Authors: Jennifer Richter, Fernando Bussiman, Jorge Hidalgo, Vivian Breen, Ignacy Misztal, Daniela Lourenco
Categories: Animal Genetics and Genomics, accuracy, broiler survival, genetic correlations, livability, AcademicSubjects/SCI00960
Source: Journal of Animal Science
Doi: 10.1093/jas/skae190
Mortality is an economically important trait usually handled as a discrete outcome from hatch time until selection in most broiler breeder programs. However, in other species, it has been shown that not only does the genetic component change over time, but also there are maternal genetic effects to be considered when mortality is recorded early in life. This study aimed to investigate alternative trait definitions of mortality with varying models and effects. Three years’ worth of data were provided by Cobb-Vantress, Inc. and included 2 mortality traits. The first trait was binary, whether the bird died or not (OM), and the second trait was a categorical weekly mortality trait. After data cleaning, 6 wk of data for the 2 given mortality traits were used to develop 5 additional trait definitions. The definitions were broiler mortality (BM), early and late mortality (EM & LM), and 2 traits with repeated records as cumulative or binary (CM and RM, respectively). Variance components were estimated using linear and threshold models to investigate whether either model had a benefit. Genomic breeding values were predicted using the BLUP90 software suite, and linear regression validation (LR) was used to compare trait definitions and models. Heritability estimates ranged from 0.01 (0.00) to 0.16 (0.01) under linear and 0.04 (0.01) to 0.21 (0.01) under threshold models, indicating genetic variability within the population across these trait definitions. The genetic correlation between EM and LM ranged from 0.48 to 0.81 across the different lines, indicating they have divergent genetic backgrounds and should be considered different traits. The LR accuracies showed that EM and LM used together in a 2-trait model have comparable accuracies to that of OM while giving a more precise picture of mortality. When including the maternal effect, the direct heritability considerably decreased for EM, indicating that the maternal effect plays an important role in early mortality. Therefore, a suitable approach would be a model with EM and LM while considering the maternal effect for EM. Single nucleotide polymorphism effects were estimated, and no individual SNP explained more than 1% of the additive genetic variance. Additionally, the SNP with the largest effect size and variance were inconsistent across trait definitions. Chicken mortality can be defined in different ways, and reviewing these definitions and models may benefit poultry breeding programs.
Keywords: accuracy, broiler survival, genetic correlations, livability
Genetic selection in broiler chickens has revolutionized growth, enabling them to attain market weight swiftly (Havenstein et al., 2003; Zuidhof et al., 2014). However, this rapid growth is coupled with significant mortality concerns (Meseret, 2016). Modern commercial strains of broiler chickens often struggle to perform natural behaviors and may experience lameness, worsening with their increasing body weight (Bassler et al., 2013). The accelerated growth rate is directly linked to cardiovascular diseases, leg disorders, bone deformities, and other issues resulting in mortality and, thus, economic losses in the industry (Kalmar et al., 2013; Zhang et al., 2018a; Hartcher and Lum, 2020).
Mortality is influenced by genetic factors but also heavily influenced by environmental factors. Long et al. (2007) reported that most of the variability in an early mortality study (0 to 14 d) between chicks was due to non-identifiable environmental factors (e.g., residual effect). Other studies have shown that first-week mortality is higher when personnel walks through the house more quickly (Cransberg et al., 2000). In contrast, others have shown that leg health is one of the most prevalent causes of culling and late mortality during the grow out of heavy broilers (Cransberg et al., 2000; Oviedo-Rondón, 2008). The literature has shown that mortality has a low heritability, making it a trait that takes several generations and lots of data to show a significant response to selection (González-Recio et al., 2008; Zhang et al., 2018b; Bermann et al., 2021). Additionally, when working with mortality, extra complexity comes from the fact that there are many reasons for death.
