nature neuroscience

Review article https://doi.org/10.1038/s41593-024-01814-0

Implementation and validation of single-cell

genomics experiments in neuroscience

Received: 7 November 2023 Marco Colonna 1 , Genevieve Konopka 2 , Shane A. Liddelow 3,4,5,6 ,
Tomasz Nowakowski 7,8,9,10 , Rajeshwar Awatramani 11,
Accepted: 15 October 2024
Helen S. Bateup 12,13,14, Cathryn R. Cadwell 7,9,15,16, Emre Caglayan 2,
Published online: 3 December 2024 Jerry L. Chen 17,18,19,20, Jesse Gillis 21,22, Martin Kampmann 23,24,
Fenna Krienen 25, Samuel E. Marsh 26,27,28, Michelle Monje 29,30,
Check for updates
Michael R. O’Dea 3, Rickie Patani 31,32, Alex A. Pollen 33,34,
Francisco J. Quintana 35,36, Marissa Scavuzzo37,38, Matthew Schmitz33,
Steven A. Sloan 39, Paul J. Tesar 37,46, Jessica Tollkuhn 22,
Maria Antonietta Tosches 40, Madeleine E. Urbanek41, Jonathan M. Werner 21,22
,
Omer A. Bayraktar 42 , Ozgun Gokce 43,44 & Naomi Habib 45

Single-cell or single-nucleus transcriptomics is a powerful tool for

identifying cell types and cell states. However, hypotheses derived from
these assays, including gene expression information, require validation, and
their functional relevance needs to be established. The choice of validation
depends on numerous factors. Here, we present types of orthogonal and
functional validation experiment to strengthen preliminary findings
obtained using single-cell and single-nucleus transcriptomics as well as the
challenges and limitations of these approaches.

Single-cell or single-nucleus RNA sequencing (sc/snRNA-seq) is a validation experiments, covering a variety of common scenarios and
powerful tool for identifying cell types and cell states, detecting gene methods. Although the selected examples are not exhaustive, they
expression and epigenetic changes in disease and dysfunction, inferring illustrate how different validation steps complement and confirm sc/
developmental trajectories and cell-state transitions, predicting gene snRNA-seq results.
regulatory mechanisms, and comparing evolutionary modifications Some layer of confirmation is essential, with additional validation
in specific tissues across species. The high-throughput nature of these being subjective and depending on the nuances of each biological system.
experiments and the ever-growing array of computational tools for ana- The simplest validation experiments are those which validate expression
lyzing the data they produce also increases the risk of false discoveries, of a small number of genes. Such an approach using in situ hybridization
or discoveries that do not manifest with functional phenotypes—which (ISH) with probes targeting cell-type-specific genes combined with dif-
confounds interpretation of these data1,2. To confirm preliminary find- ferentially expressed genes (DEGs) of interest can provide confirmation of
ings from single-cell genomics experiments, validation experiments the sc/snRNA-seq findings. Protein-level validation can be examined using
using orthogonal and functional methods are required (Fig. 1 and Table 1). immunostaining. However, these methods of validation are not suitable
Here we discuss several general use cases and examples of these for validation of large numbers of DEGs, or when several genes are neces-
validation approaches1 (Box 1 and Table 1). In Box 2, we discuss why sary for identification of a cellular state or multiple states (Box 1, third use
evolutionary comparisons are important for both basic and transla- case). To validate such claims, multiplex in situ methods or genome-wide
tional neuroscientists. We also discuss challenges and limitations of sc/ spatial sequencing methods may be necessary (see below). With a grow-
snRNA-seq for evolutionary comparisons (for example, disentangling ing array of spatial transcriptomics and proteomics methods, careful
homology versus convergent evolution), relating in vitro models to consideration of the pros and cons of each method is warranted.
in vivo biology and accounting for technical and biological variability. Finally, many sc/snRNA-seq findings reach beyond description
We present specific examples of sc/snRNA-seq findings that require of gene expression and infer putative function and gene regulatory

A full list of affiliations appears at the end of the paper. e-mail: [email protected]; [email protected];
[email protected]; [email protected]; [email protected]; [email protected]; [email protected]

Nature Neuroscience | Volume 27 | December 2024 | 2310–2325 2310

Review article https://doi.org/10.1038/s41593-024-01814-0

Phase 1: Sample collection and preparation

Organoid Fruit fly brain Zebrafish brain Mouse brain Human brain

Phase 2: Single-cell/nucleus sequencing and primary analysis

t-SNE2
t-SNE1

Phase 3: Data integration and meta-analysis

Reference Cell type Reference Query

Query

Phase 4: Validation approaches

Alternative sequencing (e.g., ATAC-seq, long-read sequencing) Visualization of DEGs (e.g., IHC/IF, single-molecule FISH, spatial transcriptomics)

IHC FISH (single molecule)

ATAC-seq Long-read sequencing

DNase
Long reads

TF
Reference
genome Spatial transcriptomics MERSCOPE (multiplexed)

MNase

DNase
ATAC Disease gene

Validation in vivo and across species or disease models

Different species Different disease models
Functional validation (e.g., phagocytosis, electrical activity and other
plate-based high-throughput in vitro assays)

Phagocytosis assay Electrical activity Plate-based high-throughput

in vitro assays
% Phagocytosis

1 2 3 4 5 6 7 8 9 10 11 12
A
B
C
10 mV

D
E
50 ms F
G
H

Control Test

Fig. 1 | Considerations for orthogonal and functional validation of sc/snRNA- analysis; Phase 3, integration with other datasets (across disease models, species
seq data. sc/snRNA-seq data, regardless of the tissue originally collected from, or laboratories); Phase 4, alternative sequencing methods (for example, to access
require multiple validation steps to ensure their biological validity (see also chromatin accessibility, or long-read sequencing to detect isoform abundance),
refs. 1,2). In addition to ensuring proper powering of cell types of interest, visualization (using IHC, FISH or spatial transcriptomics), functional validation
additional steps should be applied for best practice, including several phases: to ensure subtypes or substates of cells are terminal and not transitory and,
Phase 1, experimental design (for example, choice of model species) and sample finally, cross-species validation (of particular importance when using animal
preparation (for example, enzymatic versus mechanical digestion—which is models of disease to ensure relevance to humans). One or all of these methods,
important for collection of immune cells153 or fresh versus frozen samples); among others, may be required to validate a number of DEGs identified in initial
Phase 2, Single-cell or single-nucleus sequencing, quality control, and primary sc/snRNA-seq experiments.

Nature Neuroscience | Volume 27 | December 2024 | 2310–2325 2311

Review article https://doi.org/10.1038/s41593-024-01814-0

Table 1 | Overview of common validation methods

Orthogonal validation Readout What is being Pros Cons

approach validated

RNAscope ISH Molecular RNA transcripts Single-cell resolution spatial Low throughput
validation; relatively inexpensive
MERFISH (Vizgen) Molecular RNA transcripts Single-cell resolution spatial Costly; requires specialized equipment
validation and reagents
Visium (10x) Molecular RNA transcripts High-throughput and anatomical Not single-cell resolution; requires
validation specialized reagents
IHC Molecular Protein Easily accessible with no specialized Need validated antibodies; low
reagents required throughput
Flow cytometry Molecular Protein Quantitative readouts at protein level Validating translation potential, which
may be discordant from RNA findings;
requires validated antibodies
CyTOF Molecular Protein Quantification of multiple cellular Validating translation potential, which
components simultaneously (high may be discordant from RNA findings
throughput)
CRISPR knockout Functional Gene function Test necessity of candidate genes; Can be low throughput; can be costly to
relatively standardized workflows test multiple genes
across model systems
CRISPRi/a Functional Gene function Manipulate expression of endogenous Variability in degree of interference or
genes and monitor phenotypic activation from gene to gene; susceptible
consequences; can be multiplexed or to epigenetic or trans-acting regulatory
performed in pooled screens environment
Perturb-seq Functional Gene function Massively parallel functional readouts Not trivial to design, execute and
of gene perturbation phenotypes interpret; costly; may require robust
by single-cell transcriptomics and selective challenge
individual cell resolution; can be used
with traditional Cas9 or CRISPRa/i
CROP-seq Functional Gene function Massively parallel functional readouts Not trivial to design, execute, and
of gene perturbation phenotypes interpret; costly; may require robust
by single-cell transcriptomics and selective challenge
individual cell resolution; can be used
with traditional Cas9 or CRISPRa/i
ECCITE-seq Functional Gene function An extension of Perturb-seq/ Challenging to implement for
CROP-seq to multimodal readouts intracellular antigens
RABID-seq Functional connections Cell–cell interactions High-throughput approach to validate Requires specialized reagents and
physical cell–cell interactions bioinformatic pipelines
Circuit tracing Functional connections Cell–cell interactions Can be used to identify short-range May be difficult to label deep brain
and long-range neuronal connections regions
SPEAC-seq Functional connections Cell–cell interactions Allows screening for Requires specialized reagents and
non-cell-autonomous phenotypes of bioinformatic pipelines
gene perturbations in one cell type
on another cell type in individual
droplets
Physiological readouts Functional Physical properties Can match biophysical properties of Requires specialized skillsets
(calcium imaging, of cells cells to their transcriptional identities; (electrophysiology); may require live
electrophysiology, powerful tools available intact tissue sections, cell-type-specific
transporter activity) genetic labeling, or robust purification
strategies to target cell types of interest
Live imaging (migration, Functional Physical properties Can be performed in high throughput Requires specialized microscopes
proliferation) of cells (multiple cells per image); provides (two-photon, light-sheet) and live cell
input on cellular behavior labeling tools
Dye filling for Morphological Cell morphology Can provide morphological Low throughput; requires specialized
morphological readouts information that is far more detailed equipment
than IHC
Viral targeting Morphological Cell morphology High-fidelity morphological May be difficult to label deep brain
information, can provide sparse regions
labeling for ease of reconstruction
Fluorescent protein Morphological Cell morphology Can label all cells of one type or Depending on driver, labeled cell density
expression (driver line) subtype across the entire CNS could be too high to identify individual
complex cells
CRISPR, clustered regularly interspaced short palindromic repeats; CROP-seq, CRISPR droplet sequencing; ECCITE-seq, expanded CRISPR-compatible cellular indexing of transcriptomes and
epitopes by sequencing.

Nature Neuroscience | Volume 27 | December 2024 | 2310–2325 2312

Review article https://doi.org/10.1038/s41593-024-01814-0

Box 1 Box 2

Questions that can be The importance of evolutionary

addressed using sc/snRNA-seq studies
•• scRNA-seq data indicate gene X is expressed by cell type A. •• Evolution is a general biological principle. Thus, understanding
•• scRNA-seq data indicate gene Y is upregulated in cell type B the contribution of evolution to nervous system function
during disease or pathology. provides important foundational basic science knowledge.
•• scRNA-seq data indicate cell type C is composed of three In addition, understanding the evolutionary constraints and
substates characterized by expression of gene X, gene Y and opportunities that have occurred in many organisms informs our
gene Z, respectively. understanding of the relevance of these changes in humans.
•• Compositional analysis indicates cell type or state D increases or •• Understanding the similarities or differences between cell types
decreases in abundance in disease or pathology. helps us better translate our findings from one organism to
•• Trajectory inference or RNA velocity suggests that gene X is another (for example, from other mammals to humans).
upregulated or downregulated as cells differentiate or respond •• Convergent evolution informs us about the constraints that
to insult or pathology. shape brain evolution in terms of plasticity and functional
•• Gene regulatory network inference and peak-gene linkage organization of the tissue. In this manner, we can focus on the
analysis, for example, suggest TF1 (or enhancer E1/repressor R1) potential cellular and molecular mechanisms that correlate
modulates expression of gene A. with convergent behaviors (for example, direct corticospinal
•• Cell–cell communication analysis suggests cell type A connections onto lower motor neurons and fine motor control).
modulates cell type B through the interactions of ligand L1 and •• The implementation of evolutionary approaches can result in
receptor R1. adaption of new model systems that may offer some technical
advantages for studying a general problem (for example, the
evolution of sleep154).
mechanisms, as in the cases of cell–cell communication analysis and •• Evolved nervous system function may be directly linked to the
gene regulatory network inference3. While spatial visualization of gene emergence of many types of nervous system disorder in humans
expression or protein abundance is useful for demonstrating colocali- that are not observable in other species.
zation of a ligand and receptor, visualization alone is insufficient to •• It can be valuable to include multiple species (for example,
demonstrate biological function, bona fide cell–cell communication nonhuman primates or rodents) rather than simply completing
or transcription factor (TF)–enhancer–promoter interactions. Validat- pairwise comparisons, as it is important to show that gene
ing these mechanistic and functional inferences requires perturbation expression signatures track with the biology being studied and
experiments (Fig. 2) and further functional studies. do not present irrelevant species-specific signals.

