Chronic pain is increasingly understood as a disorder of the nervous system rather than a lingering signal from a damaged tissue. Rodent work shows the injured nervous system rewiring itself, and human MRI shows the cortex thinning and connections changing in people with persistent pain. These two bodies of evidence rarely meet. Rodent studies sequence the dorsal root ganglion and spinal cord and report enriched pathways. Human studies report which brain regions change shape or activity but say nothing about molecules. One established bridge exists: take a brain map from MRI, correlate it across regions with a post-mortem atlas of gene expression, and run a gene-set enrichment test on the genes that track the map. This has been done in schizophrenia but, as far as this file shows, never in pain.
We propose to build that bridge for pain. Our main contribution, though, is to keep the bridge from giving false answers. Enrichment tests on atlas-correlated genes tend to return the same generic terms (synapse, neuron, calcium signaling) for almost any cortical map, because neighbouring brain regions resemble each other and genes are co-expressed in large blocks. Our claim is that a test which controls for both effects lets us settle a specific question that pain neuroscience currently cannot settle. The data are pre-operative brain maps from a large surgical cohort, collected before the injury that later becomes chronic pain. The question is whether the regions that distinguish future chronic-pain patients preferentially express the microglial and immune program found in injured rodent spinal cord, or the inhibitory-interneuron and synaptic-connectivity program implied by cortical disinhibition.
Core Claim
Two rodent findings offer competing molecular accounts of where central pain plasticity comes from. The spinal-cord study Identification and experimental validation of the spinal cord hub genes of aging and neuropathic pain combined public expression data with network analysis and confirmed six upregulated hub genes in spared-nerve-injury rats. Five of them (PTPRC, ITGAM, CD53, CD68, CD74) are microglial-activation markers. The patch-clamp study Nerve Injury-Induced Neuropathic Pain Causes Disinhibition of the Anterior Cingulate Cortex found that nerve injury cuts connections from inhibitory to excitatory neurons in anterior cingulate cortex from 23.94% to 8.45%. Its network model identified this loss of inhibitory output as the main cause of cortical hyper-responsiveness, but the study gave no molecular annotation.
We hypothesize that if either program marks a pre-existing vulnerability in humans, it should be spatially aligned with the cortical pattern that separates future chronic-pain patients from patients who recover, as measured before surgery. That alignment should hold above what spatial smoothness and gene co-expression produce on their own. It should also replicate across two independent surgical populations.
The pre-surgical timing matters. Causal Ambiguity in Cross-Sectional Brain Research argues that single snapshots cannot tell cause from consequence. A brain pattern measured before the injury cannot be a consequence of chronic pain. Our design does not establish causation, but it removes the reverse-causation explanation that undermines the existing pain imaging literature.
Research Program
The data source is the NIH Acute to Chronic Pain Signatures cohort described in Predicting chronic postsurgical pain: current evidence and a novel program to develop predictive biomarker signatures. It plans pre-operative structural, diffusion, and functional MRI in 1,400 knee-arthroplasty and 1,400 thoracic-surgery patients, followed to six months, with a fixed 66% discovery and 33% holdout split. That paper uses imaging and omics as parallel inputs to prediction models and does not link them biologically. We take only its imaging and outcome labels.
For each surgical arm we compute regional maps of the pre-operative difference between patients who later develop chronic pain and patients who recover. The primary maps are cortical thickness and resting-state regional homogeneity, a measure of how synchronized each voxel is with its neighbours. Cortical thinning in the cingulate and insula is the structural change reported in Abnormal structural gray matter and structural covariance networks associated with biopsychosocial characteristics in children with multisite pain.
Following Abnormal eye movement, brain regional homogeneity in schizophrenia and clinical high-risk individuals and their associated gene expression profiles, we relate each map to regional expression in the Allen Human Brain Atlas using partial least squares regression. That method finds the gene-weight pattern that best covaries with the brain map across regions. We supply the method from that paper but change what is tested.
The candidate gene sets are fixed in advance and translated to human genes through orthology. The microglial set comes from the spinal hub genes and the shared enrichment reported in the spinal-cord study. The peripheral neuronal-injury set comes from the Hippo, Ras, and Rap1 pathway genes in the dorsal root ganglion transcriptome of Transcriptome analysis of non-coding RNAs and mRNAs in the dorsal root ganglion of peripheral nerve injury-induced neuropathic pain. The inhibitory-connectivity set is defined from Gene Ontology GABAergic-synapse and fast-spiking-interneuron terms. As a "generic brain injury" control we use the traumatic-brain-injury microglial pathways from Analyzing single-cell transcriptome sequencing data reveals immunological mechanisms of microglia-associated diseases. If that control aligns as well as the pain-derived microglial set, the result is about injury in general, not about pain.
The methodological core is a two-null enrichment score. Let $s_k$ be the observed alignment between brain map $M$ and gene set $k$, such as the mean weight of set $k$'s genes in the partial least squares component. We compare $s_k$ against two reference distributions. The first comes from spatially permuted maps that keep $M$'s smoothness: we rotate the cortical surface so neighbourhoods stay intact while their locations change. The second comes from random gene sets of matched size, drawn so that their within-set co-expression in the atlas matches set $k$. The reported statistic is
$$ z_k = \min\left( \frac{s_k - \mu^{\text{spatial}}_k}{\sigma^{\text{spatial}}_k},\; \frac{s_k - \mu^{\text{coexp}}_k}{\sigma^{\text{coexp}}_k} \right), $$
where $\mu$ and $\sigma$ are the mean and standard deviation of each null distribution. A set therefore counts as aligned only if it beats both explanations. The independent-genes assumption behind the hypergeometric test, flagged as a known weakness of standard overrepresentation analysis, is replaced by the co-expression-matched null. The arbitrary cutoff for which genes count as "imaging-correlated" is removed by scoring the full gene ranking.
