Ghiselin (1974) noted that “The species problem has to do with biology, but it is fundamentally a philosophical problem.” According to Shanker et al. (2017), “The earliest documented effort at a systematic classification of natural objects in ‘Western science’ is Aristotle’s principle of logical division, where every object (living or nonliving) was classified through a series of binary steps.” To Aristotle, objects can be classified because they have “essences”—properties that make them what they are, and those properties must therefore be shared by every member of the group to which they are assigned. Ghiselin (1974) continued, “… someone trained in logic should, one might think, long ago have stepped in and cleared up the confusion. Such is demonstrably not the case.” He proposed to do so by a “radical solution to the species problem”: that species should be considered individuals. Individuals do not have an essence: An individual’s parts (“members”) do not possess the same characteristics. A liver is not a brain; a leaf is not a root. So, too, the members of a species are not identical, and all need not possess the attributes that are typical of the species. Both Ghiselin (1974) and another philosopher of science, Hull (1976), contrasted species with chemical elements—Hull (1976) stated that, unlike species, “slots in the periodic table remain forever open” because “Any atom which arises with the appropriate atomic number counts as an instance of that element regardless of how, where, or when it arose.” Since the time of Darwin, however, species have been understood to be evolving lineages, and this is inconsistent with essentialism (Hull, 1976; de Queiroz, 1998; Hey, 2001). Hamilton (2012) argued that this “individuality thesis” was already central to the species concept of Hennig (1966), the founder of phylogenetic systematics (“cladistics”).

The idea of species as individuals is now widely accepted among systematists, and is consistent with most species concepts, though not with the Phenetic Species Concept, which recognizes species by overall similarity at many traits and is considered essentialist in that all organisms sharing identical characteristics would be grouped together into a phenetic species, even if they should originate convergently on a different planet (Ghiselin, 1974). There is, however, a view that species are metaphysically neither individuals nor groups, but have some elements of both, and that they are best treated as Homeostatic Property Clusters (HPCs); HPCs are marked by a set of characteristics, all of which need not be shared by all members, whose statistical correlation is due to an underlying homeostatic mechanism (Shanker et al., 2017; Casetta and Vecchi, 2021). Casetta and Vecchi (2021) pointed out that even Hull, in his 1976 “species as individuals” paper, wondered if the distinction between individuals and kinds was too crude; they noted that the cluster character criterion of HPCs avoids the problem that all members of a group defined by its essential characters must possess all of those characters, and pointed to genetic coherence as the homeostatic mechanism underlying the correlation of clustered characters. Thus, the philosophical debate continues, and yet another alternative was suggested by Shanker et al. (2017), who recommended a fuzzy set theory approach to defining species, in which different populations have varying probabilities of belonging to one or more groups.

Authors in the 1970s and 1980s referenced philosophical parallels between species and cell types, notably the problems with essentialist approaches (Rowe and Stone, 1977; Rodieck and Brening, 1983), but did not cite what became cornerstone literature in systematics concerning species as individuals (Ghiselin, 1974; Hull, 1976). Recently, Xia and Yanai (2019) and Moroz (2021) independently analogized cell types with chemical elements in the periodic table—the primary example of essentialist categories cited by Ghiselin (1974) and Hull (1976). Slater (2013), in discussing the philosophy underlying the definition of cell types, argued that essentialist definitions of cell type fail to meet three key criteria: that the defining properties be intrinsic; that all members and only members of the type possess those properties; and that the “essence” explain why all members also possess additional qualities in common. However, he also argued that cells comprising a cell type lack the spatiotemporal relationship to one another that is the major justification for viewing species as individuals (Slater, 2013). Instead, Slater (2013) argued that cell types are described best by a variant of the HPC concept, because they have metaphysical features of both individuals and kinds. In another parallel with species, Battaglia et al. (2013) suggested, in a paper that does not discuss philosophical issues, that cell types are treated best as fuzzy sets, in which individual cells have a probability of belonging to any of several well-defined archetypes. This seems consistent with the observation that “at the most fundamental level, single-cell dynamics is probabilistic” (Teschendorff and Feinberg, 2021); fuzzy clustering is also mentioned by Yuste et al. (2020).

Given the apparently innate human desire—perhaps “compulsion” would be a better word—to classify and to name, it is not surprising that there should be parallels between systematics and cell biology. This is particularly true because the philosophical options have generally been portrayed as binary, with essentialism losing out. But consider the following quote from the journal Biology and Philosophy (Williams, 2018):

If a philosophical consensus that species and cell types are HPCs develops, might this influence theory and practice in both systematics and cell biology?

The diversity of living organisms presents a pattern or organization that we understand to have been formed by evolutionary processes, for which systematists are interested in developing theories; systematists also seek useful criteria for classifying the products of those processes. Unfortunately, as Hull (1997) wrote in a paper titled “The ideal species concept—and why we can’t get it,” “Applicability and theoretical significance tend to be in opposition to each other. The more theoretically significant a concept is, the more difficult it is to apply.” A solution to this problem decoupled theory from practice, adopting the stance that although there are many competing definitions of species, and thus many criteria for recognizing them, there might be a primary underlying species concept (de Queiroz, 1998, 2007). What, exactly, that concept is remains debatable, but there is widespread agreement that it is based on genetic lineage, such as some version of Simpson’s (1951) Evolutionary Species Concept (Table 1; Mayden, 1997; Padial et al., 2010; Freudenstein et al., 2016). Speciation is a continuous process, and does not occur in discrete steps that are uniform across different taxonomic groups (Stankowski and Ravinet, 2021). de Queiroz (1998, 2007) contended that much of the confusion surrounding species stems from the fact that criteria such as reproductive behavior, ecology, or the fixation of morphological or molecular characters of populations (which are lineages at a particular point in time) are met in different species at different times and in different sequences, creating a “gray zone” between what all observers would agree is one species and what all would agree is two (Figure 2). Much of the species debate, therefore, involves arguments over the primacy of different criteria for recognizing species, rather than on the fundamental nature of what a species is. According to this view, from a philosophical standpoint species concepts are generally monistic, whereas criteria for recognizing them are pluralistic.