Primary orthogonal validation approaches

Validation of clusters versus validation of individual DEGs allows visualization of transcriptional discernment between cells in
By facilitating simultaneous comparison of transcriptomic profiles each group: the best clustering illustrates a diagonal line where each
from a collection of cells, sc/snRNA-seq can define clusters based on transcript has relatively uniform enrichment across all cells in a cluster,
DEGs4–6. However, algorithms used to process and analyze these data- but it is lowly expressed in all other cells across other clusters. High
sets are designed to detect transcriptional differences and, depending expression of cluster-enriched genes outside the candidate cluster
on the resolution specified, they will continue to subset cells even if indicates overclustering and too high resolution—suggesting clusters
the differences are so minute that they constitute noise7. It is essential likely represent the same cells or subtle changes between cell states.
to set several resolutions in clustering followed by iterations of com- When sub-clustering a cell type into distinct cell states, differences
putational post hoc validation before biological validation (see ref. between clusters are expected to be more subtle, reflecting shared
2 on sequencing technologies). The inherent features of clustering and distinct expression programs between cell states. An emerging
algorithms that provide modularity of the system also reinforce the alternative approach to identify these subtle continuous differences
importance of post hoc validation methods to ensure clusters reflect in gene expression that drive cell states is to identify gene programs
biology. coexpressed in subsets of cells (for example, by topic modeling12). It
may, therefore, be best to initially set the resolution high to overclus-
Integration and clustering across studies. Integration is a powerful ter before modifying the resolution, lowering it until robust cluster
tool for cross-dataset comparison, allowing joint assessment of cells separation is visualized.
in large-scale atlases. While this does remove cross-dataset independ- An additional computational method to ensure clustering is not
ence, it is frequently worth that price to jointly apply cell-type calling an algorithmic artifact is to pulse in a known disparate cell type to see
methods. These tools can be either classical integration methods with whether clusters collapse. For example, many groups subcluster cells
external dataset(s) of interest (to understand whether the cell type from larger datasets to allow identification of more subtle cellular
in question has an exact correspondence to published cell types), or subtypes or states that are otherwise hidden in larger datasets owing
transcriptomic similarity approaches (to understand whether the to divergent cell types, or missing from smaller datasets owing to
newly identified cell type is similar to a previously described cell type). underpowering of low-abundance populations. While this method is
The latter can be accomplished by using AddModuleScore() in Seurat8, standard and acceptable, adding in a different cell type is a stringent
Celltypist9 or CellHint10. Further computational methods to validate method to determine whether these subtypes or substates are truly
clustering numbers include visualization of the most highly enriched different. Unlike cell types, distinguishing between cell states is more
transcripts per cluster (top 10 or top 100) and projection across all cells complex, as cell-state transition might require only subtle differences
by cluster to create a feature plot of gene expression11. This method in expression programs and involve a continuous transition between

Nature Neuroscience | Volume 27 | December 2024 | 2310–2325 2313

Review article https://doi.org/10.1038/s41593-024-01814-0

a Testable hypotheses from single-cell datasets

Regulatory DNA elements

Targets of regulation
Regulatory RNAs/proteins

DEGs Changes in cellular function

Changes in cell state

b CRISPR-based tools for perturbation

Altering DNA sequence Controlling gene expression

Repressor
CRISPRi
DNA sgRNA Cas9

DNA cleavage > repair TSS or Gene repressed

regulatory site

Base-modifying CRISPRa
Base editing Cas9 nickase Activator
enzyme
or dCas9

TSS or
Gene induced
regulatory site

Prime editing Reverse CRISPRoff DNA methylase

Cas9 nickase transcriptase

pegRNA

CpG island Gene repressed

c High-throughput strategies for evaluating perturbations

Droplet-based scRNA-seq
+ identification of sgRNA
Validation of:
• Drivers of changes in
Effect of gene perturbation on transcriptome gene expression and
Transduce with pooled cell states (regulatory
sgRNA library targeting DNA elements and
Mammalian cells
regulatory elements, Compare trans-acting RNAs/
expressing
noncoding RNAs, sgRNA proteins)
CRISPR High
or proteins frequencies • Effect of DEGs on cell
machinery
states and cell functions
FACS sort
based on
Fluorescent fluorescence Effect of gene
stain or reporter perturbation on
for expression Low phenotype
of gene of monitored
interest, by fluorescence
cellular state
or activity

Fig. 2 | Overview of perturbation-based validation approaches. a, sc/ tools. b, Examples of CRISPR-based tools to perturb genome sequence and
snRNA-seq datasets can generate different types of functional or mechanistic gene expression. c, Experimental strategies for high-throughput CRISPR-
hypothesis. Arrows mark hypotheses for causal relationships, which are based perturbation experiments to validate and test functional or mechanistic
generated from single-cell datasets and can be tested by CRISPR-based validation hypotheses from sc/snRNA-seq datasets. TSS, transcriptional start state.

states. Moreover, the correct or most appropriate resolution of the Most importantly, all clustering of sc/snRNA-seq data necessitates
diversity of cell states is less well defined, and different resolutions post hoc validation13. Single-cell or single-nucleus data can be viewed
could correctly capture true biological differences in cell states. We as a prediction of biology that allows for guided hypothesis generation
will further discuss this issue in depth below. to gain biological insights14.

Nature Neuroscience | Volume 27 | December 2024 | 2310–2325 2314

Review article https://doi.org/10.1038/s41593-024-01814-0

DEGs as markers of individual cell types and clusters. DEGs can be number and type of assays needed to validate a newly identified cell
extrapolated as rudimentary markers of cell types or an indication of type or state remain unclear, but at a minimum it is recommended that
changes in expression programs between cell states. However, tran- findings from sc/snRNA-seq be validated using at least one independ-
scriptomic profiles provide a snapshot of cellular behaviors. Given ent assay (for example, visualization using in situ or spatial transcrip-
that DEG identification is inherently comparative, additional genes tomics, single-nucleus assay for transposase-accessible chromatin
playing a pivotal role in driving cell identity may escape such analysis. (ATAC) to highlight chromatin accessibility for DEGs, or functional
For example, an essential interaction between two genes may be nec- assays). However, it is also advisable to move beyond validation of
essary for supporting a cell’s function. In the absence of one of those individual transcripts and make every attempt to validate targets at
genes, a population of cells is actually losing the combined effect of the level of proteins, cellular physiology, anatomical distribution,
both; while this first gene may appear in a DEG analysis between these developmental lineage, morphology, connectivity and/or function.
two clusters, the second gene can go unnoticed14–16. These additional phenotypic validation steps establish the robustness
Herein lies the need for orthogonal validation, as the complex of the cell type or state under investigation and provide mechanistic
interplay of DEGs remains largely hidden within transcriptomic data. understanding of its role in the nervous system.
Orthogonal validation is a crucial intermediate in determining whether
a DEG derived from a cluster is recapitulated within a cellular context Methods of validation by visualization
relevant to the biological question at hand. Depending on parameters Spatial transcriptomics. Spatial transcriptomics, in conjunction with
used for DEG discovery, as well as the condition of samples before ISH, including single-molecule ISH or RNAscope ISH, immunofluores-
sequencing, genes defining a cluster may not translate back to a tis- cence (IF) and immunohistochemistry (IHC), represents a powerful
sue setting. The statistical method applied relies on an assumption combination of techniques for comprehensive characterization of gene
of distribution of noise, and inferred DEGs could vary with the choice expression and protein localization20–22. While sc/snRNA-seq provides
of method. For example, frequently used tests such as a generalized information on transcriptomes at the single-cell or nucleus level, ISH
linear model assume a negative-binomial or Poisson distribution, while enables visualization of specific RNA molecules within intact tissues,
a Wilcoxon rank-sum test assumes a non-parametric distribution, and confirming their spatial distribution. IF and IHC allow detection and
should be followed by multiple-hypothesis correction2. The expression localization of proteins, providing additional information on cell types,
level of DEGs may also fluctuate during a cell’s normal state; thus, cell protein–protein interactions and putative cellular functions. Integra-
types or identities cannot be declared in lieu of lineage tracing and tion of these complementary techniques verifies sc/snRNA-seq findings
developmental analysis. It then becomes important to demonstrate and allows researchers to study coexpression of genes and proteins in
that any genes of interest are truly and consistently expressed at differ- the context of tissue architecture, providing a more comprehensive
ent levels between cell states of interest before moving toward mecha- understanding of cellular behavior and molecular interactions within
nistic interrogation. This orthogonal evidence includes experiments complex biological systems.
that are integral to validation of epigenomic, morphological, spatial By leveraging the strengths of spatial transcriptomics, ISH, IF
and biophysical properties or function of cells that are predicted from and IHC, researchers can unravel intricate spatial dynamics of gene
transcriptomic data13. expression and protein localization, advancing understanding of tis-
sue development and disease pathogenesis. These methods require
Prior cell classifications as a scaffold for analysis. Understand- a priori knowledge of DEGs from sc/snRNA-seq experiments. Whereas
ably, prior cell-type classification schemes have been biased toward ISH, IF and IHC have low throughput, allowing the validation of a few
molecular assays owing to the unprecedented scale and throughput genes at a time, spatial transcriptomics has a throughput of up to
of sc/snRNA-seq17. However, there are a number of issues that can arise thousands of genes. Sequencing-based methods such as Visium and
when such data are interpreted in a vacuum: (1) it is difficult to distin- Slide-seq23 enable simultaneous capture of gene expression informa-
guish molecular features that define stable cell types from transient tion from multiple spatially defined regions within a single sample.
cell states; (2) resulting cell-type atlases may vary depending on sample These genome-wide technologies enable localization of groups of DEGs
size and analytical parameters used for clustering, leading to lack of (often called ‘gene modules’) in tissue sections. By combining spatially
reproducibility with no clear ground truth; and (3) functional relevance resolved gene expression profiling at 50–100 µm resolution (updated
of molecularly defined cell types is unclear. to 8 µm resolution with Visium HD and 10 µm with Slide-tags24) with
Recent multimodal single-cell analyses call into question the high-throughput sequencing, these methods provide comprehensive
notion of discrete cell types, suggesting that continuous and cor- validation and complement sequencing data. In situ-based meth-
related variation in cellular morphology, biophysical properties and ods, such as multiplexed error-robust fluorescence in situ hybrid-
molecular features contributes substantially to cellular diversity ization (MERFISH)25, STARmap26 and in situ sequencing27, provide
within broad transcriptomic classes18. Validation of transcriptomic high-throughput direct identification of RNA transcripts at subcel-
clusters, that is, confirming they have biological relevance, requires lular resolution of panels of several hundred genes by single-molecule
orthogonal and functional validation. One major reason for requir- fluorescence in situ hybridization (FISH) with sequential imaging and
ing such validation of in silico clusters is that cell atlas studies can signal amplification techniques. See ref. 2 for a more in-depth discus-
be underpowered for rare cells or cell states, often causing cluster- sion of these methods.
ing artifacts. Some methods have been developed to overcome this
caveat, such as FIND-seq (focused interrogation of cells by nucleic Genes versus proteins. Use of imaging-based RNA visualization to
acid detection and sequencing)19—developed to study rare astrocyte validate sc/snRNA-seq results, and/or to spatially resolve sc/snRNA-seq
populations isolated on the basis of expression of a few mRNA markers. data, provides important context and orthogonal validation. However,
Additional validation steps, including alternative sequencing efforts, when inferring potential functional consequences of gene expression
multi-dataset integration and meta-analyses and visualization (for changes, it is critical to also consider protein-level validation. Although
example, in situ, MERSCOPE), are integral to validate the biological transcript and protein levels are generally correlated, there are several
truth behind computational analyses. factors that can drive dichotomy in transcript:protein ratios, including
Validation of transcriptomic clusters rests on the hypothesis that regulatory relationships and mechanisms regulating protein localiza-
bona fide cell types should form discrete entities. In other words, if a tion, activation and turnover28–30. The nonlinear relationship between
group of cells segregates as a distinct population using multiple assays, gene expression and protein levels is particularly noticeable when
this would support its designation as a valid cell type or state. The comparing sc/snRNA-seq data with TRAP-seq or proteomics data,

