Introduction

Traditional classification of canine hip dypslasia (CHD) is reliant upon phenotypic assessment, namely by radiography and Norberg-Angle measurement, of the hip joints of full-grown animals. Drawing upon the outcome of these examinations, most European breeding organizations have adopted a classification scheme established by the Fédération Cynologique Internationale (FCI) that specifies five grades of CHD severity (A-E). Class A corresponds to dogs with no signs of hip dysplasia whereas class E indicates presence of severe hip dysplasia. The classification process is laborious and costly but currently represents the only means to acquire useful information for animal selection aimed at reducing the incidence of CHD.

Breeding registers gather data from confirmed pedigrees and combine the CHD classification of individual animals with that of close relatives to yield so-called ‘estimated breeding values (EBVs)’. EBVs are specific to a phenotypic trait of interest and express the genetic merit of a given animal in units of that trait, relative to a certain base. For selection purposes, the absolute value of an EBV is usually not important but rather the differences in EBV between animals. An EBV-based selection program to combat CHD was started in 2002 by the Society for German Shepherd Dogs (Verein für Deutsche Schäferhunde, SV) for all their stud animals. Phenotype evaluation has been mandatory in the program if studbook-documented breeding was aspired to, and only dogs rated grade A to C got unrestricted breeding permission. Among the dogs enlisted in the EBV-based selection program, the prevalence of CHD class A animals had increased from 62% to 75% by the end of 2016.

Various attempts have been made to identify genetic loci underlying CHD development with the goal to develop a DNA test that would outperform, and eventually obviate, phenotype-based selection for CHD. Thus, Guo et al. [1] devised a genomic breeding value (GBV) that was similar to approaches taken in livestock breeding programs before. A learning cohort of 359 animals was characterized for CHD by way of traditional EBVs and genotyped for some 22,000 single nucleotide polymorphisms (SNPs). Based upon the SNP allele distribution and the available EBVs, a GBV formula was developed that resulted in 70% to 90% correlation between GBVs and EBVs. For validation, the formula was also applied to 38 dogs previously classified for CHD and typed at roughly 13,000 SNPs [1]. The resulting GBVs showed moderate correlation (around 50%) with CHD class, and the sensitivity, specificity, positive and negative predictive values of the GBVs all lay in the range of 70% to 75%. Notably, the correlation between GBV and EBV remained virtually unchanged when the calculations were confined to the most informative 100 to 500 SNPs. To increase CHD selection response, the authors suggested that EBV be replaced by GBV, but at the same time emphasized that progeny phenotyping and EBV calculation should continue so as to allow improvement (i.e. retraining) of the GBV formula, if required.

A different approach to CHD prediction was taken by Distl et al. [2]. Their assessment of genetic CHD predisposition relied upon only 17 polymorphisms (16 SNPs, 1 indel) previously reported as being associated with CHD in German Shepherd Dogs [3]. According to the patent (European Patent Specification EP 2 123 777 B1) filed for the marker set and for the algorithm to compute a GBV from animal-specific marker genotypes, each marker is indicative of the genetic disposition for CHD. Homozygous wild-type genotypes were ascribed a positive (beneficial) effect while homozygous mutant genotypes were ascribed a negative (deleterious) effect. Heterozygous animals are predisposed according to a marker-specific dominance effect that usually places them somewhere in between the two homozygous states. The contribution of each individual marker in the form of numerical weights is part of the patent claim. Summation of the marker-specific weights yields the GBV of a given animal which is supposed to reflect its tendency to develop and pass on CHD. The individual numerical CHD risk is determined by gauging the GBV against a reference curve. The inventors also alluded to the possibility of predicting the risk of CHD in the offspring of genotyped dogs (section [0022]) but did not make this an explicit part of patent claim. Notably, the patent specification does not include any information about a possible validation of the GBV and, to the best of our knowledge, no such validation has been reported in the scientific literature.

A procedure similar to that proposed by Distl et al. [2], but targeted at Labrador Retrievers, was recently devised by Bartolomé et al. [4] and a patent was filed for this procedure as well (EP 2619319 A2). Prediction of CHD in Labrador Retrievers is based upon only seven SNPs that were identified in the course of a genome-wide association study with the Illumina canine HD BeadChip. SNP genotypes served as independent variables in a standard logistic regression analysis, and the final regression model reportedly achieved good predictive capability in an independent validation cohort. The area under curve equaled 0.8 whilst the combination of sensitivity and specificity ranged from 93%/58% via 80%/78% to 42%/95%. The good performance of the test was attributed to the fact that the seven SNPs were closely linked to genes that are thought to directly influence hip joint development. The authors also announced further validation of the test in Labradors and other breeds.