Maternal genetics may also be an important factor to consider when assessing early mortality. In pigs, for example, the maternal effect is equally, if not more influential, than the direct effect (Leite et al., 2021). Working with chickens, Zhang et al. (2018b) showed that the additive maternal variance was not negligible for growth traits. At the commercial level, growth is measured early in life; hence, maternal effects play a role early in a chick’s life. Although the impact of maternal effect has been investigated for broiler growth, its influence on mortality remains to be observed.
In broilers, survival is typically measured as a discrete outcome, whether the bird is alive (0) or dead (1). The evaluation of mortality is commonly performed with linear models with the assumption of a continuous distribution for phenotypes. However, it has been presented that in evaluations of categorical traits, it is more suitable to use nonlinear models, such as threshold models, to better accommodate the natural distribution of the discrete outcomes (Gianola, 1982). Using genomic information, Leite et al. (2021) successfully applied threshold models to evaluate pig survival in different production phases. Nonetheless, with the growing amount of genomic information used in evaluations, threshold models have been reported to have convergence and computing-time issues, primarily due to the nonlinearity of the set of equations involved (Hidalgo et al., 2024).
Using genomic information allows not only the identification of variants/genes associated with the trait but also helps understand the underlying variability of the traits. The changes in single-nucleotide polymorphisms (SNP) effects can help infer the genetic dynamics of different trait definitions within a population. This study aimed to assess alternative trait definitions of mortality as well as the inclusion of maternal genetic effects for genomic evaluations of mortality in a broiler population. Furthermore, we investigated linear and threshold models for variance component estimation. Lastly, we explored the genetic architecture of these trait definitions by examining the behavior of SNP effects and how the SNPs influenced the additive genetic effect.
Animal Care and Use Committee approvals were unnecessary because data were obtained from preexisting databases.
Cobb-Vantress, Inc. (Siloam Springs, AR) provided data for 3 lines of broiler breeder containing approximately 3 yr’ worth of mortality records. These mortality records may have resulted from natural deaths or culling during the period when the phenotype was being recorded. Among the datasets, 2 distinct mortality traits were a binary mortality trait (OM) and a categorical weekly mortality trait (WM). The OM was assigned a value of 1 for birds alive and 2 for those deceased, recorded from hatching until the first selection point. The WM was recorded during the initial 10 wk after hatching, with the bird’s phenotype indicating the week of death. If a bird survived beyond these 10 wk, no phenotype was assigned.
To explore alternative trait definitions and models, a new mortality phenotype was created using OM and WM. Using information from OM, phenotypes for animals that survived past 10 wk were generated. Inconsistent records between OM and WM were resolved by assuming OM to be the gold standard since it had been accepted in previous studies (Zhang et al., 2018b; Bermann et al., 2021). Any remaining inconsistent records were removed from the dataset. Birds that received a WM record of 0 were considered missing because the week of death was unknown. Due to the limited number of records in weeks 7, 8, 9, and 10, they were combined with week 6, such that the phenotype of week 6 indicated birds that survived up to at least week 6. After data editing, the remaining records covered weeks 1 to 6. The pedigrees contained at least 245,000 individuals, with at least 77,000 genotyped individuals. Those individuals were genotyped with a 60K Illumina Chicken SNP BeadChip. Furthermore, genotyping was limited to animals selected after the initial selection date, resulting in all genotyped animals having a corresponding phenotypic record. After quality control, 54,305 SNP were retained for line 1 (L1), 42,156 SNP remained for line 2 (L2), and 54,518 SNP were kept for line 3 (L3). A summary of the original data are in Table 1.
The first trait definition, termed “broiler mortality” (BM), was analogous to OM. In this approach, a binary phenotype was created whether the bird survived (1) or died (2) within the first 6 wk of life. The subsequent trait definitions involved dividing the 6 wk into 2 distinct “early mortality” (EM), encompassing the first 3 wk, and “late mortality” (LM), covering the last 3 wk. Once again, both EM and LM were represented as binary phenotypes, receiving 1 if the birds survived or 2 otherwise, within the first 3 (EM) or last 3 (LM) weeks.
