2.1 Data

A cohort of 704 FT has been recruited from their trainers during 2016, 2017 and the beginning of 2018. The trainers responded to a questionnaire for each horse that included 13 trot characteristics. The questionnaire was completed for 83% of these horses (586). From this cohort, 90% (637) horses were genotyped using the Affymetrix Axiom Equine genotyping array with 670,806 single nucleotide polymorphisms (SNPs). The horses of this sample were born between 2001 and 2016, with the majority born between 2009 and 2013 (87%). Among the horses, 43% were females, 41% were geldings and 16% were males.

From earlier studies (Brard & Ricard, 2015; Ricard, 2015), a sample of 686 horses had been previously genotyped with Illumina Equine SNP50 chip (54,602 SNPs). This cohort of horses was assessed with a protocol which aimed to identify osteochondrosis markers. It consisted of 98 sires born before 2003 and 588 progeny tested for osteochondrosis and born mostly from 2002 to 2008. They were 39% females, 31% geldings and 30% males.

The final sample, genotyped or gait surveyed, contained 1,390 FT, from which 1,323 were genotyped (95%) and 586 had a completed questionnaire (42%). Ultimately, 519 horses (37%) were both genotyped and assessed with a questionnaire. Table 1 summarized the data.

LeTrot, the French association in charge of the development of trotting races in France, and of the protection of the FT breed in its specificity, provided performance data in races for FT. Performances included all results in races from 1996 to 2018. There were 252,368 FT-born from 1994 to 2015. Among them, 38% were qualified and, therefore, authorized to participate in races.

Pedigrees were provided by IFCE (Institut Français du cheval et de l’Equitation), the French technical institute in charge of the equine industry, including breeding. Over 6 generations of pedigrees of all trotters born between 1994 and 2015 represented a total of 314,396 horses.

2.2 Imputation

To examine all 1,323 genotyped horses, imputation was performed using Fimpute 2.2 software (Sargolzaei et al., 2014). First, quality control removed SNPs, which did not reach quality test requirements. The tests used included the minimum allele frequency test (MAF) ≥ 5%, the Hardy–Weinberg p-value test ≥ 10^–6, call rate ≥ 90%, a valid position on autosomal chromosome and p-value of test of different MAFs between the two chips ≥ 10^–5. Reference map used was that of Equine 3.0 reference sequence if the SNP position is available or Equine 2.0 if not. After quality control, 377,611 SNPs were retained: 0.4% solely on the Illumina Equine SNP50 chip, 89.6% on the Affymetrix Axiom Equine chip and 10.0% on both. Imputation was performed on all these SNPs, adding the information of pedigree data for over 4 generations (6,905 horses). Quality of imputation has been already checked in Chassier et al. (2018). On a slightly lower sample of the present data, the mean concordance rate per individual was 0.9903 and the mean r² was .9905. The SNP of the DMRT3 mutation was only present in the Affymetrix Axiom Equine chip, and therefore, the allele was imputed for the 687 horses genotyped with the Illumina Equine SNP chip. The results of patented genotype at DMRT3 for the 637 horses genotyped with the Affymetrix Axiom Equine chip was given to the owners thanks to the permission and collaboration with EQUIBIOGENES and CAPILET GENETICS (https://www.equibiogenes.com/ and http://www.capiletgenetics.com/sv).

2.3 Model for racing performances

Racing performances were analysed using all the data for horses born from 1994 to 2015, including the 1,390 horses of the studied sample. Racing performance traits of horses used in the study were the same as previously described (Ricard, 2015):

A mixed linear animal model was used to analyse racing performances. Fixed effects included a combination of year of birth and gender effect (44 levels) and a cohort effect (3 levels: horses recruited for osteochondrosis analysis, horses recruited for questionnaire data and other horses). Random effect included the animal genetic value. The relationship matrix between genetic values was constructed using both genealogy data and genomic data in a single-step GBLUP method (Aguilar et al., 2010; Christensen & Lund, 2010; Fernando et al., 2014; Legarra et al., 2014) and using the rules outlined by VanRaden (2008) for genomic relationship matrix. For FH3, FH4 and FH5, the performance was repeated several times. Therefore, a random permanent environmental effect was added to the model. The software used was BLUPF90 (Aguilar et al., 2018).