Nature Neuroscience | Volume 27 | December 2024 | 2310–2325 2315

Review article https://doi.org/10.1038/s41593-024-01814-0

where ribosomes and proteins may be in soma-distant processes—par- However, in vivo screens are challenging to perform, because not all
ticularly common in CNS cells. cell types are currently easy to manipulate (for example, microglia), or
One area in which protein-level validation provides critical infor- target (for example, substates of reactive cells). In vitro screens using
mation is in the inference of cell–cell communication. The power of iPS cell-derived models and primary isolated cells are an alternative,
analysis at the single-cell or nucleus level has led to a rapid expan- and have been predictive of cell states found in vivo—in particular for
sion of methods to assess putative cell–cell communication through microglia47 and astrocytes48,55. These in vitro CRISPR validation steps
ligand–receptor interactions from scRNA-seq and/or spatial tran- can predict functional consequences of changes in gene expression
scriptomics data31–33. Results from these tools, however, should be (Fig. 2).
considered hypothesis generating and not hypothesis validating.
Beyond assessing whether ligands and receptors truly interact, a lack Functional validation
of confirmation that ligands and receptors are present at the protein Functional validation studies are needed, as the physiological relevance
level and in appropriate spatial context represents a critical gap to of different cell clusters cannot be inferred from their transcriptional
fill. Expression and/or spatial colocalization can be accomplished by state, but they can serve as a cell typing guide. For cell types described
relatively straightforward techniques such as IHC or flow cytometry/ by sc/snRNA-seq, it is important to perform functional validation to (1)
cytometry by time of flight (cyTOF), and physical interactions can be ensure transcriptionally defined clusters represent a true cell type and/
examined via co-immunoprecipitation or newer techniques such as or subtype or substate, and (2) understand their properties in order to
multistage native mass spectrometry (nativeomics)34. place them into the context of the larger circuit or brain region. This is
particularly important when defining ‘states’ of cells. What ‘functional
Methods of validation by interrogation validation’ entails is still under debate. Broadly, it links the molecular
Single-cell or single-nucleus transcriptomic approaches uncover DEG profile (measured through sequencing data) with a phenotype as a way
signatures between different cell types and states. While powerful, to confirm that genes or gene modules identified in sc/snRNA-seq are
these data are descriptive and do not establish causality or provide biologically meaningful. Such inference is not trivial, and transcrip-
mechanistic insights. However, DEGs can be leveraged to generate tionally distinct clusters should not be assumed to always associate
additional hypotheses to test using perturbation-based approaches with functional differences. Indeed, owing to the heuristic nature of
to interrogate causality and link gene expression to cellular function clustering tools, some separation of cell clusters in sc/snRNA-seq data
(Fig. 2). We cover specific functional validation approaches in the next is almost inevitable and so must be interpreted with appropriate cau-
section. We will first describe additional gene editing approaches to tion in terms of its biological relevance. It is noteworthy that fleeting
validate gene expression data. It remains true, however, that even per- changes in cell state, such as stage within the cell cycle, may have greater
turbation experiments must be validated at the functional level. For transcriptomic impact than cell type or substate56. The degree of func-
instance, if a perturbation experiment removes (putatively) phagocytic tional plasticity in this context should also not be underestimated57.
pathway genes, the true final validation of this is to test the functional- These issues notwithstanding, sc/snRNA-seq has revealed relatively
ity of phagocytosis itself. consistent signatures across adult differentiated tissues, predictably
Several perturbation methods exist in cultured cells and model correlating with changes in cellular morphology58.
organisms. RNA interference technology enables knockdown of mRNAs
by synthetic short interfering RNAs or transgenically expressed short Using cell culture systems for validation
hairpin RNAs, although it suffers from off-target effects. CRISPR tech- Many protocols exist for purification and culture of CNS cells. Origi-
nology, which enables gene knockout to achieve a complete loss of nally involving enrichment (rather than purification) of rodent cells
function, has improved our ability to interrogate gene function in a from embryos or neonates and inclusion of serum in media to provide
scalable and precise manner35. CRISPR interference (CRISPRi) and trophic support, these cultures have provided understanding of basic
CRISPR activation (CRISPRa) approaches use a catalytically inactive functions of CNS cells. However, recent advances in these techniques,
Cas9 protein to recruit transcriptional regulators to genomic sites combined with expanded interest in studying disease states of indi-
of interest, enabling modulation of expression levels36 and thereby vidual cell types, requires that these methods be questioned—namely
providing a strategy to directly model changes in expression levels of as serum is not a normal component of the CNS. While methods for the
specific DEGs identified from sc/snRNA-seq. CRISPRi/a also can target culture of CNS cells in the absence of serum have emerged59–61, these are
distal regulatory elements, such as enhancers, to establish their func- not always used—raising concern about in vitro functional validation
tion in controlling gene expression37,38. experiments, in particular those of CNS glia and immune cells62. This
CRISPR-based gene perturbation can target genes of interest in does not negate the power of cell culture as a validation tool; instead it
individual experiments and in massively parallel screens using pooled highlights the importance of choosing the appropriate in vitro model
single guide RNA libraries targeting multiple genes of interest. Pooled to use for replication of populations of cells of interest identified from
screens can be conducted for a large range of phenotypes, including sc/snRNA-seq studies. Such methods have been discussed extensively
cell survival or specific cellular functions, and states can be read out for astrocytes and microglia elsewhere62, but similar points exist for
by fluorescent markers or reporters or additional sc/snRNA-seq (for all CNS cells, and will not be covered here.
example, Perturb-seq39–42 or CROP-seq41). They can also be used for Strengths of in vitro functional validation include the ability to
screening of long noncoding RNAs43 and cis-regulatory regions (for study a homogeneous and pure population of cells of a single type or
example, in induced pluripotent stem (iPS) cell-derived neurons44 state, and the ability to validate human sc/snRNA-seq data using human
and microglia44,45). As reviewed in ref. 2, the development of genome cells (Fig. 2). This involves use of (induced or embryonic) pluripotent
editing technologies, exemplified by CRISPRi/a, CRISPR deletion stem cell models63,64, two-dimensional or three-dimensional culture
and CRISPR indel screens, has facilitated large-scale perturbation systems65–67 including organoids (see below), or using primary tissue
and assessment of DNA sequences. CRISPR-based screens have been explants. Astrocyte reactive states can be used as an example, with the
successfully deployed to study cell types relevant to neuroscience, comparison of untreated pluripotent stem cell-derived astrocytes ver-
including human iPS cell-derived neurons46, microglia47, astrocytes48, sus an induced reactive state following treatment with tumor necrosis
neuron–glial co-culture systems49, brain organoids50, and in mouse factor, interleukin-1α and C1q68,69. Clusters of interest can be isolated
brains in vivo51–53. To minimize artifacts linked to some in vitro cell through live cell sorting (for example, fluorescence-activated cell
isolated systems (see below), care should be taken in choice of cell sorting or magnetic-activated cell sorting) and, after revalidation of
culture systems, or in vivo CRISPR assays should be considered54. transcriptional state, consensus homeostatic functional attributes

