general use case for scientists, who may want to compare their newly generated taxonomy to what is currently known about cell types. The ideal application for this scenario is a centralized database for taxonomy integration with an associated ontology and annotation capabilities; such a framework is well beyond the scope of this manuscript, but solutions are underway. As a starting point for this goal, Supplementary file 1 presents output files from the CCN for 18 taxonomies (including all taxonomies discussed herein; Table 4) that have been annotated with the aligned aliases from the M1 reference taxonomy presented in Figure 3. Transcriptomics-based taxonomies were collected from human, non-human primate, mouse, and reptile, and span multiple neocortical areas. In addition, several of these are matched to taxonomies collected using other modalities such as morphology, electrophysiology, epigenetics, and methylation. Such breadth provides multiple avenues of entry into this framework for annotation of novel datasets and allows for a more flexible implementation of the specific analysis workflow described in Figure 2. In particular, instead of requiring alignment of new datasets to the reference taxonomy, new datasets can be aligned with any taxonomy from Table 4, and information about cell type can then be inferred from any cell sets in any included taxonomy with a common aligned alias as the matched cell set. If this process is applied to novel taxonomies and the output files are included as supplemental materials in any resulting manuscript, then these taxonomies can be included in any future centralized database with minimal effort, providing a richer reference for further study.

Table 4
Taxonomies with applied CCN.

Table showing the set of taxonomies included in Supplementary file 1. All taxonomies include the annotated nomenclature table. Asterisk (*) and carrot (^) indicate that the updated dendrogram and cell to cell set mapping files are also included for that taxonomy, respectively. CCN202002270 is the reference taxonomy presented in Figure 3B.

Taxonomy idDescriptionReference
CCN201810310^*Mouse VISp + ALM (from the Tasic et al., 2018)Tasic et al., 2018
CCN201908210^*Human MTG (from the Tasic et al., 2018)Hodge et al., 2019
CCN201908211^*Joint mouse/human analysis (slight modification from Hodge et al., 2019)Hodge et al., 2019
CCN201912130Human M1 taxonomy using 10× dataBakken et al., 2020a
CCN201912131Human M1 taxonomy using Smart-seq and 10x dataBakken et al., 2020a
CCN201912132Marmoset M1 taxonomy using 10× dataBakken et al., 2020a
CCN202002013*Mouse MOp BICCN taxonomy using multiple RNAseq datasetsYao et al., 2020a
CCN202002270Cross species (integrated) transcriptomics taxonomyBakken et al., 2020a
CCN202002271Macaque transcriptomics taxonomy, layer 5/6 onlyBakken et al., 2020a
CCN202002272Human DNA methylation taxonomyBakken et al., 2020a
CCN202002273Human ATAC-seq taxonomyBakken et al., 2020a
CCN202002274Marmoset DNA methylation taxonomyBakken et al., 2020a
CCN202002275Marmoset ATAC-seq taxonomyBakken et al., 2020a
CCN202002276Mouse DNA methylation taxonomyYao et al., 2020a
CCN202002277Mouse ATAC-seq taxonomyYao et al., 2020a
CCN202005150^Mouse inhibitory neurons in VISp defined using electrophysiology, morphology, and transcriptomicsGouwens et al., 2020
CCN201906170Mouse neurons in VISp defined using electrophysiology and morphologyGouwens et al., 2019
CCN201805250Turtle pallium transcriptomics taxonomyTosches et al., 2018

Discussion

The complexity of cell types taxonomies and their generation now requires conventions and methodology to capture and communicate essential knowledge derived from experiments. The CCN provides a schema and workflow that allows scientists to organize their cell types within a single dataset and to link taxonomies using the aligned alias and other alias terms. However, the CCN is currently a stand-alone nomenclature schema that lacks the centralization and governance of gene-based standards proposed by the HUGO Gene Nomenclature Committee (HGNC) (Bruford et al., 2020) and does not yet have a mechanism for integrating with underlying data and metadata.

These shortcomings would be addressed through linking cell type ontology curation with corresponding databases. Ontology curation would allow users to associate data-derived cell sets to common usage terms from prior knowledge, and connect directly with the well-annotated ontology tools that are available for broader classifications (e.g., the Cell Ontology, http://www.obofoundry.org/ontology/). In addition, aligned aliases defined in reference taxonomies could represent a starting point for expansion of existing ontologies to presumptive cell types defined using other data-driven approaches (such as the terms in Table 2 for cortical neurons). Centralizing a location for taxonomies, their associated cell sets, and underlying datasets could provide a more robust ecosystem for comparing relevant nomenclature information, other metadata, and the primary data itself. Such databases can be implemented using knowledge graph-based models (Alshahrani et al., 2017; Waagmeester et al., 2020), permitting traversal across a data, information, knowledge, and wisdom hierarchy (Rowley, 2007). A potential presentation could be a ‘Cell Type Card’, instantiated as a web-accessible reference that compiles information about a specific cell set in a standardized summary. Not unlike a periodic table in structure, this concept has been implemented for genes (http://www.genecards.org), and as a prototype using transcriptomically defined cell types in mouse hippocampus and cortex (Yao et al., 2020b).

Incorporating community input on the definition and management of cell type standards will be necessary as new experiments are performed and additional evidence emerges. A cell type standards governing body would ideally be responsible for vetting ontologies for organizing data, controlled vocabulary for assigning cell type nomenclature, and will need to define a process for submission to ensure that critical data and metadata can be stored in a robust database. Deciding which taxonomies to include as reference taxonomies, along with frequency of updates, and how to address the breadth of brain regions, data modalities, cross-species reconciliation, and stochasticity of developmental and disease trajectories is essential. Organizing such a governance framework represents an important step and efforts are under way through BRAIN Initiative-funded initiatives, but is beyond the scope of work presented here.

This work presents a framework that is a modest step in a long and iterative process. With cross-disciplinary partnership and ever-increasing data, refinement of this proposed convention is expected. Together with collaborators, the Allen Institute has begun to combine ontology development, data integration, and nomenclature formalization efforts with the aim of facilitating cell type standards for the neuroscience community. Together with the goals articulated as part of the NIH BICCN and Brain Cell Data Center (BCDC) (https://biccn.org/), we seek to provide access to the diverse cell types in the human, mouse, and marmoset brain. The Allen Brain Map Community Forum (https://community.brain-map.org/c/cell-taxonomies/) has a dedicated space for discussion related to cell taxonomy refinement, to promote open and accessible opportunity for exchanging ideas and suggesting improvements. Beyond brain, whole-body projects seeking to categorize cell types, such as the NIH Common Fund-supported Human BioMolecular Atlas Program (HuBMAP, https://hubmapconsortium.org/) and the Human Cell Atlas consortium (https://www.humancellatlas.org/), will also need to leverage organizational conventions such as this, for comparable purposes that are practical and promote scientific rigor. The authors look forward to engagement with emerging communities and forums as evolution of cell classification methods continues.

Materials and methods

User-friendly executable code for applying the CCN is available on GitHub (https://github.com/AllenInstitute/nomenclature). This repository aims at providing a set of standardized terms and files that are immediately useful and also formatted to seed any future or in-process platform for cell type characterization and annotation. It is written as a user-friendly script in the R programming language (https://www.R-project.org) that includes specific details for how to apply the CCN, along with a set of example input files from a published study on cell types in human MTG (Hodge et al., 2019).

