Sensory Neurons by Their Genes: the Atlas Era

One mouse data set split the sensory neurons of the dorsal root ganglion into eleven kinds. A larger one, three years later, found seventeen. A 2024 merger of 31 data sets from six species settled on eighteen. None of the three counts is wrong, and none is final. This lesson asks what an atlas count measures, and what more it takes before a cluster of transcriptomes deserves to be called a cell type.

Touch and pain sorted afferents into Aβ, Aδ and C fibres, and sensory coding weighed labelled lines against population codes. Sequencing and genomes and recording many neurons explain how single-cell sequencing works and what a brain atlas is. Here the question is narrower: when is a cluster a type, and when is it a state, a resolution, or an artefact of how the cells were prepared?

Eleven, seventeen, eighteen: why the number moves

The dorsal root ganglion (DRG) holds the cell bodies of the neurons that innervate skin, muscle and viscera. Usoskin, Ernfors and colleagues (2015) dissociated lumbar DRGs from mice, sequenced the RNA of 622 single neurons, and let an unsupervised method group them. They found eleven classes: five neurofilament-rich classes (NF1 to NF5) with myelinated profiles, three non-peptidergic (NP1 to NP3), two peptidergic (PEP1, PEP2), and one marked by tyrosine hydroxylase (TH). Some classes suggested a job. NP3, for example, was linked by the authors to itch. Expression is a hypothesis about function, not proof of it; itch takes it further.

Zeisel and colleagues (2018) then profiled roughly half a million cells across the mouse nervous system and arranged them in a hierarchical taxonomy. Within peripheral sensory neurons they resolved eight peptidergic, six non-peptidergic and three neurofilament types: seventeen. The half million is the whole census, not a count of DRG neurons.

Think of a map. A country-level map and a county-level map of the same land are both correct; they differ in resolution. Broad classes (peptidergic, non-peptidergic, neurofilament) and fine subtypes sit at different levels of one tree. More cells, deeper sequencing or a finer clustering setting will split branches; a different preparation can also hide or reveal one.

Bhuiyan, Renthal and colleagues (2024) integrated 31 DRG and trigeminal ganglion data sets from six mammalian species (human, mouse, rat, guinea pig and two macaque species) into a single reference with eighteen DRG and fifteen trigeminal neuronal subtypes. The authors stressed that preparation, sequencing depth and analysis change how many subtypes are resolved. One of their eighteen groups is defined by injury-induced transcription factors, the Atf3 group, which is a warning that an atlas can mix types with states. They also cautioned that one molecular subtype can behave differently at different innervation targets.

Keep the fibre labels apart from the clusters. Among low-threshold mechanoreceptors (LTMRs, neurons activated by innocuous touch), C, Aδ and Aβ mean unmyelinated, lightly myelinated and heavily myelinated axons (Li and colleagues, 2011). They describe axons, not single transcriptomic clusters. And C does not mean nociceptor: in adult mouse, TH-positive neurons were identified as C-fibre LTMRs, with a mean conduction velocity of 0.58 m/s in ex vivo skin-nerve recordings from five neurons (the LTMR lesson builds the full class table).

From a cluster to a driver line

A cluster becomes a functional type only when something other than its RNA ties it to a job. The usual route is genetic access. A Cre driver expresses Cre recombinase from a marker gene’s locus, so a reporter marks cells that expressed the marker at some time; CreER is switched on by tamoxifen, so the experimenter chooses the window (tools for development covers the genetics). An intersectional strategy requires two recombinases, Cre and Flp, so that only cells that expressed both markers are labelled.

Every line has two separate quality numbers.

specificity = labelled and marker-positive cells / all labelled cells efficiency = labelled and marker-positive cells / all marker-positive cells

In a made-up validation, a line labels 200 neurons, 180 of which express the marker by in situ hybridization, and the sections hold 300 marker-positive neurons. Specificity = 180 / 200 = 90%. Efficiency = 180 / 300 = 60%. The line is good for imaging the right cells but weak for an ablation: killing 60% of a population and seeing no deficit does not show the population is unneeded.

Labelled cells 200 Of those, marker-positive (overlap) 180 All marker-positive cells in the sections 300

Lesson example: 200, 180, 300 Practice case: 150, 120, 240 Equal case: 200, 200, 200

Specificity = labelled and marker-positive / all labelled

180 / 200 = 90%

Efficiency = labelled and marker-positive / all marker-positive

180 / 300 = 60%

Reading: a line can be good for imaging the right cells and weak for an ablation, because losing part of a population (60% in the lesson’s example) and seeing no deficit does not show the population is unneeded. A CreER label records expression at the time of induction, so the same driver induced at two ages can capture two populations (TrkC-CreER at E12.5 versus P5, Bai et al., 2015).

