Towards Genetic Improvement of Social Behaviours in Large Groups of Livestock: A Case Study with Turkeys

Samenvatting (Engelstalig)

Social interactions in livestock, particularly harmful behaviours such as injurious pecking in poultry, have a significant impact on animal production and welfare. While these behaviours are heritable, the application of selective breeding is still limited due to the need for (1) adequate phenotyping methods for individuals in large groups and (2) appropriate genetic models. In this thesis, we built upon recent advancements in computer vision techniques to obtain large-scale longitudinal data on social interactions, and presented models for genetic analysis of such data, using turkeys as a case study. Thereby, we investigated the current and near-future opportunities for genetic improvement of social interactions in livestock.
In Chapter 2, we presented a method for simulating social interactions among animals using an agent-based model. This model incorporated two latent traits: the "performer’s effect (P)", representing an individual's tendency to initiate a social interaction, and the "recipient’s effect (R)", representing an individual's propensity to receive such interaction. The heritabilities of P and R were tuned to align with values found in literature, and varied across different scenarios. Within the simulation, animal movements were assumed to be completely random, resulting in variability in the number of encounters among individuals. During encounters, traits P and R were used to determine whether a social interaction occurs. Next, we presented a Generalized Linear Mixed Model to estimate genetic parameters from the resulting data, including variance components, genetic correlations and breeding values. This chapter also explored how sample size—considering both the number of individuals and interactions—impacts the accuracy and bias of breeding value estimates, yielding insights into the minimum data requirements necessary for reliable genetic analysis of social traits. Notably, we achieved a promising accuracy of 0.713 under the default scenario, which featured a population size of 2,000 individuals, heritabilities of both traits set at 0.2, and an average of approximately 100 interactions per animal—corresponding to a few weeks of observation time in practice.
Chapter 3 addressed the practical challenges of utilizing computer vision to track birds in commercial settings with large groups without the use of individual markers. This study employed a system with 34 cameras to analyse video data from a large group of ~300 turkeys in a pen of about 100 m2. We developed a markerless visual tracking system based on instance segmentation and tracking algorithms using the overlap of the field of view of different cameras. It also discussed the complexities associated with long-term tracking, particularly the challenges posed by occlusions—instances where multiple birds overlapped within the same frame. Furthermore, we highlighted the difficulties related to accurate identification and re-identification of individuals over extended periods of time. We achieved a Multi-Object Tracking Accuracy (MOTA) of 0.69 and an Identity Switch Rate (IDSR) of 23.3% after five minutes of multi-camera tracking. For practical application, where the goal is to track identified animals continuously across multiple cameras for several hours, this performance level is still insufficient.
Recognizing the difficulties in markerless tracking detailed in Chapter 3, Chapter 4 explored the use of ArUco markers to enhance tracking accuracy and facilitate long-term tracking and subsequent genetic analysis of locomotion traits. It provided a description of the marking method, detailing the process of attaching markers to the turkeys for more reliable individual identification and tracking. We utilized a similar camera system as in Chapter 3, also monitoring approximately 300 animals within a comparable pen area. The observation period extended to 12 hours per day for 7 days, providing a far larger dataset than the markerless tracking used previously. The use of marked individuals significantly reduced tracking errors resulting from occlusions and re-identification issues noted in the previous chapter. We measured four key locomotion traits: the time an animal was active, its average speed, the number of nearby individuals, and the number of individuals ahead while moving. Genetic parameters for these traits were estimated, resulting in heritability estimates of 0.15, 0.08, 0.23, and 0.23, respectively. We observed significant differences among individuals in the number of nearby individuals and the number of individuals ahead. This suggested that, in addition to the two latent traits discussed in Chapter 2 (P and R), the frequency of encounters with pen mates also plays a role in shaping social interactions in the population.
In Chapter 5, we therefore extended the genetic model introduced in Chapter 2 by including genetic variation in social tendency, i.e. an animal's tendency to encounter pen mates and the frequency of these encounters. The objective was to assess the impact of genetic variation in encounter frequency, and the need to correct for this frequency, for the accuracy of total breeding value estimates for social traits. This chapter defined a new latent trait, "social tendency (S)", and incorporated it into the simulation model. We compared two genetic analysis models: one that corrected for the number of encounters based on the actual number of encounters, and another that used the average number of encounters (which is the same constant for all observations, implying that the number of encounters is essentially ignored). The total breeding value was defined as the sum of the expected number of interactions that an individual initiates and receives. Our simulation results revealed that using the exact encounter numbers for correction yielded significantly higher accuracy in total breeding value estimation, achieving an accuracy of 0.72 ± 0.03, compared to 0.38 ± 0.02 when using average encounter numbers. These findings highlight the importance of accounting for encounter frequency in obtaining precise breeding values for social traits.
In Chapter 6, I discussed why harmful social interactions still persist in domestic animals despite centuries of selective breeding. I examined the complexities involved in the genetic improvement of behavioural traits, including challenges in causation and measurement, and proposed a potential approach that integrates environmental and social factors alongside AI technologies to enhance phenotyping and genetic modelling. The chapter highlighted the limitations of existing datasets available for training AI models and advocates the development of a dedicated dataset specifically for animals used in agriculture. Finally, it outlined a potential strategy for implementing genetic improvements in social behavioural traits in poultry.