Main

Pain is a complex, aversive perception, fundamental to adaptive survival1. Understanding the neural circuits and cell types underlying the affective and motivational features of pain experiences is crucial for advancing precision therapeutics for individuals with chronic pain conditions3,4. Current analgesic drugs, such as opioids, act on widely expressed molecular targets that reduce pain unpleasantness but also promote serious and fatal side effects. By pinpointing specific neural circuits for pain-associated aversion, which intersect with the expression of the μ-opioid receptor (MOR)5, the molecular target of morphine, a new class of effective analgesics can be developed that interfere with the unpleasantness of pain rather than pain sensation, with reduced addiction and respiratory depression side effects6,7.

Coordinated neural activity in the anterior cingulate cortex (ACC) is essential for encoding the emotional and motivational dimensions of pain8,9,10,11,12,13,14, guiding behavioural choices in real time and promoting future avoidance of harmful stimuli15,16. According to the gate control theory17 and central control models15,18, pain perception is not a passive relay of nociceptive input but is actively shaped by spinal and brain circuits that integrate sensory, cognitive and emotional information to amplify or inhibit pain signals. Within this framework, the ACC has a central role in evaluating nociceptive input in relation to valence, context and internal state. This processing supports the selection of adaptive responses, such as escape and recuperative behaviours19, which act as negative feedback to reduce further injury and promote healing. By tracking spontaneous pain-related behaviours and activity recordings in cortical neurons, we can infer the latent dynamic computations underlying pain. This approach provides a window into identifying single neurons and distributed ensembles in the ACC that represent the affective–motivational components of pain, offering candidate cellular targets for therapeutic strategies.

Perceived pain unpleasantness strongly correlates with functional magnetic resonance imaging activity in the human ACC9,19. Patients with intractable chronic pain treated with surgical cingulotomy lesions do not report changes in pain perception, intensity discrimination or reactions to momentary harmful stimuli20,21,22. Rather, their attitude towards pain is modified, dissociating the negative valence from the experience of pain. In these cases, pain becomes a sensation rather than a threat23. Similarly, in preclinical models, lesions24,25, opioids26,27,28,29 and optogenetic manipulation30,31,32,33,34,35,36 of ACC neural circuits attenuate aspects of the affective–motivational component of pain. Thus, ACC neural circuits that drive adaptive behaviour to acute noxious stimuli might also be maladaptive during chronic pain37,38. Acute pain engages MORs in the dorsal anterior cingulate and lateral prefrontal cortex, and the affective perception of pain is correlated with MOR availability in the ACC and other brain regions26,27,39,40,41. This evidence supports the role of MOR signalling within the ACC in altering the perception of pain13. Leveraging personalized deep-brain stimulation42,43,44 or cell-type-specific approaches45,46 to modulate neural activity in MOR-expressing neurons2 may mimic the prefrontal cortical actions of opioid analgesia without the associated risks of off-target pharmacotherapies1.

Mapping nociceptive and µ-opioidergic ACC neurons

Using nociceptive activity-dependent tagging (painTRAP) and immediate early gene (IEG) mapping (painFOS), we identified a ‘nociceptive hotspot’ approximately 700 μm in length located in the anterior ACC dorsal Cg1 and ventral Cg2 (Extended Data Fig. 1, Supplementary Table 1 (rows 48–49) and Supplementary Note 1). In situ hybridization revealed that approximately 30–50% of painFOS neurons and approximately 70% of Slc17a7+ glutamatergic neurons co-express Oprm1, indicating that a substantial fraction of the nociceptive hotspot expresses MORs (Extended Data Fig. 2 and Supplementary Note 1). Single-nucleus RNA sequencing of this ACC nociceptive hotspot, before and after the development of chronic neuropathic pain using the spared nerve injury (SNI) model, resolved 23 cell types. Only three glutamatergic neuronal clusters (L2/3 IT-3, L5 IT-1 and L6 CT-2) showed persistent nociceptive and neuropathic IEG signatures, all of which expressed Oprm1 (Extended Data Fig. 3 and Supplementary Note 2). Oprm1 transcript levels and distribution were unchanged during chronic neuropathic pain (Extended Data Fig. 3 and Supplementary Note 2), implying that morphine can access molecularly distinct ACC neurons that encode nociception throughout chronic pain.

