Accurate prediction of cell type-specific transcription factor binding

Keilwagen, Jens; Posch, Stefan; Grau, Jan

doi:10.1186/s13059-018-1614-y

Cited by 97 publications

(127 citation statements)

References 58 publications

Supporting

Mentioning

122

Contrasting

Order By: Relevance

“…Since auPRC is better at capturing the difference of prediction performances with imbalanced labels than auROC and recalls at different FDR levels, we choose auPRC as our primary benchmarking metric. Based on auPRC, our attention model has better performance on 69.23% (9/13) of the prediction targets than Anchor (Li, et al, 2019), 69.23% (9/13) than FactorNet (Quang and Xie, 2019), 76.92% (10/13) than Catchitt (Keilwagen, et al, 2019), and 92.31% (12/13) than Cheburashka (Lando, et al, 2016). Among all the methods, our method achieved the highest auPRC on 6 targets: CTCF/induced pluripotent stem cell (iPSC), FOXA1/liver, FOXA2/liver, GABPA/liver, HNF4A/liver, and REST/liver.…”

Section: Overall Benchmarking On Evaluation Datamentioning

confidence: 99%

“…There are 51676736 bins in total on training chromosomes in the labels, resulting in 51676736×n potential training samples for each transcription factor, where n is the number of available cell types for training. Due to limited computing capacity, we use the iterative training process s with downsampling the negatives (Keilwagen, et al, 2019;Quang and Xie, 2019). In each epoch, we first sample Nneg negative bins from all negative labels.…”

Section: Deep Neural Network Models With Attention Mechanismmentioning

confidence: 99%

“…Unsupervised methods such as hidden Markov models (Mathelier and Wasserman, 2013;Mehta, et al, 2011) and hierarchical mixture models (Pique-Regi, et al, 2011) usually classify input genomic regions into different clusters by their affinity to certain transcription factors. Supervised learning approaches require ground truth during model training and aim to precisely predict the bound/unbound status of transcription factors in specific regions, including support vector machines (Djordjevic, et al, 2003;Zhou, et al, 2015), discriminative maximum conditional likelihood (Keilwagen, et al, 2019) and random forest (Hooghe, et al, 2012;Xiao and Segal, 2009). Both types of methods usually rely on prior knowledge about sequence preference data, such as position weight matrix (Sherwood, et al, 2014), which need additional steps to generate the required features.…”

Section: Introductionmentioning

confidence: 99%

“…Both types of methods usually rely on prior knowledge about sequence preference data, such as position weight matrix (Sherwood, et al, 2014), which need additional steps to generate the required features. In addition, these features may be less reliable when no prior knowledge exists and inference based methods (such as de-novo motif discovery) are used to generate the input features (Keilwagen, et al, 2019).…”

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

Interpretable attention model in transcription factor binding site prediction with deep neural networks

Chen

Hou

Shi

et al. 2019

Preprint

View full text Add to dashboard Cite

Due to the complexity of the biological factors that may influence the binding of transcription factors to DNA sequences, prediction of the potential binding sites remains a difficult task in computational biology. The attention mechanism in deep learning has shown its capability to learn from input features with long-range dependencies. Until now, no study has applied this mechanism in deep neural network models with input data from massively parallel sequencing. In this study, we aim to build a model for binding site prediction with the combination of attention mechanism and traditional deep learning techniques, including convolutional neural networks and recurrent neural networks. The performance of our methods is evaluated on the ENCODE-DREAM in vivo Transcription Factor Binding Site Prediction Challenge datasets. The benchmark shows that our implementation with attention mechanism (called DeepGRN) improves the performance of the deep learning models. Our model achieves better performance in at least 9 of 13 targets than any of the methods participated in the DREAM challenge. Visualization of the attention weights extracted from the trained models reveals how those weights shift when binding signal peaks move along the genomic sequence, which can interpret how the predictions are made. Case studies show that the attention mechanism helps to extract useful features by focusing on regions that are critical to successful prediction while ignoring the irrelevant signals from the input.

show abstract

Section: Overall Benchmarking On Evaluation Datamentioning

confidence: 99%

Section: Deep Neural Network Models With Attention Mechanismmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Interpretable attention model in transcription factor binding site prediction with deep neural networks

Chen

Hou

Shi

et al. 2019

Preprint

View full text Add to dashboard Cite

show abstract

“…When TF ChIP-seq data is not available, TF binding motifs, used in combination with chromatin accessibility data or H3K27ac ChIP-seq data might be used to infer TF binding sites 7,20,21 . Machine learning approaches that transfer models learnt from TF ChIP-seq peaks, motifs and DNase-seq data between cell types are promising ways of imputing TF cistromes, although imputation of TF binding sites on a large scale remains to be implemented 22-27 . Computationally imputed TF binding data is expected to represent TF binding sites less accurately than TF ChIP-seq experimental data, so we sought to develop a TR prediction method that could use imputed TF cistromes effectively, along with ChIP-seq derived ones.…”

Section: Introductionmentioning

confidence: 99%

Inferring transcriptional regulators through integrative modeling of public chromatin accessibility and ChIP-seq data

