DeFine: deep convolutional neural networks accurately quantify intensities of transcription factor-DNA binding and facilitate evaluation of functional non-coding variants

Wang, Meng; Tai, Cheng Chi; Weinan, E; Wei, Liping

doi:10.1093/nar/gky215

Cited by 99 publications

(85 citation statements)

References 56 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…The detection of TADs is not as sensitive to resolution decline as algorithms for detecting TADs, we obtained roughly the same results when using the Hi-C data with various downsampling ratios [23].…”

Section: Deephic Is More Precise In Detecting Tad Boundariessupporting

confidence: 62%

DeepHiC: A Generative Adversarial Network for Enhancing Hi-C Data Resolution

Hong

Jiang

et al. 2019

Preprint

View full text Add to dashboard Cite

Hi-C is commonly used to study three-dimensional genome organization. However, due to the high sequencing cost and technical constraints, the resolution of most Hi-C datasets is coarse, resulting in a loss of information and biological interpretability. Here we develop DeepHiC, a generative adversarial network, to predict high-resolution Hi-C contact maps from low-coverage sequencing data.We demonstrated that DeepHiC is capable of reproducing high-resolution Hi-C data from as few as 1% downsampled reads. Empowered by adversarial training, our method can restore fine-grained details similar to those in high-resolution Hi-C matrices, boosting accuracy in chromatin loops identification and TADs detection, and outperforms the state-of-the-art methods in accuracy of prediction. Finally, application of DeepHiC to Hi-C data on mouse embryonic development can facilitate chromatin loop detection with higher accuracy. We develop a web-based tool (DeepHiC, http://sysomics.com/deephic) that allows researchers to enhance their own Hi-C data with just a few clicks. Author summaryWe developed a novel method, DeepHiC, for enhancing Hi-C data resolution from low-coverage sequencing data using generative adversarial network. DeepHiC is capable of reproducing highresolution (10-kb) Hi-C data with high quality even using 1/100 downsampled reads. Our method outperforms the previous methods in Hi-C data resolution enhancement, boosting accuracy in chromatin loops identification and TADs detection. Application of DeepHiC on mouse embryonic development data shows that enhancements afforded by DeepHiC facilitates the chromatin loops identification of these data achieving higher accuracy. We also developed a user-friendly web server (http://sysomics.com/deephic) that allows researchers to enhance their own low-resolution Hi-C data (40kb-1Mb) with just few clicks. The high-throughput chromosome conformation capture (Hi-C) technique [1] is a genome-wide technique used to investigate three-dimensional (3D) chromatin conformation inside the nucleus. It has facilitated the identification and characterization of multiple structural elements, such as the A/B compartment [1], topological associating domains (TADs) [2, 3], enhancer-promoter loops [4] and stripes [5] over recent decades. In practice, Hi-C data is conventionally stored as a pairwise read count matrix , where is the number of observed interactions (read-pair count) between × genomic regions and , and the genome is partitioned into fixed-size bins (e.g., 25 kb). Bin size (i.e., resolution), is a crucial parameter for Hi-C data analysis, as it directly affects the results of downstream analysis, such as predictions of enhancer-promoter interactions [6-11] or identification of TAD boundaries [6, 12-16]. Depending on sequencing depths, the size of commonly used bins ranges from 1 kb to 1 Mb. Because of the high cost of sequencing, most available Hi-C datasets have relatively low resolution [17], which limits their application in studies of genomic regulatory elements. Sequencing high-res...

show abstract

Section: Deephic Is More Precise In Detecting Tad Boundariessupporting

confidence: 62%

DeepHiC: A Generative Adversarial Network for Enhancing Hi-C Data Resolution

Hong

Jiang

et al. 2019

Preprint

View full text Add to dashboard Cite

show abstract

“…Due to the common misconception that rst convolutional layer lters learn motifs (Alipanahi et al , 2015;Kelley et al , 2016;Quang and Xie, 2016;Angermueller et al , 2016;Cuperus et al , 2017;Chen et al , 2018;Kelley et al , 2018;Bretschneider et al , 2018;Ben-Bassat et al , 2018;Wang et al , 2018;Gao et al , 2018;Trabelsi et al , 2019), deep learning practitioners continue to employ CNN architectures with large rst layer lters with the intent of capturing motif patterns in their entirety. However, we have shown that employing a large lter does not necessarily lead to whole motif representations.…”