Step-by-step application of the CCN to human MTG

Request a detailed protocol

This section addresses how to apply the CCN to an example taxonomy, from human MTG. Three inputs are required to run the CCN: (1) a cell type taxonomy (not necessarily hierarchical), (2) a cell metadata file with cluster assignments (and optionally additional information), and (3) optional manual annotations of cell sets (e.g., aliases), which typically would be completed during taxonomy generation. Example files for human MTG are saved in the repository’s data folder. Once all files are downloaded and the workspace is set up, several global variables are set, which propagate to each cell set as a starting point, and which can be updated for specific cell sets later in the process. A unique taxonomy_id of the format CCN[YYYYMMDD][#] is chosen, which will match the prefix for cell set accession IDs. To ensure uniqueness across all taxonomies, taxonomy_ids are tracked in a public-facing database, with future plans to transfer these to a more permanent solution that will also provide storage for accompanying CCN output files and relevant metadata. In addition, values for the cell set assignee, citation, structure, and ontology tag are defined, along with the prefix(es) for the cell set label. For human MTG, ‘CCN201908210’, ‘Trygve Bakken’, ‘10.1038/s41586-019-1506-7’, ‘middle temporal gyrus’, ‘UBERON:0002771’, and ‘MTG’ are used, respectively. Next, the dendrogram is read in as the starting point for defining cell sets by including both provisional cell types (terminal leaves) and groups of cell types with similar expression patterns (internal nodes). Figure 1B shows the annotated dendrogram in human MTG provided in the GitHub repository, under which are displayed the names of cell types presented in Hodge et al., 2019. These provisional cell types were named using an entirely data-driven strategy: (cell class) (L)(cortical layers of localization) (canonical marker gene) ([optional] specific marker gene), as discussed in Hodge et al., 2019.

The main script takes the preset values and dendrogram as input, assigns accession ids and labels for each cell set, and then outputs an intermediate table and a dendrogram with all CCN labels defined in the previous section (Figure 1C). By default, the provisional cell types are assigned their original name from the dendrogram as preferred alias (e.g., ‘Inh L1-2 PAX6 CDH12’), while this field is left blank for internal nodes. For all cell sets, fields for additional and aligned alias are also initially left blank. Cell set labels are formatted as the label prefix (e.g., ‘MTG’) followed by a list of the cell set labels of all included provisional cell types. For example, the ‘LAMP5/PAX6’ node in human MTG includes the first six cell types in the tree and therefore has the cell set label of ‘MTG 001–006’. The table with these CCN tags for each cell set is then written to a csv file for manual annotation, which includes two critical aspects: (1) creation of new cell sets and (2) updating CCN tags for any cell sets. Cell sets corresponding to groups of relevant cell types either based on biological relevance (e.g., LAMP5-associated cell types in MTG) or as defined using a non-hierarchical computational strategy can be added at this step. In addition, cell sets corresponding to metadata rather than cell types can also be added. For example, in human MTG, ‘CS201908210_154’ corresponds to the set of nuclei collected from neurosurgical tissue and is given a cell set label of ‘Metadata 1’ and a preferred alias of ‘Neurosurgical’.

After finalizing these cell sets, they can then be annotated to include additional aliases based on known literature (e.g., assigning ‘basket’ or ‘fast-spiking’ to relevant PVALB+ cell sets), along with the assignees and citations from which such aliases were derived (e.g., ‘Nathan Gouwens’ and ‘10.1101/2020.02.03.932244’). As another example, Inh L1-4 LAMP5 LCP2 corresponds to Rosehip cells (see Boldog et al., 2018) and therefore an additional alias for this cell type is ‘Rosehip’. The structures and associated ontology tags could also be updated at this stage. For example, previous studies in mice suggest that most non-neuronal and GABAergic cell types are conserved across cortical areas (Tasic et al., 2018; Yao et al., 2020b). Although not done here, relevant cell sets could be generalized to an anatomic structure such as ‘Neocortex’ (UBERON:0001950). A final component of manual annotation is to update relevant cell sets with an aligned alias (e.g., a common usage term), which is critical for comparison of taxonomies in the CCN. In this case, aligned aliases for all cell sets were assigned by comparison with the human M1 reference (Bakken et al., 2020a), as shown in Figure 4 for glutamatergic neurons and as described in Use case 1. It is important to note that this step requires a previous computational alignment (or some other strategy to match cell sets) to use as evidence prior to assignment of the aligned alias; cell set alignment itself is not performed as part of the CCN.

After completing the manual annotations, the updated table is read back into R for additional dendrogram annotation and for mapping of cells to cell sets. Dendrograms are revised to include the new cell sets and annotations, and then output in a few standard formats (.RData, .json, and .pdf) for ontology construction and other downstream uses. Individual cells are then mapped to cell sets using the cell metadata table, which includes a unique cell identifier, provisional cell type classification, and other optional metadata. Cells are then mapped to cell sets representing one or more provisional cell type using the annotated dendrogram and/or the updated nomenclature table using the cell set label tag. Finally, cells are mapped to remaining cell sets (if any) using custom scripts. This results in a table of binary calls (0 = no, 1 = yes), indicating exclusion or inclusion of each cell in each cell set (Figure 1D), which is written to another csv file as part of the process. This format is designed to allow for probabilistic mapping of cells to cell sets, which is beyond the scope of this manuscript. These output files are intended to be directly included as supplemental materials in manuscripts performing cell type classification in any species. In addition, the GitHub repository will be updated to include conversion functions to allow input into future community-accepted cell type databases, as such resources become available. Supplementary file 1 includes a table of applied nomenclature for all taxonomies discussed in this manuscript, along with cell to cell set mappings for a few example taxonomies.

Data availability

This work describes the creation of a cell type nomenclature convention that will, with adoption by the community, become a standard. The data cited is open data available though the Allen Institute web portal, https://brain-map.org. An open Forum is available to engage the community in further development, at https://portal.brain-map.org/explore/classes/nomenclature. Data referenced in this study is also made available according the terms of NIH's Brain Research through Advancing Innovative Neurotechnologies (BRAIN) Initiative - Cell Census Network (BICCN), through the Brain Cell Data Center portal, https://biccn.org/ and https://biccn.org/data.