Baselines
The naive reference point is the published schizophrenia pipeline applied unchanged to pain maps: partial least squares against the atlas, selection of significant genes, and overrepresentation analysis on Gene Ontology and KEGG against a whole-genome background. That paper reports 1,639 imaging-correlated genes in first-episode schizophrenia and enrichment at "all corrected p < 0.05" in synaptic, calcium, and oxytocin pathways. This is the strongest existing method for exactly this mapping and the one a pain group would use tomorrow. We run it on identical maps, atlas processing, and gene sets, so that only the statistical test differs.
The second baseline adds tissue-aware background filtering in the style of TS-GOEA: a web tool for tissue-specific gene set enrichment analysis based on gene ontology, which restricts the background to brain-expressed genes to suppress generic false positives. It is the cheapest serious attempt at the specificity problem.
Two ablations isolate our ingredients. The first keeps only the spatial null. The second keeps only the co-expression null. A third ablation replaces the pre-operative contrast with a post-hoc cross-sectional contrast where follow-up imaging exists, to show what timing contributes.
No reported result was found on Lacuna for false-positive rates or replication of atlas-based enrichment in pain, or in any setting covered by the file. Beating the baselines is therefore defined operationally on three criteria. First, a nominal 5% test must stay near 5% on null maps where the baselines exceed it. Second, a much smaller fraction of terms must be shared with a panel of unrelated cortical maps. Third, the leading gene sets must agree more closely between the knee and thoracic arms, measured by rank correlation of set scores, than they do under the baselines.
Why not just…?
Why not just run the schizophrenia pipeline on pain maps and report the pathways? This is the obvious paper. It would almost certainly produce significant synaptic and neuronal terms. The problem is that we expect it to produce the same terms for a random smooth map. Our first experiment tests this directly by running both pipelines on 1,000 spatially permuted maps and on unrelated maps such as cortical myelin and age effects. If the naive pipeline's pain enrichments cannot be distinguished from that background, the obvious paper would be uninterpretable. If they can be distinguished, our method still adds calibrated error rates.
Why not just use a richer knowledge source such as DeepSyn? Deep functional synthesis: a machine learning approach to gene functional enrichment builds a literature-mined knowledge graph and reports beating baselines by at least 18% at recovering Gene Ontology categories. That addresses missing and context-blind annotations. It does not address the spatial and co-expression dependence that inflates imaging-transcriptomic tests, because both problems sit in the data, not in the annotation. DeepSyn-derived pain-context gene sets can be fed into our scoring as additional candidates. That would be a complement, not a rival.
Why not just predict chronification from imaging and skip genes? A2CPS is already doing this, and a classifier can be accurate without explaining anything. The schizophrenia paper reached 92.3% discrimination from regional homogeneity yet still needed the atlas to say anything about mechanism. We do not compete on prediction. As a secondary check, we will report whether projecting maps onto the winning gene set changes holdout prediction, and we will not claim improvement if it does not.
Why not just sequence rodent anterior cingulate cortex after nerve injury? This is a real threat. For mechanism it is arguably the better experiment, since it directly fills the molecular gap left by the disinhibition study. Our analysis answers a different question: which rodent program has any human counterpart before injury. If the calibrated test cannot separate the microglial, interneuron, and generic-injury sets in either surgical arm, we will say so. In that case direct rodent cortical sequencing is the better investment, and our negative result is what justifies it.
Proposed Experiments
The first experiment is calibration. We generate null brain maps with realistic smoothness and score every candidate and Gene Ontology set under the naive, tissue-filtered, single-null, and two-null tests, recording the empirical false-positive rate and how many terms recur across unrelated maps.
The second experiment is the preregistered contest. On discovery-set maps we compute $z_k$ for the microglial, inhibitory-connectivity, peripheral neuronal-injury, and generic brain-injury sets. We then confirm the ranking on the 33% holdout and in the other surgical arm.
The decisive outcomes are as follows. Support for the glial-vulnerability account requires the microglial set to exceed both nulls in both arms and to beat the brain-injury control. Support for the disinhibition account requires the same pattern for the inhibitory set. A tie, or failure in either arm, counts as a negative result.
A third experiment tests generality. It applies the frozen pipeline to the pediatric multisite-pain thickness maps, asking whether the same set leads in a cross-sectional chronic-pain map, which would be consistent with a stable regional vulnerability. White-matter questions such as the divergent anisotropy findings in spinal cord injury and bowel disease are left out. The atlas samples white matter too sparsely for this method to say anything there.
References
- Identification and experimental validation of the spinal cord hub genes of aging and neuropathic pain
- Nerve Injury-Induced Neuropathic Pain Causes Disinhibition of the Anterior Cingulate Cortex
- White matter microstructural alterations in patients with neuropathic pain after spinal cord injury: a diffusion tensor imaging study
- Predicting chronic postsurgical pain: current evidence and a novel program to develop predictive biomarker signatures
- Abnormal structural gray matter and structural covariance networks associated with biopsychosocial characteristics in children with multisite pain
- Abnormal eye movement, brain regional homogeneity in schizophrenia and clinical high-risk individuals and their associated gene expression profiles
- Transcriptome Analysis of Non-coding RNAs and mRNAs in the Dorsal Root Ganglion of Peripheral Nerve Injury-Induced Neuropathic Pain
- Analyzing single-cell transcriptome sequencing data reveals immunological mechanisms of microglia-associated diseases
- TS-GOEA: a web tool for tissue-specific gene set enrichment analysis based on gene ontology
- Deep functional synthesis: a machine learning approach to gene functional enrichment