Deep learning based methods for estimating distribution of coalescence rates from genome-wide data

Demographic and population structure inference is one of the most important problems in genomics. Population parameters such as effective population sizes, population split times and migration rates are of high interest both themselves and for many applications, e.g. for genome-wide association studies. Hidden Markov Model (HMM) based methods, such as PSMC, MSMC, coalHMM etc., proved to be powerful and useful for estimation of these parameters in many population genetics studies. At the same time, machine and deep learning have began to be used in natural science widely. In particular, deep learning based approaches have already substituted hidden Markov models in many areas, such as speech recognition or user input prediction. We develop a deep learning (DL) approach for local coalescent time estimation from one whole diploid genome. Our DL models are trained on simulated datasets. Importantly, demographic and population parameters can be inferred based on the distribution of coalescent times. We expect that our approach will be useful under complex population scenarios, which cannot be studied with existing HMM based methods. Our work is also a crucial step in developing a deep learning framework which would allow to create population genomics methods for different genomic data representations.