Animal breeders often face an awkward problem: the trait they truly want to improve is expensive, difficult or impossible to measure on every animal, while cheaper proxy traits are scattered unevenly across the population. A study published in Genetics Selection Evolution now offers a way out. Researchers Alice Markey and Nicolas Gengler, working in Belgium, developed a deregression-based pooling method that compresses information from several correlated traits into a single pseudo-phenotype of the target trait, which can then be fed into a standard single-trait genetic evaluation. They tested the approach on boar taint in terminal crossbred pigs, a trait whose chemical basis involves multiple compounds and whose direct sensory measurements are costly.
The core idea builds on deregression, a technique widely used when original phenotypic records are unavailable and only published breeding values remain. Deregression converts estimated breeding values into pseudo-records, called deregressed proofs, that represent the non-parental part of an animal’s breeding value, essentially the Mendelian sampling term. Each proof carries a weight, here the deregressed effective record contribution, which quantifies how much information supports it. In this study, the authors applied deregression specifically to the breeding values of the target trait obtained from a multitrait model, so that the resulting pseudo-phenotype implicitly carries the genetic information contributed by the correlated traits skatole and androstenone.
Classical deregression, however, hides two pitfalls that the authors set out to correct. First, deregressed proofs are predicted values, and the variance of predictions is always smaller than or equal to that of the underlying variable they represent. The proofs are therefore too strongly regressed, and breeding values regenerated from them come out shrunken. Second, the weights attached to the proofs must accurately reflect their information content, yet the deregression step is tuned to the single-trait model that will consume the proofs, not the multitrait model that generated the breeding values. This mismatch distorts the genetic variance and leaves reliabilities incoherent with the breeding values they accompany.
To fix both problems, the researchers introduced two iterative adjustment factors. An expansion factor, borrowed conceptually from work by VanRaden and colleagues, stretches the deregressed proofs so that the standard deviation of the regenerated breeding values matches that of the original multitrait evaluation. A reliability-alignment factor, defined as the ratio of the standard deviation of multitrait reliabilities to that of single-trait reliabilities, rescales the weights so that reliabilities line up as well. Both factors are recomputed and reapplied within each iteration until the standard deviations converge and the relative squared change in correlations between successive iterations falls below a strict threshold. The deregression step was implemented in SAS 9.4, while the mixed model evaluations ran through the BLUPF90+ program with EM-REML, with convergence declared at a tolerance of ten to the power of minus twelve.
The testing ground was boar taint in Piétrain x Landrace crossbred pigs. In the real dataset, sensory scores were available for 2169 animals, while chemical quantifications of skatole and androstenone existed for only 388 of them, measured on neck fat collected at slaughter and analysed by an ISO17025-accredited laboratory using UHPLC-MS/MS. The three-point sensory scale was converted into a Snell score to approximate a continuous distribution, the chemical measurements were log-transformed, and all three traits were standardized. A three-generation pedigree covering 2672 animals supported the variance component estimation that anchored the rest of the study.
Because the real dataset was small and its missing-data pattern fixed, the team built an extended simulated dataset from the full Walloon pig genetic evaluation pedigree, covering 9590 crossbred animals tested since 2007 and their ancestors, 11,615 animals in total. Following a strategy adapted from Van Vleck, they simulated genetic and residual effects from multivariate normal distributions and iteratively adjusted the covariance structures until the re-estimated variance components matched those of the real data, with Frobenius norms of the relative differences all below one percent. From this synthetic population they carved out three data-availability scenarios: all three traits recorded on all animals; the target trait recorded on 90 percent of animals and the proxies on only 15 percent; and the extreme case in which the target trait is entirely absent and only skatole and androstenone are measured.
The results were strikingly consistent. Across every dataset and scenario, unadjusted deregression compressed the breeding values, with regression slopes on the multitrait reference ranging from 0.781 to 0.883 and standard deviations systematically lower, for example 0.119 versus 0.150 for phenotyped animals in the real dataset. After adjustment, the slopes returned to nearly one with intercepts near zero, and standard deviations recovered to multitrait levels, such as 0.149 in the real data and 0.170 to 0.198 across the simulated cases. The adjustment factors themselves converged quickly, with the expansion factor for the most extreme scenario declining asymptotically from 0.82 to 0.53 and the reliability-alignment factor settling at 0.85, and all datasets reaching convergence within about ten rounds, with the target-absent case needing fourteen.
Accuracy metrics reinforced the picture. Treating reliabilities as the squared correlation between true and estimated breeding values, the multitrait model and the adjusted deregression produced essentially equivalent correlations, for instance 0.609 in the target-absent scenario, while unadjusted deregression yielded a slightly lower 0.601. More tellingly, the gap between reported average reliabilities and theoretical expectations was small for the multitrait and adjusted systems, 0.001 and 0.031 respectively, but systematically larger for unadjusted deregression at 0.069, exposing its biased accounting of information content. Mean reliabilities told the same story: in the real dataset, the multitrait mean of 0.291 for phenotyped animals dropped to 0.282 under classical deregression but returned to 0.292 after adjustment, with parallel figures of 0.411, 0.405 and 0.416 for sires.
The authors describe the resulting property as extended reversibility: pseudo-records generated by the adjusted process, when re-evaluated with the single-trait model, reproduce both the breeding values and the reliabilities of the original multitrait evaluation. They are careful to note the method’s inherent constraint, namely that the pooled phenotype can only recover information already present in the underlying multitrait prediction. When the target trait is missing, usefulness depends on the genetic correlations with the available proxies, which were moderate to high in this study. For traits with low heritability and weak proxy correlations, prediction accuracy of the missing target would be expected to fall substantially, and the adjustment factors cannot conjure information that the initial evaluation never contained.
The framework opens several practical avenues. It could be adapted to sire-level implementations, in the manner of Multiple Across Country Evaluation results recovered from INTERBULL, provided double-counting of progeny and sire information is avoided. It may extend to high-dimensional phenotyping from sensors and multi-omic platforms, and to integrating historical pseudo-records into routine evaluations, including single-step genomic BLUP, with source-specific fixed effects distinguishing pooled from directly observed records. The method has not yet been formulated within a genomic evaluation framework, which the authors flag as future work, along with simulations varying heritability, genetic correlations and recording intensity to map the conditions under which pooling pays off. For now, the study demonstrates that a carefully adjusted deregression can let breeders fold sparse, heterogeneous trait information into one usable number, even when the trait they care about was never measured at all.
Subject of Research: A deregression-based method that pools multitrait information into a single pooled phenotype for genetic evaluation of boar taint in pigs
Article Title: A deregression-based approach to integrate multitrait information into a single pooled phenotype for genetic evaluation
Article References: Markey, A., & Gengler, N. (2026). A deregression-based approach to integrate multitrait information into a single pooled phenotype for genetic evaluation. Genetics Selection Evolution, 58(1), Article 76. https://doi.org/10.1186/s12711-026-01096-z
Image Credits: AI Generated
DOI: 10.1186/s12711-026-01096-z
Keywords: genetic evaluation, deregression, pooled phenotype, multitrait model, boar taint, pigs, breeding values, reliability, animal breeding, crossbreeding, simulation study, deregression-based
News Source: Juliet Wilcox. (October 27, 2026). New pooling method turns scattered pig traits into one breeding score. Scienmag.