In developing a definition of neuron types, Zeng and Sanes (2017) briefly explored what they considered to be the “three general schemes for defining species” in systematics: the biological species, based on reproductive isolation; phylogenetic relationships of lineages; and “a third school of systematics, known as typological, taxonomic or phenetic systematics, which groups individuals into species according to their similarity of genotype and/or phenotype.” They dismissed the first as inapplicable, and after noting several problems with a lineage-based approach for cells, adopted a transcriptomic similarity approach as being most useful for their primary purpose, classifying cell types for cell atlases (e.g., Callaway et al., 2021). Accordingly, they defined a neuronal cell type as “a population of neurons with properties that are homogeneous within the population but differ from those of other neurons.” This operational definition is both provisional and explicitly pluralistic (Zeng and Sanes, 2017); for example, Yao et al. (2021) referred to “transcriptomic cell types” as only one of various ways that cells could be classified.

A different thread in the neuronal cell type literature is that of Arendt (2008) and Arendt et al. (2016, 2019). These authors are particularly interested in the process by which cell types originate and evolve across species, and they defined cell type as “a set of cells in an organism that change in evolution together, partially independent of other cells, and are evolutionarily more closely related to each other than to other cells” (Arendt et al., 2016). Zeng and Sanes (2017) cited this definition in their brief discussion of phylogenetic, lineage-based approaches, and rejected it as being impractical, because the data needed for employing it are generally unavailable. Instead, they argued, “It may be more realistic to find ways to classify types within a species and then use that classification to launch an evolutionary inquiry.” In other words, although they are certainly interested in evolution, their approach prioritized pattern over evolutionary process. For their part, Arendt et al. (2019) rejected the Zeng and Sanes (2017) definition even as an operational one, because they believed that although such approaches “can provide useful classifications for neuron types within one organism, they are problematic for comparing across species. In particular, phenotypic definitions fail to distinguish two key types of evolutionary changes: the phenotypic alteration of the same cell type existing in two compared species and the origination of entirely new cell types”—precisely the process-oriented questions of greatest interest to these authors. Instead, they proposed their own operational definition, discussed below under the heading of “Phenetics and cladistics.”

The tension between interest in pattern vs. process, emphasizing diagnosis and theory, respectively, exists for both species and cell types. Might wider recognition of this tension by cell biologists bring these two approaches into harmony as has been done at least to some degree in systematics? As with species, different attributes of cells can be distinguished—morphology, physiology, gene expression—and it has long been known that these can be discordant (Tyner, 1975; Vickaryous and Hall, 2006). Other species concepts and criteria beyond the few cited by Zeng and Sanes (2017; Table 1) may be useful to consider. Could the insight that the order in which different attributes arise varies in the evolution of different species be applied to the order in which transcriptomes and other phenotypes appear during the development of different cell types (Figure 2)? Is there perhaps a single underlying theoretical basis for recognizing cell type, comparable to the role genetic lineage plays for many species concepts?

Lineages of genes, individuals, populations, and species all are important in systematics and evolutionary biology, and have complex relationships to one another (Figure 3). Despite, or perhaps because of, the central role organismal lineage plays in the species debate, the precise definition varies (Freudenstein et al., 2016); a common one is that of Simpson (1961): an ancestor-descendant series. If, as is thought, all living organisms have a single evolutionary origin, then all lineages trace back to this ancestor and all individuals are members of a single clade (an ancestor and all of its descendants) and thus are related to all other individuals to varying degrees. How should individuals be grouped meaningfully? The overall structure (topology) of this comprehensive clade of organisms is visualized differently by systematists working on different groups of organisms. Those who study multicellular eukaryotes generally refer to it as the “Tree of Life,” despite the common occurrence of hybridization and introgression in many groups, notably plants (Mallet et al., 2016), which create reticulate, non-treelike patterns (networks; Figure 3C). Paleontologists add a temporal dimension to the problem by including fossils and explicitly considering extinction (Marshall, 2017). Systematists who study unicellular organisms are necessarily conscious of reticulate relationships, to the extent that if there is a tree at all (O’Malley et al., 2010), the overall picture is that of a “cobweb of life” in which many limbs are connected by extensive horizontal transfer (Ge et al., 2005). The genealogical relationships of individuals in sexually reproducing species are also fundamentally reticulate—tokogenetic as opposed to phylogenetic (Figure 3A); such species comprise one or more lineages and represent spatiotemporally limited segments of an overall metapopulation consisting of geographically separate but genetically connected Mendelian populations. Species of sexually reproducing organisms reside at the boundary between tokogeny and phylogeny (Figure 3A).

Evolution occurs by a combination of mutation, recombination, natural selection, and genetic drift; drift is a particularly powerful force in multicellular eukaryotes, which typically have small effective population sizes, and provides a neutral explanation for many phenotypes, including cellular ones (Lynch, 2007, 2018, 2020; Lynch and Trickovic, 2020). Divergence of lineages occurs in response to both abiotic and biotic forces that restrict or promote genetic exchange. Evolution operates at the level of individuals within populations, and the results are seen in the structure of individual genomes. Gene lineages are embedded within organismal histories—the species tree shapes the gene trees of its individual members, and gene lineage phylogenies can be discordant with the organismal phylogeny (Figure 3B; Maddison, 1997; Degnan and Rosenberg, 2006). With assumptions of neutral evolution and minimal gene flow between the units of evolution (species, generally as defined in the Evolutionary Species Concept), the multispecies coalescent (MSC; Rosenberg and Nordborg, 2002; Kubatko, 2019) has been widely adopted from population genetics as a unifying statistical model for studying the pattern and process of species divergence, revolutionizing the way its adherents use molecular data to reconstruct phylogenies (Bravo et al., 2019). Moreover, as Kubatko (2019) has pointed out, “Within the last decade, methods for species delimitation have increasingly cast the problem in the framework of the multispecies coalescent.” But this has sparked debate about whether what is recognized by MSC-based delimitation methods are “real” species as opposed to simply “lineages” that comprise portions of species (Jackson et al., 2017; Sukumaran and Knowles, 2017; Leaché et al., 2018; Chambers and Hillis, 2019; Sukumaran et al., 2021). Debate about the relationships among lineages, populations, and species, involves models of speciation … which in turn requires a definition of “species.”

