Primary sensory neurons in DRGs undergo a massive change in gene expression in experimental models of neuropathic pain, changes that undoubtedly contribute to altered somatosensory signal processing in the event of neuropathy [2–11]. In our material, >11% of all expressed genes, 2552 on average, were found to be significantly up- or down-regulated in the L5DRG 3 days following transection of the L5 spinal nerve. This number is larger than prior estimates probably because of the proximal location of the lesion, and the improved protocol for detecting change. In common with prior estimates, however, it highlights a fundamental challenge. The sheer number of regulated genes means that additional methods are needed to identify which individual transcripts play an important role in pain sensation versus other processes induced by axotomy . Using correlational analysis we considerably shortened the list of candidates. In contrast, we found that very few genes were regulated in the L4DRG following transection of the L5 spinal nerve. This is surprising in light of functional studies that have argued that abnormal activity in "uninjured" L4DRG nociceptors may be important for the development of tactile allodynia in the SNL model . Perhaps the relevant changes occur in sensory signaling processes not related to altered gene expression in the L4DRG. The observations on the L4DRG confirm that the massive regulation seen in the L5DRG was in fact related to axotomy.
We recognize that the shortened gene list still includes many transcripts unrelated to neuropathic pain. Assuming random assortment, 5% of transcripts are expected to correlate with pain phenotype at p ≤ 0.05 by chance alone. However, despite residual contamination transcripts whose regulation is functionally related to pain behavior are expected to be retained at higher frequency than pain-neutral transcripts, yielding significant enrichment. At the most optimistic, assuming that all functionally relevant genes were captured by our correlation procedure, at least 95% enrichment is expected. Thus, while the lists of correlated transcripts remain too long to permit direct selection of candidate pain genes, they constitute a valuable and relatively bias-free look-up table, a tool for screening candidates derived from other approaches. The list, however, is clearly not complete. In addition to pain-correlated genes missed, there is no inherent biological contradiction to the possibility that some genes are functionally related to pain phenotype but that their regulation does not contribute importantly to strain differences.
Might the gene lists be further enriched by increasing the stringency of the selection criterion from p ≤ 0.05 to, say, p ≤ 0.01? While not impossible, given the intrinsic noise present in the phenotyping process we believe that this step would be more likely to eliminate functionally significant transcripts. This is because within the enriched list, genes with the highest correlation coefficients are not much more likely to be functionally related to pain than those with lesser, but still high correlation coefficients. Further vetting requires additional biological information. For example, there is a clear benefit to increasing the number of data points used to generate correlations. We accomplished this in three of the five strains studied by collecting phenotypic data (on tactile allodynia) from individual mice rather than relying on strain means. This resulted in a reduction in the number of genes with significant correlation coefficients, presumably by removing more false positive results than true positive results. Related approaches are to increase the number of mouse strains studied and the behavioral diversity among them.
Two other enrichment approaches have been attempted in the past. In one, lists of regulated transcripts were assembled using a variety of different rodent models of painful neuropathy [11, 20, 21]. As expected, for each model hundreds of transcripts were significantly regulated in the axotomized DRGs. Lists were shortened by identifying genes similarly regulated in more than one model. This strategy is likely to reduce random noise, but it is also likely to capture regulated genes that are not related to pain. Transcripts related to regeneration and apoptosis, for example, would all be positively selected for. The approach also makes the risky assumption that different pain models (diagnoses) share the same underlying pathophysiology. If this is not the case then pain-related genes would be systematically excluded.
A second approach compared genes regulated in the DRG following nerve injury in two rat strains that had a consistent difference in pain phenotype in the SNL neuropathy model . Although closer in concept to our study, the rat strains used were not congenic, or even particularly close in genetic background. For this reason observed differences in gene regulation may well have reflected pain-neutral differences in genetic background. Correlations based on only two strains provide zero degrees of freedom. We used five strains, and alternatively 26 individual mice (from three strains) to generate correlations. In this context it is worth noting that not only does baseline gene expression level within a particular tissue (e.g. DRG) vary among mouse strains, but it also varies among tissues and among cell types within tissues [22, 23]. Our in situ hybridization data affirm this to be the case with respect to DRG neurons of varying size, at least for some transcripts. An example is up-regulation of Trpv1 in CBA mice and down-regulation in B6 mice, and the de novo appearance of Trpv1 in DRG cells of medium and large size, but only in CBA mice.
An intrinsic limitation of microarray based studies, including ours, is that massive multiple hypothesis testing undermines the ability to establish statistical significance for any given transcript identified. This limitation is only partly eased using FDR analysis . Ultimately, our enriched list of pain-related genes needs to be subjected to secondary screens. In situ hybridization is one example. Although this method is too resource intensive to be used to screen large numbers of transcripts, it represents an independent implementation of correlational analysis. The overall pattern of gene regulation seen in the arrays was reiterated in all six candidate genes subjected to in situ analysis. However, the across-strains pattern was reproduced in only two of the six. There are several possible reasons for this difference between microarray and in situ measurements. Microarray (and TaqMan) analysis integrates over the entire ganglion, including neurons, glia and other resident cells, and includes cells with low expression levels. The cellular source of the mRNA is not identified. Likewise, the method does not take into account the possibility of up-regulation in one cell population balanced by down-regulation in another. Finally, stable expression levels in many cells may mask significant regulation in a small but important subpopulation.
Interestingly, the degree of regulation correlated significantly with pain phenotype for two of the six candidate genes studied with in situ hybridization (33%;Scn11a and Trpm8). This yield was much higher than for the array analysis which considered all regulated genes (p < 0.001). This outcome, which presumably reflects the additional information that underlay the choice of the six candidates, lends validation to the correlational approach. Regulation of the Na+ channel α subunit Scn11a correlated with levels of spontaneous pain behavior, and regulation of the cool receptor Trpm8 correlated with heat hypersensibility. For Scn11a, expression was reduced in all strains, with the degree of reduction minimal in animals that exhibited a high level of ongoing pain behavior and maximal in strains with minimal ongoing pain. A good deal of evidence links Na+ channels, including Scn11a, to the emergence of ectopic afferent hyperexcitability after nerve injury and consequent spontaneous firing and spontaneous neuropathic pain [1, 16, 25]. The observed correlation is therefore consistent with pain-protected mouse strains showing the greatest level of Scn11a down-regulation. We note, however, that mice with null mutation of Scn11a continue to show tactile allodynia in neuropathy models [26, 27]. These knockout mice have not been checked for autotomy.
For Trpm8, down-regulation occasioned reduced heat allodynia. Trpm8 functions as a cold transducer in primary afferent nociceptors . The relation to thermal sensation is obvious, but the link to altered heat response requires further investigation. Cold allodynia in the SNL model has not been compared in our series of mouse strains. In intact mice, however, sensitivity to cold is genetically correlated with sensitivity to heat . We stress that failure to find a significant correlation between gene regulation and pain phenotype in the other four transcripts tested using in situ hybridization does not mean that the corresponding genes, or their regulation following nerve injury, is not important for pain phenotype. Each of these genes may well play an essential role in pain physiology. However, the result indicates that their degree of regulation does not contribute much to across strain variability, at least in the five strains examined.