ACC MORs mediate morphine analgesia

To determine the role of cortical MORs in morphine-mediated analgesia, we genetically deleted MORs selectively in ACC neurons (Extended Data Fig. 4, Supplementary Table 1 (row 50) and Supplementary Note 3). In Oprm1**fl/fl mice given AAV9–hSyn–Cre in the ACC versus controls, systemic morphine (0.5 mg kg−1) no longer reduced affective–motivational pain responses to noxious stimuli despite intact sensory thresholds (Extended Data Fig. 4f–j and Supplementary Table 1 (rows 51–57)). Conversely, AAV9–hSyn–FLEx–OPRM1 re-expression of MORs exclusively in the ACC of global MOR knockout mice restored morphine analgesia (Extended Data Fig. 4k–r and Supplementary Table 1 (rows 58–65)). Together, these loss-of-function and gain-of-function experiments demonstrate that ACC MORs are both necessary and sufficient for the affective pain relief produced by clinically relevant morphine doses, identifying this cortical ensemble as a key target for μ-opioid-mediated analgesia.

A deep learning system for pain behaviour analysis

The complexity of pain cannot be fully captured by reflexive withdrawal responses (the current standard for preclinical analgesic evaluation), which assess only evoked responses and fail to reflect the continuing affective experience most relevant to patients with chronic pain47,48,49. Although assays such as conditioned place preference or aversion measure memory-based responses to prior pain or relief24,50,51, they do not capture the dynamic, moment-to-moment motivational behaviours driven by spontaneous or continuing pain1,52.

To address these limitations, we developed Light Automated Pain Evaluator (LUPE; Fig. 1), a behavioural analysis platform designed to resolve fine-scale, naturalistic pain-related behaviours across several timescales. LUPE enables a nuanced and translationally relevant assessment of affective–motivational pain states in freely moving mice. Named after the Greek daemon of pain and suffering (Lýpē; λῡ́πη), LUPE provides a standardized, dark environment optimized for the behaviour of nocturnal prey animals. Critically, it eliminates the presence of the human experimenter both as a looming threat and as a subjective observer, allowing for objective quantification of spontaneous nocifensive behaviours.

Fig. 1: Deep learning analysis of natural behaviour reveals how pain and opioids shape internal affective–motivational states.

a, Schematic of the standardized LUPE chamber. b, A 20-body point DLC pose-tracking model was built from male and female mouse pain behaviour. c, Behaviour segmentation models trained iteratively on supervised annotations of behaviours with A-SOiD, followed by unsupervised sub-clustering with B-SOiD. d, Motion energy heat maps illustrating spatial trajectories and intensity distributions for the six primary behavioural repertoires. e, Temporal probability plots for the six primary behaviour repertoires in 1-min bins, comparing uninjured mice to mice with left hindpaw injections of 1% formalin, 5% formalin or capsaicin. f, Raster plots of behaviour transitions within a 30-s window. g, Procedure for behavioural state inference from statistical structure of spontaneous behaviour. h, Left, model centroid transition matrices characterizing each of six inferred states. Right, comparing the fraction occupancy of mice in each state between uninjured (grey), formalin (magenta) and capsaicin (cyan) pain models (n = 20 per group; one-way analysis of variance (ANOVA); Tukey correction: Pstate1 = 0.0007, Pstate3 = 0.0078 and Pstate4 < 0.0001). i, Two-dimensional (2D) visualization of PCA of state occupancies across pain models. j, Magnitude of coefficients of each state in each PCA. k, Scores of each animal along PC1 (top) and PC2 (bottom) across pain models (n = 20 per group; one-way ANOVA; Tukey correction: PPC1 = 0.0027 and PPC2 = 0.0082). l, Dose–response of morphine on PC1 (top) and PC2 (bottom) scores in uninjured, formalin-administered and capsaicin-administered mice (n = 20 per group and dose; one-way ANOVA; Tukey correction: PPC1 uninjured < 0.0001, PPC1 formalin < 0.0001, PPC1 capsaicin < 0.0001, PPC2 uninjured < 0.0001, PPC2 formalin < 0.0001 and PPC2 capsaicin < 0.0001). ⋆P < 0.05. Bars are mean; dots are individual animals; vertical lines and shaded areas are s.e.m. See Supplementary Table 1 (rows 1–14) for statistics. NS, non-significant. Scale bar, 2.5 s (f).