References

  1. Preprint
    1. Adkins RS
    2. Aldridge AI
    3. Allen S
    4. Ament SA
    5. An X
    6. Armand E
    7. Ascoli GA
    8. Bakken TE
    9. Bandrowski A
    10. Banerjee S
    11. Barkas N
    12. Bartlett A
    13. Bateup HS
    14. Margarita Behrens M
    15. Berens P
    16. Berg J
    17. Bernabucci M
    18. Bernaerts Y
    19. Bertagnolli D
    20. Biancalani T
    21. Boggeman L
    22. Sina Booeshaghi A
    23. Bowman I
    24. Bravo HC
    25. Cadwell CR
    26. Callaway EM
    27. Carlin B
    28. O’Connor C
    29. Carter R
    30. Casper T
    31. Castanon RG
    32. Castro JR
    33. Chance RK
    34. Chatterjee A
    35. Chen H
    36. Chun J
    37. Colantuoni C
    38. Crabtree J
    39. Creasy H
    40. Crichton K
    41. Crow M
    42. D’Orazi FD
    43. Daigle TL
    44. Dalley R
    45. Dee N
    46. Degatano K
    47. Dichter B
    48. Diep D
    49. Ding L
    50. Ding S-L
    51. Dominguez B
    52. Dong H-W
    53. Dong W
    54. Dougherty EL
    55. Ecker JR
    56. Eichhorn SW
    57. Fang R
    58. Felix V
    59. Feng G
    60. Feng Z
    61. Fischer S
    62. Fitzpatrick C
    63. Fong O
    64. Foster NN
    65. Galbavy W
    66. Gee JC
    67. Ghosh SS
    68. Giglio M
    69. Gillespie TH
    70. Gillis J
    71. Goldman M
    72. Goldy J
    73. Gong H
    74. Gou L
    75. Grauer M
    76. Halchenko YO
    77. Harris JA
    78. Hartmanis L
    79. Hatfield JT
    80. Hawrylycz M
    81. Helba B
    82. Herb BR
    83. Hertzano R
    84. Hintiryan H
    85. Hirokawa KE
    86. Hockemeyer D
    87. Hodge RD
    88. Hood G
    89. Horwitz GD
    90. Hou X
    91. Hu L
    92. Hu Q
    93. Josh Huang Z
    94. Huo B
    95. Ito-Cole T
    96. Jacobs M
    97. Jia X
    98. Jiang S
    99. Jiang T
    100. Jiang X
    101. Jin X
    102. Jorstad NL
    103. Kalmbach BE
    104. Kancherla J
    105. Dirk Keene C
    106. Kelly K
    107. Khajouei F
    108. Kharchenko PV
    109. Kim G
    110. Al K
    111. Kobak D
    112. Konwar K
    113. Kramer DJ
    114. Krienen FM
    115. Kroll M
    116. Kuang X
    117. Kuo H-C
    118. Lake BB
    119. Larsen R
    120. Lathia K
    121. Laturnus S
    122. Lee AY
    123. Lee C-T
    124. Lee K-F
    125. Lein ES
    126. Lesnar P
    127. Li A
    128. Li X
    129. Li X
    130. Ye L
    131. Li Y
    132. Li Y
    133. Lim B
    134. Linnarsson S
    135. Liu CS
    136. Liu H
    137. Liu L
    138. Lucero JD
    139. Luo C
    140. Luo Q
    141. Macosko EZ
    142. Mahurkar A
    143. Martone ME
    144. Matho KS
    145. McCarroll SA
    146. McCracken C
    147. McMillen D
    148. Miranda E
    149. Mitra PP
    150. Miyazaki PA
    151. Mizrachi J
    152. Mok S
    153. Mukamel EA
    154. Mulherkar S
    155. Nadaf NM
    156. Naeemi M
    157. Narasimhan A
    158. Nery JR
    159. Ngai J
    160. Nguyen TN
    161. Nickel L
    162. Nicovich PR
    163. Niu S-Y
    164. Ntranos V
    165. Nunn M
    166. Olley D
    167. Orvis J
    168. Osteen JK
    169. Osten P
    170. Owen SF
    171. Pachter L
    172. Palaniswamy R
    173. Palmer CR
    174. Pang Y
    175. Peng H
    176. Pham T
    177. Pinto-Duarte A
    178. Plongthongkum N
    179. Poirion O
    180. Preissl S
    181. Purdom E
    182. Qu L
    183. Rashid M
    184. Reed NM
    185. Regev A
    186. Ren B
    187. Ren M
    188. Rimorin C
    189. Risso D
    190. Rivkin AC
    191. Muñoz-Castañeda R
    192. Romanow WJ
    193. Ropelewski AJ
    194. de Bézieux HR
    195. Ruan Z
    196. Sandberg R
    197. Savoia S
    198. Scala F
    199. Schor M
    200. Shen E
    201. Siletti K
    202. Smith JB
    203. Smith K
    204. Somasundaram S
    205. Song Y
    206. Sorensen SA
    207. Stafford DA
    208. Street K
    209. Sulc J
    210. Sunkin S
    211. Svensson V
    212. Tan P
    213. Tan ZH
    214. Tasic B
    215. Thompson C
    216. Tian W
    217. Tickle TL
    218. Tieu M
    219. Ting JT
    220. Tolias AS
    221. Torkelson A
    222. Tung H
    223. Vaishnav ED
    224. Van den Berge K
    225. van Velthoven CTJ
    226. Vanderburg CR
    227. Veldman MB
    228. Vu M
    229. Wakeman W
    230. Wang P
    231. Wang Q
    232. Wang X
    233. Wang Y
    234. Wang Y
    235. Welch JD
    236. White O
    237. Williams E
    238. Xie F
    239. Xie P
    240. Xiong F
    241. Yang W
    242. Yanny AM
    243. Yao Z
    244. Yin L
    245. Yu Y
    246. Yuan J
    247. Zeng H
    248. Zhang K
    249. Zhang M
    250. Zhang Z
    251. Zhao S
    252. Zhao X
    253. Zhou J
    254. Zhuang X
    255. Zingg B
    256. BRAIN Initiative Cell Census Network (BICCN)
    (2020) A multimodal cell census and atlas of the mammalian primary motor cortex
    bioRxiv.
    https://doi.org/10.1101/2020.10.19.343129
  2. Conference
    1. Bakken TE
    2. van Velthoven CTJ
    3. Menon V
    4. Hodge RD
    5. Yao Z
    6. Nguyen TN
    7. Graybuck LT
    8. Horwitz GD
    9. Bertagnolli D
    10. Goldy J
    11. Garren E
    12. Parry S
    13. Casper T
    14. Shehata SI
    15. Barkan ER
    16. Szafer A
    17. Levi BP
    18. Dee N
    19. Smith KA
    20. Sunkin SM
    21. Bernard A
    22. Phillips JW
    23. Hawrylycz M
    24. Koch C
    25. Murphy G
    26. Lein E
    27. Zeng H
    28. Tasic B
    (2020b)
    Single-cell RNA-seq uncovers shared and distinct axes of variation in dorsal LGN neurons in mice, non-human primates and humans
    Cold Spring Harbor Laboratory.
    1. Betz W
    (1874)
    Anatomischer nachweis zweier gehirncentra
    Centralblatt Für Die Medizinischen Wissenschaften 12:578–599.
  3. Book
    1. Cajal R
    (1899)
    La Textura Del Sistema Nerviosa Del Hombre Y Los Vertebrados
     Nicolas Moya.
  4. Book
    1. Gala R
    2. Gouwens N
    3. Yao Z
    4. Budzillo A
    5. Penn O
    6. Tasic B
    7. Murphy G
    8. Zeng H
    9. Sümbül U
    (2019)
    A coupled autoencoder approach for multi-modal analysis of cell types
    In: Wallach H, Larochelle H, Beygelzimer A, Fox E, Garnett R, editors. Advances in Neural Information Processing Systems. Curran Associates, Inc. pp. 9267–9276.
  5. Book
    1. Jarvis ED
    (2009) Evolution of the Pallium in Birds and Reptiles
    In: Binder M. D, Hirokawa N, Windhorst U, editors. Encyclopedia of Neuroscience. Springer. pp. 1390–1400.
    https://doi.org/10.1007/978-3-540-29678-2_3165
  6. Conference
    1. Kozareva V
    2. Martin C
    3. Osorno T
    4. Rudolph S
    5. Guo C
    6. Vanderburg C
    7. Nadaf N
    8. Regev A
    9. Regehr W
    10. Macosko E
    (2020)
    A transcriptomic atlas of the mouse cerebellum reveals regional specializations and novel cell types
    Cold Spring Harbor Laboratory.
  7. Book
    1. Lees JR
    2. Azimzadeh AM
    3. Ding Y
    4. Webb TJ
    5. Bromberg JS
    (2015)
    Cells of the immune system
    In: Li CX, Jevnikar A. M, editors. Transplant Immunology. John Wiley & Sons, Ltd. pp. 25–47.
  8. Book
    1. Rakic P
    (1984) Organizing Principles for Development of Primate Cerebral Cortex
    In: Sharma S. C, editors. Organizing Principles of Neural Development. Springer. pp. 21–48.
    https://doi.org/10.1007/978-1-4684-4802-3_2
    1. Szentágothai J
    2. Arbib MA
    (1974)
    Conceptual models of neural organization
    Neurosciences Research Program Bulletin 12:305–510.

Decision letter

  1. Catherine Dulac
    Senior Editor; Harvard University, United States
  2. Genevieve Konopka
    Reviewing Editor; University of Texas Southwestern Medical Center, United States
  3. Genevieve Konopka
    Reviewer; University of Texas Southwestern Medical Center, United States
  4. Fenna Krienen
    Reviewer; Harvard Medical School, United States
  5. Joseph D Dougherty
    Reviewer; Washington University in St. Louishington University, United States

In the interests of transparency, eLife publishes the most substantive revision requests and the accompanying author responses.

Acceptance summary:

Defining and naming brain cell types has been a long-standing challenge in neuroscience. New high-throughput techniques have facilitated the generation of many large datasets that detail multi-modal information about cell types. This manuscript presents a system for a generalizable nomenclature that can be applied to the mammalian brain. The system will permit researchers to compare their own cell-type data with other published datasets and facilitate consistent cell-type naming.

Decision letter after peer review:

Thank you for submitting your article "Cell type nomenclature for the mammalian brain: Development and application of a systematic, extensible convention" for consideration by eLife. Your article has been reviewed by three peer reviewers, including Genevieve Konopka as the Reviewing Editor and Reviewer #1, and the evaluation has been overseen by Catherine Dulac as the Senior Editor. The following individuals involved in review of your submission have agreed to reveal their identity: Fenna Krienen (Reviewer #2); Joseph D Dougherty (Reviewer #3).

