Although the genetic model findings from this study have not been replicated using identical SNPs in an independent sample, they do generate some interesting hypotheses and provide evidence for the possible role of the axon guidance pathway in PD. As such, our study represents a paradigm shift from single-gene to pathway analyses of a complex disease. Replication of genetic association findings can be challenging, particularly when the effect sizes are small [18
]. Although a second whole-genome association study of PD was recently published and has data that could be similarly mined to test the hypothesis that the axon guidance pathway is associated with PD, that dataset included only 16% of the SNPs included in the dataset that we utilized, precluding an exact replication of our models [11
]. Nevertheless, for the 32,192 SNPs common to the two studies, we identified three SNPs that were associated with PD in all meta-analysis strategies. One of the SNPs was in the axon-guidance pathway gene UNC5C,
which encodes a netrin receptor [28
Furthermore, within the secondary whole-genome association dataset [18
], we were able to construct predictive models for PD susceptibility, survival free of PD, and predicted age at onset of PD. All three models in the secondary dataset were also highly significant. This is remarkable, as the two datasets employed independent samples and different study designs. Namely, the study of Fung et al. employed fewer participants, unrelated controls, and more and different SNP markers compared to the study of Maraganore et al. [11
]. This would further attest to the robustness of our predictive models. For validation purposes, we restricted the analyses of the secondary dataset to the genes that were represented by SNPs in each of the three predictive models for the primary dataset. This biases the findings for the secondary dataset conservatively, as it is possible that additional axon-guidance pathway genes contribute to the PD outcomes in that study sample. A more conservative approach would have been to limit the analysis of the secondary dataset to either the same SNPs or to SNPs from the same linkage disequilibrium bins, as for the first study. However, because there was only limited overlap in SNPs between the two studies (16%), and because many of the SNPs employed by the two studies are not included in a common haplotype map, this would not be possible for all of the genes in the models. Nevertheless, future studies of the two whole-genome association datasets may wish to consider additional meta-analytic strategies [28
We note that the magnitudes of the genetic effects were sizeable. Within our primary whole-genome association dataset [11
], persons with high versus low model scores differed in susceptibility by 90-fold and in age at onset by more than 20 years, and polymorphism in the axon guidance pathway accounted for nearly 70% of the age at onset variance of PD. Similarly large effects were observed for the secondary dataset [18
]. The findings of a study are more likely to be true when the effect sizes are large; however, we acknowledge that limited sample size, multiple tested relationships, and flexible analyses may have reduced the positive predictive value of the large effects that we observed [29
]. Although the p
values observed for our predictive models were small, these were the values determined by the actual data in independent datasets, and bootstrapping resulted in CIs smaller than 1 × 10−27
for each outcome in the primary dataset.
Within our primary whole-genome association dataset [11
], models constructed from thousands of random selections of genomic SNPs predicted all three PD outcomes poorly by comparison to models constructed using axon guidance SNPs. Our models of survival free of PD and age at onset of PD did not predict age at study in sibling controls (nor for unrelated controls in the secondary dataset). This would again argue against a chance association of the models with PD outcomes within the samples. We note that the models constructed using axon guidance SNPs were robust and were not crippled by the elimination of SNPs. For example, in our primary dataset, even after removing the first ten SNPs to enter the models (via the stepwise selection process), the overall model p
values were smaller than 1 × 10−15
for each of the outcomes (unpublished data). This indicates that the results were not driven by a few SNPs and that the joint effects of multiple SNPs were important.
Our analyses of an available gene-expression profiling dataset revealed that axon-guidance pathway genes were differentially expressed in the brains of PD cases as compared to controls, including several genes that were predictive of PD in the models constructed from the primary whole-genome association dataset (convergence of findings) [11
]. Not all genes included in the models constructed from the primary whole-genome association dataset were abnormally expressed in the gene-expression profiling dataset, and not all axon-guidance pathway genes that were abnormally expressed in the gene-expression profiling dataset were predictive of PD outcomes in that whole-genome association dataset (unpublished data). This may be because: (1) for some genes, the whole-genome association dataset had no data (no informative SNPs); (2) for some genes, the gene-expression profiling dataset had no data (no informative probe sets); (3) the PD brains included in the gene-expression profiling dataset were from end-stage patients, and it is possible that for some axon-guidance pathway genes differential expression occurs only earlier in life; (4) the genes included in the predictive models may have been associated with PD via a mechanism other than differential expression (i.e., qualitative rather than quantitative effects); (5) there are multiple other possible limitations intrinsic to the whole-genome association and gene expression profiling studies, which may have led to false positive and false negative results (divergence of findings) [11
Although we observed differential expression of axon-guidance pathway genes in an available gene-expression profiling dataset, we have not as yet demonstrated the biological effects of specific SNPs included in the predictive models in either human gene transcription or translation datasets or in experimental cell or animal models (construct validity). We emphasize that the SNPs that were independent predictors of susceptibility, survival free of PD, or age at onset of PD in either whole-genome association dataset may be only indirect markers of the functional variants within the genes. Fine mapping of these gene loci in diverse populations and experimental studies in transfected cell lines or in transgenic animals may help to identify functional variants within the genes and to elucidate the pathogenic mechanism of the axon-guidance pathway variability that we observed [7
]. It would be of interest to genotype a large postmortem cohort for the SNPs that were predictive of PD outcomes and to directly correlate the ORs of those SNPs to the fold changes for matching RNA fragments (probe sets).
