We previously reported that thoroughbred racehorses in Great Britain are continuing to get faster (Sharman and Wilson 2015), and also showed that the rate of contemporary phenotypic improvement in GB is greatest for sprint-distance races. Here we show that the population harbours statistically significant amounts of genetic variance for speed over sprint, middle and long distances. However, our estimates of heritability are low over all distances, particularly over long-distances. We also show that the previously reported phenotypic improvement is underpinned by genetic improvement in all three distance categories. The estimated rates of improvement decrease as race distance increases. For all race distance categories, improvement rates are greater than can be reasonably be explained by drift, leading us to conclude that we are observing a selection response.
Our primary conclusion, that genetic improvement in thoroughbred racehorse speed is ongoing, contrasts qualitatively with earlier studies that argued speed was no longer increasing because the trait had reached a selection limit imposed by some form of genetic constraint (Denny 2008; Desgorces et al. 2012). Thus, while the net selection imposed by racehorse breeders is difficult to quantify (discussed further below), it is – to some extent – producing faster racehorses. Rates of genetic improvement as estimated by univariate analyses under Model A represent 60, 55 and 17% of the total phenotypic improvement over short, middle and long distances respectively. Importantly though, as noted earlier, the contemporary rates of total phenotypic improvement are themselves low. For instance, mean sprint speed is changing at an estimated rate of just +0.09% per annum relative to the observed 1997 phenotypic mean (Model 2, Sharman and Wilson 2015). This means our estimate of βG = 0.009 yards.sec−1.year−1 is a genetic improvement rate of just 0.05% relative to mean sprint speed in 1997. The corresponding values for middle and long distances are 0.04 and 0.006% per annum (relative mean speeds in 1997).
Thus, while genetic improvement consistent with selection response is apparent, the estimated rates of improvement are low. As with heritability estimates, previous estimates of genetic trends for racing performance traits vary considerably. Direct comparisons are also hampered by differences in running distance, population studied, statistical methodology, modelling decisions and – perhaps especially – the performance trait modelled. Nonetheless, a study of thoroughbred racing in Japan yielded estimated rates of genetic improvement for finishing times of 1600m races in the region of 0.01–0.02% year−1 (Oki and Sasaki 1996) which are lower than our estimates over sprint and middle distances. Both stallions and mares were being imported to Japan at this time with the aim of improving the population (Oki and Sasaki 1996). Similar estimates were also obtained for finishing times in the Brazilian thoroughbred population (da Mota et al. 2005). This study also found genetic improvement rates were lower with increasing race distance.
Conversely, much higher rates of improvement than detected here have been reported in some previous studies. For instance, in Quarter horses, a breed which races over distances of 301–402 m, genetic improvement rates as high as 0.4% year−1 have been estimated (Faria et al. 2019). A previous study of thoroughbreds running predominantly in GB and Ireland generated a point estimate of the genetic improvement rate (across all distances) of about 1% year−1 from 1952–77 using PBV for Timeform handicap ratings (Gaffney and Cunningham 1988). This corresponds to an increase in speed of about 0.1% year−1 (Hill 1988). No measure of uncertainty around this trend was presented, and we note this estimate is actually greater than the total phenotypic improvement in speed as averaged across distances over historical (1850–2012) and recent (1997–2012) periods (Sharman and Wilson 2015). While noting that the authors did check the sensitivity of their trend to the assumed value of heritability, we consider it likely that both h2 and rate of improvement were upwardly biased by common environment effects in that case (a possibility also suggested by others: e.g., Hill 1988; Langlois 1996). Here we have tried to minimise this risk by including race level effects but also modelling any trainer influence to account for offspring of ‘better’ bred horses going to ‘better’ trainers (Schulze-Schleppinghoff et al. 1985; Schulze-Schleppinghoff et al. 1987; Hill 1988; Preisinger et al. 1989; Preisinger et al. 1990). For example, in the sprint distance REML analysis under model A, we find trainer identity explains more variance than additive genetic effects and omitting it results in estimated heritability of speed rising from 0.124 to 0.244 (full results not shown). Moreover, LRT comparison showed model A was a significantly better fit of the sprint subset when trainer was included (χ20,1 = 8335.02, P < 0.001). This strongly suggests that trainer effects, if not modelled, will be a source of common environment variance that can upwardly bias VA and inflate estimates of genetic change.