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

We would like to draw your attention to changes in our revision policy that we have made in response to COVID-19 (https://elifesciences.org/articles/57162). Specifically, when editors judge that a submitted work as a whole belongs in eLife but that some conclusions require a modest amount of additional new data, as they do with your paper, we are asking that the manuscript be revised to either limit claims to those supported by data in hand, or to explicitly state that the relevant conclusions require additional supporting data.

Our expectation is that the authors will eventually carry out the additional experiments and report on how they affect the relevant conclusions either in a preprint on bioRxiv or medRxiv, or if appropriate, as a Research Advance in eLife, either of which would be linked to the original paper.

Summary:

All reviewers were in agreement that this paper presents some potential ways of tackling an important problem. However, we think there are some aspects of the paper that could be improved for clarity and to make it accessible to a broad audience. The new policy at eLife is to write a summary of essential revisions for the authors when a revised manuscript is warranted and not necessarily send the authors the full reviews.

Our essential revisions include: (1) more details on the immediate use of the system and potentially a step-by-step guide-it definitely seems like this system needs to be much more user friendly even for programmers; (2) how this approach would work without a reference set and ultimately the plan for oversight of such a reference; and (3) how to apply this across species, regions, and modalities.

Even though the new policy is not to send the full reviews, the reviewers each had some useful ideas and nuances about the essential revisions so we are also attaching them here. We do not expect you to address each and every one of these points/comments but rather take them into consideration as you address the essential revisions.

Reviewer #1:

This paper provides an important foundation to develop a universal nomenclature system for vertebrate cell types in single-cell sequencing studies. Similar to efforts to establish gene nomenclature guidelines, this resource is necessary to improve comparisons between datasets and species. As the authors note, any useful system will need to be widely agreed upon and adopted by scientists, and this paper is a good step in initiating that process. I have some general comments on the manuscript and the system that could be implemented. However, I imagine that there might be further modifications that would occur after publication of this manuscript.

1) The glutamatergic nomenclature scheme is neocortical-centric. The "layer" designation should really be a regional specification. For example, "L2/3 IT 4" could have a prefix indicating neocortex like "N L2/3 IT 4." This could allow for expansions including layers in non-neocortical regions, like hippocampus (e.g., "H L2"), or non-layered regions of the pallium like the claustrum (e.g., "C IT"). This would also enhance applicability for non-mammalian brains lacking layered organizations.

2) Glutamatergic cell types across broad vertebrate taxa (e.g. reptiles, birds, mammals) cannot be universally compared to layer-specific nomenclature in mammals because neocortical organization is unique to mammals. Therefore, a universal reference taxonomy must be supplied independently for each major grouping, and homologies can be suggested by, but not defined from, nomenclature. In general, it would be best to emphasize that similarities in cell types are not necessarily indicative of homologies for those cell types. A better description of how one might deal with cell types that diverge across species might be warranted.

3) At some point, the integration of spatial information (other than layers, such as dorsal or ventral) into single-cell sequencing experiments may become commonplace. This nomenclature scheme should be able to incorporate a spatial component if this information proves meaningful, similar to how the authors describe the use of electrophysiological data.

4) The taxonomy ID format CS[YYMMDD][#] is generally fine but note that a four-digit year notation would remove date ambiguities and is in line with universal date formats (ISO 8601).

5) "Cell set label" seems superfluous when "Cell set accession ID" can sufficiently identify each unique cell group. The distinction between identifying "neuron" versus "non-neuron" does not seem to provide enough meaningful information to warrant a separate identity. Table 1 already mentions that this label may be redundant.

6) How/ what steps will be taken to vet the data to include in the taxonomy? As the authors mention there are so many groups generating single cell data but not all of these datasets are of great quality – especially for a naming convention. Will BICCN do this? I imagine the HCA folks might only take care of human data or certain model systems? Who will be in charge of defining the reference cell types?

7) How will the batch differences (lab, sequencing method, machine) be handled?

8) It was mentioned that the same mapping and clustering technique will be applied for all datasets. -Is it possible that certain datasets might be more amenable to "tweaks" in a pipeline rather than a one size fits all approach?

Reviewer #2:

This manuscript presents a proposal for a generalizable cell type nomenclature convention system for the mammalian brain. How to define and name brain cell types is a longstanding issue; there is currently no standard convention. The recent explosion of large, single cell datasets based on molecular measurements (RNA, epigenetics) means there is both an opportunity to gain detailed and often multi-modal information about cell types, as well as a need to reconcile nomenclatures across studies. As such, this article presents a thoughtful, implementable system for a standardized nomenclature, as well as a discussion of some infrastructure and governance considerations that would facilitate widespread community adoption.

1) I read this project as having two components: (A) an immediately implementable nomenclature system (with associated code for end-users to run on their own data), and (B) a proposal outlining possible infrastructures that would support 'alignment' of community data to reference(s) and linking cell type information across studies (including the need for computational infrastructure, governance). I think there should be better motivation for why (A) should be adopted without (B). The authors state that the schema will be "immediately useful" but an end-user may not see the practical advantage over their own in-house conventions (unless/until there is a reference, governance, ontology with controlled values, etc).

2) Practically speaking, the article (or at least the github repo) should also clearly state what best-practices adoption of (A) would look like before a data repository is established, i.e. which outputs or terms are used in figures and tables, should full outputs be included as supplementary data (spreadsheets) in manuscripts etc. The schema introduces quite a few new terms and conventions and I think you have to be more explicit for how end-users should incorporate it in their own work. This could be achieved by more detailed examples as well as guidance in the github (note I ran the code using the supplied MTG dataset but did not try applying it to new data. Authors might consider adding vignettes that take in common algorithm outputs, e.g. output from Seurat, 10X cell ranger pipeline etc).

3) Multiple examples of applying the nomenclature schema to published datasets are given. I wonder whether it would be more effective to focus on one dataset. Figure 1 (human transcriptomic MTG data) and Figure 2 (mouse transcriptomic VISp data) largely do the same work, though they are displayed and formatted differently, which is a bit confusing. Figure 5 presents an example of creating a reference from the data in Figures 1 and 2, but several aspects are not clear: in (B), how are the preferred aliases named, (C) does "Human types" = cell set labels in Figure 1B, (D) how are the lines that visually link the modalities derived, and how are they formally represented in the nomenclature system. The final example (multi-modal, multi species comparison of cortical area M1) is also complex as it contains multiple datatypes as well as a derived 'reference', but as currently presented is not very effective in showing how the nomenclature is applied or how taxonomies are linked. I wonder whether it would be more effective to focus either on the taxonomies presented in Figures 1, 2 and 5, or alternatively on this large M1 study, and fully unpack how to apply and visually represent the schema with just one of these. Alternatively, one could start with toy examples that illustrate the process before applying to real data (again one might be better than several).

4) Creating or using a reference is not explicitly part of the proposed nomenclature schema, but clearly has great utility in terms of linking taxonomies (Figure 5 and 6). The authors could consider describing in more detail the considerations they have made in forming such cross dataset or cross species references. How do user-generated references fit in the proposed system – can the same classification system be applied (i.e. each reference has a taxonomy ID, each cluster has a cell set accession ID) or are there other metadata that should be included? Authors might consider a separate figure devoted to applying the nomenclature schema to a derived reference (e.g. unpacking something like Figure 5B).

5) If I understand correctly, cell set aliases can be based on seemingly very different types of evidence, including (1) quantitative alignment to a reference, (2) user inference based on observation of shared features such as marker genes (i.e., both datasets have Chodl+ cluster, so it is probably the same), and (3) inference based on prior knowledge (an ME cell set that has a location and electrophysiological profile consistent with chandelier cells is aliased to a transcriptomic cell set named "chandelier"). These are really different types of evidence and perhaps should be controlled or distinguished in the system.

Reviewer #3:

This paper is a thoughtful contribution to a tough problem and represents a reasonable step in the right direction. I think it would fit well with eLife and form the basis for beginning of better cross-paper curation of scRNAseq data and other related datasets. It is not the full solution, but is careful in its claims and I think will be an important part of the conversation towards those larger solutions. I have some moderate recommendations for revisions.