Source data

In addition to the standardized chamber and high-speed infrared videography recorded from below a glass floor (Fig. 1a and Supplementary Fig. 1), behavioural classification was driven by a multilayered analysis pipeline. Using DeepLabCut (DLC)53 to track 20 body key points, LUPE extracted detailed posture dynamics that were processed through both semi-supervised (A-SOiD54) and unsupervised (B-SOiD55) algorithms to identify six holistic behavioural repertoires: still, walk, rear, groom, lick left hindpaw and lick right hindpaw (Fig. 1b,c and Supplementary Fig. 1). These repertoires were assembled from sub-second behavioural syllables and allowed quantitative analysis of transitions across time. Motion energy plots visualized the displacement of tracked body points, defining each behaviour and distinguishing similar actions such as grooming and paw-directed licking (Fig. 1d).

To evaluate the sensitivity of LUPE to dynamic changes in pain-related behaviour, we applied the formalin and capsaicin models of acute pain56. Male and female C57Bl/6J mice were habituated to the LUPE chamber for two consecutive days and then injected in the left hindpaw with 1% or 5% formalin or 2% capsaicin, or left uninjured as controls (Fig. 1e). LUPE computed the behavioural probabilities for all six repertoires over 30-min sessions for all 60 mice in under 2 h, compared with 50–150 min for manual57 scoring of one behaviour in one mouse (with the upper bound equivalent to 54,000 min for full dataset scoring; Supplementary Fig. 1k). By automating behaviour classification, LUPE increases the speed, rigor and reproducibility of preclinical pain behaviour analysis. It also generates archival-quality datasets that include video logs and computer-scored results, facilitating transparent cross-laboratory comparison, long-term record keeping and future reanalysis.

LUPE identified a low dose of morphine (0.5 mg kg−1) that reduced licking of the injured hindpaw following both formalin-induced and capsaicin-induced injury that did not affect walking (Extended Data Fig. 6a–g and Supplementary Table 1 (rows 75–87)). Therefore, LUPE provides a sensitive measure of ethologically relevant affective–motivational pain behaviour that can identify translationally relevant analgesic doses.

Discovery of morphine-sensitive latent pain states

Inferring internal states, such as pain, from sparse spontaneous behaviour is a central challenge in ethological neuroscience. An injury to the left hindpaw may or may not elicit licking at a given moment, although the animal may still be experiencing pain. We therefore tested whether latent cognitive–affective pain states could be determined from LUPE-scored behaviour. From 58 mice (formalin, n = 19; capsaicin, n = 20; and SNI, n = 19; Fig. 1g and Extended Data Fig. 5), we modelled behavioural transitions as Markov processes using 30-s sliding windows to produce per-animal transition matrices (Fig. 1g and Extended Data Fig. 5a). Matrices were clustered by k-means (k = 6; 100-fold cross-validation; Fig. 1g and Extended Data Fig. 5b). Classification of animals to these six clusters exceeded chance (Euclidean distance between real versus shuffled cluster centroids; Extended Data Fig. 5c–h and Supplementary Table 1 (rows 66–72)). No single behaviour drove the clustering; systematic removal of individual behaviours disrupted classification less than expected by chance (Extended Data Fig. 5c).