Based on WM, 2 other approaches were formulated. The first followed a cumulative strategy (CM), resulting in 6 repeated records for each animal. Starting at 1 (first week), every week the animal survived, the phenotype increased by 1; if the animal was dead, the phenotypic record remained the same until the sixth week of life. For example, if death occurred in week 1, the animal would receive the phenotypic record 1 for all 6 wk, but if it survived week 1 and died at week 2 its records would be all 2. Using the same reasoning, if the animal survived week 2, its phenotype would change to 3 and either remain constant (death at week 3) or be increased by 1 every week, meaning it survived up to 6 wk. The second approach treated weekly mortality as a binary, repeated (RM) measurement. Once again, using 1 (survived) or 2 (died) for every week, but contrary to CM, no additional record was attributed when the animal died. For example, if a bird was dead at week 3, it would have 3 repeated records, being 1 for the first 2 wk and 2 for the third. A summary of the new trait definitions is in Table 2.
Variance components were estimated for all different trait definitions using 2 different scenario 1 (linear models) or scenario 2 (threshold models) with only pedigree information. Three different models were
where: y is the vector of phenotypic records (OM, WM, and BM in [1]; CM and RM in [2]; EM and LM in [3]); β is the vector of fixed/systematic effects (contemporary group and/or age), assumed as β∼N{0,Iσβ2} (only for threshold models); u is the vector of animal additive genetic random effects (u∼N{0,Aσu2}.); p. is the vector of permanent environmental random effects (p∼N{0,Iσp2}. and Cov(u,p′)=0); ud and um are the vectors of additive direct and additive maternal random effects, respectively, assumed as (ud∼N{0,Aσud2}. and um∼N{0,Aσum2}); e is the vector of random residuals (e∼N{0,Iσe2}); X and Z1, and Z3 are incidence matrices for fixed/systematic and random effects, respectively. Single-trait models were implemented for OM, WM, BM, CM, and RM, whereas a 2-trait model (with maternal genic effects) was implemented for EM and LM. Maternal genetic and permanent environment effects were originally included in the model. However, the variance due to the maternal permanent environmental effect approached zero, so it was removed from the model.
Variance components were estimated using BLUPF90 + (scenario 1) and GIBBSF90 + (scenario 2) software (Misztal et al., 2014b). Since Bayesian inference was used for the threshold models, a single-chain of 300,000 samples was generated, assuming a burn-in of 100,000 and thinning interval of 100. Thus, inferences about the variance components were made over 2,000 samples from the posterior distribution. Convergence evidence was assessed graphically and by applying Geweke’s diagnostic test (Geweke and In, 1995), implemented in the POSTGIBBSF90 software (Tsuruta and Misztal, 2006). Variance components from scenario 2 were transformed from the liability to the observed scale following the study by Dempster and Lerner (1950) and Gianola (1979).
All breeding values were predicted using linear models through ssGBLUP employing iteration on data with preconditioned conjugate gradient using the BLUPIOD2OMP1 software from the BLUPF90 suite of programs (Misztal et al., 2014). The variance components estimated using linear models were used directly in the linear models under scenario 1, while those estimated using threshold models were transformed to the observed scale and then used in the linear model under scenario 2. Due to the number of genotyped animals, APY Misztal et al. (2014a) was used to calculate the inverse of genomic relationship matrix (G). The number of core animals was defined as the number of eigenvalues explaining at least 98% of the variation in G (Pocrnic et al., 2016). For L1, L2, and L3, the number of core animals was 5,800, 6,100, and 9,700, respectively. Core animals were kept consistent throughout each model’s prediction of breeding values. The implemented statistical models were as before, except for replacing A^−1^ by H^−1^ (Aguilar et al., 2010), the inverse of the realized relationship matrix in ssGBLUP.