1) For cell set accession IDs, they may want to include a '.' between the CS191012 and the unique number for the cell set, and then just iterate the numbers up (.1, .2, .3, … .12 … .10000). The current scheme will max out at 1000 cell sets. That might seem like a huge number now, but someone soon will do 10x on the whole body in one paper and need more than 1000 cell sets.

2) I don't have the bandwidth at the moment to do this as a reviewer, but I would recommend they consider approaching 2-3 labs outside of their group (i.e. external Beta testers) and have them try to enter on of their datasets into this structure using the GitHub code and see how it goes, using only this paper and the associated materials as instructions and iron out any wrinkles or misunderstandings that emerge. If you want this to roll out smoothly, you want researchers' first experiences trying it to be positive to help promote wide adoption.

3) I would recommend adding a section (or perhaps a supplement) that is a clear checklist of what to do as an end researcher who might want to adopt this. If you've convinced me to do this with my data, what are the explicit and actionable recommendations for what I should do? Is this meant to be like submitting your data to GEO? Where any paper publishing a scRNAseq dataset will adopt this standardized approach to naming as the simultaneously upload their taxonomy to a particular database in a standardized format, and put a link in their Materials and methods section? I feel like this is not quite proposing that (as no such database was highlighted, though they highlight the need for one). Or rather is the hope that anyone who generates a scRNAseq dataset will provide their taxonomy in a standard file format as a supplement to their paper? If so, defining a file type (a .txon file or something?) that you are recommending everyone generate and add as supplement might be what you are championing. Explicitly naming that filetype(s) and making that recommendation might help. (If so, is that something that could be rolled into standard analysis packages – e.g. Seuret? That would lower the barrier to adoption) Or is this more like just trying to have everyone agree to use standardized gene names when they mention them? But not necessarily provide supplementary files. Like just being careful how you format your writing and figures like how I should use Pvalb for mouse genes and PVALB for human and PVALB for the protein, and not PV, PVA, etc? Anyway, I just wanted more concrete recommendations of what our expectations as authors (and as reviewers) ought to be for adoption of this standard.

As a starting point, perhaps just recommending a defined file type generated by the code (the .txon file) be included as a supplement is a reasonable recommendation at this time.

Basically, overall I think this paper is making an important and timely contribution. It did a good job of explaining their solution to addressing some of the challenges for annotating these datasets, but stopped just short of a concrete guide on how one could implement it in the near term.

https://doi.org/10.7554/eLife.59928.sa1

Author response

Summary:

All reviewers were in agreement that this paper presents some potential ways of tackling an important problem. However, we think there are some aspects of the paper that could be improved for clarity and to make it accessible to a broad audience. The new policy at eLife is to write a summary of essential revisions for the authors when a revised manuscript is warranted and not necessarily send the authors the full reviews.

Our essential revisions include: (1) more details on the immediate use of the system and potentially a step-by-step guide-it definitely seems like this system needs to be much more user friendly even for programmers; (2) how this approach would work without a reference set and ultimately the plan for oversight of such a reference; and (3) how to apply this across species, regions, and modalities.

We have reorganized our manuscript to include sections addressing each of these requests. The section titled, “Step by step application of the CCN to human MTG,” walks through the process of applying the CCN to an example data set in human middle temporal gyrus using the GitHub code. In addition, the GitHub repository itself has improved annotation, so that it is accessible as an application independent of this manuscript. After the proposal of a cell typing convention for the mammalian cortex, we present the concept of a reference atlas, and complementarity to the CCN. While there is nothing in the schema itself distinguishing a reference taxonomy from any other taxonomy, the reference taxonomies are primarily used to define aligned aliases and to facilitate alignment of other data sets. We present a specific example reference taxonomy created from the primary motor cortex (M1) in mouse, marmoset, and human, which replaces the human MTG / mouse VISp taxonomy presented in the initial submission. Governance of such a reference is beyond the scope of this manuscript, and we have substantially cut the sections related to governance and potential for future work for this reason. That being said, we provide some evidence for why our current reference is a good reference and present a collection of 18 taxonomies that use this integrated taxonomy as their reference framework.

Finally, we added a section, “Applying the CCN to existing and new datasets,” which presents five examples for how the CCN can be used right now, in the absence of a centralized repository for storing taxonomies and cell sets. This section reframes some of the use cases presented in the initial submission, and also includes some new use cases that together address how to “apply this across species, regions, and modalities.” For example, there are use cases on using the CCN to apply nomenclature to a non-mammalian brain structure, the turtle pallium, in the context of the presented reference taxonomy; and, adding aligned aliases to human MTG based on alignment with human M1; and, on mapping a taxonomy defined using cell morphology and biophysics to the cell type reference taxonomy, based on transgenically defined cell markers. The final two use cases explore biologically relevant inferences about a specific cell type (Sst-Chodl) based on aligned aliases, and on comparison of a novel data set to a set of 18 compiled taxonomies. Together, we feel that these changes address the reviewer concerns, leading to much stronger manuscript and nomenclature schema overall.

Even though the new policy is not to send the full reviews, the reviewers each had some useful ideas and nuances about the essential revisions so we are also attaching them here. We do not expect you to address each and every one of these points/comments but rather take them into consideration as you address the essential revisions.

Reviewer #1:

This paper provides an important foundation to develop a universal nomenclature system for vertebrate cell types in single-cell sequencing studies. Similar to efforts to establish gene nomenclature guidelines, this resource is necessary to improve comparisons between datasets and species. As the authors note, any useful system will need to be widely agreed upon and adopted by scientists, and this paper is a good step in initiating that process. I have some general comments on the manuscript and the system that could be implemented. However, I imagine that there might be further modifications that would occur after publication of this manuscript.

1) The glutamatergic nomenclature scheme is neocortical-centric. The "layer" designation should really be a regional specification. For example, "L2/3 IT 4" could have a prefix indicating neocortex like "N L2/3 IT 4." This could allow for expansions including layers in non-neocortical regions, like hippocampus (e.g., "H L2"), or non-layered regions of the pallium like the claustrum (e.g., "C IT"). This would also enhance applicability for non-mammalian brains lacking layered organizations.

This is an excellent point and was actually a primary focus of the workshop, “Defining an Ontological Framework for a Brain Cell Type Taxonomy: Single-Cell -omics and Data-Driven Nomenclature,” held in Seattle WA, June 2019, and from which many ideas is this manuscript originated. Unfortunately, there was very little consensus at this workshop on specifically what we should call these cell types and what level of “regional” resolution (e.g., “middle temporal gyrus” vs. “neocortex” vs. “brain”) should be used. For example, some cell types are likely region-specific (e.g., glutamatergic types, as mentioned) while others (e.g., glia) likely span multiple regions, and in many cases the regional resolution may change depending on the algorithms used for cell typing. In this case, we have attempted to bypass this issue somewhat by dividing the CCN into a schema component (which works with any taxonomy) and a specific naming convention that is limited to mammalian neocortex. Within neocortex, we further address this by implementing a proposed solution from the workshop: to define two distinct ontologies for each cell set. In additional to the cell type ontology that would be mostly region-agnostic, we provide two slots to assign a specific brain (or body) structure (“cell_set_structure”) and an associated tag a cross-species structural ontology (“cell_set_ontology_tag”) such as UBERON or a single-species ontology (e.g. from http://atlas.brain-map.org/) if higher resolution cell types are needed.

2) Glutamatergic cell types across broad vertebrate taxa (e.g. reptiles, birds, mammals) cannot be universally compared to layer-specific nomenclature in mammals because neocortical organization is unique to mammals. Therefore, a universal reference taxonomy must be supplied independently for each major grouping, and homologies can be suggested by, but not defined from, nomenclature. In general, it would be best to emphasize that similarities in cell types are not necessarily indicative of homologies for those cell types. A better description of how one might deal with cell types that diverge across species might be warranted.