Cluster centroids define the mean transition matrices of six distinct behaviour states (Fig. 1h): (1) stillness, walking, rearing and grooming; (2) stillness, walking and rearing; (3) state 1 plus licking the injured paw; (4) all behaviours except licking the uninjured paw; (5) all behaviours except stillness; and (6) stillness, walking, rearing and licking the injured paw. States evolve over seconds to minutes and show conserved dynamics across pain models (Extended Data Fig. 5i–k and Supplementary Table 1 (rows 73–74)). Behaviour states distinguished injured from uninjured animals but did not separate injury types (Fig. 1h and Supplementary Table 1 (rows 1–6)). Uninjured mice predominantly occupied states 1 and 2; capsaicin and formalin increased occupancy of states 3 and 4 and reduced time in state 1. Notably, state 4 was uniquely and dose-dependently suppressed by morphine, indicating a selectively opioid-sensitive spontaneous-pain dimension (Extended Data Fig. 6g and Supplementary Table 1 (rows 97–99)), although morphine modulated all states dose-dependently across conditions (Extended Data Fig. 6g and Supplementary Table 1 (rows 88–105)). Thus, latent affective–motivational states inferred from spontaneous behaviour track pain and analgesia.

Numeric pain index tracks injury and analgesia

To compress the diverse effects of pain and morphine across all six states, we applied principal component analysis (PCA) to the fraction of time each mouse spent in each state across pain conditions (Fig. 1i). This revealed two principal axes of variation in behaviour. Both capsaicin and formalin shifted scores along these axes, reducing the first component and increasing the second, regardless of injury model (Fig. 1k and Supplementary Table 1 (rows 7–8)). The first principal component (PC1), driven primarily by states 1 and 2, reflects a baseline behavioural structure disrupted by both injury and high-dose morphine and is termed the general behaviour scale (Fig. 1j,l (top) and Supplementary Table 1 (rows 9–11)). The second component (PC2) was weighted by states 2 and 4, selectively increased by injury and dose-dependently suppressed by morphine (Fig. 1l (bottom) and Supplementary Table 1 (rows 12–14)), capturing the presence and relief of affective pain. Because PC2 responds bidirectionally to injury and analgesia, we define it as the affective–motivational pain scale (AMPS), a data-driven, continuous index of pain-related behavioural states.

Licking as a structured motivated response to pain

The gate control theory asserts that volitional behaviours, such as rubbing or licking injured tissue, act as antinociceptive responses by recruiting touch afferents that inhibit spinal nociceptive signalling (Extended Data Fig. 7a). Consequently, the unpleasantness of pain drives motivated licking, which then reduces pain, forming a negative feedback loop. Thus, motivated licking is expected to increase with affective pain and decline as analgesia—or recuperation—is achieved.

Our analysis treated latent behavioural states as Markovian processes, in which behavioural probabilities are stable within a state (Extended Data Fig. 7b (top)). We therefore tested whether licking dynamics followed theoretical predictions by measuring the probability of each behaviour as a function of elapsed time within pain state 4, a latent state consistently enhanced by injury and dose-dependently suppressed by morphine (Fig. 1h).

Across injury models, injured-paw licking showed a reproducible temporal profile within pain state 4: near zero at state onset, accumulating in the latter half and declining just before the state transition (Extended Data Fig. 7b,c and Supplementary Table 1 (rows 106–109)). This temporal structure was not seen for other behaviours in pain state 4 or for licking pooled across all states (Extended Data Fig. 7d–f and Supplementary Table 1 (row 110)). These results indicate that paw licking is not merely a reflexive nocifensive action but an innate affective–motivational response engaged to negatively modulate pain, consistent with the gate control theory and its role as a motivated antinociceptive behaviour.