Nature Neuroscience | Volume 27 | December 2024 | 2310–2325 2316

Review article https://doi.org/10.1038/s41593-024-01814-0

can be interrogated (for example, glutamate reuptake, synaptogenesis Similarly, functional validation of astrocytes in different reactive
capacity, neurotrophic capacity) and separately the gain of entirely and disease states has been successful at adding biological value to
new functions (for example, neurotoxicity). transcriptomically defined populations. Starting with transcriptomic
Deploying these approaches means generating and/or isolating data as a roadmap for reactive astrocyte substates in vivo, research-
cell cultures with particular transcriptomic signatures that are viable ers can isolate primary rodent or iPS cell-derived human astrocytes,
for such functional characterization assays, which may present chal- recapitulate the original gene expression signature, then continue
lenges for different cell types and/or states. Single genes or gene sets with functional validation to determine whether any loss-of-function
implicated by statistical prioritization (for example, by significance and or gain-of-function changes are present across substates. Several
effect size) or necessity and sufficiency for a particular function (or set pipelines exist for producing high-throughput and controllable cell
of functions) can be genetically interrogated using these approaches. culture platforms for validating astrocyte disease biology55,68,69,87,88.
However, if a functional difference is not detected in vitro, the negative
predictive value is somewhat limited because there may be specific Cross-species comparisons
functions that are difficult to reproduce in in vitro conditions. sc/snRNA-seq provides a molecular common language definition of cell
Additional hypotheses and validation can be seeded by results types across any species with a quality genome. While the definition of
from methods for the study of cell–cell interactions. Emerging tech- a cell type typically invokes other features (for example, connectivity,
nologies that have recently been applied to astrocytes and other glial function, morphology), these features can be difficult or impossible to
cells include RABID-seq (rabies barcode interaction detection followed acquire across species17,89. Moreover, cell-autonomous gene expression
by sequencing70), SPEAC-seq (systematic perturbation of encapsulated programs are the foundation on which many (but not all) structural
associated cells followed by sequencing), which enables droplet-based and functional features of a cell are built90. This foundation of shared
culture of putative interactive cells42 and LIPSTIC (labeling immune gene expression programs and functional properties across species
partnerships by SorTagging intercellular contacts71) for cell–cell enables an inference process termed homology mapping. Properties
interactions (first used to uncover interactions between T cells and such as connectivity and physiology are far easier to study in genetically
dendritic cells). tractable and experimentally accessible animal models (for example,
Mus musculus or Drosophila), and then can be transferred by anchoring
Functional validation at the organismal level to homologous cell types in other species, including human91.
Here, we aim to provide general workflows to test the functional rel- Identifying homologous cell types across species ideally involves
evance of molecularly identified cell clusters. To perform functional identifying sets of cells in each species that access similar regulatory
validation studies in animal models, it may be necessary to generate programs for their differentiation92. Single-cell or single-nucleus
genetic reporter animals that label a given cell type based on its expres- sequencing combined with lineage tracing or fate mapping is pow-
sion of a marker/DEG72,73. Genetic reporter mice can also be useful for erful for reconstructing developmental histories of cell types93, and
mapping the in vivo properties of cell types in the nervous system hence their relationships across species. We still have a fragmented
including their activity patterns and contributions to behavior. This understanding of the lineages of transcriptionally defined cell types
strategy has been valuable to catalog properties of CNS neurons74 (with in any one species94–97. Few comparative studies have matched pro-
the caveat that some features may be species specific (for example, genitor classes across species91,98; although, even if such information
human) and thus escape validation in a different species (for example, is available, shared developmental history is neither necessary nor
a mouse model)). In many cases, cell types cannot be selectively labeled sufficient to infer cell-type homology. The challenge with matching
by expression of one marker gene alone, as other cell types may also homologous types from adult data alone is distinguishing shared
express that gene. In that case, it can be useful to use an intersectional evolutionary history from phenotypic convergence94,99, but with large
strategy whereby cells expressing two genes (or expressing one gene enough sets of species this can be alleviated100. Transcription factors
but not another) can be selectively targeted75. This strategy has been (TFs) potently specify cell-type identity, suggesting that prioritizing
used to study different types of neuromodulatory neurons, includ- TFs in cross-species cell-type mapping may improve homology assign-
ing dopamine and serotonin neurons, which have been shown by sc/ ments. A single TF or small set of factors can be sufficient to switch the
snRNA-seq to be highly heterogeneous76–79. fate and ultimate cellular identity. However, TFs are developmentally
Functional validation of neuron types has been performed in regulated and may not be conserved between the early stages of speci-
several large-scale cellular taxonomy papers80. A more general strat- fication and adulthood.
egy moving forward could be to start with cell types defined by sc/ Across species, cell types change abundance, gene expression pro-
snRNA-seq, then perform spatial transcriptomics to demonstrate files and spatial distribution. Each of these differences carries its own
anatomical localization, then record from cells to examine intrinsic challenges for cross-species comparisons. In general, cell-type simi-
membrane (or synaptic) properties and perform morphological recon- larity decreases, and homology mapping becomes less accurate, with
struction (for example, measure properties of dendrites, spines or increasing evolutionary distance91,100–104. Remarkably, sc/snRNA-seq
axons), or perform Patch-seq to pair electrophysiological recordings can be used to reveal conservation and novelty of brain cell types across
with gene expression data from the same cell81,82. A further step includes 500 million years of evolution, suggesting that core transcriptional
generating a reporter animal that labels the population of interest, then programs that define cell types are often deeply conserved99,100,105,106.
testing in vivo circuit connectivity and behavioral relevance of that Integrating transcriptional profiles from single cells or nuclei across
specific cell population. This strategy can also be applied for functional species can be difficult, because approaches may rely on assump-
validation of cell types in brain organoids. tions of 1:1 orthologous genes107, and gene duplications, losses and
Recent work in human iPS cell-derived organoids containing both sequence-level divergence increase with evolutionary distance. Evolu-
neurons and astrocytes provided functional validation of astrocytes tionary modifications of cell types may result from neutral drift, physi-
from organoids at several developmental stages, performing a variety cal constraints associated with brain reorganization or new functional
of assays that probe known astrocyte functions83. In brain organoids, requirements. Overall patterns of transcriptional divergence have been
cell morphology, protein levels, progenitor differentiation potential linked to neutral drift among primates, coupled with stabilizing selec-
and intrinsic physiological properties can be measured functionally. tion over longer timescales at the cell type or tissue level108.
Recently, in vivo imaging of neuronal activity in the mouse brain has Understanding the neuroanatomical and/or physical constraints
been combined with spatial transcriptomics84–86 to functionally vali- that drive evolutionary features such as proportional shifts in cell
date molecularly defined cell types. types across species remains challenging. Increases in cell-type

Nature Neuroscience | Volume 27 | December 2024 | 2310–2325 2317

Review article https://doi.org/10.1038/s41593-024-01814-0

abundance across species can be intuitively linked to species-specific These observations are in line with an evolutionary definition of
adaptations, such as the proportional increase in a retinal ganglion cell cell type, whereby homologous cells share expression of TFs that estab-
subtype in primates that may relate to neocortically driven adapta- lish and maintain their genetic identity114. Comparing TF expression can
tions to high visual acuity100. However, sometimes the mechanism help disambiguate homology and convergent evolution in cell-type
driving the differential modification of each cell type is difficult to comparisons of distantly related vertebrate species99,102,105,115–117. For
resolve. For example, reduced proportions of subcortically projecting example, distinct classes of cortical GABAergic interneurons in amphib-
cortical neurons in larger mammalian brains109 may relate to functional ians, reptiles, birds and mammals express the same TFs defining class
requirements to maintain scaling relationships between upper and identity; however, expression of certain effector genes, such as the
lower motor neurons despite disproportionate cortical expansion, or calcium-binding protein parvalbumin, which marks a class of mamma-
to shifts in the migration of homologous types during development, lian GABAergic interneurons, is not conserved across species99,102,105,115.
with either mechanism leading to different anatomical distributions.
Another example is the recently observed relative increase in the Are all genes equally informative?
proportion of oligodendrocyte progenitor cells compared to mature Whether and how to give different weights to genes for homology
oligodendrocytes in the human versus nonhuman primate brain110. inference remains an open question both conceptually and algorith-
This difference might support enhanced neuronal or myelin plasticity mically. While TFs seem to carry a higher weight for homology infer-
in humans, or some other phenotype that has not yet been linked to ence, TF combinatorial codes may themselves drift (for example,
oligodendrocyte function. Increased sampling through sc/snRNA-seq by paralog switching118). This is particularly relevant for comparing
across species and individuals within a species helps to distinguish distantly related species. It should be noted that many comparative
processes of drift and selection, while further analysis of scaling rela- studies are carried out in lower-quality assemblies, which may not have
tionships and functional changes are required to resolve contribution completed enough annotation for the 3′-biased data of sc/snRNA-seq.
of physical constraints and new cellular specializations. Arbitrating Standard approaches use one-to-one orthologs, assuming that genes
between such possibilities improves understanding of the targets of carry similar functions, but this assumption cannot be made for para-
evolutionary selection. Finally, issues such as overfitting to a single logs, as gene duplication may be followed by sub-functionalization
species reference dataset, and differences in genome quality across or neo-functionalization. However, limiting analysis to one-to-one
species, add technical complexity to cross-species comparisons. orthologs filters out a considerable fraction of the transcriptome when
the species compared are separated by large phylogenetic distances.
When and how does validation fail? Computational solutions to solve this problem have been proposed
Validation may fail owing to poorly designed initial sequencing experi- recently107,119.
ments. This can be due to contamination of cell types of interest in bulk Finally, limiting cross-species comparisons to TFs comes with the
sequencing efforts, or, more common with the uptick in sc/snRNA-seq risk of providing an oversimplified representation of cellular diversity.
studies, a failure to properly power the study for the cell type of interest As the field defines subtler distinctions of subtypes within given cell
(see ref. 2 for a review of study design considerations). Similar arti- classes, TF-level gradients and/or post-translational modifications
facts arise when insufficient biological replicates are sequenced—often may be identified in eliciting distinct transcriptional programs, mak-
owing to high costs of sc/snRNA-seq experiments111–113. Although chal- ing cross-species comparisons more complicated. Potential develop-
lenges exist, emerging technologies such as RNA editing and cell-based mental regulation of TF expression makes their use to define identity
assays offer promise for improved validation. Careful consideration challenging from a temporal perspective, even within a species. For
and exploration of validation options are necessary to ensure the reli- example, the comparison of TFs may not be powered to identify cell
ability and robustness of sc/snRNA-seq findings. A brief overview of types that have diverged recently (sister cell types114), which, by virtue
common validation difficulties from a computation standpoint are of their recent diversification, share a large fraction of their transcrip-
covered elsewhere1; here we cover other validation considerations. tomes. It should be noted that such post-translational modification
considerations are not unique to TFs, and should be considered for
Conceptual limitations of transcriptomic-based homology all proteins.
inference
Homology describes phylogenetic relationships and is not a synonym Are cellular transcriptomes enough to infer neuronal
of similarity. Cell types may have similar transcriptomes because they homologies?
descend from a common ancestor, or because they acquired these prop- As described above, homology inference becomes harder with increas-
erties by convergence after evolving under similar selective pressures ing phylogenetic distance, especially when there are large branch
(Fig. 3). The ideal way to discriminate between these two possibilities lengths between clades with no extant species, for example, compar-
is to sample many species and reconstruct ancestral states using the ing mammals to reptiles. Natural selection acts on the output of brain
principle of parsimony (convergent characters tend to lack phyloge- activity, that is, the ability of the brain to support adaptive behaviors in
netic continuity). Because this is not always feasible for detailed sc/ an organism’s environment. The substrate of selection is the frequency
snRNA-seq characterization, here we highlight some observations of allelic variants in the population, yet resulting changes in gene tran-
that may help with comparing transcriptomic data across species. scription that do not lead to functional changes at the individual cell
Data from neuronal cell types for which homologies were established level would not be under selection pressure. This mapping between
by independent criteria (morphology, input–output connectivity) offer genotype and phenotypes under selection in the brain is nontrivial:
two key insights. First, although transcriptomic divergence of homologous genes do not control function directly (with a few exceptions); rather,
neurons is generally a function of phylogenetic distance100,104,109, rates of they affect behavior by instructing cell-type identity, neuronal wiring,
transcriptomic divergence are cell-type specific: in primate cerebral cor- activity, spatial allocation and several other complex biological vari-
tex, for example, nonneuronal cells diverged more rapidly than neurons104. ables. Transcriptomes alone might be insufficient to infer homology
Second, transcriptomic divergence is not even across gene families. TFs when we do not know the traits under selection or how genes are related
known for specifying cell identity have conserved expression in homolo- to those traits. Other comparisons that can help with homology infer-
gous neuron types, whereas the expression of terminal markers or effector ence are defining the developmental origin and neuronal connectiv-
genes may switch more rapidly104. This indicates that homologous neurons ity of a given cell, although these criteria have their own caveats. The
may acquire species-specific functions, such as new electrophysiological concordance of developmental origin, transcriptomic similarity and
properties109, without changes to their core genetic identity. input–output connectivity is ideal for a solid homology inference.