As in species biology, “lineage” has more than one meaning in describing relationships of cells to one another (Figure 4). Just as individual organisms trace their ultimate origin to the common ancestor of all life, cells of an individual multicellular organism belong to a common historical lineage, beginning with the single cell of the zygote. Cell fusion occurs (outside of fertilization) in some animal organs (e.g., skeletal muscle) but in plants does not contribute to the generation of mature cell types (Brukman et al., 2019). Thus, for the most part, the cell lineage of a multicellular eukaryote, even more than is true of organismal lineages, can be described as a tree, and such cell lineage trees can now be reconstructed with increasing precision (Wagner and Klein, 2020). Slater (2013) noted that “Unlike species, cells do not fit into a single phylogenetic tree. Rather, development in each organism defines its own local tree,” but although this may cause philosophical concerns, the problem is minimized if, as here, the focus is on the cells of an individual in a single species. Rodieck and Brening (1983) suggested that “It might appear that the lineage of ganglion-cell types, expressed during development, would provide a causal agent as useful as the role genealogy plays in the classification of species.” However, they noted that despite the strong superficial resemblance between lineage trees of cells and organismal phylogenies, there are fundamental differences (Rodieck and Brening, 1983). With species, a lineage (however, defined) can be related ultimately to mutational variation at the genomic level. In contrast, all cells of an individual organism possess fundamentally the same genome (ignoring somatic mutation and phenomena such as endopolyploidy), and the process of differentiation that is analogous to speciation occurs through epigenetic mechanisms (Figure 1). The immediate ancestor of a species cannot produce a species that belongs, genetically, to a different clade, but a multipotent stem cell is theoretically capable of producing any mature cell type. The connection between phenotype (the mature cell’s transcriptome, morphology, etc.) and epigenotype for cells is thus looser than a whole organism’s phenotype to genotype relationship. Rodieck and Brening (1983) concluded that “if there proves to be little or no correspondence between the lineage map and its phenotypic expression, then it is unlikely that the lineage map will become a useful and widely accepted ordering of ganglion-cell types.”

Currently, the transcriptome is the most studied “phenotypic expression” of cells and variation in the full transcriptome or in the expression of specific sets of genes is the major source of data for defining cell states and types (Almeida et al., 2021; Sacher et al., 2021); as such, it has much the same relationship to cell types as the genome does to species (Figure 1). Gene expression changes as cells differentiate, and this can be visualized as a landscape (“manifold”) in pseudotime (Figures 4A–F). Depending on the method used to create the manifold from high dimensional transcriptome data, the manifold can be an acyclic graph (tree) or cyclic graph (network, Figures 4A–F; Wagner and Klein, 2020). Mani and Tlusty (2021) simulated millions of developmental programs in model-generated “organisms” and characterized their graph topologies; surprisingly, tree-like topologies were rare, and in most cases multiple cell lineages converged on the same terminal cell type. Wagner and Klein (2020) portrayed mitotic lineages as included within expression manifolds (Figures 4C,F). Alternatively, expression fate can be shown as character changes on mitotic lineage trees; discordance between expression state and mitotic lineage then appears as convergent change or state reversals, like any character showing homoplasy on a phylogenetic tree (Figures 4A,D). The potential for discordance between phenotype (expression manifold) and cell mitotic lineage is what led both Zeng and Sanes (2017) and Arendt et al. (2016, 2019) to reject using cell lineage for defining mature cell types. For example, in blood development (hematopoiesis), different cell lineages can produce functionally similar cells (Yáñez et al., 2022). Nevertheless, transcriptomic state cannot be assumed to be more important than lineage; in hematopoiesis, Weinreb et al. (2020) found that “sister cells tended to be far more similar in their fate choice than pairs of cells with similar transcriptomes” and concluded that chromatin structure might provide information that neither lineage nor transcriptome reveal.

The concept of homologous cell type lineages (Vickaryous and Hall, 2006), as distinct from mitotic lineages or developmental lineages, has been explored and developed by Arendt et al. (2016, 2019). It transcends the individual or species and considers the evolution of cell types over phylogenetic timescales rather than individual lifetimes (Figure 4G). Cell types in this sense resemble phenotypic character trees for cladistic analysis (Pogue and Mickevich, 1990; Musser and Wagner, 2015) in being embedded in organismal phylogenies (“species trees” in population genetic terminology; e.g., Nei, 1987), but perhaps have their closest analog in gene family trees, which have complex homology and functional relationships due to duplication (e.g., Glover et al., 2019). The potential for homologous cell type lineage phylogenies to provide a unifying, objective criterion for defining cell type (Almeida et al., 2021) is discussed in more detail below.

The biology of both species and cells involves more than one kind of lineage, whose relationships to one another are complex, may be nested, and can be incongruent (Osumi-Sutherland et al., 2021). Homoplasy, non-homology generated by parallelism and convergence, reveals gaps in our knowledge (Nixon and Carpenter, 2012) and thus is of fundamental interest in understanding the evolutionary process (e.g., Wake et al., 2011). Similarly, exploration of the complex connections between mitotic cell lineage and cell state, often involving state convergence, is an exciting area in cell biology (Battaglia et al., 2013; Konstantinides et al., 2018; Wagner and Klein, 2020; Osumi-Sutherland et al., 2021; Yáñez et al., 2022).