In this sketch: made-up counts of cells in sections, with in situ hybridization as the reference for “marker-positive”; these are not cells in an animal, and no cluster count and no physiology follows from these two numbers.

Specificity and efficiency share the same numerator, the labelled and marker-positive cells, but divide by different totals, so a line can score high on one and low on the other.

Timing is part of the definition. With TrkC-CreER intersected with a Ret reporter (Ret-f(CFP)), Bai and colleagues (2015) found that 3 mg tamoxifen at E12.5 labelled Merkel-associated Aβ SA1 endings, while 0.1 mg at P5 favoured Aβ field-LTMR circumferential endings. Age and dose both differed, so this is not an age-only comparison, but the same driver gave two populations. A CreER label records expression history, not a fixed identity.

The mouse DRG toolkit of Qi, Iskols, Sharma, Ginty and colleagues (2024) gathered such lines for the main subtypes: Th2A-CreER for C-LTMRs, TrkB-CreER for Aδ-LTMRs, Ret-CreER for Aβ rapidly adapting LTMRs, and new knock-ins such as Trpm8-T2a-FlpO. An intersection of Smr2-Cre with Calca-FlpE restricted labelling to an SMR2- and CGRP-positive population with free endings. The team then imaged calcium in the L4 DRG of anaesthetized young mice while indenting the skin with forces from 1 to 75 mN, heating it or cooling it. The authors reported that several genetically defined groups were not uniform: within a named subtype, members varied in mechanical and thermal tuning. A type, on this evidence, can contain graded physiology.

Read such imaging with its limits. In adult mouse DRG imaged in vivo under urethane, the GCaMP6 signal had a typical rise of about 200 ms and a decay time constant of about 700 ms (Chisholm and colleagues, 2018). Single action potentials were detected in most cases, but spikes closer together than those times summate, so a single transient does not give a spike count and its amplitude tracks firing rate only roughly. The firmest bridge between RNA and function is a measurement in the same cell. Patch-seq records a neuron with a patch electrode, then aspirates its contents for sequencing, so that physiology and gene expression come from the same cell; Cadwell and colleagues (2016) demonstrated it in mouse neocortex, not DRG.

How identity is built, and how injury takes it apart

DRG neurons arise from the neural crest (the neural crest). Sharma, Ginty and colleagues (2020) sequenced mouse DRG neurons at E11.5, E12.5, E15.5, P0, P5 and P28 to 42. Newborn sensory neurons coexpressed transcription factors that later became restricted to particular subtypes. Unspecialized here means unspecialized as to subtype; the cells were already neurons. Identity narrowed over weeks, like a sculpture cut from a shared block.

Does the target help cut? NGF from targets keeps TrkA-expressing neurons alive (Trk and p75). An Ngf knockout alone would therefore lose cells, and any change in the survivors could be selection rather than altered identity. Deleting Bax, which the neurons need to die by apoptosis, removes that confound. At P0, Ngf-null;Bax-null DRGs had fewer transcriptionally distinct populations than Bax-null controls. NGF contributes to subtype specialization; the experiment does not show that NGF alone sets every fate, and it is a mouse developmental result.

Identity can also be taken apart. Renthal and colleagues (2020) injured peripheral nerves in mice, mostly males aged 8 to 12 weeks, and sequenced single DRG nuclei. Injury suppressed the expression that marks each subtype. Over a time course associated with axons regrowing to their targets, identity re-emerged, and Atf3 was required for the full injury reprogramming. “Associated with” is the right verb: the timing fits reinnervation, but timing alone does not prove that reinnervation restores identity, and an apparent recovery could also come from which neurons survived or were sampled. Lineage tracing of injured neurons is what separates those readings, and the study used reporters for this: the share of neurons labelled by an Atf3 injury reporter was unchanged from one week to two months, which argues against selective death, and Mrgprd reporter tracing fit retained identity.

The lesson for atlases is direct. Tissue from an injured, inflamed or diseased ganglion will contain states. An Atf3-rich cluster is a condition that existing neurons enter and can leave, not a new permanent type, a point that matters for chronic pain.

