Section: Application Domains

Population Genetics

Work in this application domain started recently, with two main lines of research : dimension reduction of genetic datasets and prediction tasks using genetic data (such as the prediction of past human demography).

• Flora Jay collaborated with Kevin Caye and colleagues (TIMC-IMAG, Grenoble) who developed an R package for inferring coefficients of genetic ancestry, using matrix factorization, alternating quadratic programming and projected least squares algorithms [4]. The extension of ancestry inference and visualization methods to temporal data (for paleogenetics applications) remains to be done.

• The demographic history of one or several population (of any organism) can be partially reconstructed using modern or ancient genetic data. A common approach in the population genetics field is to simulate pseudo-datasets for which the demographic parameters are known and summarize them into handcrafted features. These features are then used as a reference panel in an Approximate Bayesian Computation (likelihood-free) framework. Flora Jay has been developping such methods for the application to whole-genome data [14][60].

• An untackled challenge in the field is to skip the summary step and directly handle raw data of genetic variations. Théophile Sanchez, who did a 6 month internship in TAU, started his PhD in October 2017 and is currently designing deep learning architectures that are suitable for multi-genome data [33]. In particular these networks should be invariant to the permutation of indiviual genomes and flexible to the input size (see Section 7.2.7).