Nature Neuroscience | Volume 27 | December 2024 | 2310–2325 2318

Review article https://doi.org/10.1038/s41593-024-01814-0

a Homology b Convergent evolution c Innovation

TFs TFs TFs

Independent Independent New cell type

Last origin origin
common
ancestor

d Classes of telencephalic GABAergic e Neocortex and DVR f Primate TAC3 striatal interneurons
interneurons

All genes
Marmoset
MC
D2 MSNs DMC Correlation
Mouse
pDC 0.21
Turtle

aDC
CGE PT 0
aLC
pLC
aDVR −0.2
CGE Pallidal/septal/ D1 MSNs pDVR

NGC cholinergic
TFs
MC
DMC
pDC
Correlation
CHAT + PTHLH +
0.55
PVALB +
Turtle

aDC
MGE PT
aLC 0
pLC
aDVR CCK + VIP+
−0.41
Olfactory bulb pDVR
Olfactory bulb
PHG.cos

CgGp.s
CgGp.i

CgGf.s

COMA
CgGf.i
PHG.l

BMA
CA4
CA3
CA2

Sub

BLA
CA1

ATZ
SIG
LIG
DG

OL

Pir
LA
TL
PL
FL
Cl

(glomerular)
(granule)
Mouse TAC3+
TH+
Turtle Hippocampus Neocortex Pallial amygdala + + +
OB neuroblasts SST NPY NOS1
Finch

Fig. 3 | Illustration of cell-type homology, convergence and innovation. ridge; MGE, medial ganglionic eminence; MSN, medium spiny neuron; NGC,
a–c, Schematics of cell-type evolution. Circles indicate TFs. d–f, Examples neurogliaform cell; OB, olfactory bulb; TF, transcription factor. Panel d adapted
from the literature using single-cell genomics to address each type of cell-type with permission from ref. 115, AAAS; panel e adapted with permission from ref.
evolution. CGE, caudal ganglionic eminence; DVR, dorsodorsal ventricular 102, AAAS; panel f reproduced from ref. 101, Springer Nature Limited.

Consideration for modeling function using in vitro systems comparisons typically use a singular in vivo dataset as a reference,
Inaccessibility of neural tissues in humans and other species can limit which ignores potential variability within the reference data that
the types of functional validation that can be performed, making organoids may not recapitulate. Especially with sparse and noisy sc/
in vitro models necessary tools for validating sc/snRNA-seq data snRNA-seq data, any individual dataset carries error. A neural organoid
generated from nonexperimental species120,121. The experimental model that recapitulates signal from a single in vivo dataset may in fact
tractability of in vitro models makes them an attractive model for the be a poor general model if the reference signal is of low quality and fails
functional characterization of transcriptomics states and changes to replicate. Therefore, to avoid overfitting, it is useful to incorporate
through perturbation experiments (Fig. 2). Organoids121 are experi- cross-validation of in vivo signal among in vivo datasets140. Quantify-
mentally tractable systems to model the cell-type heterogeneity121–127 ing DEG statistics across cell types and collating the P values and fold
and spatial organization128,129 of neural tissues. A striking example of changes of genes derived from individual in vivo datasets establishes
functional investigations made possible include the generation of a benchmark of reference signal for interpreting organoid differential
human neural organoids with or without a single amino acid change expression statistics.
found in Neanderthals, enabling study of the neurobiological con- For cross-species comparisons, sc/snRNA-seq dissection of
sequences of genetic variation found in an extinct species130,131. neural organoids can resolve key developmental differences across
Single-cell or single-nucleus dissection of neural organoids can species, such as molecular mechanisms underlying neural progeni-
provide cell-type resolution of developmental trajectories132,133, tor variation across human and primate organoids126,132. However,
enable perturbation of dynamic gene regulatory networks134, model observations are robust only to the class of variability sampled, and
neurological disease mechanisms135–137 and support neurodevelop- assessments should be applied to diverse genetic backgrounds (cell
mental cross-species comparisons126,132,138. The inaccessibility of lines) or differentiation protocols to identify signals that are not
non-postmortem human neural tissues positions organoids and/or specific to an individual cell line or protocol. As examples, different
monoculture systems (often) as the only available option for func- organoid protocols aiming to derive similar neural lineages (cortical
tional investigations, demanding that the limitations of these models organoids) have reported that biases in differentiation patterns139
be acutely understood. and cell-line-specific effects141,142 can obscure disease phenotypes
sc/snRNA-seq comparisons demonstrate the capacity of neu- in organoid models. Sampling increased genetic and/or technical
ral organoids to model broad in vivo neural cell types across numer- variability increases the likelihood of replicable signal and may buffer
ous genomic modalities122–124,133,139. However, these single-cell data against overfitting.

Nature Neuroscience | Volume 27 | December 2024 | 2310–2325 2319

Review article https://doi.org/10.1038/s41593-024-01814-0

Species 1 Species 2 Potential pitfalls

Cell or nucleus
Droplet
Bead
• Doublet clusters can be misinterpreted as
Doublets
real and species specific

Native mRNA

• Empty droplets can be misinterpreted as real

Ambient RNA
• Ambient RNA contamination imbalance can
contamination
appear as differential expression

Ambient mRNA

Comparable • Regional differences can be misinterpreted as

brain regions species-specific differences

Gene expression
Gene expression

Species 1
Comparable • Demographics-related differences can be
demographics misinterpreted as species-specific
Species 2 differences

Age Species 1 Species 2

Fig. 4 | Illustration of technical and biological artifacts. Schematics of how could also occur in comparison of other variables: samples of different ages, at
evolutionary comparisons using single-cell genomics could be vulnerable to different stages of disease, across different CNS regions or following different
misinterpretations due to either biological (for example, brain region selection/ treatment paradigms. As always, analysis should be considered in the context of
dissection or demographics such as age) and/or technical (for example, doublet the underlying biology being interrogated.
or ambient RNA) artifacts. It should be noted that these comparisons and errors

Technical limitations of transcriptomic-based homology The challenges in using sc/snRNA-seq approaches to study cell
inference types across species are multiplicative. Even with reliable in vivo data,
While we have attempted to provide conceptual criteria above, there spatiotemporal context and biological variation must be considered
are computational challenges to defining cell-type homologies because when modeling homology. In vitro studies have the same challenges
there are no formal or uniformly accepted criteria. Multiple methods amplified: cell-type distributions are untethered to the anatomy that
predict homologous cell types107,143–148, but integrating across species is reproducibly generated in vivo, with the added concern that the
can be difficult because identification of homologous cell types often observed cell states approximate those seen in vivo, heavily layered
relies on heuristics such as shared nearest neighbors and nonlinear with various sources of technical variation. Despite these challenges,
data transformations, rather than formal models of gene expression existing data and tools wielded with perspicacious judgment have ena-
divergence and cell-type evolution149. As such, inclusion or exclusion bled the discovery of new cell types and shared features and principles
of cell types within a given dataset can alter which cell types appear of vertebrate brain development and function.
homologous. For example, a putative primate-specific cell type thought
to be most similar to other striatal interneurons101 was determined to Technical and biological artifacts
actually be more similar to diencephalic neurons when such cell types As evolutionary findings may be challenging to experimentally vali-
were further included in the analysis150. This issue is a potential caveat date, it is important to consider experimental factors that could lead to
for any type of comparison, whether it is between species, regions or erroneous interpretations. Some of these are pertinent to evolutionary
developmental time periods—and is particularly important for the comparisons, but most are generalizable to other types of comparison
deployment of functional orthogonal validation experiments. Com- (and can be mitigated by careful orthogonal validation). Recent stud-
positional concerns are especially pressing in the context of in vitro ies have shown that technical artifacts such as doublets and ambient
studies, in which different iPS cell lines respond divergently to pattern- RNA contamination can lead to misinterpretations151. This issue is
ing factors and generate cultures with variable compositions. It is also exacerbated when datasets are compared without properly adjust-
important to consider that conserved populations can be repurposed ing for sequencing artifacts. For example, if datasets for one species
to different brain structures over development. Recent work highlights contained more artifacts (for example, higher doublet rate, greater
that classes of inhibitory neurons that migrate to rodent olfactory bulb ambient RNA contamination), the result could be misinterpreted as
have been redirected to the expanded primate white matter98, and a species-specific effect (Fig. 4). Equally, it is crucial to obtain demo-
a mammalian conserved interneuron type is most numerous in the graphically and spatiotemporally similar brain tissues from all species
mouse hippocampus but more abundant in the primate neocortex101. for a proper evolutionary comparison (or samples for within-species

Nature Neuroscience | Volume 27 | December 2024 | 2310–2325 2320

Review article https://doi.org/10.1038/s41593-024-01814-0

comparisons). If regional boundaries are not rigorously considered 13. Sokol, L. et al. Prioritization and functional validation of target
during dissection, it is possible to compare improperly matched brain genes from single-cell transcriptomics studies. Commun. Biol. 6,
regions, which can lead to misclassification of region-specific cellular 648 (2023).
and molecular features as species- or sample-specific results. Although 14. Lahnemann, D. et al. Eleven grand challenges in single-cell data
spatial transcriptomics may alleviate this problem for species with small science. Genome Biol. 21, 31 (2020).
brain sizes. In addition to brain regions, developmental time points 15. Hao, Y. et al. Integrated analysis of multimodal single-cell data.
should be matched to prevent misinterpreting age-specific effects as Cell 184, 3573–3587 (2021).
species-specific effects (Fig. 4). However, matching developmental 16. Ding, J. et al. Systematic comparison of single-cell and
time points in distantly related species might be impossible, and het- single-nucleus RNA-sequencing methods. Nat. Biotechnol. 38,
erochrony should also be considered as a mechanism for evolutionary 737–746 (2020).
change. Finally, it is important to consider that age matching often 17. Zeng, H. & Sanes, J. R. Neuronal cell-type classification:
depends on an estimate based on life history traits, and some cell types challenges, opportunities and the path forward. Nat. Rev.
may be more sensitive to the effects of age than others (for example, Neurosci. 18, 530–546 (2017).
glia change more than neurons in very old age152). Thus, interpretation 18. Scala, F. et al. Phenotypic variation of transcriptomic cell types in
should consider the age bracket of the samples. mouse motor cortex. Nature 598, 144–150 (2021).
This paper applies Patch-seq to explore the connection
Conclusion between single-cell gene expression, morphology and
We have highlighted the importance of validating sc/snRNA-seq electrophysiological properties in neurons of the mouse motor
data using in-depth data analysis, functional characterization, cortex, illustrating the complex relationship between a cell’s
cross-validation, multi-omics integration and follow-up validation transcriptional subtype and its functional properties.
experiments. We also emphasize the need for specific practices to 19. Clark, I. C. et al. Identification of astrocyte regulators by nucleic
handle confounds in cross-system analyses, such as sampling broadly acid cytometry. Nature 614, 326–333 (2023).
within each system, measuring variance, assessing similarity without 20. Jung, N. & Kim, T. K. Spatial transcriptomics in neuroscience.
merging and reporting robustness with effect sizes (see also refs. 1,2). Exp. Mol. Med 55, 2105–2115 (2023).
These practices can help avoid overfitting and bias, provide meaning- 21. Fangma, Y., Liu, M., Liao, J., Chen, Z. & Zheng, Y. Dissecting the
ful cross-system assessments, and reveal the molecular mechanisms brain with spatially resolved multi-omics. J. Pharm. Anal. 13,
of brain evolution, disease responses and adaptive phenotypes. By 694–710 (2023).
applying sc/snRNA-seq approaches with careful consideration of 22. Hartman, A. & Satija, R. Comparative analysis of multiplexed
their inherent challenges and limitations, researchers can advance in situ gene expression profiling technologies. eLife 13, RP96949
our understanding of cellular heterogeneity and evolution across dif- (2024).
ferent biological systems. 23. Rodriques, S. G. et al. Slide-seq: a scalable technology for
measuring genome-wide expression at high spatial resolution.
References Science 363, 1463–1467 (2019).
1. Adameyko, I. et al. Applying single-cell/nucleus genomics to 24. Russell, A. J. C. et al. Slide-tags enables single-nucleus barcoding
studies of cellular heterogeneity and cell fate transitions in the for multimodal spatial genomics. Nature 625, 101–109 (2024).
nervous system. Nat. Neurosci. https://doi.org/10.1038/s41593- This paper describes a new approach for spatial single-nucleus
024-01827-9 (2024). genomic profiling in frozen tissue sections using spatial
2. Bonev, B. et al. Opportunities and challenges of single-cell and barcoding, which can be applied to multiple modalities
spatially resolved genomics methods for neuroscience discovery. including snRNA-seq, ATAC-seq and T cell antigen receptor
Nat. Neurosci. https://doi.org/10.1038/s41593-024-01806-0 (2024). sequencing. This tool can be applied for multimodal, orthogonal
3. Luecken, M. D. & Theis, F. J. Current best practices in single-cell discovery and validation experiments.
RNA-seq analysis: a tutorial. Mol. Syst. Biol. 15, e8746 (2019). 25. Chen, K. H., Boettiger, A. N., Moffitt, J. R., Wang, S. & Zhuang, X.
4. Trapnell, C. Defining cell types and states with single-cell RNA imaging. Spatially resolved, highly multiplexed RNA profiling
genomics. Genome Res. 25, 1491–1498 (2015). in single cells. Science 348, aaa6090 (2015).
5. Kiselev, V. Y. et al. SC3: consensus clustering of single-cell 26. Wang, X. et al. Three-dimensional intact-tissue sequencing of
RNA-seq data. Nat. Methods 14, 483–486 (2017). single-cell transcriptional states. Science 361, eaat5691 (2018).
6. Hao, Y. et al. Dictionary learning for integrative, multimodal and 27. Ke, R. et al. In situ sequencing for RNA analysis in preserved tissue
scalable single-cell analysis. Nat. Biotechnol. 42, 293–304 (2024). and cells. Nat. Methods 10, 857–860 (2013).
7. Nayak, R. & Hasija, Y. A hitchhiker’s guide to single-cell 28. Liu, Y., Beyer, A. & Aebersold, R. On the dependency of cellular
transcriptomics and data analysis pipelines. Genomics 113, protein levels on mRNA abundance. Cell 165, 535–550 (2016).
606–619 (2021). 29. Wegler, C. et al. Global variability analysis of mRNA and protein
8. Satija, R., Farrell, J. A., Gennert, D., Schier, A. F. & Regev, A. concentrations across and within human tissues. NAR Genom.
Spatial reconstruction of single-cell gene expression data. Bioinform 2, lqz010 (2020).
Nat. Biotechnol. 33, 495–502 (2015). 30. Buccitelli, C. & Selbach, M. mRNAs, proteins and the emerging
9. Dominguez Conde, C. et al. Cross-tissue immune cell analysis principles of gene expression control. Nat. Rev. Genet. 21,
reveals tissue-specific features in humans. Science 376, eabl5197 630–644 (2020).
(2022). 31. Almet, A. A., Cang, Z., Jin, S. & Nie, Q. The landscape of cell-cell
10. Xu, C. et al. Automatic cell-type harmonization and integration communication through single-cell transcriptomics. Curr. Opin.
across Human Cell Atlas datasets. Cell 186, 5876–5891 (2023). Syst. Biol. 26, 12–23 (2021).
11. Menon, V. Clustering single cells: a review of approaches on 32. Dimitrov, D. et al. Comparison of methods and resources for
high-and low-depth single-cell RNA-seq data. Brief. Funct. cell-cell communication inference from single-cell RNA-seq data.
Genomics 17, 240–245 (2018). Nat. Commun. 13, 3224 (2022).
12. Pancheva, A., Wheadon, H., Rogers, S. & Otto, T. D. Using topic 33. Armingol, E., Officer, A., Harismendy, O. & Lewis, N. E.
modeling to detect cellular crosstalk in scRNA-seq. PLoS Comput. Deciphering cell-cell interactions and communication from gene
Biol. 18, e1009975 (2022). expression. Nat. Rev. Genet. 22, 71–88 (2021).

