When Pain Persists

If you silence a small group of amygdala neurons in a mouse with a nerve injury, the paw still pulls away from a touch, yet the escape and avoidance fall. So where is the pain? Persistent pain is not one thing stored in one place. It is a set of changes at several levels, each measured in a particular animal or patient, and the skill this lesson trains is keeping each result at its own level.

Touch and pain taught sensitization, gate control and descending control, and spinal pain circuits taught the dorsal-horn cell types and the KCC2 shift. This lesson starts past both: what stays changed after injury, which cells hold the change, and what the human evidence shows.

Four words for pain, and why they are not four boxes

The International Association for the Study of Pain (IASP) names three mechanistic descriptors. Nociceptive pain arises from actual or threatened non-neural tissue damage through nociceptor activation. Neuropathic pain is caused by a lesion or disease of the somatosensory nervous system. Nociplastic pain arises from altered nociception without clear evidence of either of those causes. The descriptors can coexist in one patient: nociceptive and nociplastic pain, for instance, can both be present.

A fourth word, inflammatory, comes from a different classification. Woolf describes inflammatory pain as pain caused by immune-system activation after tissue injury or infection. It is a statement about a process, not a fourth IASP box, so a single patient can fit a process label and a mechanistic descriptor at once. “Chronic” is a statement about duration, and none of these words is.

The practical point for what follows: when a paper says “neuropathic” it should mean a documented lesion or disease, and when it says “nociplastic” it is making a claim that rests partly on exclusion, and the proposed clinical criteria also ask for signs of pain hypersensitivity in the painful region. The animal models below are nerve-injury models, so they speak most directly to neuropathic mechanisms, and their extension to the other descriptors is a hypothesis.

The ganglion that will not stop firing

Start with the first neuron. North and colleagues studied dorsal root ganglia surgically excised from human donors who were undergoing tumour surgery, some with radicular or neuropathic pain and some without. They dissociated the neurons, kept them in culture for more than 24 hours, and recorded them by patch clamp, and they sequenced RNA from the same tissue. Spontaneous action potentials, firing with no applied stimulus, were recorded in 20 of 149 sampled neurons, and that spontaneous firing was associated with the donors’ radicular or neuropathic pain. Two limits matter. It is an association, and the neurons came from 17 patients with spinal tumours, 10 of whom contributed a spontaneously active cell, so 20 of 149 is not the fraction of human sensory neurons that misfire.

What drives such firing can be tracked in rats. Li and colleagues ligated the L5 spinal nerve in male Sprague-Dawley rats and measured Nav1.7 in the L4 and L5 ganglia seven days later. Nav1.7 mRNA and protein fell in the injured L5 ganglion and rose in the adjacent, uninjured L4 ganglion. Read that twice, because “injury raises Nav1.7” is a tempting summary and it is wrong. In this model the sign of the change depends on which ganglion you sample (in a second nerve-injury model the same authors saw a rise in both), and an expression change is not the same as a change in firing. The sodium channel lesson covers what Nav1.7 does when it is present or absent.

A worked problem: what a firing periphery does to the cord

Chen and colleagues recorded in vivo, by whole-cell patch clamp, from lamina I and II neurons of anaesthetized male Sprague-Dawley rats aged 5 to 6 weeks (Nature Communications, 2018). The mean spontaneous excitatory postsynaptic current (EPSC) frequency was 9.0 Hz (SEM 1.3, n = 20 neurons) in sham-operated rats and 15.9 Hz (SEM 1.6, n = 20 neurons) seven days after spared nerve injury.

1. Absolute change: 15.9 Hz minus 9.0 Hz is 6.9 Hz.
2. Ratio: 15.9 divided by 9.0 is 1.77, a 1.77-fold rise.
3. Percentage increase: 100 times 6.9 divided by 9.0 is 76.7 percent. The base is the sham value, so 1.77-fold and a 76.7 percent increase are the same fact.
4. Spread: an SEM of 1.3 Hz with n = 20 means a standard deviation of about 1.3 times 4.47, or 5.8 Hz, and 1.6 gives about 7.2 Hz. Neurons vary widely, so the groups overlap even though the means differ. As a rough size of the difference, the standard error of the difference is the square root of 1.3 squared plus 1.6 squared, or 2.06 Hz, and 6.9 divided by 2.06 is about 3.3. That is a back-of-envelope figure, not the authors’ test, and neurons recorded from the same animal are not independent.

