Investigating causal biological relationships between reproductive performance traits in high-performing gilts and sows1

Abstract

Efficient management of swine production systems requires understanding of complex reproductive physiological mechanisms. Our objective in this study was to investigate potential causal biological relationships between reproductive performance traits in high-producing gilts and sows. Data originated from a nutrition experiment and consisted of 200 sows and 440 gilts arranged in body weight blocks and randomly assigned to dietary treatments during late gestation at a commercial swine farm. Reproductive performance traits consisted of weight gain during late gestation, total number born and number born alive in a litter, born alive average birth weight, wean-to-estrous interval, and total litter size born in the subsequent farrowing. Structural equation models combined with the inductive causation algorithm, both adapted to a hierarchical Bayesian framework, were employed to search for, estimate, and infer upon causal links between the traits within each parity group. Results indicated potentially distinct reproductive networks for gilts and for sows. Sows showed sparse connectivity between reproductive traits, whereas the network learned for gilts was densely interconnected, suggesting closely linked physiological mechanisms in younger females, with a potential for ripple effects throughout their productive lifecycle in response to early implementation of tailored managerial interventions. Cross-validation analyses indicated substantial network stability both for the general structure and for individual links, though results about directionality of such links were unstable in this study and will need further investigation. An assessment of relative statistical power in sows and gilts indicated that the observed network discrepancies may be partially explained on a biological basis. In summary, our results suggest distinctly heterogeneous mechanistic networks of reproductive physiology for gilts and sows, consistent with physiological differences between the groups. These findings have potential practical implications for integrated understanding and differential management of gilts and sows to enhance efficiency of swine production systems.

Keywords: hierarchical Bayesian models; structural equation model; structure learning; swine reproductive physiology.

Figure 1.

Scatterplot matrices depicting the empirical distributions and marginal associations between reproductive performance traits in sows (left panel) and in gilts (right panel). For each panel, histograms are shown along the main diagonal, whereas bivariate scatterplots are shown on the lower triangle and estimated Pearson correlation coefficients are presented on the upper triangle. Asterisks indicate that the coefficient is significantly different from zero (***P-value < 0.0001, **P-value < 0.01, *P-value < 0.05). GAIN = female weight gain during late gestation; TB = total number born in a litter; BA = number born alive in a litter; BABW = born alive average body weight; WEI = wean-to-estrous interval; SuTB = total number born in the subsequent gestation.

Figure 2.

(A) Undirected graph of reproductive performance traits in sows detected by the inductive causation algorithm implemented with 80% highest posterior density intervals. (B) Plausible structures within the equivalence class defined by traits connected in Panel (A). Links without arrowheads represent associations between traits; links with arrowheads represent causal effects from the trait on the arrow tail to the trait on the arrowhead. GAIN = female weight gain during late gestation; TB = total number born in a litter; BA = number born alive in a litter; BABW = born alive average body weight; WEI = wean-to-estrous interval; SuTB = total number born in the subsequent gestation.

Figure 3.

(A) Partially oriented graph of reproductive performance traits in gilts detected by the inductive causation algorithm implemented with 80% highest posterior density intervals. (B) Fully oriented graph obtained after incorporating additional temporal information to (A). Links without arrowheads represent associations between traits; links with arrowheads represent causal effects from the trait on the arrow tail to the trait on the arrowhead. GAIN = female weight gain during late gestation; TB = total number born in a litter; BA = number born alive in a litter; BABW = born alive average body weight; WEI = wean-to-estrous interval; SuTB = total number born in the subsequent gestation.

Figure 4.

Links and posterior means of structural coefficients between reproductive performance traits in sows (A) and in gilts (B) learned using a mixed-models adapted inductive causation algorithm implemented with 80% highest posterior density intervals. Refer to Table 2 for further details. GAIN = female weight gain during late gestation; TB = total number born in a litter; BA = number born alive in a litter; BABW = born alive average body weight; WEI = wean-to-estrous interval; SuTB = total number born in the subsequent gestation.

Figure 5.

Stability analysis of the learned network of reproductive performance traits for sows (A) and gilts (B) using a leave-one-block-out Jackknife resampling approach. Values (%) indicate the percentage of resampled datasets for which each link was present in the learned network structure. GAIN = female weight gain during late gestation; TB = total number born in a litter; BA = number born alive in a litter; BABW = born alive average body weight; WEI = wean-to-estrous interval; SuTB = total number born in the subsequent gestation.

Figure 6.

Network structures learned from 5 subsets (labeled panels A–E) created from the gilt dataset by random sampling without replacement, to mimic the sow dataset in size and structure. GAIN = female weight gain during late gestation; TB = total number born in a litter; BA = number born alive in a litter; BABW = born alive average body weight; WEI = wean-to-estrous interval; SuTB = total number born in the subsequent gestation.