Nature Neuroscience | Volume 27 | December 2024 | 2310–2325 2321

Review article https://doi.org/10.1038/s41593-024-01814-0

34. Gault, J. et al. Combining native and ‘omics’ mass spectrometry 55. Hasel, P., Rose, I. V. L., Sadick, J. S., Kim, R. D. & Liddelow, S. A.
to identify endogenous ligands bound to membrane proteins. Neuroinflammatory astrocyte subtypes in the mouse brain.
Nat. Methods 17, 505–508 (2020). Nat. Neurosci. 24, 1475–1487 (2021).
35. Wang, J. Y. & Doudna, J. A. CRISPR technology: a decade of 56. Buettner, F. et al. Computational analysis of cell-to-cell
genome editing is only the beginning. Science 379, eadd8643 heterogeneity in single-cell RNA-sequencing data reveals hidden
(2023). subpopulations of cells. Nat. Biotechnol. 33, 155–160 (2015).
36. Gilbert, L. A. et al. Genome-scale CRISPR-mediated control of 57. Velten, L. et al. Human haematopoietic stem cell lineage
gene repression and activation. Cell 159, 647–661 (2014). commitment is a continuous process. Nat. Cell Biol. 19, 271–281
37. Kearns, N. A. et al. Functional annotation of native enhancers with (2017).
a Cas9-histone demethylase fusion. Nat. Methods 12, 401–403 58. Crow, M., Paul, A., Ballouz, S., Huang, Z. J. & Gillis, J.
(2015). Characterizing the replicability of cell types defined by single cell
38. Thakore, P. I. et al. Highly specific epigenome editing by RNA-sequencing data using MetaNeighbor. Nat. Commun. 9, 884
CRISPR-Cas9 repressors for silencing of distal regulatory (2018).
elements. Nat. Methods 12, 1143–1149 (2015). 59. Brunner, G., Lang, K., Wolfe, R. A., McClure, D. B. & Sato,
39. Adamson, B. et al. A multiplexed single-cell CRISPR screening G. H. Selective cell culture of brain cells by serum-free,
platform enables systematic dissection of the unfolded protein hormone-supplemented media: a comparative morphological
response. Cell 167, 1867–1882 (2016). study. Brain Res. 254, 563–575 (1981).
40. Dixit, A. et al. Perturb-seq: dissecting molecular circuits with 60. Foo, L. C. et al. Development of a method for the purification and
scalable single-cell RNA profiling of pooled genetic screens. culture of rodent astrocytes. Neuron 71, 799–811 (2011).
Cell 167, 1853–1866 (2016). 61. Bohlen, C. J., Bennett, F. C. & Bennett, M. L. Isolation and culture
41. Datlinger, P. et al. Pooled CRISPR screening with single-cell of microglia. Curr. Protoc. Immunol. 125, e70 (2019).
transcriptome readout. Nat. Methods 14, 297–301 (2017). 62. Guttenplan, K. A. & Liddelow, S. A. Astrocytes and microglia:
42. Wheeler, M. A. et al. Droplet-based forward genetic screening of models and tools. J. Exp. Med. 216, 71–83 (2019).
astrocyte-microglia cross-talk. Science 379, 1023–1030 (2023). 63. Hall, C. E. et al. Progressive motor neuron pathology and the role
43. Wu, D. et al. Dual genome-wide coding and lncRNA screens in of astrocytes in a human stem cell model of VCP-related ALS. Cell
neural induction of induced pluripotent stem cells. Cell Genom. Rep. 19, 1739–1749 (2017).
2, 100177 (2022). 64. de Majo, M. et al. Granulin loss of function in human mature brain
44. Cooper, Y. A. et al. Functional regulatory variants implicate organoids implicates astrocytes in TDP-43 pathology. Stem Cell
distinct transcriptional networks in dementia. Science 377, Rep. 18, 706–719 (2023).
eabi8654 (2022). 65. Cuomo, A. S. E. et al. Single-cell RNA-sequencing of
45. Yang, X. et al. Functional characterization of Alzheimer’s disease differentiating iPS cells reveals dynamic genetic effects on gene
genetic variants in microglia. Nat. Genet. 55, 1735–1744 (2023). expression. Nat. Commun. 11, 810 (2020).
46. Tian, R. et al. CRISPR interference-based platform for multimodal 66. Virdi, G. S. et al. Protein aggregation and calcium dysregulation
genetic screens in human iPSC-derived neurons. Neuron 104, are hallmarks of familial Parkinson’s disease in midbrain
239–255 (2019). dopaminergic neurons. NPJ Parkinsons Dis. 8, 162 (2022).
47. Drager, N. M. et al. A CRISPRi/a platform in human iPSC-derived 67. Szebenyi, K. et al. Human ALS/FTD brain organoid slice cultures
microglia uncovers regulators of disease states. Nat. Neurosci. 25, display distinct early astrocyte and targetable neuronal
1149–1162 (2022). pathology. Nat. Neurosci. 24, 1542–1554 (2021).
This paper uses a multiplex CRISPRi and CRISPRa platform using 68. Liddelow, S. A. et al. Neurotoxic reactive astrocytes are induced
iPS cell microglia to identify genes that alter microglial cell by activated microglia. Nature 541, 481–487 (2017).
states and functions when perturbed. It also uses scRNA-seq and One of the first studies to develop an in vitro functional testing
dataset integration to demonstrate that microglial states in this platform to interrogate and validate a transcriptomically defined
in vitro system replicate states observed in postmortem human reactive astrocyte substate.
brain tissue. 69. Barbar, L. et al. CD49f is a novel marker of functional and reactive
48. Leng, K. et al. CRISPRi screens in human iPSC-derived astrocytes human iPSC-derived astrocytes. Neuron 107, 436–453 (2020).
elucidate regulators of distinct inflammatory reactive states. 70. Clark, I. C. et al. Barcoded viral tracing of single-cell interactions
Nat. Neurosci. 25, 1528–1542 (2022). in central nervous system inflammation. Science 372, abf1230
49. Li, E. et al. CRISPRi-based screens in iAssembloids to elucidate (2021).
neuron-glia interactions. Preprint at bioRxiv https://doi.org/ 71. Pasqual, G. et al. Monitoring T cell-dendritic cell interactions in vivo
10.1101/2023.04.26.538498 (2023). by intercellular enzymatic labelling. Nature 553, 496–500 (2018).
50. Esk, C. et al. A human tissue screen identifies a regulator of ER 72. Gerfen, C. R., Paletzki, R. & Heintz, N. GENSAT BAC
secretion as a brain-size determinant. Science 370, 935–941 cre-recombinase driver lines to study the functional organization
(2020). of cerebral cortical and basal ganglia circuits. Neuron 80,
51. Jin, X. et al. In vivo Perturb-seq reveals neuronal and glial 1368–1383 (2013).
abnormalities associated with autism risk genes. Science 370, 73. Gong, S. et al. A gene expression atlas of the central nervous
eaaz6063 (2020). system based on bacterial artificial chromosomes. Nature 425,
52. Wertz, M. H. et al. Genome-wide in vivo CNS screening identifies 917–925 (2003).
genes that modify CNS neuronal survival and mHTT toxicity. 74. Gertler, T. S., Chan, C. S. & Surmeier, D. J. Dichotomous
Neuron 106, 76–89 (2020). anatomical properties of adult striatal medium spiny neurons.
53. Ramani, B. et al. Scalable, cell type-selective, AAV-based in vivo J. Neurosci. 28, 10814–10824 (2008).
CRISPR screening in the mouse brain. Preprint at bioRxiv 75. Fenno, L. E. et al. Targeting cells with single vectors using
https://doi.org/10.1101/2023.06.13.544831 (2023). multiple-feature Boolean logic. Nat. Methods 11, 763–772 (2014).
54. Wheeler, M. A. et al. Environmental control of astrocyte 76. Kramer, D. J. et al. Generation of a DAT-P2A-Flpo mouse line
pathogenic activities in CNS inflammation. Cell 176, 581–596 for intersectional genetic targeting of dopamine neuron
(2019). subpopulations. Cell Rep. 35, 109123 (2021).