Much has been accomplished in systematics by employing models of the coalescent process to infer species relationships from the gene lineages embedded in them (Figure 3B). In the coalescent approach, incongruence between two types of lineages—species and gene—is a key source of data, rather than a problem. It also can be invoked to account for phenotypic homoplasy (hemiplasy; Guerrero and Hahn, 2018). Can the cell-level models that are being developed (Weinreb et al., 2020; Teschendorff and Feinberg, 2021) in an analogous way harness the discordance between the mitotic and transcriptional lineages of cells to define cell types?

Accommodating variation at different organizational levels, from genes to genomes to populations, is a challenge for species definitions both in theory and in practice. Mutation, which includes not only base substitution but also insertions, deletions, transpositions, chromosome structural changes (inversions, translocations), and recombination, ensures that individuals are rarely genetically identical. An appreciation of copy number variation in multicellular eukaryotes has led to the adoption of the pangenome concept from bacterial genomics; it is now recognized that no single individual plant or animal genome provides a complete picture of gene content in its species (e.g., Gao et al., 2019; Miga and Wang, 2021).

The pangenome concept captures one aspect of genetic variation among related individuals, which is often structured at the level of populations. The potential for “over-splitting,” particularly when the distinguishing characters are microscopic or otherwise cryptic, is not a new concern, nor is it confined to any particular species concept or recognition criterion—there has always been debate between lumpers and splitters. Taxonomic ranks, both formal (subspecies) and informal (variety), have been used to designate groupings that do not rise to the level of differentiation considered to merit species status; even many systematists who hold that species are “real” natural entities consider these categories to be artificial constructs, as is also true of genera, families, and higher taxonomic ranks.

Variation also occurs over the course of development, and individuals can appear very different at different stages of their lives. An acorn does not look like the oak tree that produced it; larval and pupal stages do not look like the butterfly they will become. Yet in both cases the individual at each different stage represents a single species, and for the purposes of defining that species and reconstructing its phylogenetic relationships individuals at any life stage are “character-bearers”—what the founder of cladistics, Willi Hennig, termed “semaphorants” (Figure 3A; Havstad et al., 2015). A complete description of a species includes all of its semaphorants; species can be compared, and their phylogenies reconstructed, from any homologous characters gleaned from comparable semaphorants in different species. The importance of semaphorants diminished with the reliance on molecular data for phylogeny reconstruction, for which individuals are typically the units of gene or genome sampling (Freudenstein et al., 2016). This certainly is true in the concatenation paradigm, where the sequences of multiple genes sampled from an individual form a single row in the data matrix that is then aligned with the aggregated sequences from each other individual. However, methods based on the MSC, though they sample individuals, do so as representatives of a species (Bravo et al., 2019), so each individual is again potentially a semaphorant, bearing a subset of the characters and character states found in the species.

Like conspecific individuals, cells of the same type are not identical (Cembrowski and Menon, 2018; Usaj et al., 2021). At the transcriptomic level, a source of cell-to-cell variation is dropout “due to low amounts of mRNA in individual cells and inefficient mRNA capture, as well as the stochasticity of mRNA expression” (Qiu, 2020). From a practical perspective, dropout is similar to the longstanding problem of missing data in phylogenetic data matrices (Xi et al., 2015). Biologically, transcriptional bursting (Tunnacliffe and Chubb, 2020) means that the full picture of the transcriptome of a cell type cannot be obtained from any single cell, much as the pangenome of a species cannot be inferred from a single genotype. Consequently, in practice, much of “single cell” biology involves clusters of cells with similar transcriptomes or other -omic phenotypes (e.g., Coate et al., 2020); thus, Arendt et al. (2019) recommended using small groups of cells with very similar expression (“metacells”), rather than individual cells, to reconstruct cell type trees. Because technical causes of dropout are expected to be roughly the same for all cell types in a given experiment on a per cell basis, any differences in the pattern of dropout among populations of cells should be due to biological causes. Qiu (2020) found that “dropout pattern in scRNA-seq data is as informative as the quantitative expression of highly variable genes” and suggested “embracing” dropout by using binarized data rather than transcript counts instead of ignoring it. Bouland et al. (2021) found that a similar approach did indeed capture biologically meaningful variation in single cell transcriptomes. Nevertheless (Jiang et al., 2022), criticized methods that use binarized gene expression data, because they ignore information from differential expression of the same genes in different cell types.

Regardless of the lineage to which a particular cell belongs, its expression changes as it makes the transition from stem cell to its mature cell state. Single cell or single nucleus RNA-seq experiments produce a “snapshot” that includes mature cells, the stem cells destined to give rise to them, and cells in transitional states. The picture is tremendously rich in detail (Teschendorff and Feinberg, 2021), and includes information that can be used to identify the position of cells in “pseudotime” along a differentiating cell lineage (Campbell and Yau, 2018), and to predict the future states of cells using information on spliced vs. unspliced mRNA molecules (RNA velocity; La Manno et al., 2018). Nevertheless, because it is a snapshot involving multiple cells, Weinreb and Klein (2020) noted that single cell transcriptomic approaches alone cannot determine, for any particular cell, when “progenitor cells become committed to one or more fates or how many distinct paths might lead cells to the same end states.” The question of fate determination is relevant not only for defining cell type but also for such issues as determining whether plants have a committed germline (Lanfear, 2018; Burian, 2021). Sagar and Grün (2020) concluded that emerging data are “challenging the classical view of cell fate commitment as a discrete binary decision process where immature multipotent progenitors become lineage restricted in a stepwise fashion”; instead, differentiation may occur probabilistically “in a continuous transcriptional and chromatin landscape.” Other authors also have emphasized the continuous nature of cellular commitment (Xia and Yanai, 2019; Quake, 2021).

