Olumide Victor Onabanjo’s doctoral research shows that artificial intelligence may help explain why genetic variation does not disappear after long-term selection.
For many years, breeding programmes have made livestock increasingly productive. Dairy cows produce more milk, laying hens lay more eggs, and broiler chickens grow faster than before. If the best animals are selected in every generation, why does the genetic variation on which breeding depends not eventually disappear?
Modern animal breeding research uses computer simulations to predict how a population will develop over many generations. These models are important tools when researchers and breeding organisations plan future breeding programmes. However, there is a problem: traditional models often predict that genetic variation will gradually be depleted, whereas reality suggests otherwise.
Although livestock populations have been subjected to intensive selection for many years, genetic progress has not come to a halt.
Genes interact
One possible explanation lies in the way genes interact. Most classical models assume that each gene contributes independently of the others. In reality, however, genes operate in complex networks in which the effect of one gene may depend on which other gene variants are present in the individual.
To investigate the importance of such genetic interactions, Onabanjo applied deep learning, the same type of artificial intelligence used in applications such as image analysis . This approach is particularly well suited to detecting complex patterns in large datasets.
The aim was to determine whether these models could describe the relationship between genotype and phenotype, that is, between genes and traits, more realistically than current models. The models were used to learn the relationship between the animals’ genetic code and their traits without being explicitly programmed.
The results suggest that genetic variation can be maintained far better than traditional simulations indicate. This is because variation embedded in interactions between genes may gradually be converted into a form that selection can continue to exploit.
Choice of model influences how genetic development is described
In one of Onabanjo’s studies, the models were used to simulate the development of genetic variation over 20 generations. Traditional models retained 55 to 64 per cent of the original variation, whereas the simplest deep-learning model lost all variation after only five generations. The most advanced models, by contrast, produced very high levels of genetic variation, and one model showed an almost four-fold increase after 20 generations. These results demonstrate that the choice of model can have a major influence on how genetic development is described.
Where does the increased genetic variation come from?
To better understand the source of this increased variation, the simulations were extended to 40 generations.
The results showed that variation was converted to additive genetic variation. In the deep-learning models, additive genetic variation increased substantially, while variation associated with gene interactions declined.
This suggests that genetic variation may be hidden in interactions between genes and gradually become available for selection. In the traditional models, no comparable transformation was observed. Both additive variation and interaction-related variation declined throughout the simulations.
Potential for more realistic breeding models
Onabanjo’s doctoral research demonstrates that deep-learning models can be used to describe the relationship between genes and traits in a more sophisticated way than traditional methods. The findings provide a possible explanation for why real breeding populations appear to retain genetic variation for longer than current simulations predict.
The studies were, however, based on simulations that included several simplifications. For example, only one trait was considered at a time, inbreeding was not controlled, and gene effects were assumed to remain constant across all generations. Further development and validation of the models are therefore required before the results can be applied directly in practical breeding programmes.
Olumide Victor Onabanjo will defend his doctoral thesis, "Modeling the genotype-phenotype map by deep learning to predict long-term selections" at Norwegian University of Life Sciences (NMBU) on 15 October.
Professor Peer Berg, NMBU, has been the principal supervisor. Co-supervisors have been Professor Theodorus Meuwissen, NMBU, and Professor Hans Magnus Gjøen, NMBU.