We find it plausible that polymorphism in the axon-guidance pathway genes could lead to aberrant trajectory of the ascending dopaminergic pathway during embryonic brain development and in turn cause congenital deficiency in nigrostriatal dopamine. In other words, some persons could be “wired differently” from birth and could thus be at greater risk to develop PD and at an earlier age. Animal experiments provide support for this pathogenic mechanism. The study of embryonic brain development in knockout mice has suggested a critical role for axon guidance proteins, including semaphorin and slit proteins, in the guidance of the ascending trajectory of dopamine pathways from cell bodies in the substantia nigra to the ipsilateral striatum [13
]. By the 16th day of embryogenesis, mutant mice deficient in semaphorin and slit proteins exhibit a loss of midline repulsion of ascending dopamine pathways, so that the fibers ascend aberrantly to the contralateral striatum. This results in a visible reduction in the dopamine innervation of the striatum in semaphorin and slit deficient mice. Striatal dopamine deficiency is the neurochemical hallmark of PD [32
Similarly, transgenic mice with disrupted ephrin signaling exhibit a 40%–50% quantitative reduction in nigrostriatal innervation and exhibit classical behavioral model characteristics of PD [33
]. There is also some indirect support of this pathogenic “miswiring” mechanism in humans. Recently, a rare sequence variation in a single axonal-dendritic development gene was associated with Tourette syndrome, another disorder involving dopamine neurons [35
]. Furthermore, developmental anatomic asymmetries in nigrostriatal innervation resulting from polymorphism in axon-guidance pathway genes could account for motor asymmetries later seen in patients with PD [36
Another possibility is that axon-guidance pathway molecules contribute to the pathogenesis of PD via other mechanisms. For example, these proteins also play an important role in axonal repair and in apoptotic signaling. Polymorphism within the pathway may make neurons more or less vulnerable to endogenous or exogenous toxins that trigger the cell death response [37
]. The continued expression of some axon-guidance pathway genes, such as netrin-1
in nigral and striatal neurons throughout the mammalian life cycle (including embryogenesis and adulthood) would support multiple possible functions, and it has been suggested that netrin or netrin-like molecules may be of use in understanding and treating PD [10
]. Prior knowledge of many biological mechanisms by which the axon guidance pathway may contribute to the pathogenesis of PD increases the probability that our genetic findings may be true (warranting validation studies) [29
Significance and Innovation
In summary, we employed a genomic pathway approach to determine whether polymorphism in a candidate pathway (axon guidance) predisposed to a complex disease (PD). We found that multiple SNPs in axon-guidance pathway genes were strong predictors of PD susceptibility, survival free of PD, and age at onset of PD in two independent whole-genome association datasets; and many axon-guidance pathway genes were differentially expressed in PD. Although an exact replication of the predictive genetic models remains to be done (same SNPs, independent samples), we consider these findings timely and potentially important. To date, predictive biomarkers for PD are lacking. Our genetic findings for the axon guidance pathway and PD might suggest a new environmental focus on exposures that occur during intrauterine life (miswiring hypothesis).
Our findings would also provide evidence that complex diseases such as PD can be due to the joint effects of many genes that, taken singly, would show only small effects (additive effects model and epistasis) [40
]. While the available datasets were exploratory and may have been underpowered to detect the main effects of single SNPs or probe sets, our findings are consistent with recent studies that report greater statistical power to detect the joint effects of multiple loci than the main effects of single loci [44
]. By contrast to familial aggregation and twin studies, our findings demonstrate that a largely sporadic disease such as PD can in fact have a strong genetic component [4
] and suggest that a similar genomic pathway approach might provide insights into several other complex diseases. Finally, these findings provide important insights regarding the molecular mechanisms that may control dopamine circuit formation and programmed cell death in healthy versus diseased individuals. This in turn may facilitate the development of treatments for axonal synaptic regeneration and repair and for neuroprotection in PD [17
]. We envision that a systematic study of all defined genomic pathways will yield additional important findings for PD, even in the absence of prior evidence of strong single-gene association.