Tasic (2018) noted that whereas the cellular equivalents of phyla, such as the “cardinal classes” of neurons described by Fishell and Kepecs (2020), might be readily identifiable, the same is not true of lower hierarchical levels that correspond to genera or species. This is due to continuous variation, even within some narrowly defined cell types (Cembrowski and Menon, 2018; Efroni, 2018; Quake, 2021; Shojaee et al., 2021), and to the associated problem of distinguishing cell type from cell state. A single cell type can differ in essential characteristics over the course of its life—what is considered the same cell type can have different states due to development, environment, treatment, or location. Some of the authors of the Clevers et al. (2017) poll of definitions of cell type appear to subscribe to the nominalist position that “cell type” is an arbitrary designation, and that only cell states exist. This echoes the position of some systematists that lineages are real, whereas species are arbitrary (Vaux et al., 2016; Mishler, 2021). But most cell biologists see a real distinction between cell types, which are “hard-wired,” and cell states, which are “soft-wired” (Arendt et al., 2016, 2019; Morris, 2019). However, the distinction between hard- and soft-wired can vary over the course of a cell’s developmental trajectory (Fishell and Kepecs, 2020). Cembrowski and Menon (2018) and Tasic (2018) have suggested that cell states are reversible, whereas cell types are not, at least under standard conditions.

In the analogy of cell types with elements in the periodic table, cell states are like isotopes (Xia and Yanai, 2019; Moroz, 2021). Alternatively, extrapolating from Tasic’s (2018) comparison with taxonomic categories, if cell types are analogous to species, then cell states might be analogous to subspecies or varieties. A different approach adopted from systematics would consider cell states as semaphorants—different manifestations of the same biological entity, united by some core features but bearing a unique set of characters depending on their stage of development and physiology. But this approach requires that the entity to which semaphorants belong—species or cell types—first be defined.

A number of cell biologists have recognized the parallels between the issue of how broadly a cell type or a species should be defined, referring explicitly to “lumpers and splitters” (Rodieck and Brening, 1983; Armañanzas and Ascoli, 2015; Tasic, 2018; Yuste et al., 2020). Decisions about how many species to recognize are often guided by the kinds of variation available to the taxonomist, particularly whether the characters that distinguish taxa are easily discernible. For example, the Eastern spring beauty (Claytonia virginica) comprises several well-differentiated chemical and chromosomal lineages with distinctive geographical ranges, but morphological variation among plants from these groups is cryptic (Doyle, 1984), and therefore the groups have not been named formally. This type of subjective practical decision is specific to each organismal group, presumably because each group has a different sequence of character evolution (Figure 2; de Queiroz, 1998, 2007), and is captured at a particular time point in its evolutionary trajectory—systematists cannot see its future, and in all but rare cases lack a detailed enough fossil record to know its past.

In contrast, the snapshot available to cell biologists provides a much more comprehensive sample of differentiation, from stem cells to mature cells, for an individual at the developmental stage at which it is studied. Moreover, this process is expected to be similar across all species and at all stages in the lives of individuals. Of course, because an individual sampled at a specific stage of its life is itself a semaphorant for its species, a complete picture of cell types even of a single species may not be obtainable without greater sampling. But the questions troubling cell biologists about state vs. type seem more tractable than those involving species because there is more hope that a shared set of fundamental rules exist to be discovered. Will such rules reinforce the legitimacy of “cell type” as a theoretical concept as well as a practical category? Regardless of that answer, what characters could be used to define cell types?

The late plant taxonomist Arthur Cronquist is reputed to have said that “a good species is what a good taxonomist says it is”; Mayden (1997) calls this the Morphological Species Concept (Table 1). By a “good” taxonomist, Cronquist meant one with a keen eye, familiar enough with the taxa in question to be able to discern the key characters by which meaningful groupings of individuals could be discriminated from one another, filtering out polymorphisms, plasticity, and other uninformative variation. This subjective approach to taxonomy was challenged in the 1960s and 1970s by “numerical” taxonomists, who instead embraced variation, using large numbers of characters with minimal a priori filtering in “phenetic” analyses that identified clusters based on overall similarity rather than on criteria from evolutionary theory. Phenetic groupings were dependent on the algorithms used, and recognition of species under the Phenetic Species Concept was strictly operational and based on arbitrary similarity cutoffs (Mayden, 1997); the trees (dendrograms) produced were meant to portray similarity rather than genealogy or evolution. Genome clustering, also an operational approach (Hull, 1997), continues to be used in bacteria, with delimitation of species being proposed for genomes with Average Nucleotide Identity (ANI) scores >95% (Konstantinidis et al., 2006; Jain et al., 2018) on the theory that in nature there is a discontinuity in genetic variation at this level; the approach is controversial (Murray et al., 2021; Rodriguez et al., 2021).

The cladistic approach largely supplanted phenetics in eukaryotic systematics by the 1980s, after considerable and often vitriolic debate. The original name of cladistics—“phylogenetic systematics”—clearly rooted its approach in the evolutionary process, with the goal of identifying clades comprising ancestors and their descendants defined by shared-derived characters (synapomorphies). In cladistic analysis, the principle of parsimony is used to select among phylogenetic trees (cladograms) whose topologies depict the relationships of species and higher taxa. The use of cladistic parsimony methods for inferring organismal (as opposed to gene) relationships below the species level is more controversial and led to various versions of the Phylogenetic Species Concept (Table 1; Mayden, 1997; de Queiroz, 2007; Freudenstein et al., 2016). One version holds that phylogeny ends at the species level; cladistic approaches and terminology should not be applied to within-species tokogenetic relationships (Figure 3), and species should instead be defined by characters or combinations of characters fixed in individuals and populations. Other versions require species to be monophyletic, forming a clade in which individuals share at least one synapomorphy, so that the species is defined by at least one autapomorphy (a derived character unique to that species and not shared with other species). In paleontology, this consideration is relevant to the issue of when to recognize new species that differ from modern descendants, given an incomplete fossil record. In a model where anagenetic change occurs—change without the formation of new clades by cladogenesis, presumably due to speciation—morphological changes are presumed to be autapomorphies, and thus could mark new “chronospecies” (Figure 5; Silvestro et al., 2018; Marshall, 2019).