To detect significant SNPs, single-trait analysis was performed using previously estimated variance components (Ricard, 2015). GWAS was performed with back solutions of SNP effect, and p-value was calculated as described in Aguilar et al. (2019). Significant SNPs were identified using p-values. Genomic control (Devlin & Roeder, 1999) was applied to the p-values to prevent incorrect distribution of the test statistics. The effective number of independent tests in our data was calculated as the inverse of the mean of linkage disequilibrium (r²) between all available pairs of SNPs by chromosome (Li et al., 2012). Linkage disequilibrium between chromosomes was assumed to be negligible. The genome-wide significance was then set at 1% and divided by the number of independent tests. Significant SNPs were grouped according to the distance between them to define a lower number of significant regions. As long as distance between successive significant SNPs is <1,000 kb, they were assumed to be part of the same region. In these regions, the highest significant SNP was retained defining the quantitative trait nucleotides (QTNs).

The back solution of the SNP effect only provides the effect of allele. For the QTNs, a supplementary analysis was performed to estimate the effects of the 3 possible genotypes. The objective was to detect a possible dominance effect for the heterozygote type. For this analysis, genotypes were considered as fixed effects and only phenotypes for genotyped animals were included in the analysis. To consider all performances, phenotypes were computed as deregressed proofs calculated from the estimated breeding values and reliabilities obtained with the first full model without genomic data. Proofs were calculated according to Ricard et al. (2013). Then, genomic best linear unbiased prediction (gBLUP) (VanRaden, 2008) was used with ASReml software (Gilmour et al., 2014), which provided the Wald F statistics for the fixed effects, that is the genotypes. To perform gBLUP with ASReml, the inverse of the genomic matrix calculated with the BLUPF90 program was used. Results are the solutions of fixed effects and Wald F statistics for all QTNs and traits.

A cross-validation was used to compare different strategies of genetic evaluation to predict racing performances. The training set included all horses born until 2010 (n = 192,167) and their performance results until 2012 (first age for qualification of horses born in 2010). The validation set included all genotyped horses born from 2011 (n = 446) and their performance results until 2018. Three genetic evaluations were calculated as follows:

For the cross-validation, multitrait analyses were used. To reduce the size of the system of equations, several multitraits were used, including the following: (a) Q, LnEH3, LnEH4, LnEH5-10; (b) LnEH3, FH3, SH3; (c) LnEH4, FH4, SH4; (d) LnEH5-10, FH5-10, SH5-10; and (e) LnEH3, LnEH4, LnEH5-10, LnES3, LnES4, LnES5-10.

2.4 Model for trotting technique questionnaires

The questionnaire used included 13 questions about each of the horses. Questionnaire responses were converted into binary variables (0/1), with 0 representing the absence of default or characteristic and 1 representing its presence. They were as follows:

A mixed linear animal model was used to estimate heritability and the environmental and genetic effects on questionnaire responses. Fixed environmental effects included gender, birth year and trainer effect (23 levels). Random effects included the animal's genetic value. The relationship matrix between genetic values was constructed using genealogy data (9,092 ancestors over 6 generations). Probit transformation was used to transform binary variables. The software used was BLUPF90 (Aguilar et al., 2018). From this analysis, we constructed residual trotting technique variables equal to probit variable response to questionnaire, minus the estimated effect of gender, year of birth and trainer.

Principal component analysis was performed on the residual trotting technique variables using the FactoMineR package (Lê et al., 2008). The results led to the creation of three synthetic trotting technique variables to summarize the characteristics of each horse's trotting technique as measured by the questionnaire. The synthetic variable was constructed from the grouping of elementary variables. The synthetic variable was the sum of elementary binary variables capped at 1, 2 and 3 respectively. Elementary variables were regrouped as follows:

Synthetic variables were used for GWAS analysis with the same methodology described for racing performance variables. A mixed linear animal model was used. Fixed environmental effects included gender, birth year and trainer effect (23 levels). Random effects included the animal's genetic value. The relationship matrix between genetic values was constructed using genealogy data, and genomic data in a single-step GBLUP method and back solutions of SNP effect and p-value were calculated as described in Aguilar et al. (2019). Significant SNPs were identified using p-values after genomic control.

