Upper Limits - The Doctor

Research into the genetic basis of racing performance in the thoroughbred population has found a surprisingly high level of genetic variation, which is the raw material on which selection acts to deliver improvement to the breed. Despite the abundance of variation and selection for the best runners, though, thoroughbreds seem to have reached a performance limit. Records simply aren’t broken frequently enough to make sense of the genetic result. In this post I’ll be reviewing the scientific evidence for genetic variation in thoroughbreds and offering some of the attempts to resolve this puzzle. I’ve been watching in amazement as record after record falls at these summer Olympics. It’s enough to make one wonder if we’ll ever reach the upper limit to athletic performance in our species, or if such a thing even exists. In contrast, some have wondered if the thoroughbred has reached the upper limit of its athletic performance because, for instance, the same march of progress that can be seen for the 800m record in humans is not found in the progression of records in the mile and a half in horses (Secretariat 2.24:00). Why the difference? Why is it that humans keep setting new records while thoroughbreds don’t? There is certainly a financial incentive to put together a horse that can set new race records. So, it’s not that. And arranged matings, which would be unethical in humans, can only give horses an advantage in the pursuit of breaking records. As with many unsettling trends in thoroughbreds, breeding has been offered as the culprit. The claim is that the gene pool in thoroughbreds has been drained of genetic variation. In this post, I’ll be reviewing the scientific evidence for genetic variation in thoroughbreds. I’ll show that a study that attempts to measure genetic variation for performance finds it in abundant supply, creating a seeming paradox: plenty of variation, plenty of selection, but no new records. Finally, I’ll see if I can dig us out of that paradox. In order to measure the amount of genetic variation for a trait of interest, you can follow these three steps. First (1), you need to measure the trait on a large number of individuals. You are almost certain to find that these measures differ, and so you’ll have trait variation. Next (2), you need to obtain their pedigree. And then (3), you essentially ask: how accurate is it to use one individual’s trait as a prediction of his/her relative’s trait? When this is a reliable predictor, you’ve got genetic variation for the trait. When this is not a reliable predictor, then you have some other source of trait variation that is not due to genes. In a classic study by Gaffney and Cunningham (1988), the trait that was measured was TIMEFORM rating. The TIMEFORM rating is assigned to a horse at the end of a season and is an estimate by informed handicappers of how much weight the horse should carry in a hypothetical free handicap race. The better the horse, the higher the TIMEFORM score. This measure is subjective but permits comparison between horses across different tracks, different distances, different competition, and different conditions, etc. The other appeal of this measure for Gaffney and Cunningham was that it offered them a lot of performance data (over 30,000 3-year-olds with TIMEFORM ratings, over 2000 sires with TIMEFORM ratings, plus the TIMEFORM ratings from most of the dams went into this study). In addition, pedigrees for thoroughbreds are known and are easy to obtain. The data on parentage is presumably accurate and goes back centuries to the beginning of the Stud Book. With these data, Gaffney and Cunningham had Steps (1) and (2) in hand. Gaffney and Cunningham then performed Step (3) by comparing TIMEFORM ratings of offspring with those of their sires and dams and also by comparing the ratings of half-siblings (presumably in the traditional, geneticists’ sense of the term – not in the horse-breeding sense of the term, which only counts individuals that share a mother as half-sibs). What Gaffney and Cunningham found was surprising. The heritability of performance – or, in other words, the proportion of total trait variation that can be chalked up to genetic variation – was ~36%. Heritability, loosely speaking, can be used to measure how much response you get in the next generation when you perform selection. To predict the amount of change in one generation, one multiplies the heritability of a trait with the selection differential of that trait. For instance, if the heritability is 36% and the average TIMEFORM rating of the selected parents is 10 pounds more than the population average, the offspring will have an average TIMEFORM rating that is (0.36)*(10) = 3.6 pounds heavier than the previous generation’s average. Gaffney and Cunningham also had the ability from the data on pedigree to determine the selection strength in the thoroughbred population. In males, selection is equivalent to breeding only from the top 6% of runners and in females, selection is equivalent to breeding only to the top 52%. Transforming these selection intensities to selection differentials and combining them with the heritability gives us an estimate of a change of 0.92 TIMEFORM units per year. In other words (and in case some of that derivation was opaque), based on the amount of genetic variation and the strength of selection, the thoroughbred population should improve, on average, by 0.92 TIMEFORM units per year. In the same issue of the journal in which Gaffney and Cunningham published their study, Hill offered a “back-of-the-napkin” calculation that extends this result, which I adapt here. In a mile and a half race, a one-pound handicap is roughly equal to 0.7 lengths. So each year, Gaffney and Cunningham predict that the breed improves by ~0.64 lengths. Taken over 35 years, this would mean an improvement of 22.5 lengths, or ~60 yards, which would shave almost 4 seconds off of the time in a mile and a half race. So why haven’t we seen Secretariat’s 1973 mile and half bested by 60 yards in the past 35 years? Why no 2:20 in the Belmont? Something has to give. There are potentially two ways out of this seeming paradox. Gaffney and Cunningham point out that the response to selection that they predict is a change to the average of the population, which says nothing about the extremes of the population. See, the average may have very well changed by 4 seconds, but we can’t tell from this what happened to the extremes, which is what we focus on when we discuss records. Thus, our formulas for heritability and selection strength are ill-equipped to make statements about how much longer Secretariat’s record is expected to stand. It is mistaken to think that the record will progress by 4 seconds over the course of 35 years just because the average is predicted to do so. Even still, if the average of the population were expected to move up by almost 4 seconds, one would have to have a large change to the distribution of all mile and a half times in order to explain how all of the horses since Secretariat have failed to come in under 2:24. While Gaffney and Cunningham are correct to point out that our formulas deal with averages – not extremes – it requires special pleading to use this as the way out of the seeming paradox. An alternate way out of the paradox is to show that the methods for measuring heritability are biased upwards and overestimate the actual extent of genetic variation. This may well be the case because there is likely a resemblance between relatives for the kind of training and care they receive. Expensive stud fees are usually paid by those who can afford the best trainers, too. Thus quality can be passed from generation to generation but not along the genes, an example of shared environmental variation. We would mistakenly measure this component of resemblance as genetic, however. It’s not entirely clear how performance will respond to selection if this shared environmental variance is a major contributor to familial resemblance. In conclusion, maybe there is little variation in the gene pool of thoroughbreds. And maybe this explains the stagnation of racing records in the breed. Or maybe there is genetic variation in horses and our assumptions about the genetics of performance are wrong, with the gene variants of interest not being amenable to the kinds of selection we perform. Perhaps our breeding schemes are wrong for the kind of genetic variation that exists. Regardless, it doesn’t add up. Other domesticated species and agricultural species produce much greater returns for selective breeders. Sorting out the puzzle of why horses fail to do so will be of interest not only to horse breeders but also to the agricultural community as a whole.