A general framework for predicting the transcriptomic consequences of non-coding variation

Abstract

17Genome wide association studies (GWASs) for complex traits have implicated thousands of genetic 18 loci. Most GWAS-nominated variants lie in noncoding regions, complicating the systematic translation 19 of these findings into functional understanding. Here, we leverage convolutional neural networks to 20 assist in this challenge. Our computational framework, peaBrain, models the transcriptional machinery 21 of a tissue as a two-stage process: first, predicting the mean tissue specific abundance of all genes and 22Most reported disease-associated variation for complex traits lies in non-coding regions of the genome 1 . 34Despite advances in discovery and annotations of functional non-coding elements across the genome 2-35 5 , characterising the consequences of non-coding variants remains a major challenge in human genetics. 36Prediction of the transcriptomic consequences of non-coding variation represents one solution [6][7][8][9][10] . 37Current methods of variant-expression prediction can be broadly divided into two classes: (a) methods 38 that predict alterations in epigenetic and transcription factor binding sites (TFBS), such as DeepSEA 8 39 and Basset 10 ; and (b) methods that directly predict RNA abundance from genotype or sequence data, 40 such as PrediXcan 6 and TWAS 9 . Methods in the former category do not capture differences in transcript 41 expression as a result of genotypic variation 8,10 and are relatively poor predictors of alterations in the 42 histone code 8 ; methods in the latter category are not able to identify which of the variants detected 43 within an eQTL association locus are functional 6,9 . 44 45To address these concerns, here, we introduce a single framework, called promoter-and-enhancer-46 derived abundance (peaBrain) model, which consolidates both of these approaches. Within the 47 class-C models, we added additional channels corresponding, for those tissues where such data were 87 available, to the consolidated epigenomes from the Epigenomics Roadmap, including tissue-specific 88 peaks from H3K4me1, H3K4me3, H3K9ac, H3K9me3, H3K27me3, and H3K36me3 ChIP-seq 89 experiments, and experimentally-derived DNase hotspots 17 . 90 91We observed that DNA-only (class-A) models captured nearly a fifth of the variance in mean gene 92 abundance across all GTEx tissues (10-fold cross-validated median out-of-sample-r 2 [oos-r 2 ] values 93 across all tissues = 17%). Addition of non-specific regulatory annotations (class-B models) markedly 94 6 improved model performance across all tissues (median cross-validated oos-r 2 = 45%; Figure 1). (We 95 average the oos-r 2 across all 10-folds within a tissue and use the median across all tissues to assess 96 global performance; see Online Methods.) For example, for EBV-transformed lymphocytes, the 10-97 fold cross-validated average oos-r 2 is 56% for the class-B model compared to the 15% in the 98 corresponding class-A model. Addition of tissue-specific annotations further improved model 99 performance, such that class-C models captured more than half t...

Predicting mRNA Abundance Directly from Genomic Sequence Using Deep Convolutional Neural Networks

“…Next, we compared our best baseline and Xpresso models with the reported results from existing models described in the literature, delineating five categories based upon the types of features used either as input data or as intermediate training stages: (1) those using nothing more than sequence features, which included our method and three others (Abdalla et al, 2018;Bessiè re et al, 2018;McLeay et al, 2012); (2) those using MPRAs to measure promoter activity (van Arensbergen et al, 2017;Cooper et al, 2006;Landolin et al, 2010;Nguyen et al, 2016); (3) those using the binding signal of TFs at promoter regions, as measured by ChIP (Cheng et al, 2011(Cheng et al, , 2012McLeay et al, 2012;Ouyang et al, 2009;Zhou et al, 2018); (4) those using the signal of histone marks, such as H3K4me1, H3K4me3, H3K9me3, H3K27Ac, H3K27me3, and H3K36me3 at promoters and gene bodies, as measured by ChIP (Abdalla et al, 2018;Cheng et al, 2011;Dong et al, 2012;Karli c et al, 2010;McLeay et al, 2012;Schmidt et al, 2017;Zhou et al, 2018); and (5) those using the DNase hypersensitivity signal at promoters and nearby enhancers (Dong et al, 2012;Duren et al, 2017;McLeay et al, 2012;Schmidt et al, 2017;Zhou et al, 2018) (Figure 4E). Many of these models were trained and tested on cell lines, such as K562, GM12878, and mESCs, for which ChIP data are available for a multitude of histone marks and TFs.…”

Predicting mRNA abundance directly from genomic sequence using deep convolutional neural networks

“…Next, we compared our best baseline and Xpresso models to existing models described in the literature, categorizing the types of features used as input in each model into five categories: i) those using nothing more than sequence features, which included our method and two others 14,42 , ii) those using MPRAs to measure promoter activity 23,[43][44][45] , iii) those utilizing the binding signal of transcription factors (TFs) at promoter regions, as measured by ChIP 11,13-15 , iv) those utilizing the signal of histone marks such as H3K4me1, H3K4me3, H3K9me3, H3K27Ac, H3K27me3, and H3K36me3 at promoters and gene bodies, as measured by ChIP 11,12,14,16,17,42 , and v) those utilizing DNase hypersensitivity signal at promoters and nearby enhancers 12,14,16,46 ( Figure 4E ).…”