ImaGene: a convolutional neural network to quantify natural selection from genomic data

Abstract

BackgroundThe genetic bases of many complex phenotypes are still largely unknown, mostly due to the polygenic nature of the traits and the small effect of each associated mutation. An alternative approach to classic association studies to determining such genetic bases is an evolutionary framework. As sites targeted by natural selection are likely to harbor important functionalities for the carrier, the identification of selection signatures in the genome has the potential to unveil the genetic mechanisms underpinning human phenotypes. Popular methods of detecting such signals rely on compressing genomic information into summary statistics, resulting in the loss of information. Furthermore, few methods are able to quantify the strength of selection. Here we explored the use of deep learning in evolutionary biology and implemented a program, called ImaGene, to apply convolutional neural networks on population genomic data for the detection and quantification of natural selection.ResultsImaGene enables genomic information from multiple individuals to be represented as abstract images. Each image is created by stacking aligned genomic data and encoding distinct alleles into separate colors. To detect and quantify signatures of positive selection, ImaGene implements a convolutional neural network which is trained using simulations. We show how the method implemented in ImaGene can be affected by data manipulation and learning strategies. In particular, we show how sorting images by row and column leads to accurate predictions. We also demonstrate how the misspecification of the correct demographic model for producing training data can influence the quantification of positive selection. We finally illustrate an approach to estimate the selection coefficient, a continuous variable, using multiclass classification techniques.ConclusionsWhile the use of deep learning in evolutionary genomics is in its infancy, here we demonstrated its potential to detect informative patterns from large-scale genomic data. We implemented methods to process genomic data for deep learning in a user-friendly program called ImaGene. The joint inference of the evolutionary history of mutations and their functional impact will facilitate mapping studies and provide novel insights into the molecular mechanisms associated with human phenotypes.Electronic supplementary materialThe online version of this article (10.1186/s12859-019-2927-x) contains supplementary material, which is available to authorized users.

Deep learning for population size history inference: design, comparison and combination with approximate Bayesian computation

“…However, the discrepancy between ABC and SPIDNA reconstructions in the last 10,000 years might also be due to the sensitivity of ANNs to overfitting and to mispecifications in the model generating training data. For example, decrease in performances due to demographic mispecification has already been shown for selection inference based on ANNs (Torada et al, 2019). In our case, model mispecification arises because cattle breeds are subjected to strong artificial selection pressures, with few males contributing to the next generations, which is a clear violation of the coalescent assumptions underlying our training simulated set.…”

“…The fourth baseline is a newly design CNN with rectangular shaped filters that cover more than one haplotype (see Methods). This “custom CNN” is a first step towards an architecture tailored to raw genomic data, because spatial information is preserved as for recent ANNs applied to population genetics (Chan et al, 2018, Flagel et al, 2018, Torada et al, 2019), but also because asymmetrical filters of various sizes account for the heterogeneous entities of axes (haplotype versus SNP, rather than pixel versus pixel). Finally we adapted and re-trained four networks among the top-ranked CNNs proposed by Flagel et al (2018) so that they could reconstruct a 21-epoch model of instantaneous effective population size rather than the three-epoch model initially investigated by the authors, and for practicability we called them Flagel CNNs.…”

“…Although the custom CNN network above cannot be guaranteed to be exactly invariant to the haplotype order, it can approximately learn this specificity. To avoid wasting training time to learn that there is no information in the row order, it has been proposed to systematically sort the haplotypes according to a predefined rule (Flagel et al, 2018, Torada et al, 2019). Because there is no ordering in high dimensional space that is stable with respect to perturbations (Qi et al, 2017), we chose yet another alternative and enforced our network to be permutation-invariant by design.…”

“…A considerable shift occurred when several studies applied ANN directly on genomic data instead of using summary statistic. Various CNN architectures processing SNP matrices were proposed to infer recombination rates along the genome (Chan et al, 2018, Flagel et al, 2018), selection (Flagel et al, 2018, Torada et al, 2019), introgression (Flagel et al, 2018) and three-step population size histories (Flagel et al, 2018). The CNN implemented by Chan et al (2018) and based on Deep Sets (Zaheer et al, 2017) is invariant to haplotype permutation (i.e to the permutation of lines in the SNP matrix) thanks to convolution filters that treat each haplotype in an identical way.…”

“…The CNN implemented by Chan et al (2018) and based on Deep Sets (Zaheer et al, 2017) is invariant to haplotype permutation (i.e to the permutation of lines in the SNP matrix) thanks to convolution filters that treat each haplotype in an identical way. The other approaches proposed instead to sort haplotypes by similarity before processing them with filters sensitive to the haplotype order (Flagel et al, 2018, Torada et al, 2019). More recently, Recurrent Neural Networks (RNN) were applied to estimate the recombination rate (Adrion et al, 2019), and Generative Adversarial Networks (GAN) to learn the distribution of genomic datasets and to generate artificial genomes (Yelmen et al, 2019).…”