Improving selection performance in dairy cattle through phenomic selection

Background and challenges

Since the 2000s, breeding programmes have used performance predictions to complement data on non-evaluated individuals. These predictions are based on information from the genome of the individuals: genotyping data. This strategy, known as genomic selection, has significantly increased the efficiency of breeding programmes for many animal and plant species and has become a reference method in genetic improvement.

However, genomic selection has one drawback: the need to have genotyping data, which in some cases is too expensive to obtain (e.g. for field crop species for which thousands of candidates are produced each year, or for orphan species for which no efficient genotyping tool exists).

Phenomenal selection: a promising new alternative?

One alternative is to use phenomic selection, recently introduced by Rincent et al. (2018), which consists of making performance predictions from phenomic data obtained by spectroscopy, rather than from genomic data. Spectroscopy has the advantage of being inexpensive, non-destructive, and already routinely implemented, both in breeding programmes for many plant species (to assess product quality) and in some animal species, notably in milk improvement programmes.

The prediction performances obtained for different study cases are similar to those obtained with genomic prediction models.

This very recent method has never yet been evaluated in an animal model and needs to be more widely tested and optimised.

Goals

The DeepPhenomic project proposes a first application of phenomic selection to an animal model: the method will be tested in dairy cattle, in a large-scale system (several tens of thousands of animals with mid-infrared spectra on milk, of which approximately 8,000 are genotyped).

The results of the phenomic predictions will be compared with those of a classical genomic evaluation.

The project also plans to optimise the exploitation of spectral data with functional methods on the one hand and neural networks on the other:

• functional analysis will be specifically tested in a multi-environment context, where the prediction of unobserved spectra could increase the accuracy of phenomic prediction.
• Neural networks will be used to test the interest of artificial intelligence methods in the context of phenomic selection, thanks to the very broad scope of the experiment.

If successful, this work could have important implications for dairy cattle improvement, and would constitute a proof of concept for many other animal and plant species.

Contact - coordination

Units involved and partners

INRAE participants

Partners