Sequential regulatory activity prediction across chromosomes with convolutional neural networks.
Kelley DR, Reshef YA, Bileschi M, Belanger D, McLean CY, Snoek J. · 2018
Read Sequential regulatory activity prediction across chromosomes with convolutional neural networks. by Kelley DR, Reshef YA, Bileschi M, Belanger D, McLean CY, Snoek J. (2018) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc5932613.txt.
Introduction
Kelley and co-authors predict regulatory activity from DNA sequence with neural networks. Their article appeared in Genome Research. Their system reads sequence alone to forecast cell profiles. It predicts epigenetic marks and gene expression by cell type. Convolutional networks spot promoters and distant control elements. The system combines those signals into expression forecasts.
Predictions for variant effects matched known causal variants, they report. Those variants came from human expression studies. The model can suggest how a disease-linked variant acts. That helps narrow which variant in a locus is causal. Noncoding variants drive most trait links found by genome scans. A sequence-only model gives those links a testable story.
Catalog record
| CREATOR | Kelley DR, Reshef YA, Bileschi M, Belanger D, McLean CY, Snoek J. |
|---|---|
| YEAR | 2018 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 61 KB |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Sequential regulatory activity prediction across chromosomes with convolutional neural networks By David R Kelley, Yakir A Reshef, Maxwell Bileschi, David Belanger, Cory Y McLean, Jasper Snoek Published in: Genome Research DOI: 10.1101/gr.227819.117 ABSTRACT Abstract Models for predicting phenotypic outcomes from genotypes have important applications to understanding genomic function and improving human health. Here, we develop a machine-learning system to predict cell-type–specific epigenetic and transcriptional profiles in large mammalian genomes from DNA sequence alone. By use of convolutional neural networks, this system identifies promoters and distal regulatory elements and synthesizes their content to make effective gene expression predictions. We show that model predictions for the influence of genomic variants on gene expression align well to causal variants underlying eQTLs in human populations and can be useful for generating mechanistic hypotheses to enable fine mapping of disease loci. Although many studies show strong relationships between variation in genotype and phenotype across a range of human diseases and traits, the mechanisms through which this relationship operates remain incompletely understood (Boyle et al. 2017). Noncoding variation has especially stifled progress; most genomic loci statistically associated with phenotypes via genome-wide association studies (GWAS) do not alter coding sequence, but mechanisms for only a rare few have been thoroughly described (e.g., Claussnitzer et al. 2015). Numerous lines of evidence suggest that many noncoding variants influence traits by changing gene expression (Maurano et al. 2012; Albert and Kruglyak 2015; Finucane et al. 2015; O'Connor et al. 2017). In turn, gene expression determines the diversity of cell types and states in multicellular organisms (Roadmap Epigenomics Consortium et al. 2015). Thus, gene expression offers a tractable intermediate phenotype for which improved modeling would have great value. Large-scale consortia and many individual laboratories have mapped the epigenetic and expression profiles of a wide variety of cells (The ENCODE Project Consortium 2012; Forrest et al. 2014; Roadmap Epigenomics Consortium et al. 2015). Further, it has recently become appreciated that many of these data can be accurately modeled as a function of underlying DNA sequence using machine learning. Successful predictive modeling of transcription factor (TF) binding, accessible chromatin, and histone modifications has provided mechanistic insight and useful interpretation of genomic variants (Ghandi et al. 2014; Alipanahi et al. 2015; Whitaker et al. 2015; Zhou and Troyanskaya 2015; Kelley et al. 2016). In particular, the expansive training data available from the 3 billion–nucleotide (nt) human genome has enabled deep learning methods with huge numbers of parameters to make significantly more accurate predictions on held-out test data than previous approaches (Zhou and Troyanskaya 2015; Kelley et al. 2016). Despite this progress, models to predict cell-type–specific gene expression from DNA sequence have remained elusive in complex organisms. Existing models use experimental annotations as input (e.g., peak calls for various known regulatory attributes), allowing them to shed light on the relationships between these annotations (Cheng et al. 2012; González et al. 2015) but not analyzing the causal role of the underlying sequence in