Using ssGBLUP with APY, GEBV were predicted for each trait definition with each model. Then, SNP effects were estimated by backsolving GEBV for each model according to Bermann et al. (2022) taking advantage of the equivalence between APY ssSNP-BLUP and APY ssGBLUP:
where: a^ is the vector of SNP effects, f(ai) is the allelic frequency of the ith SNP, Zc is a centered SNP content matrix for core animals, u^c is the vector of GEBV for core animals, and Gcc−1 is the inverse of the genomic relationship matrix among core animals, with G constructed
where: α is the blending parameter (5%) to avoid singularity problems in G (Vitezica et al., 2011); a and b are tuning parameters to ensure compatibility between G and A22 (VanRaden, 2008), with other elements previously described. Each trait definition and model were used to perform a genome-wide association study (GWAS) to investigate the underlying complex genetic architecture of mortality. Additionally, the percentage of additive genetic variance explained by each SNP was also calculated, as well as the P-value associated with each SNP following Bermann et al. (2022) and Leite et al. (2023).
The linear regression (LR) method (Legarra and Reverter, 2018) was employed to assess the performance of each model and trait definition with respect to OM. This approach compares genetic evaluations, including whole and partial datasets, based on differences in means and covariances using a set of focal individuals. The number of focal animals changed from line to line since they were selected as young, genotyped animals in the partial data. Generally, 1,914 (L1), 709 (L2), and 4,376 (L3) animals were used for validation. The partial dataset was created by removing phenotypes from the focal animals, their siblings, and contemporaries. Let the partial dataset be denoted by subscript p, while the whole dataset is denoted by subscript w, then validation statistics
where: u^w and u^p are the vectors of GEBV for focal animals from whole and partial datasets, respectively, F¯ is the average inbreeding coefficient of validation animals, σu2 is the additive genetic variance; and u^¯w and u^¯p are the GEBV averages from whole and partial data, respectively.
Posterior means and standard deviations of heritability estimates from scenarios 1 and 2 are presented in Tables 3 and 4, respectively. Direct heritability’s ranged from 0.01 to 0.16 for OM, WM, BM, CM, and RM in scenario 1 and from 0.11 to 0.21 for OM, WM, and BM in scenario 2, indicating a low but existing genetic variability across trait definitions. Scenario 2 produced higher estimates of heritability in the liability scale, but when converted to the observed scale (data not shown), variances were reduced, and heritability estimates were also reduced almost to the same value estimated using the linear model. The heritability estimates for WM under the scenario 1 are much larger than those for the other trait definitions. This larger heritability could be due to the structure of WM, where there are several more levels of responses, thus approaching normality. The heritability estimates for OM and BM were very similar across lines, most likely due to OM and BM being very similar traits. Heritability estimates for the CM and RM were consistently slightly lower than those for OM, WM, and BM, with RM having lower estimates than CM. When comparing overall estimates across lines, L2 showed higher heritability estimates among these trait definitions than L1 and L3, with L3 consistently having lower estimates. This may suggest that environmental factors have a more considerable impact on mortality for L3.
Our heritability estimates agreed with those reported in the literature. Zhang et al. (2018b) reported heritability for a general chicken mortality trait to be 0.14 on the liability scale using a threshold model with sex as a systematic effect and contemporary group as random effect. Bermann et al. (2021) also reported heritability for mortality to be 0.14 on the liability scale using threshold models. These heritability estimates were estimated using very similar models as the model used to estimate OM. However, in a small experimental group of chickens, Pakdel et al. (2005) reported heritability of 0.32 for a total mortality trait that spanned the first 7 wk of the bird’s life using linear models with additional fixed effects. This higher heritability estimate may be due to the number of records, size of population, or the differences in the statistical model. Differences between the estimates from this study and the estimates reported in literature could be due to several factors, such as strategies used for estimation, differences in measurement of the trait, different trait definitions or differences in sample size (Rekaya et al., 2013).
