Can Deep Learning Improve Genomic Prediction of Complex Human Traits?

Abstract

The genetic analysis of complex traits does not escape the current excitement around artificial intelligence, including a renewed interest in "deep learning" (DL) techniques such as Multilayer Perceptrons (MLPs) and Convolutional Neural Networks (CNNs). However, the performance of DL for genomic prediction of complex human traits has not been comprehensively tested. To provide an evaluation of MLPs and CNNs, we used data from distantly related white Caucasian individuals (n 100k individuals, m 500k SNPs, and k = 1000) of the interim release of the UK Biobank. We analyzed a total of five phenotypes: height, bone heel mineral density, body mass index, systolic blood pressure, and waist-hip ratio, with genomic heritabilities ranging from 0.20 to 0.70. After hyperparameter optimization using a genetic algorithm, we considered several configurations, from shallow to deep learners, and compared the predictive performance of MLPs and CNNs with that of Bayesian linear regressions across sets of SNPs (from 10k to 50k) that were preselected using single-marker regression analyses. For height, a highly heritable phenotype, all methods performed similarly, although CNNs were slightly but consistently worse. For the rest of the phenotypes, the performance of some CNNs was comparable or slightly better than linear methods. Performance of MLPs was highly dependent on SNP set and phenotype. In all, over the range of traits evaluated in this study, CNN performance was competitive to linear models, but we did not find any case where DL outperformed the linear model by a sizable margin. We suggest that more research is needed to adapt CNN methodology, originally motivated by image analysis, to genetic-based problems in order for CNNs to be competitive with linear models.

Learning polygenic scores for human blood cell traits

“…SNPs in the context of this study; the last layer outputs the final result of the model and the layer(s) in between are called hidden layer(s). A function node in hidden and output layers typically transforms the inputs from the previous layer with a weighted linear sum followed by an activation function 23 . For example, f 1 in Figure 2 can be represented as:…”

“…As this work is, to our knowledge, the first attempt to employ MLPs and CNNs for genomic prediction of blood cell traits, there was no prior information that could be used for the design of network architecture for this task. Therefore, similar to the previous work 23 , we used a genetic algorithm to search for the optimal MLP and CNN architectures as well as other hyperparameters, e.g. the number of layers, the number of neurons at each layer, activation functions, optimizers, dropouts, etc., on the train set, in which 10% of the samples were used as a validation set.…”

“…Machine learning techniques have demonstrated superiority in constructing polygenic scores for several traits and diseases, largely immune-related, e.g. celiac disease, type 1 diabetes, and Crohn's disease [19][20][21][22][23] . Owing to their simplicity and efficiency, multivariate linear models, such as elastic net and support vector machines, are one of the most widely used categories of machine learning methods for PGS construction.…”

Using the structure of genome data in the design of deep neural networks for predicting amyotrophic lateral sclerosis from genotype

“…Here we observe some marked improvements of Promoter-CNN + ALS-Net over logistic regression. An explanation might be that Bellot et al (2018) use a (substantially) simpler (nondeep) network architecture and make a pre-selection of genetic features based on linear models, and hence overlook non-additive interactions already in the pre-selection step.…”

“…A few studies have considered using deep learning for genotype-phenotype association studies. Most approaches first reduced the number of variants included in the model either by selecting variants that were known to be associated with disease (Uppu and Krishna, 2017;Hess et al, 2017), or by preselecting those variants that showed a sufficiently strong correlation with phenotype in a regular GWAS (Montañez et al, 2018b;Bellot et al, 2018). Two studies combine the latter strategy with the use of autoencoders for further dimensionality reduction (Montañez et al, 2018a;Fergus et al, 2018).…”