There is no requirement that the characters by which species are recognized be responsible for causing their divergence from their progenitor, or even that they be adaptive. However, “speciation genes” that could drive divergence, for example by leading directly to reproductive isolation (Figure 1), remain a topic of interest in the evolution literature (Burdon and Marshall, 1981; Wang and Hahn, 2018).

In the 1990s, the cladistic parsimony approach to phylogeny reconstruction was challenged by maximum likelihood and Bayesian methods employing explicit models of molecular evolution. Character change is treated probabilistically and not as a source of discrete apomorphies as in cladistics. Model-based approaches are now the mainstay of phylogenomics; like the MSC methods that underpin the species tree paradigm (Bravo et al., 2019), they are rooted in population genetics.

One of the Clevers et al. (2017) authors (Allon Klein) wrote, “No single attribute has served for cell type classification. Yet ‘we know it when we see it.”’ This echoes the Morphological Species Concept. But, he continued, “We are left with a functional but flawed taxonomy: functional, because it provides a language to describe biology; yet flawed, because it lacks consistency.” The search for objective criteria has led to the recognition of “transcriptomic cell types” by the neuron community (Zeng and Sanes, 2017; Cembrowski and Menon, 2018; Yuste et al., 2020; Callaway et al., 2021; Yao et al., 2021). Like phenetic clustering approaches to species recognition, this approach does not make strong theoretical claims—its primary goal is classification (Zeng and Sanes, 2017). It also has some of the same problems as phenetics. Breaking up continuous variation can involve subjective decisions (e.g., Cembrowski and Menon, 2018; Sagar and Grün, 2020), and the identification of biologically meaningful clusters from the high dimensional data produced by expression of 20,000 or more genes (Efroni, 2018; Quake, 2021) makes clustering more an art than a science (Kobak and Berens, 2019; Capdevila et al., 2021). Indeed, Chari et al. (2021) referred to single cell genomics as a “specious art” based on their finding that commonly used unsupervised clustering approaches for dimensionality reduction do a poor job of grouping adjacent cells in this complex parameter space. Northcutt et al. (2019) also noted limitations of transcriptomics but expressed optimism that the approach “can partially indicate identity, particularly once supervised methods incorporating known cell identification are employed.”

What of cladistic approaches? Vickaryous and Hall (2006) reconstructed most parsimonious trees from a matrix of 19 biochemical, physiological, and morphological characters in their study of neuron diversity. However, they did not use this approach to define cell types, but rather to apportion predetermined cell types into groups in a hierarchical classification. Raj et al. (2018) and Jones et al. (2020) also used parsimony methods for reconstructing cell lineage phylogenies, but not to identify cell types. Yuste et al. (2020) considered trees generated from (phenetic) clustering methods to be in “the historical tradition of using cladistics to classify organisms, assuming common ancestors in their evolution and synapomorphies (shared derived traits) among related clades”; this conflation of phenetic dendrograms with cladograms would horrify any cladist!

In contrast, the sister cell type theory of Arendt (2008) and Arendt et al. (2016, 2019) is explicitly phylogenetic, driven by the concept of homology as similarity due to common descent. A cell type is defined operationally as “a set of cells accessing the same regulatory program driving differentiation” (Arendt et al., 2019). Each cell type is characterized by the presence of a unique Core Regulatory Complex (CoRC; Figure 6), defined as “A protein complex composed of terminal selector transcription factors that enables and maintains the distinct gene expression program of a cell” (Arendt et al., 2016). The concept of terminal selector genes (TSGs), as high level cooperating regulators, generally transcription factors (TFs), that act to specify and maintain cell identity by controlling downstream selector and effector genes and by repressing other identities, has been developed by Hobert (2008; 2016; 2021); Patel and Hobert (2017); Hobert and Kratsios (2019); Sun and Hobert (2021). The CoRC includes not only TSGs but also “more general cofactors” (Arendt et al., 2016). The CoRC model requires “that cells of the same type implement the same hard-wired differentiation program, using the same transcription factors, regulatory elements, microRNAs, and so on” (Arendt et al., 2019). New cell types arise from an existing cell type by modification of the CoRC, for example by duplication and divergence of TSGs. During eukaryotic diversification “cell typogenesis” led both to the evolution of new functions and the partitioning of existing functions of already complex cell types, analogous to subfunctionalization of duplicated genes (Arendt, 2008; Arendt et al., 2019).

Cell type evolution can be represented as a phylogenetic tree of homologous cell types, hierarchical in nature like a species cladogram (Figure 4G; Arendt et al., 2016, 2019). The CoRC thus corresponds to the autapomorphy that defines the monophyletic species in some versions of the Phylogenetic Species Concept (de Queiroz, 2007) and should be shared by all members of the cell type “species” regardless of their state. It should, in theory, be a better apomorphy than the cell type marker genes used in many single cell studies to identify cell clusters, particularly in cross-species comparisons (Liang et al., 2018), since it is upstream of their expression. It is also of interest that this apomorphy is causative: Its formation creates and maintains as well as defines a cell type. This distinguishes the CoRC apomorphy from the characters that define species, whose adaptive value is unknown and which often may be neutral.