The intention of this nomenclature schema was to provide guidance for typing of cells in the mammalian cortex, as indicated in the title of our manuscript. One of the primary reasons for restricting our naming conventions to mammals is for the precise reason mentioned: that they may break down in other species. However, the schema itself is agnostic to species (or organ system) and can be run independently of a universal reference taxonomy or can still be linked to such a taxonomy using the aligned, preferred, and additional alias tags. For example, we now present as a use case the example of a turtle pallium which was aligned with mouse neocortex (to the extent possible) in the initial publication (Tosches et al., 2018). In this case, non-neuronal cells and GABAergic interneurons could be matched to some degree between species, and this information was captured by assigning aligned aliases of Pvalb, Sst, OPC, microglia, etc. to the appropriate cell sets in turtles. For glutamatergic neurons, only superficial vs. deep layer neurons could be distinguished, and these show an inversion between clades. This information is captured by the additional alias tag, and this entire taxonomy with applied CCN is included as part of Additional File 1 in our updated submission. Finally, it is worth reiterating that data-driven strategies for linking (or failing to link) such cross-species taxonomies are assumed to have already occurred prior to applying the CCN and that the goal of the CCN is to capture these existing links. It is also worth noting that no links to a reference taxonomy are required for application of the CCN to be of some use. We have updated the text of the manuscript to address these points.

3) At some point, the integration of spatial information (other than layers, such as dorsal or ventral) into single-cell sequencing experiments may become commonplace. This nomenclature scheme should be able to incorporate a spatial component if this information proves meaningful, similar to how the authors describe the use of electrophysiological data.

As mentioned in #1 above, we now add a “cell set structure” tag and an associated “cell set ontology tag” to our nomenclature schema, which should address integration of spatial information.

4) The taxonomy ID format CS[YYMMDD][#] is generally fine but note that a four-digit year notation would remove date ambiguities and is in line with universal date formats (ISO 8601).

Our taxonomy ID was intended to strike the balance between a long but highly informative name (e.g., “MTG-20201020-Transcriptomics”) and a short generic name (e.g., “t251”), while ensuring that all taxonomies remain unique in the short term. However, we recognize this weakness, along with the weakness of having a maximum of 1000 cell sets (as mentioned below). Given the reviewer feedback, we have made the following changes, which are now implemented in the manuscript and associated CCN scripts:

– Taxonomy_id = CCN[YYYYMMDD]#

– Cell_set_id = CS[YYYYMMDD]#_#

5) "Cell set label" seems superfluous when "Cell set accession ID" can sufficiently identify each unique cell group. The distinction between identifying "neuron" versus "non-neuron" does not seem to provide enough meaningful information to warrant a separate identity. Table 1 already mentions that this label may be redundant.

As multiple reviewers (and co-authors) agree with this assessment, we have dramatically reduced discussion of cell_set_label from the nomenclature schema in this version of the manuscript. The tag itself is important for the GitHub scripts to work properly and so the term remains, but it is only discussed in this context.

6) How/ what steps will be taken to vet the data to include in the taxonomy? As the authors mention there are so many groups generating single cell data but not all of these datasets are of great quality – especially for a naming convention. Will BICCN do this? I imagine the HCA folks might only take care of human data or certain model systems? Who will be in charge of defining the reference cell types?

The advantage of this nomenclature schema is that it could be run on any taxonomy, regardless of data modality or data quality. For example, a small lab could perform a clustering and integration analysis (e.g., with a pre-existing reference), and then name cell types using this schema in such a way that cell sets are linked with the reference. In this case, the reference would not be changed and the quality of the single cell data does not matter. The issue of governance is a challenging one that is beyond the scope of this manuscript, and therefore we have reduced our discussion and speculation on this topic. That being said, we feel that the new reference taxonomy presented here (cross-species primary motor cortex) represents a good starting reference taxonomy for mammalian cortex because it represents a well-curated taxonomy presented as the output of a multi-lab collaboration through the BICCN. To aid in its use we present all a collection of 18 taxonomies which use the aligned alias terms introduced in this taxonomy as Additional File 1.

7) How will the batch differences (lab, sequencing method, machine) be handled?

Batch differences should be handed prior to application of any nomenclature schema. Strategies for doing so are beyond the scope of this manuscript.

8) It was mentioned that the same mapping and clustering technique will be applied for all datasets. -Is it possible that certain datasets might be more amenable to "tweaks" in a pipeline rather than a one size fits all approach?

Clustering and alignment should be handed prior to application of any nomenclature schema. Strategies for doing so are beyond the scope of this manuscript.

Reviewer #2:

[…] 1) I read this project as having two components: (A) an immediately implementable nomenclature system (with associated code for end-users to run on their own data), and (B) a proposal outlining possible infrastructures that would support 'alignment' of community data to reference(s) and linking cell type information across studies (including the need for computational infrastructure, governance). I think there should be better motivation for why (A) should be adopted without (B). The authors state that the schema will be "immediately useful" but an end-user may not see the practical advantage over their own in-house conventions (unless/until there is a reference, governance, ontology with controlled values, etc).

We view (A) as containing two distinct components which are applicable in different (but overlapping) use cases, but otherwise agree with this assessment. First, the immediately implementable nomenclature system revolving around the idea of cell sets and taxonomies. Second, a specific convention for naming mammalian neocortical cell types, which presents a starting proposal for a common language to allow linking of cell types between studies. With this in mind, we have attempted to improve on our explanation for why (A) should be adopted without (B). This can be found in the section titled “Applying the CCN to existing and new datasets”, which includes five use cases for immediate application of the schema. Of particular note, we now present a series of taxonomies (including a reference taxonomy) on which the CCN has already been applied and to which other groups can immediately compare their results (see use case #5 and Supplementary file 1). This being said, we recognize that having an ontology with controlled values, a centralized database, and governance could increase utility and are actively pursuing this line of work. Such topics are challenging and beyond the scope of the manuscript and we therefore have reduced our discussion and speculation on these topics.

2) Practically speaking, the article (or at least the github repo) should also clearly state what best-practices adoption of (A) would look like before a data repository is established, i.e. which outputs or terms are used in figures and tables, should full outputs be included as supplementary data (spreadsheets) in manuscripts etc. The schema introduces quite a few new terms and conventions and I think you have to be more explicit for how end-users should incorporate it in their own work. This could be achieved by more detailed examples as well as guidance in the github (note I ran the code using the supplied MTG dataset but did not try applying it to new data. Authors might consider adding vignettes that take in common algorithm outputs, e.g. output from Seurat, 10X cell ranger pipeline etc).

This is a great suggestion. Our goal for this manuscript is to have a system that could be implemented by a computational or non-computational researcher by following step-by-step instructions. To this end the GitHub repo has been updated with more complete step-by-step instructions for implementation (along with corresponding code), and these instructions have been described in the manuscript in more detail than in the initial analysis. We have also made some edits to the CCN (described below) and have changed the GitHub code accordingly. As part of these updates, we have included text regarding what to do after implementation as related to proposed best practices for publication with the goal of allowing future intake into a centralized database directly from manuscript Supplementary files. The idea of providing vignettes for applying the CCN as part of a Seurat or 10x cell ranger pipeline is a great one that we hope to revisit at a later time.

3) Multiple examples of applying the nomenclature schema to published datasets are given. I wonder whether it would be more effective to focus on one dataset. Figure 1 (human transcriptomic MTG data) and Figure 2 (mouse transcriptomic VISp data) largely do the same work, though they are displayed and formatted differently, which is a bit confusing. Figure 5 presents an example of creating a reference from the data in Figures 1 and 2, but several aspects are not clear: in (B), how are the preferred aliases named, (C) does "Human types" = cell set labels in Figure 1B, (D) how are the lines that visually link the modalities derived, and how are they formally represented in the nomenclature system. The final example (multi-modal, multi species comparison of cortical area M1) is also complex as it contains multiple datatypes as well as a derived 'reference', but as currently presented is not very effective in showing how the nomenclature is applied or how taxonomies are linked. I wonder whether it would be more effective to focus either on the taxonomies presented in Figure 1, 2 and 5, or alternatively on this large M1 study, and fully unpack how to apply and visually represent the schema with just one of these. Alternatively, one could start with toy examples that illustrate the process before applying to real data (again one might be better than several).