Nature Neuroscience | Volume 27 | December 2024 | 2310–2325 2322

Review article https://doi.org/10.1038/s41593-024-01814-0

77. Poulin, J. F. et al. Mapping projections of molecularly defined 99. Woych, J. et al. Cell-type profiling in salamanders identifies
dopamine neuron subtypes using intersectional genetic innovations in vertebrate forebrain evolution. Science 377,
approaches. Nat. Neurosci. 21, 1260–1271 (2018). eabp9186 (2022).
78. Ren, J. et al. Single-cell transcriptomes and whole-brain 100. Hahn, J. et al. Evolution of neuronal cell classes and types in the
projections of serotonin neurons in the mouse dorsal and median vertebrate retina. Nature 624, 415–424 (2023).
raphe nuclei. Elife 8, e49424 (2019). 101. Krienen, F. M. et al. Innovations present in the primate interneuron
79. Okaty, B. W. et al. A single-cell transcriptomic and anatomic atlas repertoire. Nature 586, 262–269 (2020).
of mouse dorsal raphe Pet1 neurons. Elife 9, e55523 (2020). 102. Tosches, M. A. et al. Evolution of pallium, hippocampus, and
80. BRAIN Initiative Cell Census Network. A multimodal cell census cortical cell types revealed by single-cell transcriptomics in
and atlas of the mammalian primary motor cortex. Nature 598, reptiles. Science 360, 881–888 (2018).
86–102 (2021). 103. Peng, Y. R. et al. Molecular classification and comparative
81. Cadwell, C. R. et al. Electrophysiological, transcriptomic and taxonomics of foveal and peripheral cells in primate retina.
morphologic profiling of single neurons using Patch-seq. Cell 176, 1222–1237 (2019).
Nat. Biotechnol. 34, 199–203 (2016). 104. Jorstad, N. L. et al. Comparative transcriptomics reveals
A technical improvement is provided with the advent of human-specific cortical features. Science 382, eade9516 (2023).
PATCH-seq—enabling the electrophysiological recording and 105. Hain, D. et al. Molecular diversity and evolution of neuron types in
transcriptomic measurement in individual neurons. the amniote brain. Science 377, eabp8202 (2022).
82. Fuzik, J. et al. Integration of electrophysiological recordings This paper shows that generating single-cell genomics brain
with single-cell RNA-seq data identifies neuronal subtypes. atlases from non-model organisms (that is, an exemplar lizard)
Nat. Biotechnol. 34, 175–183 (2016). and comparing to mammals can identify conserved neuronal
83. Sloan, S. A. et al. Human astrocyte maturation captured in 3D cell types throughout the brains of both species across hundreds
cerebral cortical spheroids derived from pluripotent stem cells. of million years of evolution while also showing how evolution
Neuron 95, 779–790 (2017). has modified this conservation based on connectivity needs.
84. Xu, S. et al. Behavioral state coding by molecularly defined 106. Lust, K. et al. Single-cell analyses of axolotl telencephalon
paraventricular hypothalamic cell type ensembles. Science 370, organization, neurogenesis, and regeneration. Science 377,
eabb2494 (2020). eabp9262 (2022).
85. Condylis, C. et al. Dense functional and molecular readout of a 107. Tarashansky, A. J. et al. Mapping single-cell atlases throughout
circuit hub in sensory cortex. Science 375, eabl5981 (2022). metazoa unravels cell type evolution. Elife 10, e66747 (2021).
86. Bugeon, S. et al. A transcriptomic axis predicts state modulation 108. Chen, J. et al. A quantitative framework for characterizing the
of cortical interneurons. Nature 607, 330–338 (2022). evolutionary history of mammalian gene expression. Genome
87. Guttenplan, K. A. et al. Knockout of reactive astrocyte activating Res. 29, 53–63 (2019).
factors slows disease progression in an ALS mouse model. 109. Bakken, T. E. et al. Comparative cellular analysis of motor cortex in
Nat. Commun. 11, 3753 (2020). human, marmoset and mouse. Nature 598, 111–119 (2021).
88. Guttenplan, K. A. et al. Neurotoxic reactive astrocytes induce cell This paper used multimodal single-cell approaches (snRNA-seq,
death via saturated lipids. Nature https://doi.org/10.1038/s41586- snATAC-seq, single-cell methylation, morphology and
021-03960-y (2021). Patch-seq) across mammalian species to determine the
89. Zeng, H. What is a cell type and how to define it? Cell 185, species-specific adaptiations of cell types in one brain region
2739–2755 (2022). and focuses on differences in Betz cells.
90. Yuste, R. et al. A community-based transcriptomics classification 110. Caglayan, E. et al. Molecular features driving cellular complexity
and nomenclature of neocortical cell types. Nat. Neurosci. 23, of human brain evolution. Nature 620, 145–153 (2023).
1456–1468 (2020). This paper demonstrates the power of comparing proportions
91. Hodge, R. D. et al. Conserved cell types with divergent features in of nonneuronal cell types (that is, the oligodendrocyte lineage)
human versus mouse cortex. Nature 573, 61–68 (2019). across closely related primate species and the integration of
92. Arendt, D., Bertucci, P. Y., Achim, K. & Musser, J. M. Evolution of ancient human genomes to identify modern human-specific
neuronal types and families. Curr. Opin. Neurobiol. 56, 144–152 variants associated with human-specific chromatin
(2019). accessibility.
93. Wagner, D. E. & Klein, A. M. Lineage tracing meets single-cell 111. Jeon, H. et al. Statistical power analysis for designing bulk,
omics: opportunities and challenges. Nat. Rev. Genet. 21, 410–427 single-cell, and spatial transcriptomics experiments: review,
(2020). tutorial, and perspectives. Biomolecules 13, 221 (2023).
94. Bandler, R. C. et al. Single-cell delineation of lineage and genetic 112. Svensson, V. et al. Power analysis of single-cell RNA-sequencing
identity in the mouse brain. Nature 601, 404–409 (2022). experiments. Nat. Methods 14, 381–387 (2017).
95. Cadwell, C. R. et al. Cell type composition and circuit 113. Su, K., Wu, Z. & Wu, H. Simulation, power evaluation and sample
organization of clonally related excitatory neurons in the juvenile size recommendation for single-cell RNA-seq. Bioinformatics 36,
mouse neocortex. Elife 9, e52951 (2020). 4860–4868 (2020).
96. Mayer, C. et al. Developmental diversification of cortical inhibitory 114. Arendt, D. et al. The origin and evolution of cell types. Nat. Rev.
interneurons. Nature 555, 457–462 (2018). Genet. 17, 744–757 (2016).
97. Nowakowski, T. J. et al. Spatiotemporal gene expression 115. Colquitt, B. M., Merullo, D. P., Konopka, G., Roberts, T. F. &
trajectories reveal developmental hierarchies of the human Brainard, M. S. Cellular transcriptomics reveals evolutionary
cortex. Science 358, 1318–1323 (2017). identities of songbird vocal circuits. Science 371, eabd9704
98. Schmitz, M. T. et al. The development and evolution of inhibitory (2021).
neurons in primate cerebrum. Nature 603, 871–877 (2022). 116. Kebschull, J. M. et al. Cerebellar nuclei evolved by repeatedly
This paper uses comparative single-cell genomics in duplicating a conserved cell-type set. Science 370, abd5059
mammalian brains over developmental stages, identifying the (2020).
evolutionary changes in developmental specification of a class 117. Lamanna, F. et al. A lamprey neural cell type atlas illuminates the
of primate-specific interneurons. origins of the vertebrate brain. Nat. Ecol. Evol. 7, 1714–1728 (2023).