We tested the capacity of the trotting technique questionnaire to predict the genotype of the horse for the DMRT3 mutation. We used the random forest classification analysis using all questionnaire binary variable and the genotype at DMRT3 on the 519 horses, with both questionnaire and genotype using the “randomForest” function of the “randomForest” package in R (Breiman, 2001; Liaw & Wiener, 2018). The default parameters were used, that is the number of trees was 500 and the number of variables tried at each split was (). The training set included two-thirds of the sample chosen at random, and the remainder was used as the validation set. We performed 100 resamplings.

We tested the capacity of trotting technique questionnaire to predict racing performances. We used the SAS software SAS/GLM (SAS/STAT, 2019) to perform general linear model analysis on racing performances. The dependent variables were residuals of performances, that is racing performance variables, minus the estimated effect of the combination of the year and gender and of the cohort (see above for the full model used to analyse racing performance). The independent variables were the three synthetic trotting technique variables defined after the principal component analysis.

3.1 General statistics on the data

Genotyped horses were significantly more frequently qualified compared with the whole FT population: 88% versus 38%. Their racing performances were slightly higher than those of all racing horses: z-score around 0.6 for earnings (LnEH2, LnEH3, LnEH4, LnEH5-10, LnES3, LnES4, LnES5-10) but close to zero for all other traits (FH3, FH4, FH5, SH3, SH4, SH5). Horses with questionnaire were almost all qualified (98%). The superiority for racing performances of this surveyed sample was in the same order as that of the genotyped sample (z-score for earnings around 0.7 and close to 0 for other traits).

For DMRT3 mutation, the proportion of AA genotype in the sample of the 1,323 genotyped horses was 59%, and there were 37% CA and 4% CC, with A allele frequency of 77%. The A allele frequency increased with time (Figure 1).

3.2 GWAS on racing performances

The mean of linkage disequilibrium over the genome was determined to be 0.0005920. The number of independent tests was then estimated to be 1/0.0005920 = 1689. Then, the p-value threshold was set to 1%/1689 = 5.9 × 10⁻⁶ or −log(p-value) = 5.2. There were 195 SNPs which exceeded the threshold. These SNPs belonged to 25 regions on the genome (Table 2). There were 19 regions with only one significant SNP and 6 with more than one. Five regions with more than one SNP covered from 11,181 to 1,126,145 bp. The 6th region was a large region near DMRT3 involving 146 SNPs ranging from −2.2 to +3.6 Mb around the position 23:22,999,655. The DMRT3 mutation was not the highest p-value (−log(p-value) = 14.68) in the region but close.

3.3 Genotype effects of DMRT3 and QTNs on racing performance

The effects of DMRT3 genotypes on all racing performances are presented in Figure 2 expressing the difference with respect to the AA genotype. For qualification status (Q), the AA was significantly the most favourable genotype, followed by CA and then CC. There were no horses with CC genotype in races at 2 years old, and AA was the best genotype at this age compared with CA for earnings. Then, for earnings, from 3 to 5–10 years old, the advantage of AA compared with CA was not significant and the CA performed better than AA after 4 years, but with no significant effect, except for racing under saddle. The CC genotype was associated with the worst earning results, except for the LnES5-10, but this result was not significant. For the proportion of finished races, the genotype AA demonstrated significantly higher results at all ages compared with CC. Further, the probability of participating in a race was highest for the AA genotype.