Despite the fact that the marker set and GBV algorithm proposed for German Shepherd Dogs by Distl et al. [2] obviously have never been validated, the procedure was proposed as the long sought-after DNA-based CHD test for this breed. Therefore, the Verein für Deutsche Schäferhunde e.V. (German Society for German Shepherd Dogs, SV) initiated a prospective study among their members to systematically investigate the validity of the claims made in connection with patent EP 2 123 777 B1 [2]. The study was carried out between 2012 and 2015 under the auspices of the SV. Here, we report on the design, results and implications of these investigations. CHD is morphologically similar to another detrimental bone condition, namely canine elbow dysplasia (CED). Therefore, an explorative evaluation of the 'predictive capability' of the investigated GBVs for CED was attached to the CHD study as a negative control not hitherto linked to the SNPs in question.

The present study was devised to validate the patented marker set and GBVs according to their intended use in individual dogs, as laid down in the patent claims. An originally planned multi-generational follow-up of the included dogs was abandoned as a consequence of the poor performance of the test in the present study. In fact, section [0022] of the patent contains a description of how to use the markers and GBVs to predict CHD in the prospective offspring of pairs of genotyped dogs as well. To this end, the patent proposes calculation of the probabilities of all possible offspring genotypes, assuming Mendelian inheritance, followed by subjecting the potential offspring genotypes to the same CHD risk assessment as described for experimentally determined genotypes. The weighted average of these genotype-specific risks is then taken as the overall CHD risk of prospective offspring. Since the same method of CHD risk assessment is used in both instances, however, the feasibility of risk calculation for prospective offspring would require, at the very least, the feasibility of risk calculation for singletons.

Materials and methods

Markers

PCR primers for the amplification of 17 CHD-associated markers were designed according to the specification of patent EP 2 123 777 B1 (S2 Table). However, allelic and flanking DNA sequence information was adopted only if it could be confirmed experimentally in our own laboratory. In fact, it turned out that the DNA sequence at four of the markers (TiHo1, TiHo18, TiHo19 and TiHo35) had to be revised because the information given in the patent specification was wrong (see Table 1 for the necessary adjustments).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Revised DNA sequence context of four CHD markers (based upon own experiments).

https://doi.org/10.1371/journal.pone.0182093.t001

Results

CHD and CED status

A total of 935 dogs were assessed in terms of their CHD status (Table 2). For 928 of these, diagnostic information was also available for CED. Most animals were affected by neither disease (phenotype class 1). More specifically, 657 dogs (70.3%) were free of CHD whilst 714 (76.9%) lacked signs of CED. Of the 287 animals suffering from CHD, 179 (62.4%) showed a mild phenotype (class 2) whereas 25 (8.7%) were diagnosed with severe CHD (class 5). The remainder had a mild to moderate phenotype (class 3: 53, 18.5%; class 4: 21, 7.3%). In the case of CED, 70 of 214 affected dogs (32.7%) suffered from a mild form (class 2) whereas 37 (17.3%) had severe CED (class 5). Some 39 (18.2%) and 68 (31.8%) animals were diagnosed with mild to moderate CED (class 3 or 4, respectively). The two conditions were found to be significantly associated with one another in that 525 dogs (56.6%) had neither condition whilst 82 (8.8%) had both (OR: 1.73, 95% CI: 1.25–2.38).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Distribution of canine hip dysplasia (CHD) and elbow dysplasia (CED).

https://doi.org/10.1371/journal.pone.0182093.t002

Phenotype prediction

As was to be expected owing to the order-preserving nature of standardization transformations, no notable differences in terms of GBV or GBV(S) were observed between affected and non-affected animals for either disease. Thus, the median GBV value was equal to 1.018 (inter-quartile range IQR: 4.055) among dogs diagnosed with CHD, compared to 1.442 (IQR: 3.346) in the unaffected group. Contrary to expectation, the genomic breeding value GBV, which is supposed to be positively correlated with CHD risk, was therefore decreased rather than increased in affected animals, although the difference between the two groups was not statistically significant (Kruskal-Wallis χ² = 1.170, 1 d.f., p>0.2). Virtually identical results were obtained for GBV(S) (data not shown). In the case of CED, the results were also negative as expected in that the median GBV was 1.378 (IQR: 3.090) for affected and 1.320 (IQR: 3.670) for unaffected animals (Kruskal-Wallis χ² = 0.970, 1 d.f., p>0.3).

When the predictive utility of GBV for CHD was assessed by way of logistic regression analysis, an OR increase per unit GBV increase of 0.98 emerged that was not however significantly different from unity (95% CI: 0.93–1.03). Moreover, the AUC of the predictive test based upon the logistic-regression-derived CHD risk amounted to a minuscule 0.523.

