New phenotypes from milk MIR spectra: challenges to obtain reliable predictions

In the recent years, the research aiming to predict new phenotypes from the FT-MIR analysis of milk was very active. Models were developed to predict phenotypes such as fine milk composition, cow health and environmental impact or technological properties of milk. Those models could be of great interest in order to perform genetic studies as they could allow generating large amount of data at large scale and with reasonable cost. To achieve this, it is nonetheless necessary to insure that the models provide reliable predictions when applied on the large diversity of spectral data met on real field conditions. The robustness of models -its capacity to be ‘all terrain’ and provide good results in various conditions- is therefore essential to ensure reliability of predictions. Robustness could be estimated by evaluating the error in external validation (RMSEP), the reproducibility of predictions between instruments and the ability of the calibration dataset to cover the variability of routine field data. However, in current literature, the model robustness is often omitted. Models are frequently developed on reduced dataset, with limited number of herds, breeds and diets. Additionally, models are evaluated by looking to the statistical performances, through the R2 and the standard error (RMSE or SEC), while the robustness is rarely assessed. Finally, only a limited number of models is used in routine and faces the large variability of real field conditions to provide phenotypes for management of cows or genetic studies. The objective of this work is consequently to evaluate the impact of different factors influencing robustness on prediction quality. The impact of sampling scheme (oriented vs random), and model development are investigated. Effect of inclusion of variability in the model by adding countries, breeds, MIR instruments and days in milk are also investigated. The obtained results encourage for international collaborations in order to constitute large and robust datasets and enable the use of models in routine conditions.