New insights for enhancing feed efficiency
- Autores: Mónica Mora Fenoll
- Directores de la Tesis: Mercedes Piles Rovira (dir. tes.), Agustín Blasco (tut. tes.)
- Lectura: En la Universitat Politècnica de València ( España ) en 2024
- Idioma: inglés
- Tribunal Calificador de la Tesis: Clara Díaz Martín (presid.), Francisco Maroto Molina (secret.), Filippo Biscarini (voc.)
- Programa de doctorado: Programa de Doctorado en Ciencia y Tecnología de la Producción Animal por la Universitat Politècnica de València
- Materias:
- Enlaces
- Tesis en acceso abierto en: RiuNet
- Resumen
The thesis comprises a series of studies aimed at improving genetic selection programs to increase feed efficiency at two stages of the productive life of pigs: lactation and growth. In Chapters 2 to 5, various methodologies and sources of information (genetics, genomics, and phenomics) were explored to achieve this goal.
Feed efficiency during lactation is a complex system of interrelated phenotypic traits. In Chapter 2, structural equation models were used to infer the causal associations between daily feed intake, sow weight balance, litter weight gain, backfat thickness balance, and sow metabolic body weight. Data from 1,100 sows from a divergent selection experiment for efficiency in growing pigs were analysed. The inductive causation algorithm was used to define the structural equation model. Results indicated that daily feed intake and daily litter weight gain influenced daily sow weight balance, with average structural coefficients of 0.12 and -0.03, respectively. Subsequently, daily sow weight balance impacted daily backfat thickness balance and sow metabolic body weight, with structural coefficients of 0.70 and -1.22, respectively. Different selection strategies were evaluated. Selection strategies that did not rely on feed intake records proved equally effective, although increasing litter weight gain while reducing or maintaining constant feed intake without deteriorating the sow's body condition seems challenging.
In Chapter 3, the benefits of multi-output and stacking methods to improve the accuracy of predicting feed efficiency from the genotype of growing pigs were explored. Data from 5,828 pigs and 45,610 SNPs were analysed, implementing four different strategies. Random forest and support vector regression models were trained with different subsets of predictor variables (from 200 to 3,000 of the most informative SNPs). The standard strategy, which used only genotypes as predictors, showed the best performance. With the random forest algorithm and the top 1,000 SNPs, the average Spearman correlation was 0.23 (0.04).
Regarding high-throughput phenotyping, the use of accelerometers and cameras to obtain information on animal behaviour was proposed. In Chapter 4, 16 rabbits were equipped with accelerometers and filmed for 2 weeks. The time series were segmented into time windows of different sizes (1, 3, 5, 7, and 9 seconds), and 41 features were calculated per window. Random forest, support vector machine, and gradient boosting models were trained to classify active/inactive states. The random forest model, with a 9-second window and the 5 most important features, achieved a median precision of 1 and a median recall of 0.93.
Finally, in Chapter 5, a tracking-by-detection algorithm was combined with a radio frequency identification system to track twelve pigs in a pen, three of which had unique coat markings that facilitated visual distinction. The You Only Look Once (v8) algorithm was implemented for detection and the BoT-SORT algorithm for tracking. Pigs with similar coat markings were tracked for an average of 49 minutes, while pigs with unique markings were tracked for 152 minutes, with a 90% or higher match with the ground truth. This combination of devices represents a significant advancement in the identification, tracking, and traceability of animals in groups.
