Cell Type Nomenclature - v1

Overview

Naming Cell Types

Example:  Human brain - middle temporal gyrus & cell type aliases

The upper right of this diagram shows an annotated dendrogram, or hierarchical organization of cell types defined for a specific taxonomy, in human middle temporal gyrus, as described in this recent study in Nature¹.  Below the dendrogram, the names of cell types that presented in this publication are given.  On the left is an inset with the provisional cell type alias, label, and accession IDs for a subset of cell types, as they would be defined in the current nomenclature schema.  Although not shown, Inh L1-4 LAMP5 LCP2 corresponds to Rosehip cells (described in this recent study in Nature Neuroscience²) and therefore a cell type alternative alias for this type would be "Rosehip" as other nodes highlighted in blue might also have colloquial aliases.

Naming Cell Sets

Example:  Human brain - middle temporal gyrus & cell sets

This diagram shows example descriptors for three cell sets, shown on the same dendrogram presented earlier (human brain -middle temporal gyrus & cell type aliases), with example nodes in blue.  In this case, only one of the three nodes has a cell type alias, which refers to expression of specific canonical genes.  In all three cases, the cell set label indicates the cell types included within that node in the tree.

Usage:  Application of nomenclature conventions to novel or published data sets

Any nomenclature schema is only useful if it is adopted by many scientists studying cell types across multiple institutes and organ systems.  Therefore, we have aimed to make this schema easy to understand, lightweight, and easy to apply to novel or published data sets, and have provided both tools and examples for this nomenclature can be applied to any cell-level data.  Code for applying this nomenclature is available in the `allen_institute_nomenclature` github repository.  An example of how this schema can be applied to an existing data set is shown below. 

In a study of mouse primary visual cortex using ~1700 cells⁴, a total of 42 neuronal and 7 non-neuronal cell types were proposed, and cell type-specific mRNA processing and genetic access to these transcriptomic types was accomplished through the use of many transgenic Cre lines.   This work was published in Nature in 2016⁴, and the data can be browsed with an interactive navigation application.  Above, we have taken the hierarchical organization of cell types and cell type aliases as described and have applied the nomenclature schema to append cell type labels and accession IDs for each cell type (right column).  We have also annotated a subset of the internal nodes with a cell set alias, label, and accession ID, as described.  Note that by default, node aliases are left blank (e.g. nodes C and F), but nodes representing a useful collection of cell types can be manually tagged with an alias (e.g., node "B" contains all Pvalb types and no others).  Details and code for reproducing this analysis are in the `allen_institute_nomenclature` github repository. 

Nomenclature and Reference Data Sets

Defining and refining reference cell types

A general workflow is presented below, to assist with defining and refining reference cell types in a generalizable way. This workflow accomodates for methodological differences in cell type definitions that will likely vary between labs and change as new methods are developed.  Note that this workflow makes some major assumptions about cell type ontology, data visualization, and governance that will be discussed in the next section.  

This workflow can be broken down into four broad stages:

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.  The example workflow reflects the current state of the field.  We propose a formalization to the nomenclature, which can (and should!) be applied to each data set independently.

2. Second, an initial reference cell type classification will be defined by taking the results from one or more (ideally validated) data sets and integrating these data together in a single analysis.  Transcriptomics is well suited to allow quantitative comparisons across species, across developmental time, and between brain and other organs, and our goal is for reference cell types to be defined using this modality. For example, scAlign and Seurat have been used to align data from human and mouse cortex into a single consensus cell type classification (see below).  Once reference cell types are defined, features discriminating cell types should be calculated separately in each data set so that canonical discriminating features can be separated from ones represented in only a subset of data sets.  These features, along with additional meta-data available for these cell types from any of the integrated data sets will be used to match cell types to existing ontologies where possible, and to update ontology terms as needed.  Finally, reference cell types will be named as described in this nomenclature convention, with some additional constraints discussed below.  Note: while we present a reference classification here as an integration of multiple data sets, it is perfectly valid to have a single data set stand alone as a reference data set.

3. Third, this reference cell type classification will now be used as a comparator for any related data sets, providing a mechanism for transferring prior knowledge about cell types across data sets.  Existing data can be renamed by mapping cell sets onto the reference classification and then updating the cell set primary alias to match the reference.  Similarly, new data sets can either be clustered independently and mapped to the reference or directly integrated with the reference.  In either case the resulting nomenclature will be informed by the reference data set.

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

Reference cell type nomenclature

In summary, we propose a formalization to cell type nomenclature that can be applied to (ideally) all cell type classifications.  The term "provisional cell type" is used to represent quantitatively-derived data clusters defined by running a specific computational algorithm on a specific data set (here called an "taxonomy").  All cell sets (including provisional cell types) are assigned multiple identifier tags, which are used for different purposes.  Cell set accession IDs can track unique cell sets across the entire universe of taxonomies.  Cell set aliases can provide both data-driven and common-usage cell type names for public consumption (e.g., in manuscripts).  Cell set labels provide general tags that allow easy tracking of groups of cell types within a single taxonomy that could span many organ systems.  These terms apply to reference cell type classifications as well.  In addition, we define a reference cell type classification as an taxonomy applied to the results of a reference cell type analysis, and expand the scope of a primary alias.

In a reference cell type classification, the primary alias is designed to allow tracking of the same cell type across multiple taxonomies defined using data from multiple modalities.  For this reason, it should match (directly map to) cell types defined in the relevant ontology (i.e., common usage types in the Neuron Phenotype Ontology).  Either this, or other cell set aliases should include prior knowledge (provenance), canonical discriminating genes, and/or information from other modalities (such as electrophysiological properties, if available) to provide the best data to match cell types in the reference with cell types defined previously, or in future taxonomies from any modality. 

Example:  Human & mouse cortex cell type nomenclature

Above we present cell types defined in human middle temporal gyrus using snRNA-Seq (A; Hodge, et al.¹) and in mouse primary visual cortex (VISp) and ALM using scRNA-Seq (B; Tasic, et al. ³).  These studies identified ~75-100 types per cortical area per species from ~15,000 cells, which each could be defined using one (or a combination of) robust marker genes. In addition, we characterized these cell types by associated relevant meta-data from the assigned cells, including things like cortical layer of dissection, brain region, alignment statistics, and (in mouse) projection targets of a subset of cells.  We then performed data integration on these two data sets using scAlign and including only cells from V1 in mouse, and arrived at ~40 reference cell types (C).  The resulting data set included some one-to-one matches (e.g., a single mouse cell type matches a single human data set), with many of the remaining reference types matching to internal nodes of the tree (e.g., cell sets that are not defined as putative cell types).  From this data,  primary aliases were assigned using a combination of (i) robust gene markers from the literature, (ii) highly discriminating gene markers in these data, (iii) projection targets in mouse, (iv) historical names based on cell shape, and (v) broad cell class names (that directly map to ontologies).  In principle, this homologous cell type classification could be used as a cross-species reference cell type classification; however, now that data on additional cortical areas is available using multiple RNA sequencing platforms, it seems prudent to await data integration that will provide a more comprehensive reference.