In this study, the comparison of heritability estimates shows that these estimates can vary depending on the statistical model used for analysis. When modeling traits with the addition of other random effects, such as the maternal genetic effect in our case, the direct heritability may be reduced. Without including this effect in the model, the maternal genetic variance may be absorbed into the additive genetic and the residual variances. This is observed in the estimates of genetic parameters from both linear and threshold models. In the case of EM and LM, where variance components were estimated using 2-trait models, the heritability estimates are higher on average due to the increased information provided by genetic correlations between the traits.
For lines L1 and L3, heritabilities for EM were slightly lower than for LM whereas for L2, LM was lower than EM. This was true for both scenarios 1 and 2. González-Recio et al. (2008) reported a heritability of 0.02 from linear models for a “late mortality” trait (14 to 42 d of age) in chickens. The authors stated that it seems unlikely that the true value of this parameter is higher than 0.05 due to the amplitude of their posterior intervals. Our results from scenario 1 agree with those from González-Recio et al. (2008), most likely because when modeling a binary trait as a linear output, a small variance is captured by the model. Gianola (1982) also stated that discrete traits, when analyzed as continuous, may have underestimated heritabilities, especially when incidences are low. However, other studies found ascites-related mortality, heart-failure mortality, and heart-lung failure mortality to have higher heritabilities ranging from 0.06 to 0.15 (de Greef et al., 2001). These conditions contributing to mortality are typically found to cause issues later during the chicken grow out period. Therefore, there may be more prominent underlying genetic factors for conditions leading to mortality during the later periods of grow-out, making heritabilities larger. Another reason may be caused by how their definition of mortality was developed in that the heart-lung failure mortality encompassed that of the other 2 mortality traits. In addition, LM may have higher heritability estimates compared to EM due to selective phenotyping. If an animal dies in EM, it will have a missing record for LM. Conversely, if an animal survives EM, it should have a record for LM.
Genetic correlations between EM and LM in the 2-trait models are in Table 5. The values ranged from 0.48 to 0.81 in the scenario 1 (among the 3 lines) and from 0.52 to 0.64 in scenario 2. These genetic correlations suggest that EM and LM may be considered different traits. In other species, such as pigs, it has been shown that preweaning and postweaning mortality are 2 genetically different traits with low correlations and sometimes even divergent. Leite et al. (2021) reported a strong, positive correlation of 0.83 between nursery and finishing survival traits, while Dufrasne et al. (2014) reported a moderately positive correlation of 0.59. Due to the lower correlations in the latter, mortality at different phases was considered different traits, and both were included in the evaluation. It may also be necessary to include both traits in the selection index for poultry to make better-informed decisions. Across the lines, differences in genetic correlations may be related with differences in genetic background or selection within each population. Each line may have their own levels of pleiotropy or linkage disequilibrium that plays a role in the degree of genetic correlation.
When including the maternal genetic effect into scenario 1 with EM and LM, heritability estimates for the direct effects of EM decreased by 50% for L1, 88% for L2, and 60% for L3, whereas the decrease for LM was 33% for L1 and L2, and no change was observed for L3. Under scenario 2, including the maternal genetic effect caused a decrease of 60% for L1, 81% for L2, and 50% for L3 for EM. Meanwhile, for LM, the direct heritability decreased only by 27% for L1 and 33% for L2 and increased by 7% for L3. Maternal heritability estimates for EM (EMm) were 0.01 (L1), 0.02 (L2), and 0.01 (L3) for scenario 1, and 0.04 (L1), 0.06 (L2), 0.03 (L3), for scenario 2. As for LM maternal effects (LMm), those accounted for 0.01 (L1) and 0.00 (L2 and L3) of the total variance for scenario 1 and 0.02 (L1) and 0.01 (L2 and L3) for scenario 2.
