DeepSelectGene: Deep Learning for Genotype Data and its Application in Genomic Selection

Summary :

Deep learning (DL) methods are beginning to be used as predictive models for phenotypes based on genotype data in the context of human diseases and production traits for genomic selection in domestic animals. These models necessitate training with numerous data sets (genotype > phenotype pairs), which may not always be feasible for certain species genotyped with only a few thousand animals.

The project of this thesis will address this limitation by successively using two type of DL methods: a first generative DL method, for example, "generative adversarial neural networks" (GANs), will allow the simulation of genotype data from training on a small but qualitatively representative real data set. Thus, we will artificially increase the size of the database necessary for the proper training of a second predictive DL model for phenotype prediction from a genotype (50-800K SNP). This second DL model, adopting a simpler structure for predictions, will need to be optimized based on our initial exploratory studies on the subject during the GenIALearn project on bovine genomic selection (MP DIGIT-BIO 2022-2024).

In summary, this thesis work will propose a novel method for simulating genotype data that is beneficial for i) enhancing understanding of the genetic determinism involved in phenotype formation; ii) generating additional quasi-real data essential for training the DL prediction model. Therefore, this DL model for phenotype prediction can be applied with just a few thousand genotype-phenotype data sets and will subsequently be improved as the database is progressively enriched. This topic represents a highly interdisciplinary field in genomic modeling, encompassing genetics, genomics, statistics, and data science, and is at the forefront of AI applications in genomics.

Contact :

• Sihan Xie
• Linkedin