determining those annotations. Even with intra-experiment training data to infer the relevant sequence motifs, the complexity of distal regulation where enhancers can interact with promoters across hundreds of thousands of nucleotides challenges the current generation of methods (Levine 2010; Long et al. 2016). However, well-established gene regulation principles from inquiry into enhancer biology and 3D chromosomal contact domains have yet to be fully incorporated into expressive machine-learning models (Mifsud et al. 2015; Dekker and Mirny 2016). Modeling larger sequences and diverse experimental data offers a path forward to improve predictive accuracy. More effective models would enable researchers to profile one instance of a tissue or cell type and project that profile to individuals with varying genomic sequence. Here, we use novel machine-learning algorithms to learn to predict thousands of epigenetic and transcriptional profiles from hundreds of human cell types using only DNA sequence as input. By using the model, we predict the difference between the two alleles of genomic variants for these thousands of data sets, focusing particularly on predicted changes to gene expression. We demonstrate the considerable potential value of this observation to identify likely causal variants and mechanisms within GWAS loci. RESULTS Basenji In previous work, we introduced a deep convolutional neural network approach named Basset for modeling “peak”-based chromatin profiles, focusing particularly on DNase I hypersensitivity (Kelley et al. 2016). Given an input sequence of 500–1000 bp, the model makes a single binary prediction for the sequence's activity in any training data set provided. Here, we modify the Basset architecture to (1) model distal regulatory interactions and (2) predict finer-resolution, quantitative (as opposed to binary) genomic profiles that are more appropriate for the dynamic range of gene expression (Fig. 1). As a related approach, but one begging for metaphor to a more nimble and far-sighted hound, we refer to this method as Basenji. [Figure: Sequential regulatory activity prediction. DNA sequences come in to the model one hot encoded to four rows representing A, C, G, and T. The annotations are fabrications to help convey the reasons for the various elements of the architecture. We apply several layers of convolution and max pooling, similar to previous methods (Kelley et al. 2016), to obtain representations that describe 128-bp bins. To share information across large distances, we apply several layers of dilated convolutions. The purple squares indicate the columns that the convolution directly sees; the teal shade is drawn proportional to the number of operations performed on that column with respect to the center position. Dilated convolution layers are densely passed on to the final prediction layer, where a width-one convolution layer makes predictions across the sequence. We compare these predictions to the experimental counts via a Poisson regression loss function and use stochastic gradient descent with back propagation to fit the model parameters.] The model accepts much larger (217=) 131-kb regions as input and, similar to Basset, performs multiple layers of convolution and pooling to transform the DNA to a sequence of vectors representing 128-bp regions (Methods). To share information across long distances, we then apply several layers of densely connected dilated convolutions (Methods). After these layers, each 128-bp region aligns to a vector that considers the relevant regulatory elements across a large span of sequence. Finally, we apply a final width-one convolutional layer to parameterize a multitask Poisson regression on normalized counts of aligned reads to that region for every data set provided (Hashimoto et al. 2016). That is, the model's ultimate goal is to predict read coverage in 128-bp bins across long chromosome sequences. Modeling count data, as opposed to peak data, required careful preprocessing beyond that performed in the standard pipelines of genomics consortium projects. For example, processed consortium data discard multimapping read alignments, which leaves half the genome incompletely annotated, despite the substantial evidence that repetitive sequence is critical to gene regulation (Feschotte 2008). Thus, we downloaded the original sequencing reads for 593 ENCODE DNase-seq, 1704 ENCODE histone modification ChIP-seq, 356 Roadmap DNase-seq, 603 Roadmap histone modification ChIP-seq, and 973 FANTOM5 CAGE experiments. These experiments represent 529 unique cells/tissues profiled by DNase-seq, 1136 unique cells/tissues profiled by ChIP-seq, and 595 unique cells/tissues profiled by CAGE. We processed these data with a pipeline that includes additional computation to make use of multimapping reads and to normalize for GC bias