Application of machine learning and deep learning to improve animal genomic selection

Background and challenges

The most commonly used models in genomic prediction (additive genetic model such as GBLUP) are very efficient in predicting the genetic value of animals on a few genetically correlated traits.  On the other hand, this type of model does not allow the integration of a very large number of heterogeneous measurements, nor does it predict many output traits without knowing their genetic correlations. Moreover, this model is limited in its ability to take into account the many non-linear interactions that occur between regions of the genome or environmental factors.

In order to overcome these obstacles, we propose statistical learning (machine learning) and deep learning methods, derived from AI, to process both additive genetic information and non-linear genetic information present in massive genotyping data.

Goals

The GenIALearn project proposes to evaluate the performance of statistical and deep learning methods for the joint prediction of multiple complex traits, by integrating massive genotyping data. Two main families of methods will be compared altogether and versus the reference method GBLUP:

• On the one hand, ensemble learning methods (random forests, gradient boosting), coupled with a learning step to represent the input data, in order to propose reference prediction levels;
• On the other hand, deep learning methods of different architectures (neural networks), coupled with learning step on massive data base,  which should  produce predictive models  adapted for animal genomic selection.

Contacts/coordination:

Partnerships

INRAE participants

Partners