Many people, scientists and non-scientists alike, object to what they perceive as genetic determinism. This is often a reaction to geneticists apparently over-reaching and claiming that some trait or condition is “caused by” a single gene. A common rejoinder is that any biological process obviously involves many hundreds of gene products, interacting with each other in complex ways and so it is nonsense to say that the trait is determined by a single gene. That is absolutely true, if you are using the word “gene” purely in the molecular biology sense – as a piece of DNA that encodes a particular product (usually a protein). But geneticists also use it in the original sense, as a unit of heredity – a genetic variant or mutation that can be passed on across generations and that influences some phenotype.
Genetics is not about how a characteristic arises, it is about how variation in that characteristic arises. For example, when you are describing someone, you might say: “She’s got blue eyes”, but you probably wouldn’t say: “She’s got two eyes”. Both characteristics are determined by the genetic program, but only one is affected by genetic variation. Eye colour is therefore a trait, because it varies across the population and that variation is due to genetic differences. Having blue eyes, insofar as it necessarily involves having eyes in the first place, is obviously not caused by a single piece of DNA – it takes thousands of gene products to build eyes, blue or otherwise. But having eyes that are blue, as opposed to brown, can be due to a single genetic variation.
As it happens, though, eye colour is also a good example of genetic interactions. Because, while it’s true that a single mutation can explain the difference in eye colour between some people, it’s also true that many people carry more than one such mutation in any of several different genes. The ultimate colour that emerges is thus often determined by interactions amongst multiple genetic variants.
|Distribution of heights in female and male college students (source)|
This is even more true for traits like height or IQ which differ in a quantitative way across the population. The differences between any two individuals for traits like these are typically not caused by a single genetic variant, but by many. When we talk of genetic interactions, we are asking how the effects of such mutations combine. Do their effects simply sum up or do they interact in a more complex way?
This is an important question because it affects our ability to discover the contributing mutations in the first place and, crucially, to predict any particular individual’s phenotype from their genotype.
Height provides a good example. Genome-wide association studies with huge numbers of people (both subjects and authors) have identified variable positions in the genome that show a statistical association with height. While the sequence at most positions in the genome is the same in most people in the world, around one in a thousand positions comes in different flavours – at such positions, the DNA might be an “A” in some people, but a “T” in others. By looking at millions of such sites, researchers have found 180 where the average height of people with one version, say the “T”, is very slightly greater than the average of those with the “A”.
For any individual, you can then count how many “tall” variants they have across all these sites. If you plot the distribution of this score across the population you can see how it correlates with height. It turns out this relationship is remarkably linear. As you increase the number of tall variants, the average height continues to increase at the same rate – people don’t suddenly start to get much more tall with each new variant and they also don’t reach a point where adding more tall variants starts to have a smaller effect. The exact same linear relationship is seen for genetic variants affecting body-mass index.
|From Speliotes et al. (This graph is kind of misleading because it makes the relationship look very predictive by plotting a single value for mean height in each bin. For any particular number, there will still be a very wide range of heights – the average is just slightly different. This is similar to the effect of the Y chromosome – the average height of men is greater than the average height of women but if all you know about someone is their sex you have effectively no predictive power of their specific height).|
To me, this is a genuinely surprising result. It seems to go against expectations from experimental genetics, where non-linear (or “epistatic”) interactions between mutations are the norm – ubiquitous really. It is quite common, for example, for two mutations, in two different genes, to have no effect singly but a drastic effect when combined. Or for a mutation to have very different effects on different genetic backgrounds (this is true for disease-causing mutations in humans as well as in animals).
By contrast, the results from genome-wide association studies in humans seem to suggest that it is simply the number of such variants that matters in any individual and that the precise combination has little effect. Indeed, that is precisely how it has been interpreted by people who suggest that the value of such a score could be used to predict an individual’s phenotype.
The problem with such an interpretation is that genome-wide association studies give us only an average effect of each variant across the population – i.e., they measure the statistical effect of having one version of a particular variant versus another, averaging all other genetic variation out. Each such number is computed independently. If there are many variants involved and some of them show non-additive interactions, we would never see that because the number of individuals sharing both those particular variants is such a small percentage of our sample and the background of additional variants will be so diverse that any non-additive interactions will tend to average out. Indeed, it has been shown that you can mathematically treat the interactions as additive across the population, whether they are or not in individuals – at least for the purposes of identifying variants affecting a trait.
But that doesn’t mean they actually are additive and this remains a crucial question for our ability to extrapolate population averages to make predictions about individuals. The importance and ubiquity of such non-additive interactions is revealed by a powerful approach that is possible in animals.