Nature Neuroscience | Volume 27 | December 2024 | 2310–2325 2323

Review article https://doi.org/10.1038/s41593-024-01814-0

118. Shafer, M. E. R., Sawh, A. N. & Schier, A. F. Gene family evolution 141. Czerminski, J. T., King, O. D. & Lawrence, J. B. Large-scale
underlies cell-type diversification in the hypothalamus of teleosts. organoid study suggests effects of trisomy 21 on early fetal
Nat. Ecol. Evol. 6, 63–76 (2022). neurodevelopment are more subtle than variability between
119. Tarashansky, A. J., Xue, Y., Li, P., Quake, S. R. & Wang, B. isogenic lines and experiments. Front. Neurosci. 16, 972201
Self-assembling manifolds in single-cell RNA sequencing data. (2022).
Elife 8, e48994 (2019). 142. Lee, J. H. et al. Cell-line dependency in cerebral organoid
120. Agboola, O. S., Hu, X., Shan, Z., Wu, Y. & Lei, L. Brain organoid: induction: cautionary observations in Alzheimer’s disease
a 3D technology for investigating cellular composition and patient-derived induced pluripotent stem cells. Mol. Brain 15, 46
interactions in human neurological development and disease (2022).
models in vitro. Stem Cell Res. Ther. 12, 430 (2021). 143. Butler, A., Hoffman, P., Smibert, P., Papalexi, E. & Satija, R.
121. Lancaster, M. A. et al. Cerebral organoids model human brain Integrating single-cell transcriptomic data across different
development and microcephaly. Nature 501, 373–379 (2013). conditions, technologies, and species. Nat. Biotechnol. 36,
122. Amiri, A. et al. Transcriptome and epigenome landscape of 411–420 (2018).
human cortical development modeled in organoids. Science 362, 144. Haghverdi, L., Lun, A. T. L., Morgan, M. D. & Marioni, J. C. Batch
eaat6720 (2018). effects in single-cell RNA-sequencing data are corrected by
123. Bhaduri, A. et al. Cell stress in cortical organoids impairs matching mutual nearest neighbors. Nat. Biotechnol. 36, 421–427
molecular subtype specification. Nature 578, 142–148 (2020). (2018).
124. Camp, J. G. et al. Human cerebral organoids recapitulate gene 145. Johansen, N. & Quon, G. scAlign: a tool for alignment, integration,
expression programs of fetal neocortex development. Proc. Natl and rare cell identification from scRNA-seq data. Genome Biol.
Acad. Sci. USA 112, 15672–15677 (2015). 20, 166 (2019).
125. Gordon, A. et al. Long-term maturation of human cortical 146. Polanski, K. et al. BBKNN: fast batch alignment of single cell
organoids matches key early postnatal transitions. Nat. Neurosci. transcriptomes. Bioinformatics 36, 964–965 (2020).
24, 331–342 (2021). 147. Stuart, T. et al. Comprehensive integration of single-cell data.
126. Pollen, A. A. et al. Establishing cerebral organoids as models of Cell 177, 1888–1902 (2019).
human-specific brain evolution. Cell 176, 743–756 (2019). 148. Welch, J. D. et al. Single-cell multi-omic integration compares and
127. Velasco, S. et al. Individual brain organoids reproducibly form contrasts features of brain cell identity. Cell 177, 1873–1887 (2019).
cell diversity of the human cerebral cortex. Nature 570, 523–527 149. Mostajo-Radji, M. A., Schmitz, M. T., Montoya, S. T. & Pollen, A. A.
(2019). Reverse engineering human brain evolution using organoid
128. Qian, X. et al. Sliced human cortical organoids for modeling models. Brain Res. 1729, 146582 (2020).
distinct cortical layer formation. Cell Stem Cell 26, 766–781 150. Krienen, F. M. et al. A marmoset brain cell census reveals regional
(2020). specialization of cellular identities. Sci Adv. 9, eadk3986 (2023).
129. Xiang, Y. et al. Fusion of regionally specified hPSC-derived 151. Caglayan, E., Liu, Y. & Konopka, G. Neuronal ambient RNA
organoids models human brain development and interneuron contamination causes misinterpreted and masked cell types in
migration. Cell Stem Cell 21, 383–398 (2017). brain single-nuclei datasets. Neuron 110, 4043–4056 (2022).
130. Pinson, A. et al. Human TKTL1 implies greater neurogenesis in 152. Soreq, L. et al. Major shifts in glial regional identity are a
frontal neocortex of modern humans than Neanderthals. Science transcriptional hallmark of human brain aging. Cell Rep. 18,
377, eabl6422 (2022). 557–570 (2017).
131. Trujillo, C. A. et al. Reintroduction of the archaic variant of NOVA1 153. Norimoto, H. et al. A claustrum in reptiles and its role in slow-wave
in cortical organoids alters neurodevelopment. Science 371, sleep. Nature 578, 413–418 (2020).
eaax2537 (2021). 154. Marsh, S. E. et al. Dissection of artifactual and confounding glial
132. Kanton, S. et al. Organoid single-cell genomic atlas uncovers signatures by single-cell sequencing of mouse and human brain.
human-specific features of brain development. Nature 574, Nat. Neurosci. 25, 306–316 (2022).
418–422 (2019).
133. Uzquiano, A. et al. Proper acquisition of cell class identity in Acknowledgements
organoids allows definition of fate specification programs of the We thank I. Adameyko, T. Bakken, A. Bhaduri, B. Bonev, G.
human cerebral cortex. Cell 185, 3770–3788 (2022). Castelo-Branco, F. Chen, C. Chhatbar, S. Codeluppi, R. Corces, J. Fan,
134. Fleck, J. S. et al. Inferring and perturbing cell fate regulomes in M. G. Filbin, D. Gate, G. Green, M. Heiman, K. Harris, H. Hochgerner,
human brain organoids. Nature https://doi.org/10.1038/s41586- F. Inoue, M. Kellis, C. N. Kim, J. Krull, G. L. Manno, A. Levine, Q. Li,
022-05279-8 (2022). S. Linnarsson, M. Lotfollahi, C. Luo, Q. Ma, E. Macosko, C. Mayer,
135. Bershteyn, M. et al. Human iPSC-derived cerebral organoids K. R. Maynard, V. Menon, P. Nano, M. Nitzan, M. Prinz, S. Quake, V.
model cellular features of lissencephaly and reveal prolonged Ramani, R. Satijia, L. Schirmer, Y. Shen, N. Sun, F. Theis, C. Walsh,
mitosis of outer radial glia. Cell Stem Cell 20, 435–449 (2017). X. Wang, J. D. Welch and J. Yang for the insightful feedback on this
136. Trevino, A. E. et al. Chromatin accessibility dynamics in a model work. M.C. thanks A. U. Antonova for helpful additional discussions.
of human forebrain development. Science 367, eaay1645 (2020). G.K. is a Jon Heighten Scholar in Autism Research and Townsend
137. Li, C. et al. Single-cell brain organoid screening identifies Distinguished Chair in Research on Autism Spectrum Disorders at UT
developmental defects in autism. Nature 621, 373–380 (2023). Southwestern, and was partially supported by the James S. McDonnell
138. Mora-Bermudez, F. et al. Differences and similarities between Foundation 21st Century Science Initiative in Understanding Human
human and chimpanzee neural progenitors during cerebral Cognition Scholar Award (220020467), Simons Foundation for
cortex development. Elife 5, e18683 (2016). Autism Research Award (947591) and National Institutes of Health
139. Tanaka, Y., Cakir, B., Xiang, Y., Sullivan, G. J. & Park, I. H. Synthetic (NIH) grants (NS126143, HG011641, MH126481, NS115821). S.A.L. is
analyses of single-cell transcriptomes from multiple brain supported by the Carol and Gene Ludwig Family Foundation, The
organoids and fetal brain. Cell Rep. 30, 1682–1689 (2020). Cure Alzheimer’s Fund, the National Multiple Sclerosis Society and
140. Werner, J. M. & Gillis, J. Preservation of co-expression defines NIH grants (R01EY033353, R03NS127079). C.R.C. is supported by
the primary tissue fidelity of human neural organoids. Preprint at the Weill Neurohub, the Shurl and Kay Curci Foundation, the CURE
bioRxiv https://doi.org/10.1101/2023.03.31.535112 (2023). Epilepsy Taking Flight Award and grants from the NIH (K08NS126573,

Nature Neuroscience | Volume 27 | December 2024 | 2310–2325 2324

Review article https://doi.org/10.1038/s41593-024-01814-0

U01NS132353). E.C. is a Neural Scientist Training Program Fellow in vivo screening methods. All other authors declare no competing
in the Peter O’Donnell Brain Institute at UT Southwestern. J.L.C. is interests.
supported by NIH grant U01MH10907. J.G. is supported by NIH grant
R01MH113005. M.S. is supported by a HHMI Hanna H. Gray Fellowship, Additional information
a SFARI Pilot Award and a NYSCF Druckenmiller Fellowship. J.W. is Correspondence and requests for materials should be addressed
supported by NIH grant R01MH113005. M.E.U. is supported by the to Marco Colonna, Genevieve Konopka, Shane A. Liddelow, Tomasz
National Science Foundation Graduate Research Fellowship Program Nowakowski, Omer A. Bayraktar, Ozgun Gokce or Naomi Habib.
and a Hertz Foundation Fellowship Program.
Peer review information Nature Neuroscience thanks Renzo Mancuso
Author contributions and the other, anonymous, reviewer(s) for their contribution to the
All authors contributed to the writing of this Review. peer review of this work.

Competing interests Reprints and permissions information is available at

M.C. is a member of the Scientific Advisory Board of Vigil, NGMBio, www.nature.com/reprints.
Cartesian and Halyardtx. M.C. receives research support from Ono
Pharmaceutical, is a consultant for CST and has patents pending on Publisher’s note Springer Nature remains neutral with regard to
LILRB4 and TREM2. S.A.L. maintains a financial interest in AstronauTx jurisdictional claims in published maps and institutional affiliations.
and Synapticure and is on the Scientific Advisory Board of the Global
BioAccess Fund. S.A.L. is an inventor on US Patents WO2018081250A1 Springer Nature or its licensor (e.g. a society or other partner) holds
and WO2022187517A1. M.K. is a co-scientific founder of Montara exclusive rights to this article under a publishing agreement with
Therapeutics and serves on the Scientific Advisory Boards of Engine the author(s) or other rightsholder(s); author self-archiving of the
Biosciences, Casma Therapeutics, Cajal Neuroscience, Alector and accepted manuscript version of this article is solely governed by the
Montara Therapeutics, and is an advisor to Modulo Bio and Recursion terms of such publishing agreement and applicable law.
Therapeutics. M.K. is an inventor on US Patent 11,254,933 related to
CRISPRi and CRISPRa screening, and on a US Patent Application on © Springer Nature America, Inc. 2024

1
Department of Pathology and Immunology, Washington University School of Medicine in St. Louis, St. Louis, MO, USA. 2Department of Neuroscience,
Peter O’Donnell Jr. Brain Institute, UT Southwestern Medical Center, Dallas, TX, USA. 3Neuroscience Institute, NYU Grossman School of Medicine, New
York, NY, USA. 4Department of Neuroscience & Physiology, NYU Grossman School of Medicine, New York, NY, USA. 5Department of Ophthalmology,
NYU Grossman School of Medicine, New York, NY, USA. 6Parekh Center for Interdisciplinary Neurology, NYU Grossman School of Medicine, New
York, NY, USA. 7Department of Neurological Surgery, University of California, San Francisco, San Francisco, CA, USA. 8Eli and Edythe Broad Center for
Regeneration Medicine and Stem Cell Research, University of California, San Francisco, San Francisco, CA, USA. 9Weill Institute for Neurosciences,
University of California, San Francisco, San Francisco, CA, USA. 10Department of Anatomy, University of California, San Francisco, San Francisco, CA, USA.
11
Department of Microbiology and Immunology, Northwestern University, Chicago, IL, USA. 12Department of Molecular and Cellular Biology, University of
California, Berkeley, Berkeley, CA, USA. 13Department of Neuroscience, University of California, Berkeley, Berkeley, CA, USA. 14Chan Zuckerberg Biohub,
San Francisco, CA, USA. 15Department of Pathology, University of California, San Francisco, San Francisco, CA, USA. 16Kavli Institute for Fundamental
Neuroscience, University of California, San Francisco, San Francisco, CA, USA. 17Department of Biomedical Engineering, Boston University, Boston,
MA, USA. 18Center for Neurophotonics, Boston University, Boston, MA, USA. 19Department of Biology, Boston University, Boston, MA, USA. 20Center for
Systems Neuroscience, Boston University, Boston, MA, USA. 21Department of Physiology and Donnelly Centre for Cellular and Biomolecular Research,
University of Toronto, Toronto, Ontario, Canada. 22Cold Spring Harbor Laboratory, Cold Spring Harbor, NY, USA. 23Institute for Neurodegenerative
Diseases, University of California, San Francisco, San Francisco, CA, USA. 24Department of Biochemistry and Biophysics, University of California, San
Francisco, San Francisco, CA, USA. 25Princeton Neuroscience Institute, Princeton University, Princeton, NJ, USA. 26F.M. Kirby Neurobiology Center, Boston
Children’s Hospital, Boston, MA, USA. 27Harvard Medical School, Boston, MA, USA. 28Stanley Center for Psychiatric Research, Broad Institute of MIT
and Harvard, Cambridge, MA, USA. 29Department of Neurology and Neurological Sciences, Stanford University, Stanford, CA, USA. 30Howard Hughes
Medical Institute, Stanford University, Stanford, CA, USA. 31Department of Neuromuscular Disease, UCL Queen Square Institute of Neurology, London,
UK. 32The Francis Crick Institute, Human Stem Cells and Neurodegeneration Laboratory, London, UK. 33Eli and Edythe Broad Center of Regeneration
Medicine and Stem Cell Research, University of California, San Francisco, San Francisco, CA, USA. 34Department of Neurology, University of California,
San Francisco, San Francisco, CA, USA. 35Ann Romney Center for Neurologic Diseases, Brigham and Women’s Hospital, Harvard Medical School, Boston,
MA, USA. 36Broad Institute of MIT and Harvard, Cambridge, MA, USA. 37Department of Genetics and Genome Sciences, Case Western Reserve University
School of Medicine, Cleveland, Ohio, OH, USA. 38Institute for Glial Sciences, Case Western Reserve University School of Medicine, Cleveland, OH, USA.
39
Department of Human Genetics, Emory University School of Medicine, Atlanta, GA, USA. 40Department of Biological Sciences, Columbia University,
New York, NY, USA. 41Biomedical Sciences Graduate Program, University of California, San Francisco, San Francisco, CA, USA. 42Wellcome Sanger
Institute, Cambridge, UK. 43Department of Old Age Psychiatry and Cognitive Disorders, University Hospital Bonn, Bonn, Germany. 44German Center for
Neurodegenerative Diseases (DZNE), Bonn, Germany. 45Edmond & Lily Safra Center for Brain Sciences, The Hebrew University of Jerusalem,
Jerusalem, Israel. 46Present address: Institute for Glial Sciences, Case Western Reserve University School of Medicine, Cleveland, OH, USA.
e-mail: [email protected]; [email protected]; [email protected]; [email protected];
[email protected]; [email protected]; [email protected]

Nature Neuroscience | Volume 27 | December 2024 | 2310–2325 2325