What the 76.7 percent is: a rise in the frequency of synaptic events onto these neurons. What it is not: a 76.7 percent rise in how unpleasant the animal finds anything. The next two sections say why those must stay apart.

Immune routes that differ by sex and by compartment

The spinal cord holds a second store of change, in cells that are not neurons. Tsuda and colleagues, working in rats after nerve injury, found that induced P2X4 receptors, ATP-gated channels, sat on reactive spinal microglia and not on neurons or astrocytes. Lowering P2X4 levels suppressed the tactile allodynia, and microglia that were stimulated with ATP in culture and transferred to naive rats produced allodynia. That is converging evidence, in rats, that stimulated microglia can be necessary and sufficient for the behaviour as measured.

How the signal reaches neurons is a separate paper. Coull and colleagues, working in rats, showed that ATP-stimulated microglia release BDNF, that this depolarizes the anion reversal potential of lamina I neurons in spinal slices, and that blocking BDNF-TrkB signalling reversed both the shift and the allodynia. The spinal pain circuits lesson works through the chloride mechanism, and Trk and p75 covers the receptor. Merging the 2003 and 2005 papers into one discovery is the commonest credit error in this field.

Then came the sex question. Sorge, Mapplebeck and colleagues tested male and female mice, mostly 7 to 12 weeks old, after a nerve injury. Microglia were unnecessary for mechanical hypersensitivity in females under the conditions tested, while the hypersensitivity itself was comparable. Female mice appeared to use adaptive immune cells, likely T lymphocytes. Female mice lacking adaptive immune cells became sensitive to the glial inhibitor, so the pathway can switch. The result is conditional and in mice; the authors themselves raise the idea of treating men and women differently, but they tested no people. It does not say that women use T cells and men use microglia.

Compartment matters as well. Yu and colleagues (Nature Communications, 2020) used mice aged 4 to 12 weeks and a nerve injury that spared one branch. Depleting macrophages reduced mechanical hypersensitivity in both males and females, and macrophages in the dorsal root ganglion expanded after injury. Deleting CSF1 from sensory neurons reduced that expansion in males but not in females. The systemic depletion could not exclude a contribution from skin dendritic cells that were depleted at the same time, though the authors favoured the ganglion macrophages as the main contributor. So “immune cells drive chronic pain” splits into spinal microglia, ganglion macrophages and adaptive cells, each with its own sex and context dependence. What would settle whether this holds in people is not obvious: it would need cell-type-specific tests in human tissue or a human intervention, and none of the mouse data support a treatment claim.

Unpleasantness as an ensemble, not a centre

A withdrawal reflex is cheap to measure and easy to mistake for pain. Corder, Ahanonu and colleagues asked whether affective-motivational behaviour has its own circuit. In male mice they imaged basolateral amygdala (BLA) principal neurons in repeated sessions and identified an ensemble that encoded the negative valence of noxious stimuli. After a nerve injury, innocuous stimuli recruited the same ensemble. They then used activity-dependent genetic tagging to silence that ensemble, and affective-motivational pain behaviours fell while the measured withdrawal reflexes were preserved.

Scope matters: this is male mice, a nerve-injury model, and the behaviours the authors chose. The mouse gives no verbal report, so the result separates two measurable endpoints; it does not show that unpleasantness is made in the BLA or that silencing it abolishes sensing of the stimulus. The emotion and stress lesson places the amygdala in the wider circuit.

