Abstract 393: Predicting DNA accessibility in the pan-cancer tumor genome using RNA-Seq, WGS, and deep learning

Abstract

DNA accessibility, chromatin regulation, and genome methylation are key drivers of cancer transcription. However, there is much left to be understood about the functional implications of sequence-level data to the regulation of gene expression, especially when it comes to the noncoding genome. Recently [Kelley, D., Snoek, J., and Rinn, J., Genome Res. 2016] trained neural networks to effectively predict DNA accessibility in multiple cell types. These models make it possible to explore the impact of mutations on the predicted accessibility and thus directly link one aspect of the gene regulation puzzle all the way down to the sequence level. We present a model with improved performance on the original dataset of 164 ENCODE and Roadmap Epigenomics Consortium sample types, and then extend the method to provide predictions on any sample with RNA-Seq data without need of DNase-seq for the sample. We first demonstrate that with several model and algorithmic changes we improve performance across 164 cell types from a mean AUC of 0.895 to a mean AUC of 0.910. Unfortunately current accessibility models require DNase-seq for each new cell type. Models for detecting transcription factor binding sites, which rely on ChIP-seq for training data, also share this issue. In order to generalize sequence-based predictive models to apply to unseen cell types without requiring re-training we investigate using RNA-Seq as a proxy signature of cell type. The model aims to capture the interdependence of gene expression levels that characterize a cell with the regulatory logic in which sequence-level signatures are combined to determine accessibility without restriction to cell type. We explore the model’s performance when applied to held-out cell types in the ENCODE and Roadmap Epigenomics Consortium data as well as data from the TCGA Pan-Cancer initiative. We look for the impact of non-coding changes in whole-genome sequencing data in TCGA samples, and report on predicted differences in DNA accessibility across cancer subtypes. Citation Format: Kamil Wnuk, Jeremi Sudol, Shahrooz Rabizadeh, Patrick Soon-Shiong, Christopher Szeto, Charles Vaske. Predicting DNA accessibility in the pan-cancer tumor genome using RNA-Seq, WGS, and deep learning [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2017; 2017 Apr 1-5; Washington, DC. Philadelphia (PA): AACR; Cancer Res 2017;77(13 Suppl):Abstract nr 393. doi:10.1158/1538-7445.AM2017-393

Integrating regulatory DNA sequence and gene expression to predict genome-wide chromatin accessibility across cellular contexts

“…Overall, our best performing ResNet(+mean accessibility) model achieves an AUPRC of 0.76 while the the previous best published model in the literature i.e. the Factorized Basset model (Wnuk et al, 2017) achieves 0.69 (Fig. 1C) on a matched training/validation/test data split.…”

“…To provide the model with quantitative information on the availability of trans-regulator TFs, we follow recent work (Wnuk et al, 2017) that extended the Basset model to predict chromatin accessibility in heldout cellular contexts, using RNA-seq profiles as surrogates of cell-type specific availability and activity of trans-regulators. RNA-seq profiles have been shown to uniquely identify individual cell types while preserving biological similarity between cell types (Sudmant et al, 2015).…”

“…Residual neural networks (ResNets) (He et al, 2016) have been show to be more effective for training CNNs with a large number of layers by introducing skip connections between blocks of convolution layers to optimize gradient flow and improve learning. Utilizing these concepts, we use a ResNet architecture to extend previous models (Kelley et al, 2016;Wnuk et al, 2017). The residual network (He et al, 2016) consists of blocks in which the input is transformed through one or more convolutional layers to an intermediate output to which the input is added back.…”

“…DNA sequence coupled with RNA expression profiles. Hence, we expand upon previous work to predict genome-wide maps of chromatin accessibility using sequence and gene expression data (Kelley et al, 2016;Wnuk et al, 2017). We introduce multimodal, residual neural network architectures (He et al, 2016) that integrate cis-regulatory sequence and context-specific expression of transregulators to predict genome-wide chromatin accessibility profiles across cellular contexts.…”

“…The Basset model was recently enhanced by factorizing the convolution layers (Wnuk et al, 2017) (Factorized model). The Factorized model increases the model depth -the 3 convolution layers in Basset are replaced by 9 convolution layers.…”