There are many challenges to making the CoRC concept truly operational (Zeng and Sanes, 2017; Arendt et al., 2019), beginning with identifying CoRCs (Almeida et al., 2021). As protein complexes, CoRCs cannot be assayed simply by monitoring transcription of TF genes, since TF proteins can be transcribed in one cell type and function in others (Clark et al., 2020); advances in single cell proteomics ultimately will resolve this problem (Labib and Kelley, 2020). Xia and Yanai (2019) discussed the use of CoRC as a means of defining cell type, noting that the CoRC remains largely conceptual for most cell types, and that a proxy for the CoRC is to use expression profiles of TFs, among which are TSGs of the CoRC (see also Almeida et al., 2021). Monitoring transcription is difficult enough given that many if not most genes, including TFs, are expressed across cell types (Figure 6; Xia and Yanai, 2019; Coate et al., 2020); regulatory regions differ by the strength with which they bind TFs rather than solely by which TFs they bind (Mora-Martinez, 2021); and TF expression is quantitative, not qualitative (e.g., Fishell and Kepecs, 2020; Shojaee et al., 2021). The assembly of such complexes can also be spatially and temporally disjunct from where they function (Charest et al., 2020). At the transcriptional level, “apomeres”—modules of genes defining novel or modified functions—are the apomorphies that define a cell type (Arendt et al., 2016, 2019).

Arendt et al. (2019) particularly noted the problem of the continuous nature of cell type variation among developmental and physiological states, as discussed above, including such issues as delayed commitment of a cell lineage to a particular fate (Wagner and Klein, 2020; Weinreb et al., 2020). Much would seem to depend on when and how the CoRC is assembled (Sagar and Grün, 2020)—whether as a single step or as a sequence of additions, and if the latter, whether these intermediate states are functional, and how they correspond to stages of tissue, organ, or individual development (Figure 7). If intermediate stages of the CoRC are functional—as might be expected given the ability of each TS to activate its own regulatory network (Hobert, 2008)—then the question of defining cell types becomes similar to the issue of determining the boundaries of chronospecies from fossils (Figure 5). Shojaee et al. (2021) referred to cell state transitions being “choppy” rather than truly continuous, occurring in waves, echoing the punctuated equilibrium vs. gradualism debate in paleontology (Benton and Pearson, 2001). But in the case of cell types there is greater hope of resolution given the availability of a much more complete inventory of cells, including cells in the process of differentiating, in contrast to the fossil record, which is generally fragmentary and is dominated by extinction events (Benton and Pearson, 2001; Marshall, 2017).

Arendt et al. (2019) recommended the application of objective phylogenetic methods for constructing cell type phylogenies, among which they included both parsimony and distance methods. Although they also mentioned modeling the cellular evolution process, they did not mention any specific models of cell differentiation or evolution. The need to develop probabilistic models of neuron cell commitment has been noted in several papers (Sagar and Grün, 2020; Wagner and Klein, 2020; Yuste et al., 2020). Mukamel and Ngai (2019) proposed that “Computational and statistical modeling of transcriptomic measurements from a range of neuron types could indicate which transcription factors are the core regulators of cell type identity.” In contrast, Fishell and Kepecs (2020) rejected what they call the “classic view” of specification of cell type by combinatorial action of TFs in favor of an attractor model built on Waddingtonian concepts of developmental commitment (Trapnell, 2015; Efroni, 2018); they contended that “combinations of TFs can initialize but not realize cell fates.”

The CoRC concept has the potential to provide both a theoretical underpinning and an operational criterion for defining most if not all mature cell types in a way unthinkable for species, with their diverse modes of origin. Is the CoRC indeed a universal feature of cell biology? If so, can methods be developed to assay it as an operational criterion? Can models of cell type differentiation be developed along the lines of the model-based approaches that revolutionized systematic biology for phylogeny reconstruction, and more recently (though more controversially) for defining species? As Yuste et al. (2020) put it, “A robust statistical framework that enables a quantitative definition of cell type (or tendency to be a type) is clearly needed.” What information should such models incorporate, beyond transcriptomic data? Could chromatin criteria be incorporated (Weinreb et al., 2020; Winick-Ng et al., 2021)?

Freudenstein et al. (2016) argued that the concept of lineage that increasingly has come to dominate systematics is necessary but not sufficient for defining species. They noted that the commonly invoked Evolutionary Species Concept involves more than just history in its definition. A species is “a phyletic lineage (ancestral-descendent sequence of interbreeding populations) evolving independently of others, with its own separate and unitary evolutionary role and tendencies” (Simpson, 1951, 1961; italics added here). Freudenstein et al. (2016) summed up their thesis (italics in original):

Species, therefore, are not only historical units; they are functional entities, and their function is directly connected to their ecological niche, as emphasized by a close relative of the Evolutionary Species Concept, the Ecological Species Concept of Van Valen (1976): a species is “a lineage (or a closely related set of lineages) which occupies an adaptive zone minimally different from that of any other lineage in its range and which evolves separately from all lineages outside its range.” Phenotypic characters are proxies for the difficult-to-define ecological niche function of a species: Simpson (1961) considered “morphological resemblances and differences” of populations to be related to roles if such differences are adaptive. Geography can also play a role in species delimitation as another proxy for “role” (Shanker et al., 2017). It is becoming more feasible to identify not only characters that are under positive selection (are adaptive), but even the small number of genes that might actually drive speciation (e.g., Choi et al., 2020).

Under the Evolutionary Species Concept, and in contrast to other species definitions (Table 1), related lineages that are geographically disjunct but share the same niche are members of the same species (Freudenstein et al., 2016). Such taxa represent one of several kinds of “cryptic species,” another being the opposite condition of taxa that are genetically and morphologically similar but have differentiated ecologically (Fiser et al., 2018). Another source of cryptic taxa is clinal variation, in which characters change gradually and continuously across the range of a species (Figure 8); clines can be caused by various phenomena, including primary divergence of populations into separate species and secondary contact between fully or partially differentiated taxa (Endler, 1977).