Grandinson et al. (2005) reported a low maternal heritability for preweaning mortality (0.02), whereas Lund et al. (2002) reported a maternal heritability of 0.08 under linear models. Leite et al. (2021) reported maternal heritability from threshold models for piglet farrowing survival and piglet lactation survival to be 0.09 and 0.08, respectively, thus indicating that both piglet and sow genetics impact piglet survival. These authors also found that the maternal heritability remained constant over time. However, in this same study, the piglet additive genetic variance increased as the piglet aged, suggesting that the importance of the piglet’s own genes increases as they age. Our results agree with studies mentioned above on including the maternal genetic effects. Adding the maternal genetic effect reduced the direct heritability by an average of 0.04 for EM and 0.01 for LM under scenario 1 (across all lines), with a larger average reduction of 0.10 for EM and 0.03 for LM under scenario 2. Including the maternal genetic effect considerably decreased direct heritability for EM but not for LM under both scenarios 1 and 2, indicating the maternal effect plays a more important role earlier in the chick’s life and should be considered in the models.
Validation statistics from the LR method are presented in Figure 1 for all trait definitions except for RM and CM. Both CM and RM had comparable accuracy to other trait definitions, but the bias average standard deviation for the repeatability models was over 8-fold larger than the average of the other trait definitions. With this bias, it was determined that the repeatability models were not viable, and thus, their results were not shown. Compared to the baseline definition under scenario 1 for L1 (OM—0.43), the accuracy increased by 12% for WM (0.48), 2% for BM (0.44), and for LMd (0.46), except for EM, LM, and EMd. The accuracy remained constant between OM and LM, decreasing by 12% for EM and EMd (0.38 and 0.38).
Figure 1. Prediction accuracy, bias, dispersion breeding values from ssGBLUP evaluations between whole and partial datasets for both scenarios 1 and 2 for each trait mortality (OM), weekly mortality (WM), broiler mortality (BM), early and late mortality (EM and LM) and early and late mortality with maternal genetic effect (EM
d, LMd, EMm, and LMm) among all 3 lines.
For L2, compared to the baseline definition under scenario 1 (OM—0.53), the accuracy decreased by 66% for WM (0.18), 34% for BM (0.35), 28% for EM (0.38), 19% for LM (0.43), 32% for EMd (0.36), and 13% for LMd (0.46). Following a similar trend as L2, in L3, the comparison between the baseline definition under scenario 1 (OM—0.39), the accuracy only increased by 5% for LM (0.41). The accuracy decreased by 28% for WM (0.28), 8% for BM (0.36), 13% for EM (0.34), 15% for EMd (0.33), and 3% for LMd (0.39). When using scenario 1, OM generally produced an overall better accuracy than any of the other trait definitions explored across the 3 lines of data.
When comparing the baseline scenarios 1 to 2, the accuracy increased by 2% for L1 (0.45 vs. 0.46) but then decreased by 6% for L2 (0.53 vs. 0.50) and by 10% for L3 (0.39 vs. 0.35). When comparing the baseline definition under scenario 1 to the remaining scenario 2 trait definitions for L1, the accuracy increased by 9% for WM (0.47), by 5% for BM (0.45), by 18% for LM (0.87), and by 26% for LMd (0.54). The accuracy decreased by 12% for EM (0.38) and by 7% for EMd (0.40) for L1. For L2, the comparison between the baseline definition under scenario 1 and the scenario 2 trait definitions showed a decrease in accuracy for all trait definitions except for LMd, which showed an increase of 6% (0.56). The remaining trait definitions showed a decrease by 9% for WM (0.48), by 28% for BM (0.38), by 36% for EM (0.34), by 9% for LM (0.48), and by 23% for EMd (0.41) for L2. Furthermore, in L3, the comparison between baseline definition under scenario 1, OM, with all remaining threshold trait definitions, revealed an increase in accuracy by 51% for WM (0.59), by 3% for BM (0.40), by 28% for LM (0.50), and by 21% for LMd (0.47). The accuracy decreased by 13% for both EM (0.34) and EMd (0.34).