Qin

Fan

Zheng

et al. 2019

Preprint

View full text Add to dashboard Cite

21We developed Lisa (http://lisa.cistrome.org) to predict the transcriptional regulators (TRs) of 22 differentially expressed or co-expressed gene sets. Based on the input gene sets, Lisa first uses 23 compendia of public histone mark ChIP-seq and chromatin accessibility profiles to construct a 24 chromatin model related to the regulation of these genes. Then using TR ChIP-seq peaks or 25 imputed TR binding sites, Lisa probes the chromatin models using in silico deletion to find the 26 most relevant TRs. Applied to gene sets derived from targeted TF perturbation experiments, Lisa 27 boosted the performance of imputed TR cistromes, and outperformed alternative methods in 28 identifying the perturbed TRs. 29 30 Keywords 31 Transcription factors, gene regulation, chromatin accessibility, DNase-seq, H3K27ac ChIP-seq, 32 differential gene expression, gene set analysis 33 34 List of abbreviations 35 TF: transcription factor 36 CR: chromatin regulator 37 TR: transcriptional regulator 38 RP: regulatory potential 39 ISD: in silico deletion 40 ROC: receiver operator characteristic 41 AUC: area under curve 42 ChIP-seq: chromatin immunoprecipitation followed by DNA sequencing 43 DNase-seq: DNase I digestion followed by DNA sequencing 44 H3K27ac: histone H3 lysine 27 acetylation 45 AR: Androgen Receptor 46 ER: Estrogen Receptor 47 GR: Glucocorticoid Receptor 48 49 Introduction 50Transcriptional regulators (TRs), which include transcription factors (TFs) and chromatin 51 regulators (CRs), play essential roles in controlling normal biological processes and are frequently 52 implicated in disease [1][2][3][4] . The genomic landscape of TF binding sites and histone modifications 53 collectively shape the transcriptional regulatory environments of genes 5-8 . ChIP-seq has been 54 widely used to map the genome-wide set of cis-elements bound by trans-acting factors such as 55 TFs and CRs, which we henceforth refer to as "cistromes" 9 . There are approximately 1,500 56 transcription factors in human and mouse 10,11 , regulating a wide variety of biological processes in 57 constitutive or cell-type-specific manners, and tens of thousands of ChIP-seq and DNase-seq 58 experiments have been performed in human and mouse. We previously developed the Cistrome 59 Data Browser (DB) 12 , a collection of uniformly processed TF ChIP-seq (~11,000) and chromatin 60 profiles (~12,000 histone mark ChIP-seq and DNase-seq) in human and mouse. 62The question we address in this paper is how to effectively use these data to infer the TRs that 63 regulate a query gene set derived from differential or correlated gene expression analyses in 64 human or mouse. TR ChIP-seq data, when available, is the most accurate available data type 65 representing TR binding. ChIP-seq data availability, in terms of covered TRs and cell types, even 66 with large contributions from projects such as ENCODE 13 , is still sparse due to the limited 67 availability of specific antibodies. Although advances have been made in TR cistrome mapping 68 with the introduction of technolo...

show abstract

A preliminary computational outputs versus experimental results: Application of sTRAP, a biophysical tool for the analysis of SNPs of transcription factor‐binding sites

Moradifard

Saghiri

Ehsani

et al. 2020

Molec Gen & Gen Med

View full text Add to dashboard Cite

BackgroundIn the human genome, the transcription factors (TFs) and transcription factor‐binding sites (TFBSs) network has a great regulatory function in the biological pathways. Such crosstalk might be affected by the single‐nucleotide polymorphisms (SNPs), which could create or disrupt a TFBS, leading to either a disease or a phenotypic defect. Many computational resources have been introduced to predict the TFs binding variations due to SNPs inside TFBSs, sTRAP being one of them.MethodsA literature review was performed and the experimental data for 18 TFBSs located in 12 genes was provided. The sequences of TFBS motifs were extracted using two different strategies; in the size similar with synthetic target sites used in the experimental techniques, and with 60 bp upstream and downstream of the SNPs. The sTRAP (http://trap.molgen.mpg.de/cgi-bin/trap_two_seq_form.cgi) was applied to compute the binding affinity scores of their cognate TFs in the context of reference and mutant sequences of TFBSs. The alternative bioinformatics model used in this study was regulatory analysis of variation in enhancers (RAVEN; http://www.cisreg.ca/cgi-bin/RAVEN/a). The bioinformatics outputs of our study were compared with experimental data, electrophoretic mobility shift assay (EMSA).ResultsIn 6 out of 18 TFBSs in the following genes COL1A1, Hb ḉᴪ, TF, FIX, MBL2, NOS2A, the outputs of sTRAP were inconsistent with the results of EMSA. Furthermore, no p value of the difference between the two scores of binding affinity under the wild and mutant conditions of TFBSs was presented. Nor, were any criteria for preference or selection of any of the measurements of different matrices used for the same analysis.ConclusionOur preliminary study indicated some paradoxical results between sTRAP and experimental data. However, to link the data of sTRAP to the biological functions, its optimization via experimental procedures with the integration of expanded data and applying several other bioinformatics tools might be required.

show abstract

Accurate prediction of cell type-specific transcription factor binding

Cited by 97 publications

References 58 publications

Interpretable attention model in transcription factor binding site prediction with deep neural networks

Interpretable attention model in transcription factor binding site prediction with deep neural networks

Inferring transcriptional regulators through integrative modeling of public chromatin accessibility and ChIP-seq data

A preliminary computational outputs versus experimental results: Application of sTRAP, a biophysical tool for the analysis of SNPs of transcription factor‐binding sites

Contact Info

Product

Resources

About