Inferring past history from genetic data using ABC and Deep Learning approaches

Current methods for demographic inference rely either on the complete observed genetic dataset in the few cases where the likelihood is tractable, or on summary statistics. These handcrafted summary statistics are usually picked carefully and inform about many aspects of demographic history. Nonetheless, the reduction of whole sequence or SNP data into summary statistics is drastic and might not always be optimal. Hence, we propose to build deep neural networks that extract automatically the required information. I will briefly present our ABC approach, designed for whole-genome sequence data and applied to the inference of successive bottleneck and expansions. I will then introduce basic deep learning concepts through a few broadly known applications and present our deep neural architecture that takes as input the genetic data of several sampled individuals, predicts a vector of demographic parameters, and is trained on simulated data. We evaluated the performances on a specific model of bottleneck and expansions, and showed that it compares reasonably well with several ABC algorithms. I will discuss the required methodological improvements for our specific case, as well as some pitfalls and advantages for the field.