Upon closer inspection of the relationship between the phenotypic classification and GBV, a notably negative trend became apparent for CHD (Fig 1A), but not for CED (Fig 1B). With increasing CHD classification (from class 1 to class 5), the median GBV decreased from 1.442 via 1.134 (classes 2, 3 and 4 combined) to -1.234 (class 5). If anything, GBV as proposed in patent specification EP 2 123 777 B1 therefore provides a predictive ‘test’ of minor utility (post hoc AUC: 0.622) for severe CHD, but only after inverting the interpretation of the test result.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Relationship between genomic breeding value GBV and severity (phenotypic class) of CHD (Fig 1A) and CED (Fig 1B).

https://doi.org/10.1371/journal.pone.0182093.g001

In view of the poor performance of GBV and GBV(S), we tried to develop a post hoc predictive model for CHD, based upon the genotypes available for the 17 SNPs in question. To this end, each marker genotype was encoded by a combination of two variables, namely the dosage of the respective CHD-associated allele (0, 1 or 2) and a covariate for the dominance effect equal to 1 for heterozygotes and 0 for homozygotes.

Even the full predictive model including all additive and dominance effects of the 17 markers yielded only minor predictive utility for CHD (AUC: 0.610). Stepwise selection of predictors as implemented in SAS procedure LOGISTIC, and employing a significance threshold of 5%, the final model comprised only the allele dosage of markers TiHo25 (score χ² = 5.794, 1 d.f., p = 0.016) and TiHo26 (score χ² = 5.865, 1 d.f., p = 0.015). The estimated OR per HD-associated allele was 0.78 (95%CI: 0.61–0.95) for TiHo25 (allele G) and 1.41 (95% CI: 1.14–1.75) for TiHo26 (allele C). No other significant additive or dominance effects were observed. The post hoc predictive utility of the final model was however found to be poor (AUC = 0.567).

An analogous analysis carried out for severe CHD (class 5) also included the allele dosage of TiHo25 in the final model (score χ² = 6.525, 1 d.f., p = 0.011), with an estimated per-allele OR of 0.54 (95%: 0.33–0.88). In addition, two highly significant dominance effects were observed, namely for TiHo12 (score χ² = 17.614, 1 d.f., p<0.0001) and for TiHo18 (score χ² = 7.249, 1 d.f., p = 0.007). Taken together, the three genetic risk factors yielded moderate to high predictive utility (post hoc AUC: 0.793). The two dominance effects were indeed remarkable. Whilst none of the 21 wild-type homozygotes for TiHo12 and only 9 of the 660 homozygotes for CHD-associated allele T developed severe CHD, this was true for 16 of 254 heterozygotes. Considered post hoc, this would imply a relative CHD risk of 4.77 (95% CI: 2.13–10.65) for genotype TC. For GC heterozygotes at marker TiHo18, a relative CHD risk of 2.94 (95% CI: 1.28–6.75) became apparent.

Discussion

The genetic architecture of complex diseases like CHD differs fundamentally from that of monogenic disorders. Whereas the latter, by definition, are due to a few genetic changes with high penetrance, the former result from the interplay of a large and unknown number of environmental and genetic factors, most of which have small effects [6]. Nevertheless, given the rapid technological development in genetics research, it can be expected that the vast majority of genetic loci contributing appreciably to a complex disease, in humans as well as in non-human mammals cared for by humans, will be identified in the not too distant future [6]. Even though unraveling statistical genotype-phenotype relationships is only a first step towards a better understanding of disease etiology, such associations can nevertheless be of immediate practical benefit in that they may form the basis of diagnostic or prognostic tests. This includes the prediction of canine diseases such as CHD and CED.

A number of CHD-associated genetic loci have been identified in the past, and the 17-marker panel proposed by Distl et al. [2], for which a European patent (EP 2 123 777 B1) was filed in 2009, has been promoted as the most promising candidate for a predictive test for CHD so far. However, to the best of our knowledge, this panel has never been validated in an independent animal population. Furthermore, no studies of the practical applicability and utility of the patent claims have been published in the scientific literature so far. Remarkably, the same group has published other genetic markers of CHD [7,8], but the only reference to the patent made in these articles was mentioning Marschall & Distl [3] in the course of introducing the literature on CHD-associated markers. The patented markers themselves were not experimentally investigated further.

Under the auspices of the German Society for German Shepherd Dogs (SV), a prognostic study was therefore performed to investigate the predictive capability of the patented SNP panel for CHD in a large, independent and population-representative sample. The primary intention of the study was to assess the capability of the SNPs to predict individual CHD development. An extension covering the prediction of CHD in the offspring of genotyped dogs was scheduled for after successful completion of the first study phase, but was later abandoned.

