Updates to Growth Trait Predictions in BOLT

When the current genetic evaluation system at International Genetic Solutions (IGS), powered by BOLT software, was deployed several years ago, it was viewed as a ‘quantum leap’ forward in genetic prediction. IGS is the current service entity that runs genetic evaluation services for more than a dozen partner breeds including NALF and the Canadian Limousin Association. Single-step utilization of genomic data was the key feature of the new evaluation methodology proving EPD for growth, maternal, calving ease and carcass traits. As with any system, it is constantly evolving. Over the last year several US and Australian breed organizations have joined the partnership as well as Neogen.

The IGS science team has been busy evolving the models used to produce EPD and several genomic data processing improvements. The goal of each of the changes has been to improve the prediction of growth traits. The improvements result in more accurate EPD that do a better job of predicting progeny performance. Several the changes to the evaluation system are outlined below and are anticipated to be deployed during the summer of 2020.

Genomics tools have revolutionized livestock improvement systems around the globe. With many of IGS partner breeds now routinely submitting genotypes to the evaluation, development of a new genomics pipeline was needed. This new data pathway from genotype lab to evaluation software features a new multibreed imputation process to provide better predicted genotypes from a variety of genotype marker density products. The new process has been in service since late summer 2019 and appears to be performing very well.

Yet to be implemented in the production run is a planned update in the marker subset used in the evaluation. The subset used in production represent the first marker selection for markers that had genetic association with more than one trait in the evaluation. Not all makers of the approximately 50,000 used as the imputation target or on the 50K genotype platform influence a trait. In fact, a vast majority don’t. The new marker subset was constructed to use a reduced set where each maker had impact on at least one trait eliminating the forced correlation structure between traits. The number of markers were reduced to a minimal set to avoid overfitting the genetic model and introduction of noise and maximize prediction accuracy.

A new contemporary group definition was implemented for weaning weight and milk EPD calculations. The new grouping forces all calves born to first calf heifers (2-year-olds) into groups separate from progeny from mature cows.  Many producers manage these young cows and their calves separately but don’t always get them coded into appropriate contemporary groups reflective of that difference in management. The new system reduces that environmental noise by automatically dividing those groups.

Another improvement to the weaning weight model is a change to the genetic correlation between weaning weight direct (calf growth) and milk (weaning weight influenced by maternal environment). This correlation had been set at -0.3 since the evaluation was performed at Cornell. The new model uses a ‘0’ correlation. There exists a wide range of estimates for this correlation in the scientific literature. The science team has agreed that considering those differences the best approach is to treat them independent traits (correlation of zero). One impact of the change will be removal of downward pressure on Milk EPD for very high growth animals that is observed in the current production run. In the case of non-parents with high WW EPD, Milk EPD was being adjusted downward without any daughter calf records in the analysis.

It’s well documented that male calves grow faster than female calves. As a result, there tends to be more phenotypic variation in the weight measurements of males versus females. The updated evaluation system models the variance amongst records for the genders differently, providing better predictions.

The fifth major change is the implementation of a method to handle different methods of collecting birth weights. When the IGS science team began looking at the growth trait data they observed that not all reported data followed expected amounts of variation. Much of that difference stems from different methods used to observe birth weights.  Sometimes producers report in two- or five-pound increments due to the scale used. Other producers may use a hoof tape to estimate weights which substantially reduces variation. In some cases, the data was the clearly fabricated (all weights in the group were equal) or the result of using standardized weights when none was reported. A machine learning algorithm was developed to classify birth data contemporary groups. Once classified, the useful records are included in the evaluation accounting for the differences in variation for each classification. Clearly falsified data is omitted from the evaluation.

Lastly, the IGS science team has conducted a wide range of validation studies to evaluate the performance of the current production model and new features proposed for deployment in the upgraded evaluation system. The updates resulted in substantial improvements in the predictive capacity of the EPDs produced. One statistical approach used to evaluate the performance of the two EPD systems is the correlation between animals’ EPDs and the adjusted performance records when their data was not included in the analysis. A second criteria was an analysis to detect prediction bias. In effect the test evaluated the relationship between an animal’s adjusted phenotype and it’s EPD for that trait. A departure in the test parameter from 1 indicated a bias. The test parameter of 1 indicated a one-unit change in EPD was associated with a one-unit change in adjusted phenotype. The proposed improvements in the evaluation resulted in decreases in bias across birth, weaning and yearling weight and milk evaluations.