Mouse markers that do not transfer

Human DRG atlases are often built from nuclei or from tissue sections, because donor tissue is often frozen and large human neurons are hard to dissociate intact. Nguyen and colleagues (2021) sequenced 1,837 neuronal nuclei from lumbar DRGs of donors aged 34 to 55. Some human classes had no clear mouse equivalent in their comparisons. One, H9, coexpressed TRPM8 and PIEZO2 transcripts, which suggests a cell sensitive to both cold and touch but does not show it. MRGPRX1, though not strongly expressed, localized selectively to H10. Primate MRGPRX genes have no one-to-one mouse counterparts, so do not rename a mouse Mrgpr population after a human gene. The 2024 atlas linked mouse groups to function using rodent genetics: Mrgprd neurons to nociception, Mrgpra3+Trpv1 coexpressing neurons to itch, and Mrgpra3+Mrgprb4 coexpressing neurons to C-LTMR physiology.

Spatial data keep anatomy but add a trap. Tavares-Ferreira and colleagues (2022) used Visium, an array of barcoded spots 55 µm across, on lumbar DRG sections from eight organ donors, four women and four men, and validated markers by RNAscope in situ hybridization.

Worked problem: what does a spatial atlas count? The authors selected 4,356 barcodes (spots) overlapping single neurons and retained 3,952 after filtering. They curated twelve neuronal clusters, five annotated as C nociceptors.

1. Retention = 3,952 / 4,356 = 0.9073, or 90.73%.
2. Check: 4,356 × 0.9073 = 3,952.2, which rounds to 3,952.
3. C-nociceptor fraction of cluster labels = 5 / 12 = 0.4167, or 41.67%.
4. Is either the fraction of human DRG neurons that are C nociceptors? No. The first is a filtering statistic. The second counts labels and ignores that clusters differ in size. Even per-cluster spot counts would not be cell counts, because a 55 µm spot over a neuron also captures RNA from satellite glia and neighbours.

The same study found TAC1 and CALCA, the genes for substance P and CGRP, broadly expressed across its human nociceptor clusters, so the authors suggest the mouse split into peptidergic and non-peptidergic nociceptors may not carry over as a clean binary. Jung and colleagues (2023) compared mouse, guinea pig, macaque and human DRG with one workflow. Tafa4 was found mainly in mouse C-LTMRs, but human TAFA4 was strongest in NTRK2- and SCN5A-positive Aδ-LTMRs, confirmed by in situ hybridization. A mouse mechanism tied to a marker needs a human check of which cells carry that marker. Nociceptors and sodium channels meets the same problem with drug targets.

So what is a type? A working answer: a molecular group that recurs across data sets and is anchored by at least one independent measure (ending, physiology, target or behaviour after manipulation), read with its state and its resolution in mind. Graded physiology can remain inside it.

Design problem

Nguyen and colleagues (2021) reported human DRG classes with no clear mouse equivalent, using single-nucleus sequencing. The 2024 harmonized integration of 31 data sets found no human-specific neuronal clusters. Design the work that would decide whether one proposed human-only class is real.

One way to do it

Start with the known confound. Nguyen and colleagues noted that whole-cell sequencing of mouse DRG had found a Trpa1-rich group that a compared mouse nuclear data set missed, so a mismatch between species may partly reflect what each method captures.

1. Match the preparation. Compare human nuclei with mouse nuclei (and, where possible, human whole cells with mouse whole cells), at similar sequencing depth and with the same clustering pipeline and resolution. If the class finds a mouse partner once the methods match, it was a method difference.
2. Test robustness. Check that the class appears in several donors, not one, and survives reclustering at a few resolutions and integration methods. A class carried by one donor or one setting is fragile.
3. Validate in tissue. Pick two or three markers that define the class and do multiplex in situ hybridization on human DRG sections, counting single neurons that coexpress them and measuring their soma size. Do the same with the orthologous genes in mouse DRG. A coexpressing population present in human sections and absent from mouse ones supports a real difference in expression.
4. Tie it to function. Expression alone does not make a functional type. Record from human DRG neurons in culture or by Patch-seq, or use microneurography data where an afferent class can be matched, and ask whether cells with the marker profile share a physiology.

Expected outcomes: if the class disappears when methods are matched, it was an artefact; if it survives matching and tissue validation but has no distinct physiology, it is a real molecular difference of unknown function; only all four together would make it a functional human-mouse difference, and calling it human-specific would also need comparison with other species such as macaque.