Consistently, the accuracy for EM and EMd was lower than the other trait definitions and OM. This lower accuracy may be due to EM containing less information than the other trait definitions since it only contained records from the first 3 wk of the bird’s life. The changes in accuracy were line-specific, probably because of the amount of information available, genetic architecture of the trait, selection intensity, and/or the genotyping strategy within each line. The accuracies from the scenario 2 were generally better than those from scenario 1, indicating that scenario 2 might be more appropriate when modeling binary chicken mortality. By decomposing mortality in EM and LM, part of the accuracy is lost for EM but regained and improved for LM. Thus, considering EM and LM in a 2-trait model will be comparable, if not better, than OM. This is also true when including the maternal effect into the model with EM and LM as the accuracies further increased.
The estimated bias for all trait definitions was ranged from −0.10 to 0.10. Among trait definitions, the bias ranged from −0.05 to 0.01 for L1, −0.06 to 0.07 for L2, and 0.00 to 0.05 for L3 among scenario 1. Among scenario 2, the bias ranged from −0.03 to 0.01 for L1, −0.06 to 0.09 for L2, and −0.01 to 0.06 for L3. On average, L3 had the largest bias compared to L2, which had the lowest bias. Bias estimates were comparable from scenarios 1 and 2 with similar trends across trait definitions; however, specific trends of bias were not consistent across lines. Predictions were mostly over dispersed for all trait definitions, with a b1 ranging from 0.81 to 0.94 for L1, 0.76 to 0.95 for L2, and 0.77 to 0.86 for L3 for scenario 1. For scenario 2, dispersion estimates ranged from 0.86 to 1.02 for L1, 0.81 to 1.01 for L2, and 0.79 to 0.88 for L3. Bias and dispersion estimates were not consistent across lines, and thus, once again, these results are line specific and cannot be generalized across these lines.
Spearman’s rank correlations were calculated between OM and the remaining trait definitions from scenarios 1 and 2 (Figure 2). Line 2 had an overall smaller reranking compared to L1 and L3. Rank correlations from scenario 1 indicated a smaller reranking for BM, EM, LM, and EMm across all 3 lines, with values above 0.70. A larger reranking existed between OM and WM and LMd. For all 3 lines, the correlation between WM and the rest of the trait definitions ranged from small negative to small positive values, centering around zero, indicating that the ranking for WM is not the same as the original mortality as well as the remaining trait definitions.
Figure 2. Rank correlations of breeding values for L1, L2, and L3 from scenarios 1 and 2 for each trait mortality (OM), weekly mortality (WM), broiler mortality (BM), early and late mortality (EM and LM), and early and late mortality with maternal genetic effect (EM
d, LMd) among all 3 lines.
Rank correlations from scenario 2 followed a similar trend to scenario 1 where values indicated a smaller reranking for BM, EM, LM, and EMm and a larger reranking for WM and LMd across all 3 lines. For L2 and L3 from scenario 2, rank correlations between WM and other trait definitions were mostly smaller except for EM and LMd. Across all lines and models, rank correlation values from LMd with all other traits, evidencing that when the maternal effect is modeled with LM, there is a much larger reranking of individuals. For a breeding program, it is important to be aware of how changes in modeling may affect the ranking of selection candidates. These correlations show that for most trait definitions, reranking should be expected.
Our GWAS resulted in no SNP having significant associations (P-value > 0.05/number of SNP in each line) with any trait definitions from all available datasets. These results indicate that for these data, there are no major regions or genes greatly influencing mortality under these trait definitions. For all traits, no SNP explained more than 1% of the additive genetic variance, with the largest peak explaining not more than 0.20%. A consistent SNP peak with a linkage disequilibrium trail was found in OM, EMd, and LMd, yet it did not explain more than 1% of the additive genetic variance. Most peaks for OM, EMd, and LMd are not consistent across models. As dead animals were not genotyped in these populations, the results are one-sided in that we only have genomic information of the animals that survived. When considering the SNP effects, the magnitude was different when comparing OM to EM, as shown in Figure 3. This difference is most likely caused by changes in the GEBV distributions, as shown in Figure 4. When comparing OM to LM, the magnitude of SNP effects is more similar, just as the distribution of GEBV is more similar for these trait definitions.