ACC dynamics reflect nociception and behaviour

To link ACC activity to morphine-sensitive pain behaviour, we performed single-cell calcium imaging in freely behaving mice inside LUPE. We expressed AAV9–hSyn–jGCaMP8m and implanted 1.0-mm GRIN lenses at ACC nociceptive hotspot coordinates (n = 5 male mice; Fig. 2a,b and Extended Data Fig. 8a,b). With a head-mounted one-photon miniscope, we recorded neural activity during acute inflammatory pain (left hindpaw intraplantar injection of 2% capsaicin (10 μl)) and after morphine analgesia (0.5 mg kg−1; Fig. 2c,d). A Fisher linear decoder (100-fold cross-validation) reliably decoded spontaneous behaviours from ACC population activity across mice and sessions, independent of injury or opioid treatment (Fig. 2e, Extended Data Fig. 8g,h and Supplementary Table 1 (row 116)).

Fig. 2: Neural dynamics in ACC track acute pain and analgesia.

a, Microendoscope calcium imaging synced with LUPE behaviour tracking. b, GRIN lens implant and hSyn–GCaMP8m expression in ACC Cg1. c, From top to bottom, average and single-cell neural activity (z-score) from a representative mouse, LUPE behaviours, states inferred by our behavioural state model and probability of behaviours given states and behaviour history (binomial GLM). d, Capsaicin imaging protocol injury (intraplantar; 2%; left hindpaw) and morphine (intraperitoneal (i.p.); 0.5 mg kg−1; n = 5). e, Fisher decoder accuracies predicting behaviours from neural activity, averaged over mice (permutation test; Extended Data Fig. 8g). f, Area under the receiver operating characteristic curve (auROC) of GLMs predicting Plick from e in each animal (n = 5). g, Calcium events per second of neurons in all sessions (two-way ANOVA; Tukey correction: Pinteraction = 0.0009). h, Mean ± s.e.m. fraction of positive and negative Plick neurons during capsaicin (red outline) and capsaicin + morphine (purple outline) sessions. i,j, Calcium events per second of positive (i) and negative (j) Plick neurons in capsaicin and capsaicin + morphine sessions (two-way ANOVA; Tukey correction: positive Plick neurons, Pinteraction = 0.0004; negative Plick neurons, Pinteraction = 0.026). k, Average lick probability around lick bout onset (two-tailed unpaired t-test: Pnegative, 1–2 s = 0.0003). Light grey, 0–1 s; dark grey, 1–2 s. l,m, Left, average activity in positive (l) and negative (m) Plick neurons around lick bout onset, pooled across animals. Right, area under the curve (AUC) of lick probability from 0 to 1 s and from 1 to 2 s post- initiation (two-tailed unpaired t-test: Pnegative, 0–1 s = 0.0003). Light orange, 0–1 s; dark orange, 1–2 s (l). Light blue, 0–1 s; dark blue, 1–2 s (m). n, Behavioural probability as a function of fraction time in pain state 4 (n = 19–20 per group). o, Cumulative lick probability over fraction of time remaining in pain state 4 (two-sided Kolmogorov–Smirnov test: P = 1.1 × 10−23). p, Summary of results. ⋆P < 0.05. Bars, lines or dots are mean; error bars and shaded areas are s.e.m. (see Supplementary Table 1 (rows 15–24) for statistics). Scale bars, 1.0 mm (b (yellow bar)), 1 zS (c).