In addition to the technicalities of the CHD-associated SNPs involved, the patent specification also contains instructions for calculating a standardized genomic breeding value ‘dev’ (here abbreviated as GBV(S)) from a given SNP profile, which in turn can be used to quantify the individual risk of developing CHD by gauging GBV(S) against a reference curve that was also included in the patent description. According to the instructions given, GBV(S) is expected to range from—0.5, reflecting the most beneficial genetic disposition, to +0.5 for animals at the highest disease risk. Such a range would result if the non-standardized breeding value GBV followed a symmetrical distribution. In our study, by contrast, GBV(S) was found to range from -0.613 to +0.387 owing to the pronounced skewness of GBV as observed in the data. Most importantly, however, neither GBV nor GBV(S) was found to bear any appreciable statistical relationship to phenotypic CHD.

The utility of a predictive test with a continuous outcome such as, for example, GBV or GBV(S) can be assessed by way of the so-called ‘receiver-operating characteristic curve’ (ROC). The ROC relates the probability of true positive results to the probability of false positive results that would ensue from using distinct thresholds to dichotomize the test outcome. The AUC sensibly ranges from 0.5 (equivalent to tossing a coin, i.e. total lack of predictive capability) to 1.0 (perfect prediction) and is independent of the prevalence of the disease in question. According to Janssens et al. [9], an AUC of around 0.8 is usually regarded as both necessary and sufficient for screening purposes. However, the genomic breeding value GBV proposed by Distl et al. [2] achieved an AUC of only 0.523 for CHD in our study. Since this rather poor performance, in principle, could have reflected inadequacy of the marker weights used for the definition of GBV, we subjected the SNP genotypes to post hoc logistic regression analysis, allowing for all possible additive and dominance effects. Nevertheless, even the full regression model, including all 17 markers, yielded an AUC of just 0.610 in our study. Moreover, this result was found to hinge essentially on just two markers, namely TiHo25 and TiHo26. All other markers lacked even minuscule predictive value for CHD. Notably, when applied to CED, the 17 SNPs also lacked any appreciable predictive value as was to be expected because CED served as a kind of negative control in the present study.

With hindsight, our results may not have been surprising because, despite the uncontested prospects of genomic selection in livestock and plant breeding [10], the vast majority of genotype-phenotype relationships in complex diseases studied so far in humans as well as in non-human species have turned out to be of little, if any, diagnostic or prognostic value [11]. This is why independent validation in representative population samples should be a conditio sine qua non of any practical use of tests exploiting such relationships. While emphasizing their potential usefulness, Wray et al. [10] also highlighted the necessity of prior validation of SNP-based genomic breeding values before they can be sensibly applied for selection purposes in the context of complex traits. If this advice had been followed in the present case, the poor performance of the CHD test proposed by Distl et al. [2] would have become apparent well before its attempted commercialization.

Instead of basing CHD prediction upon a small number of pre-selected markers, Guo et al. [1] proposed the inclusion of all available genomic information in a prognostic model, an approach that recalls strategies in livestock and crop breeding programs for quantitative traits alluded to above. In their ‘agnostic’ approach, all SNPs on a given chip that turn out disease-associated in a ‘learning population’ are eventually included in the CHD risk calculated from the genetic profiles of other animals. Interestingly, Guo et al. [1] also suggested constant updating of the respective SNP panel on the basis of subsequently acquired phenotypic and CHD-specific genetic information. This approach would clearly contradict the validation requirement emphasized by Wray et al. [10] and therefore highlights the difficulties in finding a consensus on how to genetically predict complex traits.

Conclusions

In a three years prospective study of nearly 1000 animals, we have evaluated and shown invalid a prognostic, SNP-based test for canine hip dysplasia (CHD) that was patented [2] based upon data from a previous scientific publication [3]. The test has since been propagated to various potential clients as a suitable means of animal selection in practical breeding, including the German Society for German Shepherd Dogs (SV) who acted as main sponsor of the present study. In its original form, the SNP-based test turned out to be completely incapable of predicting CHD in our target population. If anything, some of the respective SNP may be predictive of severe CHD, but such a combined test would require validation in another independent study to be credible. For the time being, practical application of the SNP-based test proposed for the prediction of CHD in German Shepherd Dogs is thus not recommended.

Our assessment of an alleged genetic test for CHD is the first systematic and independent attempt to validate a DNA-based predictive procedure meant for practical use in a non-human mammal species. In this respect, we regard our work as exemplary for the caution required towards any genetic test claimed to predict complex multifactorial traits in the field. Whatever evidence apparently underpins its predictive capability in the original report, a predictive test should be considered fit for practical use only after it has been tested successfully and independently in the intended target population.

Supporting information

Acknowledgments

This study was supported by Verein für Deutsche Schäferhunde e.V. (SV).