Section: Motif Representations Are Not Very Sensitive To 1st Layer Ltmentioning

confidence: 99%

“…A common method to validate a trained CNN is to demonstrate that rst layer lters have learned biologically meaningful representations, i.e. PWM-like representations of sequence motifs (Alipanahi et al , 2015;Kelley et al , 2016;Quang and Xie, 2016;Angermueller et al , 2016;Cuperus et al , 2017;Chen et al , 2018;Kelley et al , 2018;Bretschneider et al , 2018;Ben-Bassat et al , 2018;Wang et al , 2018;Gao et al , 2018;Trabelsi et al , 2019). The few studies that perform a quantitative motif comparison of the rst layer lters against a motif database nd that less than 50% have a statistically signicant match (Kelley et al , 2016;Quang and Xie, 2016).…”

Section: Introductionmentioning

confidence: 99%

Representation Learning of Genomic Sequence Motifs with Convolutional Neural Networks

Koo

Eddy

2018

Preprint

View full text Add to dashboard Cite

Although convolutional neural networks (CNNs) have been applied to a variety of computational genomics problems, there remains a large gap in our understanding of how they work. Here we perform systematic experiments on synthetic sequences to reveal principles of how CNN architecture in uences the internal representations of genomic sequence motifs that are learned. We focus our study on representations learned by rst convolutional layer lters. We nd that deep CNNs tend to learn distributed representations of partial sequence motifs. However, we demonstrate that the architecture of a CNN can be modi ed to predictively learn more interpretable localist representations, i.e. whole motifs. We then validate that the representation learning principles established from synthetic sequences generalize to in vivo sequences.

show abstract

“…More recently, the rapid development of deep learning techniques has enabled mining in high-dimensional sequences data. Some examples include DeepSEA (Zhou & Troyanskaya, 2015), DeepBind (Alipanahi, Delong, Weirauch, & Frey, 2015), DanQ (Quang & Xie, 2016), Define (Wang, Tai, E, & Wei, 2018), and Basenji (Kelley et al, 2018). However, because data sets used for training in those algorithms vary, comparisons across different models can become a problem considering there is currently no goldstandard for evaluation (Nishizaki & Boyle, 2017).…”

Section: Introductionmentioning

confidence: 99%

Predicting functional variants in enhancer and promoter elements using RegulomeDB

Dong

Boyle

2019

Human Mutation

View full text Add to dashboard Cite

Here we present a computational model, Score of Unified Regulatory Features (SURF), that predicts functional variants in enhancer and promoter elements. SURF is trained on data from massively parallel reporter assays and predicts the effect of variants on reporter expression levels. It achieved the top performance in the Fifth Critical Assessment of Genome Interpretation “Regulation Saturation” challenge. We also show that features queried through RegulomeDB, which are direct annotations from functional genomics data, help improve prediction accuracy beyond transfer learning features from DNA sequence‐based deep learning models. Some of the most important features include DNase footprints, especially when coupled with complementary ChIP‐seq data. Furthermore, we found our model achieved good performance in predicting allele‐specific transcription factor binding events. As an extension to the current scoring system in RegulomeDB, we expect our computational model to prioritize variants in regulatory regions, thus help the understanding of functional variants in noncoding regions that lead to disease.

show abstract

DeFine: deep convolutional neural networks accurately quantify intensities of transcription factor-DNA binding and facilitate evaluation of functional non-coding variants

Cited by 99 publications

References 56 publications

DeepHiC: A Generative Adversarial Network for Enhancing Hi-C Data Resolution

DeepHiC: A Generative Adversarial Network for Enhancing Hi-C Data Resolution

Representation Learning of Genomic Sequence Motifs with Convolutional Neural Networks

Predicting functional variants in enhancer and promoter elements using RegulomeDB

Contact Info

Product

Resources

About