Source data

We next investigated whether ACC ensembles encode nociceptive signals by modality or valence. In uninjured mice, we delivered mechanical and thermal stimuli to the left hindpaw (0.16-g filament; pin prick; 30 °C water drop; 55 °C hot-water drop; 6 °C acetone) and compared these to orally consumed stimuli of opposing valence (10% sucrose versus quinine) and a 55 °C hot-water drop. ACC neurons showed greater overlap in responses to noxious stimuli across modalities (approximately 11–20% overlap among excited cells and approximately 25% among inhibited cells) than between stimuli of opposite valence. Among excited cells, only 6% responded to both 55 °C heat and sucrose versus 13% for heat and quinine; the overlap among inhibited cells was approximately 11% for both pairings (Extended Data Figs. 9a,c and 10e). Activity patterns and cross-decoding performance further separated heat-activated versus sucrose-activated neurons (Extended Data Fig. 10a–e), consistent with valence-specific encoding in ACC neural populations.

Morphine inhibits pain-tracking ACC neurons

To test whether neurons encoding lick probability are morphine-sensitive, we first estimated lick probability by aligning neural activity to behaviour and latent states and fitting a binomial generalized linear model (GLM; current state and behaviour history using two previous time steps; Fig. 2c). These state-based GLMs outperformed chance and yielded a pseudo-continuous lick-probability trace (Fig. 2c). We then trained GLMs to predict behavioural probability from the PCs capturing 80% of neural variance per animal per session (Extended Data Fig. 8c). These models exceeded performance on shuffled data (Fig. 2f, Extended Data Fig. 8d and Supplementary Table 1 (rows 111–115)). Neurons with the largest absolute weights in the top 3 PCs (P < 0.001; |z-score of coefficient | > 1.5) were labelled Plick neurons (Extended Data Fig. 8e,f).

The ACC population activity was elevated during capsaicin sessions versus baseline and was selectively suppressed by 0.5 mg kg−1 of morphine only with injury (Fig. 2g and Supplementary Table 1 (row 15)). Plick neurons split into positive (activity increased at lick onset; 16.5 ± 3.4% of cells) and negative (activity decreased at lick onset; 16.6 ± 3.4% of cells) subpopulations (Fig. 2h). Morphine inhibited these subpopulations differently; positive Plick cells were suppressed selectively during pain state 4, whereas negative Plick cells showed broader state-independent inhibition (Fig. 2i,j and Supplementary Table 1 (rows 16–17)). Together, this indicates population-dependent and state-dependent modulation of Plick neurons by morphine.

We next examined the dynamics surrounding the onset of a lick bout. Behaviourally, morphine reduced lick probability 1–2 s after lick onset, indicating impaired lick maintenance (Fig. 2k and Supplementary Table 1 (rows 18–19)). Morphine did not alter positive Plick neurons but increased inhibition of negative Plick neurons 0–1 s after lick offset, preceding the behavioural change (Fig. 2l,m and Supplementary Table 1 (rows 20–23)). These effects sharpened Plick selectivity for licking versus other behaviours (Extended Data Fig. 8i–l and Supplementary Table 1 (rows 117–119)).

Over longer pain state 4 bouts, morphine narrowed the lick-probability profile in capsaicin-treated mice by reducing early-state licking and causing greater accumulation near the state end (Fig. 2n,o and Supplementary Table 1 (row 24)). Thus, morphine relieves the affective–motivational drive to lick by suppressing spontaneous activity in positive Plick neurons and enhancing behaviour-locked inhibition of negative Plick neurons, producing delayed initiation and reduced maintenance of licking during pain state 4 (Fig. 2p).

Morphine restores chronic pain-disrupted dynamics

We performed longitudinal miniscope calcium imaging in mice expressing hSyn–GCaMP8m with GRIN lenses in ACC Cg1, recorded in LUPE 1 day before and 1 day, 7 days, 14 days and 21 days after SNI (n = 9) or in uninjured controls (n = 9; Fig. 3a). SNI produced an immediate, sustained increase in spontaneous left-hindpaw lick rate, pain-state occupancy and AMPS scores versus controls (Fig. 3b–e and Supplementary Table 1 (rows 25–27)). As in acute injury, a single 0.5 mg kg−1 of morphine dose at 3 weeks post-SNI reduced AMPS scores in mice with SNI (Fig. 3f and Supplementary Table 1 (row 28)), indicating that opioids remain effective for affective–motivational features of chronic neuropathic pain.