Our results also provide some insight into why rates of genetic improvement rates are low. In the simplest case, the univariate breeder’s equation predicts selection response as the product of heritability and linear selection differential (Lush 1937). Thus, response is limited if heritability is low and/or selection is weak. We suggest both are possible here. Conditional on fixed effects included in Model A, heritability for speed is low, particularly over long distances. The strength of selection on speed is unknown for reasons outlined above. However, by making the strong assumption that the breeder’s equation holds true, then for a trait under simple truncation selection with repeated observations per individuals, realised selection intensity i for each trait can be calculated as follows (Walsh and Lynch 2018):
$$i=\frac{{\beta }_{G}L}{h{\sigma }_{A}\surd (\frac{n}{1+(n-1)R})}$$
Where βG is the per annum rate of improvement and L is the generation time (in years), \({\sigma }_{A}\) is the additive standard deviation, h is the square root of the heritability, n is the number of observations per individual, and R is the trait repeatability. Setting L to 9.2 for sprinters, 9.5 for middle-distance and 9.8 for long-distance (mean parental age at offspring birth in each data subset), n (mean number of records per individual) to 5.96 (sprint), 5.10 (middle-distance) and 3.59 (long-distance) and letting R = (VA + VPE)/VP then substituting in our parameter estimates from univariate analyses under Model A yields values of isprint = 1.202, imiddle-distance = 0.905, and ilong-distance = 0.317. Reiterating that these are illustrative calculations made with strong assumptions, they suggest selection could be weaker than previously estimated. For example, isprint = 1.202 equates to selecting approximately 28% of the population under truncation selection. This compares to estimates of selection (across all distances) on Timeform handicap ratings of 23% (More O’Ferrall and Cunningham 1974), 32% (Field and Cunningham 1976) and 29% (Gaffney and Cunningham 1988). The lower realised selection intensities over longer races correspond to selecting >40 and >80% of the population for middle- and long-distance speed respectively.
Weak and/or inaccurate selection on speed traits may emerge cumulatively from the decision making of individual horse breeders for multiple reasons. First, speed is just one measure of performance; jockeys ride to win races, not to break records, and other phenotypic attributes contribute to a horse’s success (Langlois 1996). We have accounted for a wide range of factors in our modelling, but nuances like temperament or responsiveness to jockey controlled race tactics are unknown and unaccounted for. Nonetheless, in a racing context it is implausible that any programme of selection for increased performance (however defined) would not incorporate the aim of increasing speed. Second, there has been a general reluctance to incorporate genetic and/or genomic prediction methods (e.g., BLUP, GBLUP) in horse breeding (Hill 2016). Although low heritabilities would pose a limit to selection accuracy, such approaches still offer well documented advantages relative to selecting on phenotype. Third, since most breeders have commercial objectives, optimum sale price for resultant offspring is important. Reputation matters and ‘fashionable’ pedigrees may command higher prices regardless of actual genetic merit (Wilson and Rambaut 2008). Fourth, even given reliable information about genetic merit, cultural and economic factors limit availability of the best genes. To be allowed to race, thoroughbreds must be produced by natural matings not artificial insemination. This limits the number of offspring that can be produced each year from a given sire, while covering fees in excess of £100,000 mean leading stallions are only financially accessible to a small percentage of breeders. Fifth, selecting across multiple traits (rather than, for example, just selecting on sprint speed) is expected to reduce selection on each trait. Sixth, and rather speculatively, selection on speed may be partially countered by antagonistic selection on injury risk. It has been claimed that thoroughbreds are becoming more susceptible to injury (Drape 2008; Mitchell 2008; Gibbons 2014), perhaps as a consequence of morphology changes which have coevolved with speed (Gilbey 1903). Given additive genetic variance underpinning some thoroughbred health and conformation traits (Ibi et al. 2003; Oki et al. 2005; Oki et al. 2008; Welsh et al. 2013; Norton et al. 2016), investigation of the potential (genetic) association between injury risk and race performance would be timely. This could help to understand the evolution of speed, and may also provide tools to address ongoing welfare concerns in horseracing.
A final point emerging from our analysis is that the three distance-specific speed traits offer distinct selection targets in the sense that all pairwise genetic correlation are less than +1. We also find that the heritability of speed is lower over long-distance races, a result consistent with several other studies of finishing times (Oki et al. 1995; da Mota et al. 2005; Ekiz and Kocak 2007; Bakhtiari and Kashan 2009; Velie et al. 2015b). Thus, there is potential to improve performance across all distance categories including long distances (although there has been a commercial trend over recent decades to focus on shorter distances). The estimated correlation between sprint and long-distance performance is notably lower (rG=0.47), a result that mirrors a recent finding in Brazilian thoroughbreds (da Mota 2006). Biologically this is unsurprising; many studies have highlighted the divergence of physiological and biomechanical trait optima across running distances in human athletes (Thompson 2017). Indeed, recent work on myostatin encoding gene (MSTN) in thoroughbred racehorses has shown associations between genotype and optimal running distance that will contribute to genome-wide departures from rG = 1 across distances (Hill et al. 2010).
In summary, we show here that speed in thoroughbred horses is heritable across categories of race distance. We also show that genetic improvement attributable to selection is contributing to previously demonstrated weak – but non-zero – rates of phenotypic improvement. However, our analyses also show that selection responses are of a limited magnitude, particularly for long-distance race performance. Low heritabilities and among-distance genetic correlation structure contribute to this pattern but weaker selection than previously assumed also seems possible. Accuracy of selection may be low across all distance categories, particularly given that modern genetic tools are rarely applied in thoroughbred breeding. This obviously contrasts with most livestock species in which much more rapid selection responses are regularly achieved.