The role of species in ecosystems has become an increasingly important topic as the incidence of invasive species has accelerated due to climate change and globalization. One controversial topic has been whether species are interchangeable—whether it is the role they play, rather than their precise identity, that is critical for ecosystem function (Funk et al., 2017). Leuzinger and Rewald (2021) “argue that ecosystem function and ecosystem services need to be viewed not only through a taxonomic lens, but increasingly also through a functional, trait-based one.”

One author in the Clevers et al. (2017) opinion paper (Junhyong Kim) emphasized the cell’s “system-level roles,” noting that “A cell provides structure and it dynamically processes input environmental materials into outputs. The totality of its activity characterizes an ecological guild within the ecosystem of the organism.” He concluded that “we need an ecological definition of the cell type.” Such thinking is not new—Rowe and Stone (1977) concluded that cell type classifications should be treated as hypotheses concerning the cell’s “functional niche.” In this ecological view, each cell type is to a tissue or organ what an ecological species is to its ecosystem—each plays a particular functional role. Indeed, several other Clevers et al. (2017) authors mentioned “function” as a key criterion for defining cell type, despite the long-recognized difficulty of using function as an operational criterion (Rodieck and Brening, 1983).

Cell type classification systems prior to the availability of single cell transcriptomes were based largely on morphology, and identified a relatively small number of basic types. For plants, Xu et al. (2021) cited textbooks that list fewer than 20 cell types, most or all of which are found in more than one tissue and organ (e.g., parenchyma), though often with distinctive morphologies in different regions. For example, Arabidopsis leaf epidermal cells have a jigsaw puzzle shape, in contrast to the smoother borders of sepal epidermal cells (Xu et al., 2021). Epidermal cells presumably play the same general role in these and other organs, but the correlation of their morphological and spatial differences could indicate variation in their niche functions within their different cellular ecosystems. Identical transcriptomes could indicate identical functional roles, but the transcriptomes of epidermal cells in different plant organs have yet to be compared. How much difference between two transcriptomes would be required to indicate meaningfully different functions? And how does within- vs. between-region transcriptomic variance among epidermal cells compare?

Here again, the rich neuron literature provides some guidance. Variation among the vast diversity of neurons can often be categorized into a small number of groups (Stanley et al., 2020). According to Fishell and Kepecs (2020), 90% of cortical interneurons fall into four transcriptomic classes that they further subdivided into a detailed hierarchy, but they warn that this “comes with the clear caveat that both the percentage composed by these four cardinal classes and their relative contributions in specific areas will vary widely across the cortex.” Similarly, Osumi-Sutherland et al. (2021), working on a human cell atlas, noted that “cell types can be hierarchically classified and categorized in ever-increasing levels of resolution, from a general cell type such as an endothelial cell to more specialized types such as a liver sinusoidal endothelial cell (LSEC) and then down to highly specialized types found in specific locations such as a periportal LSEC.” Similar spatial fine structure also occurs in the mouse brain (Stanley et al., 2020).

Location is critical in cells as it is in real estate. Börner et al. (2021) stated that for the human cell atlas, “The spatial location of cell types within anatomical structures matters,” and Weiss et al. (2022) concluded that for melanomas “the anatomic position of the cell of origin endows it with a unique transcriptional state that makes it susceptible to only certain oncogenic insults.” But just as cell lineage and cell type can be incongruent, the same transcriptomic cell types can occur in disjunct regions (Raj et al., 2018; Yuste et al., 2020). For example, Yao et al. (2021) reported that although some of the 364 transcriptionally defined neuronal cell types they identified in the mouse isocortex were localized to very specific subregions or layers, most were much more broadly distributed. Discordance between location and cell type is not limited to vertebrates: Achim et al. (2018) reported transcriptomic vs. regional incongruence in annelid larvae. Should geographically disjunct but transcriptomically similar cells be split into different cell types? Xia and Yanai (2019) took the lumper approach: “the same cell type may be the integral building unit for many different tissues and/or organs” even though “such cell types that are found in multiple tissues and/or organs may not be exactly the same.”

For cell types as for species, function is more difficult to describe than are phenotypes that may serve as its proxies. If transcriptomes are not always well-aligned with cellular geography and environment, what about cell morphology? Peng et al. (2021) found that neurons of the mouse brain could be divided into 11 different “projection neuron types” many of which could be further subdivided into narrower morphological subgroups. Although the major types correlated with regional transcriptome variation reported by Yao et al. (2021), this was not true at the fine-grained scale, with the same transcriptomic cell types in different brain regions having different morphologies. Peng et al. (2021) concluded that “many aspects of morphological diversity cannot be accounted for by currently identified transcriptomic subtypes or clusters.” Because this morphological diversity is connected with the cell’s functional role in its specific niche in the brain ecosystem, their results highlight the need “to develop methods that enable complete reconstruction of morphology and in-depth gene-expression profiling to be conducted on the same cell” (Peng et al., 2021).

There are clearly parallels between a species’ role in an ecosystem and a cell’s function in a tissue, and in both cases the use of more accessible proxies for these important attributes is not foolproof. In systematics, the increasing availability of DNA sequence data has resulted in an even greater reliance on genetic lineage in defining species, but other characters clearly are also important for understanding species biology and evolution. This has led to what has been called “integrative taxonomy” (Padial et al., 2010; Fujita et al., 2012), which operationally is often an “iterative taxonomy” approach that begins with genetic lineage-based hypotheses and refines these using characters such as morphology (Yeates et al., 2011). If genetic lineages are insufficient to define species (Freudenstein et al., 2016; Sukumaran et al., 2021), then other characters must play a role analogous to supervised clustering of transcriptomic data. Ideally, such characters should be related to the ecological role of species, at least under the Evolutionary Species Concept. In cell biology, can integration of various criteria, such as transcriptome, morphology, physiology, and tissue/organ geography and ecology, be used to define and classify cell types, as envisioned by Peng et al. (2021)?