Figure 3. SNP effects for baseline mortality (OM) and early and late mortality with the inclusion of the maternal genetic effect (EM
d, LMd, EMm, and LMm) for L1. The colors show the percentage of additive genetic variance explained by each SNP.
Figure 4. Densities of genomic breeding values for baseline mortality (OM) and early and late mortality with the inclusion of the maternal genetic effect (EM
d, LMd, EMm, and LMm) for L1 for all genotyped individuals.
Tsuruta et al. (2017) reported SNP variance peaks that reached 2.5% for mortality around the location of the DGAT1 gene on BTA 14, which is also associated with milk production in dairy cattle. These results indicate that for this population, there is a relationship between cow mortality and milk production. However, Berry et al. (2010) reported no effect of DGAT1 on dairy cow longevity, which is related to mortality. In pigs, Guo et al. (2016) reported that QTL regions identified for early mortality overlapped with the regions associated with the total number born alive. Fragomeni et al. (2014) found that the variation explained by the top SNP windows in a chicken population changed over generations, indicating that SNP with large variance detected in one dataset may not be detected in the following generations. Thus, the usefulness of those SNP over time is limited. As our peaks differ depending on the trait definition, working with mortality as a single trait may not be optimal. Instead, splitting up mortality into time periods may help get a clearer understanding of the underlying genetic architecture of the different phases of mortality.
Overall, mortality is a complex trait to select against because of many differing factors, both genetic and environmental. As this trait has genetic variability, sustained improvement is possible if it is included in the index as a long-term breeding goal. By using alternative trait definitions for mortality, small increases in the accuracy may be achieved, allowing for improvements at the phenotypic level. Reverter et al. (2022) showed that even small changes in accuracy of breeding values can have a significant impact on phenotypic changes. As shown in this study, using threshold models to estimate variance components and then transforming those estimates to the linear scale allows for higher estimates of accuracy, indicating that by using this modeling strategy, there could be better improvements if the costs of using threshold models are feasible. Other methods of modeling mortality, such as survival models, have been shown to add no real benefits compared to animal models (Poulsen et al., 2022). Thus, further investigation of mortality models is needed to continue to make progress on this trait.
Diversifying trait definitions offers a promising approach to enhance the accuracy of predicting mortality GEBV and refine selection strategies for future generations of broiler breeders. Our exploration and testing of various trait definitions advocate for segmenting mortality into distinct time phases to consider the varying genetic influences on survival. Notably, incorporating maternal effects, especially during the early stages of a chick’s life when the maternal genetic impact is significant, holds the potential to improve the understanding of mortality mechanisms. This study highlights the suboptimal nature of treating mortality as a simple binary trait, clearly demonstrated by the fluctuations in variance explained by SNP based on different trait definitions as well as the fluctuations of additive genetic variances reflected by the different heritability estimates. Seeking better ways to model mortality will undoubtedly advance our understanding of its underlying genetic architecture and enable more effective breeding practices to improve overall broiler breeder survival, performance, and welfare.
This study was supported by Cobb-Vantress, Inc.
Jennifer Richter, Department of Animal and Dairy Science, University of Georgia, Athens, GA 30602, USA.
Fernando Bussiman, Department of Animal and Dairy Science, University of Georgia, Athens, GA 30602, USA.
Jorge Hidalgo, Department of Animal and Dairy Science, University of Georgia, Athens, GA 30602, USA.
Vivian Breen, Cobb-Vantress, Inc., Siloam Springs, AR 72761, USA.
Ignacy Misztal, Department of Animal and Dairy Science, University of Georgia, Athens, GA 30602, USA.
Daniela Lourenco, Department of Animal and Dairy Science, University of Georgia, Athens, GA 30602, USA.
Vivian Breen is an employee of Cobb-Vantress, Inc (Siloam Springs, AR). The authors declare no real or perceived conflicts of interest.