Effects all the 25 QTNs on all racing performances are summarized in Figure 3. Each QTN was chosen according to p-value threshold 5.9 × 10⁻⁶ established in 3.2 for one trait, but in this figure the p-values reported for all traits were an uncorrected raw p-value. All QTNS had an effect on one of earnings traits (LnEH2, LnEH3, LnnEH4, LnEH5-10, lnES3, LnES4 or LnES5-10). Among them, 13 had an effect on both earnings in harness races and races under saddle, 9 on harness races only and 3 on races under saddle only. The same genotype was always in the same direction for all earnings traits: always positive or negative with the notable exception of DMRT3. The rare allele had positive effect on earnings for 14 of the QTNs. Most (17) had also a similar effect on qualification. Few (8) had also an effect on the percentage of finished races but not at all ages, except for DMRT3. When the QTN had an effect on the percentage of finished races, it was in the same direction than earnings for 5 and in the other direction for 2 (not counting DMRT3). Significant effect on number of starts occurred principally for the trait at the age of 5 to 10 years. Other than DMRT3, 9 QTNS are concerned, 7 in the same direction than earnings and 2 in the opposite.

3.4 DMRT3 region

The DMRT3 region can be divided into two parts: SNPs located in position prior to DMRT3 mutation with a relatively high linkage disequilibrium with DMRT3 and one located after the DMRT3 mutation with less linkage (Appendix S1). Manhattan plot over the region for all racing performances is reported in Appendix S2. Many SNPs in the region were highly significant for Q. For earnings, many SNPs in the region were highly significant for LnEH2; however, from 3 years of age to 10 years, none reached the significant threshold, except for LnES5-10. The effect of many SNPs in the DMRT3 region on probability to finish a race was significant at all ages. Figure 4 plots the estimated effects of alleles of the 146 SNPs in DMRT3 region for pairs of traits related to earnings. Favourable alleles on early performance (Q/LnE2) became negative in late performance (LnH5-10 and LnS5-10). Figure 5 plots absolute effect of alleles of the 146 SNPs in DMRT3 region according to linkage disequilibrium with DMRT3 mutation. There was no clear evidence of SNPs whose effect would not be related to its linkage disequilibrium with DMRT3, except perhaps some points for LnEH3.

3.5 Questionnaire: analysis of elementary variables

The plot illustrated in Figure 6 demonstrated the variable circle of the principal component analysis for the first two dimensions. Only 32% of the variability was explained by these two first components. We needed 6 components to explain 65% of the variance meaning that each defect was rather a different trait. From the plot, we used 3 main groups of variables to qualify the trotting technique of the horse. The first main group was the ability to pace, including pacer and horses defective at pace. The second group of variables grouped horses defective at gallop, with heavy trot and a bad mouth. The third group included all other locomotion defects. In later analysis, questionnaires were then summarized with these 3 groups. The 3 synthetic variables were constructed from the grouping of elementary variables: they were the sum of elementary binary variables, capped to 1, 2 and 3, respectively. They will be named “pacer,” “heavy trot gallop,” and “other defects.”

There was a negative association between horses qualified as “pacer” and “heavy trot gallop.” There were no horses qualified as “pacer” which was also qualified as “defective gait: gallop.” There was a positive association between “other defects” and both “pacer” and “heavy trot gallop”: 77% of horses scored higher than 0 for “pacer” and 78% of scored higher than 0 for “heavy trot gallop” were scored higher than 0 for “other defects” compared with 64% and 52% for horses scored 0.

Some of the trotting technique questionnaire variables were heritable, as well as synthetic variables (Table 3.)

3.6 Questionnaire: GWAS analysis

The frequency of the DMRT3 genotypes for trained horses was very different according to responses to the trotting technique questionnaire (Table 4). However, the random forest classification based on questionnaire responses was unable to correctly predict the horses’ genotypes at DMRT3. The mean over samplings of correctly predicted genotypes was 83% for the training samples and 62% for the validation samples, but never exceeded 77%. Mostly, the model failed to identify the CA genotype, with 50% of true CA genotype assigned to AA genotype. It failed also to identify CC genotypes, which were never predicted, and in 49% of the cases, were assigned to AA. Therefore, our results demonstrate that genotyping is necessary, regardless of the knowledge of the trotting technique of the horse.