Fig. 3: Morphine targets functionally compromised ACC neurons to relieve chronic pain.

a, SNI protocol for chronic neuropathic pain. b, Log-transformed rate of licking at the injured limb in SNI (red; n = 9) or uninjured controls (grey; n = 9; two-way repeated measures ANOVA; Tukey correction: Pinteraction = 0.0036). c, Heat map of average state occupancy. d, Occupancy of pain and non-pain states in SNI and uninjured mice (two-way repeated measures ANOVA; Tukey correction: Pinteraction < 0.0001). e, AMPS score in SNI and uninjured mice (two-way repeated measures ANOVA; Tukey correction: Pinteraction = 0.0089). f, AMPS score in SNI and uninjured mice 3 weeks post-SNI and morphine (0.5 mg kg−1; intraperitoneal; two-way repeated measures ANOVA; Tukey correction: Pinjury = 0.0085; Ptreatment = 0.0041). g,h, Left: lick-evoked activity in positive (g) and negative (h) Plick neurons before (black) and after SNI (warm colour gradient; yellow = 1 day and red = 3 weeks post-SNI). Right: area under the curve of lick-evoked activity (0–1 s post-onset) in SNI and uninjured mice (two-way repeated measures ANOVA; Tukey correction: Ppositive,interaction < 0.0001; Pnegative,interaction = 0.021). i,j, Lick-evoked activity in positive (i) and negative (j) Plick neurons. Left, lick-evoked activity at baseline (black), 3 weeks post-SNI (red) and 3 weeks post-SNI + morphine (blue). Right, lick-evoked activity 3 weeks post-SNI versus uninjured mice (two-way repeated measures ANOVA; Tukey correction: Ppositive,interaction = 0.035; Pnegative,treatment < 0.0001; Pnegative,injury < 0.0001). k,l, Calcium event rate in positive (k) and negative (l) Plick neurons before and after morphine treatment (two-way repeated measures ANOVA; Tukey correction: Ppositive,treatment = 0.0023; Pnegative,treatment = 0.0124). m, Linear regression predicting lick rate (purple) and pain state occupancy (grey) from the average magnitude of lick-evoked activity 3 weeks post-SNI (top), after morphine (middle) and change between sessions (bottom; Bonferroni-corrected P values displayed). ⋆P < 0.05. Bars, lines or dots are mean; error bars and shaded areas are s.e.m. (see Supplementary Table 1 (rows 25–37) for statistics).

Source data

Using the same Plick identification shown in Fig. 2, SNI impaired decoding accuracy for sensory stimuli, behaviours and latent states relative to baseline and controls (Extended Data Fig. 11a–h and Supplementary Table 1 (rows 121–128)) and persistently blunted lick-evoked responses in both positive and negative Plick neurons (Fig. 3g,h and Supplementary Table 1 (rows 29–30)). Morphine reversed these SNI deficits at lick onset and further enhanced Plick responses in uninjured mice (Fig. 3i,j and Supplementary Table 1 (rows 31–32)). SNI reduced single-cell lick selectivity in both Plick subtypes, which morphine restored (Extended Data Fig. 11i,j and Supplementary Table 1 (rows 129–132)). At 3 weeks, morphine suppressed positive Plick activity in mice with SNI and inhibited negative Plick cells across groups, reproducing the state-dependent and state-independent effects seen in acute pain (Fig. 3k,l and Supplementary Table 1 (rows 33–34)) and shifted the proportions of negative and positive Plick neurons regardless of injury (Extended Data Fig. 11k,l and Supplementary Table 1 (rows 133–136)). The magnitude of lick-evoked responses predicted within-session lick rate and pain-state occupancy, and morphine-induced increases in these responses predicted reductions in both behaviour and time spent in pain sta