Benchmarking predictive models: evaluating parametric, ensemble, and deep learning approaches for animal phenotype prediction from genotypes. - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2024

Benchmarking predictive models: evaluating parametric, ensemble, and deep learning approaches for animal phenotype prediction from genotypes.

Jocelyn de Goër de Herve
Anne Ricard
Thierry Tribout
Joon Kwon

Résumé

Over the past two decades, advancements in DNA genotyping have revolutionized the acquisition of extensive genomic data, essential for enhancing genomic selection in animal breeding across multiple traits. Concurrently, the rapid progress of artificial intelligence has piqued interest in its application to genomic selection (1,2). Despite its potential, the utilization of ensemble methods and deep learning models for predicting livestock animal phenotypes based on genotypes lags behind parametric GBLUP mixed models, which are regarded as the current state of the art. The primary objective of this study was to conduct a benchmark analysis, comparing various parametric, ensemble, and neural network methods for predicting animal phenotypes using extensive genotyping data. The dataset employed for training and validating different models comprised 100,000 Holstein cows characterized for 33 quantitative traits (phenotypes: milk production, fertility and morphology category). Genotyping data (50K SNP/animal) were utilized to predict the 33 traits, employing parametric mixed model GBLUP, as well as gradient boosting (GB) and various deep learning models, including MultiLayer Perceptrons with various architectures (MLP), Convolutional Neural Networks (CNN), CNN Transformer, Value Imputation and Mask Estimation (VIME), and MLP Variational AutoEncoder (VAE). To assess predictive model performance, we employed the same data for training, validation, and evaluation sets, measuring accuracy through root mean square error (RMSE) and Pearson correlations (R) between actual and predicted traits. According to the correlation coefficient (R), GBLUP (R=0.38) outperformed all other models in predicting most of the traits. However, on average, MLP performed slightly less effectively (R=0.38) than GBLUP, with no significant difference (p>0.05). The remaining models (GB, CNN, CNNT, VIME+MLP, VAE+MLP) demonstrated lower efficiency. Conversely, based on RMSE, certain traits (2 to 9 traits of milk production and fertility) were more accurately predicted by GB or MLP models, never by more complex models (CNN, CNNT, VIME+MLP, VAE+MLP). Further refinement of deep learning models specifically tailored to genomic data is essential for improved prediction of traits determined by polygenic gene effects. References: 1-Bellot, P., de los Campos, G., & Pérez-Enciso, M. (2018). Can deep learning improve genomic prediction of complex human traits? Genetics, 210(3), 809–819. https://doi.org/10.1534/genetics.118.301298 2-Abdollahi-Arpanahi, R., Gianola, D., & Peñagaricano, F. (2020). Deep learning versus parametric and ensemble methods for genomic prediction of complex phenotypes. Genetics Selection Evolution, 52(1), 12. https://doi.org/10.1186/s12711-020-00531-z Key words: IA, Deep learining , SNP , genomics, animal, milk production
Fichier principal
Vignette du fichier
EMBL-IA & BIOLOGY -Heidelberg-2024.pdf (1.69 Mo) Télécharger le fichier
Origine : Fichiers éditeurs autorisés sur une archive ouverte

Dates et versions

hal-04510253 , version 1 (05-04-2024)

Identifiants

  • HAL Id : hal-04510253 , version 1

Citer

Eric Barrey, Blaise Hanczar, Julien Chiquet, Didier Boichard, Jocelyn de Goër de Herve, et al.. Benchmarking predictive models: evaluating parametric, ensemble, and deep learning approaches for animal phenotype prediction from genotypes.. AI and biology Symposium, EMBO EMBL, Heidelberg, Mar 2024, HEIDELBERG, Germany. ⟨hal-04510253⟩
0 Consultations
0 Téléchargements

Partager

Gmail Facebook X LinkedIn More