Our initial strategy was to present a number of different examples for use of this nomenclature schema, but the point about presenting a single application in more depth and focus is well taken. To address this comment we have reorganized the manuscript as described in the response to the next comment, including redefining the multi-species M1 taxonomy as the reference and removing the figures on mouse VISp (although the taxonomies are still retained in Supplementary file 1).

4) Creating or using a reference is not explicitly part of the proposed nomenclature schema, but clearly has great utility in terms of linking taxonomies (Figures 5 and 6). The authors could consider describing in more detail the considerations they have made in forming such cross dataset or cross species references. How do user-generated references fit in the proposed system – can the same classification system be applied (i.e. each reference has a taxonomy ID, each cluster has a cell set accession ID) or are there other metadata that should be included? Authors might consider a separate figure devoted to applying the nomenclature schema to a derived reference (e.g. unpacking something like Figure 5B).

Reviewer #1 had similar concerns, which are addressed above. In summary, the advantage of the CCN is that it could be run on any taxonomy, regardless of data modality, data quality, or origin species (although some modifications to the specific naming conventions will likely be needed for distant species and for structures outside of neocortex). For example, a small lab could perform a clustering and integration analysis (e.g., with a pre-existing reference), and then name cell types using this system in such a way that cell sets are linked with the reference. In addition, reference taxonomies can be treated identically to any other taxonomy in this schema, with the caveat that aligned aliases are more likely to be generated through reference taxonomies and through other taxonomies (although this is not required). We have reorganized the structure of our paper to (1) present an example of a single taxonomy run by itself (human MTG) with direct tie-in to the GitHub repo, (2) present the idea of a reference taxonomy along with an example from the primary motor cortex, and then (3) present the same human MTG example but in the context of this M1 reference taxonomy. In principle, this also addresses the issue of immediate utility as all of this is done without a governing body, and is presented as the first such use case. We don’t explicitly address the topic of metadata in the manuscript, but there are likely to be other pieces of meta-data that may be particularly relevant to different taxonomies (e.g., species). These can be provided either as separate meta-data tables or as optional additional columns in the CCN.

5) If I understand correctly, cell set aliases can be based on seemingly very different types of evidence, including (1) quantitative alignment to a reference, (2) user inference based on observation of shared features such as marker genes (i.e., both datasets have Chodl+ cluster, so it is probably the same), and (3) inference based on prior knowledge (an ME cell set that has a location and electrophysiological profile consistent with chandelier cells is aliased to a transcriptomic cell set named "chandelier"). These are really different types of evidence and perhaps should be controlled or distinguished in the system.

This is a good point. To address this, we have added tags for “cell_set_alias_assignee” and “cell_set_alias_citation” to the CCN as a mechanism for providing both credit and evidence for alias terms. While this does not distinguish different types of evidence, it does provide both a person to ask and a place to look for such evidence, and if there are multiple aliases, there can also be more than one assignee or citation. We feel that this is a good starting point at addressing evidence but are certainly open to suggestions for ways of more carefully capturing this point without increasing the complexity of the CCN to a great extent.

Reviewer #3:

This paper is a thoughtful contribution to a tough problem and represents a reasonable step in the right direction. I think it would fit well with eLife and form the basis for beginning of better cross-paper curation of scRNAseq data and other related datasets. It is not the full solution, but is careful in its claims and I think will be an important part of the conversation towards those larger solutions. I have some moderate recommendations for revisions.

1) For cell set accession IDs, they may want to include a '.' between the CS191012 and the unique number for the cell set, and then just iterate the numbers up (.1, .2, .3, … .12 … .10000). The current scheme will max out at 1000 cell sets. That might seem like a huge number now, but someone soon will do 10x on the whole body in one paper and need more than 1000 cell sets.

Reviewer 1 had related feedback and given feedback from both reviewers, we have made the following changes:

– Taxonomy_id = CCN[YYYYMMDD]#

– Cell_set_id = CS[YYYYMMDD]#_#

2) I don't have the bandwidth at the moment to do this as a reviewer, but I would recommend they consider approaching 2-3 labs outside of their group (i.e. external Beta testers) and have them try to enter on of their datasets into this structure using the GitHub code and see how it goes, using only this paper and the associated materials as instructions and iron out any wrinkles or misunderstandings that emerge. If you want this to roll out smoothly, you want researchers' first experiences trying it to be positive to help promote wide adoption.

This is a great suggestion and we are actively pursuing related collaborations. At the moment our collaborations are focused more on other consortia, where we are working with representatives from HubMAP and HCA to at least make sure that this system is compatible with similar systems being designed there. In particular, Peter Kharchenko’s group, who is developing the Cell Annotation Platform, has been providing direct feedback on the GitHub repo by running test cases. This specific suggestion of approaching external groups to Beta test is excellent and will be pursued in the near future, but is beyond the scope of the manuscript.

3) I would recommend adding a section (or perhaps a supplement) that is a clear checklist of what to do as an end researcher who might want to adopt this. If you've convinced me to do this with my data, what are the explicit and actionable recommendations for what I should do? Is this meant to be like submitting your data to GEO? Where any paper publishing a scRNAseq dataset will adopt this standardized approach to naming as the simultaneously upload their taxonomy to a particular database in a standardized format, and put a link in their Materials and methods section? I feel like this is not quite proposing that (as no such database was highlighted, though they highlight the need for one). Or rather is the hope that anyone who generates a scRNAseq dataset will provide their taxonomy in a standard file format as a supplement to their paper? If so, defining a file type (a .txon file or something?) that you are recommending everyone generate and add as supplement might be what you are championing. Explicitly naming that filetype(s) and making that recommendation might help. (If so, is that something that could be rolled into standard analysis packages – e.g. Seuret? That would lower the barrier to adoption) Or is this more like just trying to have everyone agree to use standardized gene names when they mention them? But not necessarily provide supplementary files. Like just being careful how you format your writing and figures like how I should use Pvalb for mouse genes and PVALB for human and PVALB for the protein, and not PV, PVA, etc? Anyway, I just wanted more concrete recommendations of what our expectations as authors (and as reviewers) ought to be for adoption of this standard.

As a starting point, perhaps just recommending a defined file type generated by the code (the .txon file) be included as a supplement is a reasonable recommendation at this time.

This comment is a lot to unpack, but I think is mostly addressed in responses to reviewers 1 and 2, who had similar suggestions for adjustments to the manuscript. In particular, the GitHub repo has been updated with more complete step-by-step instructions for implementation (along with corresponding code), and these instructions have been described in the manuscript in more detail than in the initial submission. As part of this, we have included steps regarding best practices for publication with the goal of allowing future intake into a centralized database directly from manuscript supplementary files. We have also reorganized the paper in a way that should make the relationships with reference taxonomies – both now and after the formation of a governing body--more clear. While we like the idea of having something that could be rolled into standard analysis packages, there is an important manual step related to annotating aligned aliases and (optionally) adding cell sets for cell types in non-hierarchical organization that would make such automation challenging. That being said, we hope that our improved readme with the GitHub repo makes implementation cleaner (and we will test this – see comment #2). Finally, we now propose including a specific file as Supplementary file 1; this is a zip file containing (1) the CCN annotation table for cell sets, (2) cell to cell set mappings, and (3) an updated dendrogram (optional).

Basically, overall I think this paper is making an important and timely contribution. It did a good job of explaining their solution to addressing some of the challenges for annotating these datasets, but stopped just short of a concrete guide on how one could implement it in the near term.