Among the studied population, therefore trained and almost all qualified (98%), we found that 100% of horses defective at pace and 94% of horses qualified as pacer were AA, even if all AA horses were not qualified as pacer or defective at pace. None of the horses with genotype CC and only 6% of CA were qualified as pacer and none CC or CA as defective at pace. AA versus CA or CC is clearly a mark of the ability to pace. In the same way, but less systematic, CA and CC horses were more frequent among horses defective at gallop: 90% of horses with defective gait at gallop were of these genotypes. Therefore, the presence of at least one C copy is a marker of inability to pace and tendency to gallop (or to have heavy trot and bad mouth). Other defect in trotting technique is distributed more equally in the 3 genotypes. Genotype AA at DMRT3 increased the probability to pace, avoid “heavy trot gallop” and had no significant effect on other defects. Heterozygote CA was closer to CC than AA.

Table 5 summarized GWAS analysis with 9 QTNs detected. For horses qualified as “pacer,” in addition to the region of DMRT3, there was a QTN on chromosome 10 and 3 QTNs on chromosome 28 representing 7 SNPs above the significant threshold. For all, the rare allele had a positive effect on the ability to pace, contrary to the effect of DMRT3 were the frequent homozygote horse (AA) paced. The 3 QTNs on chromosome 28 were distant but in linkage disequilibrium (r² = .66, .32, .46). The QTNs had no effect of the other trotting technique traits, contrary to DMRT3 which had also an effect on “heavy trot gallop.” For the trait “heavy trot gallop,” in addition to DMRT3, there were 3 QTNs: two on chromosome 6 and one on chromosome 19. For two of these three QTNs, the rare allele increased the frequency of “heavy trot gallop” horses. So, there was only one frequent allele (C, 64%) which leads to gallop: the one on chromosome 6:28,033,876. For “other defects,” there were 2 QTNs. One was the 6:28,033,876 where the frequent allele leads to more detects, as for the trait “heavy trot gallop.” And the second one on chromosome 3 was in the same direction (frequent allele, increase in defects).

DMRT3 was the only QTN detected for both racing performances and trotting technique.

3.7 Questionnaire: effect on racing performance

The frequency of horses scored 1 for “pacer” was higher at 2 years old (22%) and much lower in races under saddle (<5%, two low number of pacer horses to estimate an effect on earnings in races under saddle). The effect to be scored 1 for “pacer” was always favourable whatever the performance trait (Figure 7). Significant effect was found on earnings at 2 and 3 years old and on number of starts at 3 and 4 years old. The frequency of horses scored higher or equal to 1 for “heavy trot gallop” was higher in races under saddle (65%) and lower at 2 years old (35%). The effect to be scored higher or equal to 1 for “heavy trot gallop” was unfavourable for late performances (LnEH5, FH4, FH5-10, SH5-10), especially for the proportion of finished races. For the trait “other defects,” there were no differences in frequency of horses scored higher or equal to 1 at the different ages of performance and between harness and under saddle races. To be scored to 1 or more was unfavourable for performances at 3 years old (LnEH3, FH3, SH3), but then, there was no effect on later performances.

3.8 Genetic evaluation

Results on genetic evaluation for racing performance traits are outlined in Figure 8. Comparisons involved 4 evaluations: “QTNs,” the sum of QTN estimates obtained from ssGBLUP with QTNs as fixed effects; “genealogy,” the BLUP evaluation with a relationship matrix based on genealogy; “genomic,” the ssGBLUP evaluation with a relationship matrix based on genealogy and genomic without QTNs; and “genomic + QTNs,” the sum of QTN estimates and ssGBLUP genomic value. Results for qualification (Q) were not relevant because the sample was highly selected on this trait (98% of horses qualified in the validation set). The use of only QTN estimates yielded the lowest correlations, and sometimes, even negative correlations, except for FH4 and FH5-10. Genomic and genealogic evaluation often yielded close correlations, generally slightly higher for the genomic evaluation. The mean of absolute differences in correlation between these two types of evaluations was 0.01, with the maximum for FH4 with the correlation of 0.11 with genomic evaluation and 0.07 with genealogic evaluation. The addition of QTNs to genomic evaluation yielded lower or higher correlation in equal cases. The gain was often less important (average of +0.02) than the loss (average of −0.05).