If we have two individual organisms that differ in some trait, presumably due to the effects of multiple genetic differences, then we can imagine a thought experiment: what would happen if we could decompose these mutations – if we could look at their effects one-by-one, to see how much each one contributes to the difference and to compare this with their combined effects?
Obviously, that experiment is not possible with two individual organisms. But it is possible if we have clones of those individuals. Exactly that situation exists for lines of inbred mice.
Many different lines (or strains) of lab mice exist, most of which are completely inbred. That is, they have been backcrossed for so many generations that no genetic variation exists within the strain. Each animal within the strain is genetically identical – even the two copies that each animal possesses of each chromosome are genetically identical.
When each such line was generated, some arbitrary spectrum of genetic variants was effectively frozen in place – while there is no genetic variation left in any particular line, there is lots of genetic variation between lines.
This causes many phenotypic differences between them. These are most obvious in things like coat colour, but extend to all kinds of traits, including behavioural ones. Mice from some lines may be more active, more anxious, more sociable, more aggressive, more clever (in mouse terms, that is – the inbred ones are not the brightest at the best of times), or differ in many other behavioural tendencies.
Now, if you’re a scientist interested in the genetics of behaviour, these lines are a gold mine. If you can figure out which of the genetic differences between two lines account for some behavioural difference, you would have an entry point into the biological processes controlling that behaviour (whether it’s aggression, anxiety or mouse-smarts). Trouble is, these differences are rarely simple. In fact, they can be complicated in very unexpected ways.
First, differences in quantitative traits like behavioural tendencies rarely come down to a single genetic difference between strains. Crossing strains together (say a highly active and a less active strain) typically generates F1 hybrid offspring with a value for the trait that is somewhere between the two parental lines. By backcrossing these hybrids to either of the parental lines and correlating chromosome inheritance with the value of the trait, it is possible to map out regions of the genome influencing the trait (known as quantitative trait loci). A typical finding is that there may be anywhere from 5 to 10 mappable loci contributing to the differences in the trait between the two lines.
Now comes the unexpected part. Say you have mapped 10 loci that all seem to contribute to the difference between a high and a low activity strain. You might expect that each of these 10 different regions would contribute a small percentage to the phenotypic difference and that their effects would simply add up, quantitatively speaking. That is, if you had one locus that caused 10% of the increase in activity, and another that caused 15% of it, that if you combined both of them, you would see 25% of the increase.
How could we test these expectations? What you would like to be able to do is examine the effects of each of these loci by themselves, rather than comparing the two strains with all 10 of those differences. Exactly that kind of experiment is made possible by the generation of chromosome substitution strains. These are lines that have been generated by crossing together two strains (let’s call them A and B) and then backcrossing the hybrids to the A strain, while molecularly tracking inheritance of just one of the chromosomes from the B strain (and vice versa). Over time, this can generate a series of lines (one for each chromosome) that have the full genetic background of the A strain, with the exception of a single B-type chromosome.
Now you can ask exactly the question we wanted to ask – what is the effect on the trait of each of the individual loci in isolation? If we start with the background of the low activity A strain, and look at lines that each have one B chromosome containing a “high activity” allele, where will their phenotypes lie on the line between the parental strains?
Well, here’s the surprise. In many cases (like here and here), the phenotypic effects of these single loci are much bigger than you would expect – often explaining 50% or more of the difference between the two strains. This means that if you simply added up their effects you would get much more than the 100% of the difference you started with. In fact, the range for behavioural traits averages at ~800%, if you simply add up the effects of the decomposed individual loci. Even more remarkably, some of the individual chromosome substitution strains show a phenotypic level that is outside the range of either of the initial parents, sometimes even moving the trait distribution in the opposite direction to the “donor” parent strain.
|Spiezio et al|
These results clearly show that non-additive interactions for variants affecting quantitative traits are common, large and unpredictable. They are a ubiquitous feature of the genetic architecture of quantitative traits, whether morphological, physiological or behavioural and are seen across many different species, including worms, flies, chickens, yeast. Even if such interactions average out across all the combinations encountered in the population, so that they appear additive, statistically, this biological reality places a severe limit on our ability to predict any individual’s phenotype based purely on additive calculations.
In fact, even if we can begin to define some non-additive interactions by studying the phenotypic effects of various pairwise combinations across many people, it will still be very difficult to predict any new individual’s phenotype because, just like each of the distinct chromosome substitution strains in mice, their precise combination of all variants will never have been seen before. In a strange way, I find that comforting – we are each much more unique than a purely statistical overview would suggest.