Both other reviewers agree with this assessment, and as a step in this direction we have reorganized the structure of our paper as follows: (1) we present an example of a single taxonomy run by itself (human MTG) with direct tie-in to the GitHub repo, (2) we present the idea of a reference taxonomy along with an example from the primary motor cortex, (3) we present the same human MTG example but in the context of this M1 reference taxonomy as an example of immediate utility, and (4) present four additional use cases that can be done now (e.g., without a governing body or centralized database). We also clarify some aspects related to cross-species, cross-modal, and cross-region nomenclature.

https://doi.org/10.7554/eLife.59928.sa2

Article and author information

Author details

  1. Jeremy A Miller

    Allen Institute, Seattle, United States
    Contribution
    Conceptualization, Resources, Data curation, Software, Formal analysis, Visualization, Methodology, Writing - original draft, Project administration, Writing - review and editing
    For correspondence
    jeremym@alleninstitute.org
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-4549-588X
  2. Nathan W Gouwens

    Allen Institute, Seattle, United States
    Contribution
    Conceptualization, Resources, Data curation, Software, Formal analysis, Visualization, Methodology, Writing - original draft, Writing - review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-8429-4090
  3. Bosiljka Tasic

    Allen Institute, Seattle, United States
    Contribution
    Conceptualization, Resources, Funding acquisition, Methodology, Writing - review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-6861-4506
  4. Forrest Collman

    Allen Institute, Seattle, United States
    Contribution
    Conceptualization, Resources, Data curation, Software, Formal analysis, Validation, Visualization, Methodology, Writing - review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-0280-7022
  5. Cindy TJ van Velthoven

    Allen Institute, Seattle, United States
    Contribution
    Conceptualization, Resources, Data curation, Software, Formal analysis, Validation, Visualization, Methodology, Writing - review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-5120-4546
  6. Trygve E Bakken

    Allen Institute, Seattle, United States
    Contribution
    Conceptualization, Resources, Data curation, Software, Formal analysis, Validation, Visualization, Methodology, Writing - review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-3373-7386
  7. Michael J Hawrylycz

    Allen Institute, Seattle, United States
    Contribution
    Writing - review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-5741-8024
  8. Hongkui Zeng

    Allen Institute, Seattle, United States
    Contribution
    Resources, Funding acquisition, Writing - review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-0326-5878
  9. Ed S Lein

    Allen Institute, Seattle, United States
    Contribution
    Resources, Funding acquisition, Writing - review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-9012-6552
  10. Amy Bernard

    Allen Institute, Seattle, United States
    Contribution
    Conceptualization, Resources, Methodology, Writing - original draft, Project administration, Writing - review and editing
    For correspondence
    amyb@alleninstitute.org
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-2540-1153

Funding

Allen Institute

  • Jeremy A Miller
  • Nathan W Gouwens
  • Bosiljka Tasic
  • Forrest Collman
  • Cindy TJ van Velthoven
  • Trygve E Bakken
  • Michael J Hawrylycz
  • Hongkui Zeng
  • Ed S Lein
  • Amy Bernard

National Institute of Mental Health (U19MH114830)

  • Hongkui Zeng

National Institute of Mental Health (U01MH114812)

  • Ed S Lein

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

The authors would like to acknowledge general input and considerations on aspects of cell nomenclature from attendees and affiliates of the workshop, ‘Defining an Ontological Framework for a Brain Cell Type Taxonomy: Single-Cell -omics and Data-Driven Nomenclature’, held in Seattle, WA, June 2019, including Alex Pollen, Alex Wiltschko, Alexander Diehl, Andrea Beckel-Mitchener, Angela Pisco, Anna Maria Masci, Anna-Kristin Kaufmann, Anton Arkhipov, Aviv Regev, Becky Steck, Bishen Singh, Brad Spiers, Chris Mungall, Christophe Benoist, Cole Trapnell, Dan Geschwind, David Holmes, David Osumi-Sutherland, Davide Risso, Deep Ganguli, Detlev Arendt, Ed Callaway, Eran Mukamel, Evan Macosko, Fenna Krienen, Gerald Quon, Giorgio Ascoli, Gordon Shepherd, Guoping Feng, Hanqing Liu, Jay Shendure, Jens Hjerling-Leffler, Jessica Peterson, Joe Ecker, John Feo, John Marioni, John Ngai, Jonah Cool, Josh Huang, Junhyong Kim, Kelly Street, Kelsey Montgomery, Kara Woo, Lindsay Cowell, Lucy Wang, Luis De La Torre Ubieta, Mark Musen, Maryann Martone, Michele Solis, Ming Zhan, Nicole Vasilevsky, Olga Botvinnik, Olivier Bodenreider, Owen White, Peter Hunter, Peter Kharchenko, Rafael Yuste, Rahul Satija, Richard Scheuermann, Samuel Kerrien, Sean Hill, Sean Mooney, Sten Linnarsson, Tim Jacobs, Tim Tickle, Tom Nowakowski, Uygar Sümbül, Vilas Menon, and Yong Yao. We thank the NIH and CZI for generous co-sponsorship of this workshop. We would also like to acknowledge the many members of the Allen Institute, past and present, who contribute to or support the development of data and analysis of brain cell types - and the organization of this information, particularly Christof Koch, Kimberly Smith, Zizhen Yao, Carol Thompson, Rebecca Hodge, Jonathan Ting, Lucas Graybuck, Thuc Nguyen, Jim Berg, Staci Sorensen, Nik Jorstad, Susan Sunkin, Stefan Mihalas, Rob Young, Tim Fliss, Lydia Ng, Shoaib Mufti, and Stephanie Mok. Research and methods reported in this publication were supported by the Allen Institute, and the National Institute of Mental Health of the National Institutes of Health under award numbers U19MH114830 (to HZ) and U01MH114812 (to ESL). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. The authors would like to thank the Allen Institute founder, Paul G Allen, for his vision, encouragement and support.

Senior Editor

  1. Catherine Dulac, Harvard University, United States

Reviewing Editor

  1. Genevieve Konopka, University of Texas Southwestern Medical Center, United States

Reviewers

  1. Genevieve Konopka, University of Texas Southwestern Medical Center, United States
  2. Fenna Krienen, Harvard Medical School, United States
  3. Joseph D Dougherty, Washington University in St. Louishington University, United States

Publication history

  1. Received: June 12, 2020
  2. Accepted: December 28, 2020
  3. Accepted Manuscript published: December 29, 2020 (version 1)
  4. Version of Record published: January 7, 2021 (version 2)

Copyright

© 2020, Miller et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 2,669
    Page views
  • 355
    Downloads
  • 3
    Citations

Article citation count generated by polling the highest count across the following sources: PubMed Central, Crossref, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Download citations (links to download the citations from this article in formats compatible with various reference manager tools)

Open citations (links to open the citations from this article in various online reference manager services)

  1. Further reading

Further reading

    1. Cell Biology
    2. Neuroscience
    Francois Singh et al.
    Research Article

    Parkinson’s disease (PD) is a major and progressive neurodegenerative disorder, yet the biological mechanisms involved in its aetiology are poorly understood. Evidence links this disorder with mitochondrial dysfunction and/or impaired lysosomal degradation – key features of the autophagy of mitochondria, known as mitophagy. Here, we investigated the role of LRRK2, a protein kinase frequently mutated in PD, in this process in vivo. Using mitophagy and autophagy reporter mice, bearing either knockout of LRRK2 or expressing the pathogenic kinase-activating G2019S LRRK2 mutation, we found that basal mitophagy was specifically altered in clinically relevant cells and tissues. Our data show that basal mitophagy inversely correlates with LRRK2 kinase activity in vivo. In support of this, use of distinct LRRK2 kinase inhibitors in cells increased basal mitophagy, and a CNS penetrant LRRK2 kinase inhibitor, GSK3357679A, rescued the mitophagy defects observed in LRRK2 G2019S mice. This study provides the first in vivo evidence that pathogenic LRRK2 directly impairs basal mitophagy, a process with strong links to idiopathic Parkinson’s disease, and demonstrates that pharmacological inhibition of LRRK2 is a rational mitophagy-rescue approach and potential PD therapy.

    1. Neuroscience
    Tima Zeng et al.
    Research Article Updated

    Our experiences in the world support memories not only of specific episodes but also of the generalities (the ‘gist’) across related experiences. It remains unclear how these two types of memories evolve and influence one another over time. In two experiments, 173 human participants encoded spatial locations from a distribution and reported both item memory (specific locations) and gist memory (center for the locations) across 1–2 months. Experiment 1 demonstrated that after 1 month, gist memory was preserved relative to item memory, despite a persistent positive correlation between them. Critically, item memories were biased toward the gist over time. Experiment 2 showed that a spatial outlier item changed this relationship and that the extraction of gist is sensitive to the regularities of items. Our results suggest that the gist starts to guide item memories over longer durations as their relative strengths change.