The anterior cingulate cortex (ACC) shows a different lesson. Acuña and colleagues imaged CaMKIIa-expressing ACC neurons in male mice with a head-mounted microscope, before and after nerve injury. Population activity discriminated noxious stimuli from other sensory stimuli better than any single cell’s response identity did, and the population representation stayed stable even as single-cell selectivity changed. Nerve injury impaired the sampled population’s classification of sensory events. Coding here lives in the population, so a search for a single “pain neuron” in cortex asks the wrong question.

The ACC also reaches back down. Chen and colleagues found that ACC stimulation potentiated spinal excitatory synaptic transmission in the rodent experiments, that the facilitation persisted independently of the rostral ventromedial medulla, and that optogenetic activation of ACC neurons projecting to the spinal cord sensitized pain-related behaviour in mice. Descending control is not only the inhibitory brainstem route of touch and pain.

Predicting who stays in pain, and what tolerance is

Prediction is the human question. Baliki and colleagues followed adults with subacute back pain for a year with fMRI. In the main completed cohort (mean age 40.9 years, SEM 2.3; 20 women among 39 patients), 20 patients were classed as recovering and 19 as persisting, using a 20 percent change in pain intensity as the cutoff. Greater baseline functional connectivity between medial prefrontal cortex and nucleus accumbens predicted persistence, with an ROC area of 0.83, and 13 further patients were used for validation. The corticostriatal circuit is the one met in motivation and reward, which fits the idea that reward and motivation circuits take part in chronification.

Separation between the curves 1.35 relative units Cutoff cursor 0.70 relative units Resample 19 persisting + 20 recovering simulated patients

Area under the ROC curve

0.83

Separation

1.35

Cutoff

0.70

	Expected	Drawn
Persisting patients above the cutoff (n = 19)
Recovering patients above the cutoff (n = 20)
Pairs of 380 with the persisting patient higher
Area from this draw (pairs ÷ 380)

In this sketch: two normal shapes of equal spread (1 relative unit), independent draws, and a separation set by you, with the default chosen to give the area of 0.83 that the lesson reports; the study’s individual scores are not shown, and only the group sizes (19 persisted, 20 recovered) and the reported area are its numbers. Rule to keep: the area is a probability about pairs of patients, not an accuracy and not one patient’s risk.

Tolerance is a separate topic from chronification, and it concerns drugs. Opioid receptors signal through Gαi/o: that inhibits adenylyl cyclase, presynaptic Gβγ inhibits the opening of voltage-gated calcium channels, and postsynaptic Gβγ activates GIRK channels, so both reduce transmission (a review by Corder, Castro, Bruchas and Scherrer, 2018). In knock-in mice whose receptors lacked carboxyl-terminal phosphorylation sites, Kliewer and colleagues found much less analgesic tolerance, yet tolerance to morphine persisted even when every tested site was removed. Phosphorylation contributes to tolerance and is not the whole account. A drug that works by a different mechanism, suzetrigine, an inhibitor of the NaV1.8 sodium channel, was approved by the FDA on 30 January 2025 for adults with moderate to severe acute pain. That approval does not establish efficacy in chronic pain.

Design problem

Baliki and colleagues found that baseline medial prefrontal to nucleus accumbens connectivity predicted persistence of back pain at one year, with an ROC area of 0.83. What does that support, and what study would go further?

One way to do it

What it supports. The measurement came before the outcome, so connectivity carries information about who will persist. The area under the ROC curve, 0.83, means that if you pick one patient who persisted and one who recovered at random, the persisting patient has the higher connectivity about 83 percent of the time. It is a discrimination index.

What it does not support. It is not 83 percent accuracy, and it is not an individual’s chance of lasting pain, which would depend on the cutoff, how common persistence is and calibration. The design is observational, so it does not show that the circuit causes chronification. Sample size is also small: 39 patients in the main cohort.

A study that goes further. First, a prospective replication in an independent, larger cohort, with calibration reported (do predicted risks match observed outcomes?) and the cutoff fixed in advance. Second, for cause, a manipulation: an intervention that changes connectivity or the circuit in a randomized design, with persistence as the outcome. Third, report who dropped out, since persisting and recovering patients may leave a study at different rates.