Robust Deep Learning for Artificial Genomics and Population Genetics – RoDAPoG

Advances in sequencing technologies and active worldwide sampling strategies drastically increased the amount of genomic data. Efficiently extracting information from massive genomic data remains an open problem for which deep learning methodologies are promising candidates provided that it is clear to developers and users when, why and with what limitations these methods should be applied. In addition many human genomic sequences are not publicly available for privacy reasons. In a recent pioneer work we use generative networks to create artificial genomes and demonstrate that these could compensate for the inaccessibility of private data. To unleash the full potential of artificial genomes we will develop new methods that are robust in terms of scalability and privacy. Secondly, we will explore the interpretability and robustness of deep neural networks emerging in population genetics, a key step towards reinforcing their reliability and that can benefit other biological fields.