Chromatin interaction aware gene regulatory modeling with graph attention networks

Karbalayghareh, Alireza; Şahin, Merve; Cs, Leslie

doi:10.1101/2021.03.31.437978

Cited by 16 publications

(35 citation statements)

References 29 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Previous approaches typically modelled enhancer sequences explicitly via predefined sets of features, which were informed by prior biological knowledge 43 . In contrast, deep learning, in particular convolutional neural networks, do not require prior knowledge and can learn accurate models directly from raw data [44][45][46][47][48][49][50][51][52][53] . Once trained on raw data, these models allow the extraction and interpretation of the learned rules by novel types of tools 44,45,47,48,[54][55][56][57][58][59][60] .…”

mentioning

confidence: 99%

DeepSTARR predicts enhancer activity from DNA sequence and enables thede novodesign of enhancers

Almeida

Reiter

Pagani

et al. 2021

Preprint

View full text Add to dashboard Cite

Enhancer sequences control gene expression and comprise binding sites (motifs) for different transcription factors (TFs). Despite extensive genetic and computational studies, the relationship between DNA sequence and regulatory activity is poorly understood and enhancer de novo design is considered impossible. Here we built a deep learning model, DeepSTARR, to quantitatively predict the activities of thousands of developmental and housekeeping enhancers directly from DNA sequence in Drosophila melanogaster S2 cells. The model learned relevant TF motifs and higher-order syntax rules, including functionally non-equivalent instances of the same TF motif that are determined by motif-flanking sequence and inter-motif distances. We validated these rules experimentally and demonstrated their conservation in human by testing more than 40,000 wildtype and mutant Drosophila and human enhancers. Finally, we designed and functionally validated synthetic enhancers with desired activities de novo.

show abstract

mentioning

confidence: 99%

DeepSTARR predicts enhancer activity from DNA sequence and enables thede novodesign of enhancers

Almeida

Reiter

Pagani

et al. 2021

Preprint

View full text Add to dashboard Cite

show abstract

“…In contrast, sequence-based methods such as 1D ResNet often focus on local sequence patterns and are not good at capturing interactions of residues that are far away from each other along the primary sequence. Studies have shown that explicitly modeling long-range residue interactions can greatly improve performance in various tasks [38] in addition to functional prediction. Using predicted structural information may also improve the prediction accuracy of GO terms with high specificity.…”

Section: Explicit Structural Information Amends Function Interpretation In Longer Sequences and High-specificity Go Termsmentioning

confidence: 99%

“…Graph attention network(GAT) [36] is a type of graph neural network (GNN) that performs graph convolution with self-attention [37]. GAT and GNN are used to model gene expression and study protein structure refinement [38,39]. Unsupervised protein sequence models are used to capture inter-residue relationships for protein contact prediction and have become an integral part of many protein structure prediction methods [23,24,32].…”

Section: Introductionmentioning

confidence: 99%

Accurate Protein Function Prediction via Graph Attention Networks with Predicted Structure Information

2021

Preprint

View full text Add to dashboard Cite

Experimental protein function annotation does not scale with the fast-growing sequence databases. Only a tiny fraction (<0.1%) of protein sequences in UniProtKB has experimentally determined functional annotations. Computational methods may predict protein function in a high-throughput way, but its accuracy is not very satisfactory. Based upon recent breakthroughs in protein structure prediction and protein language models, we develop GAT-GO, a graph attention network (GAT) method that may substantially improve protein function prediction by leveraging predicted inter-residue contact graphs and protein sequence embedding. Our experimental results show that GAT-GO greatly outperforms the latest sequence- and structure-based deep learning methods. On the PDB-mmseqs testset where the train and test proteins share <15% sequence identity, GAT-GO yields Fmax(maximum F-score) 0.508, 0.416, 0.501, and AUPRC(area under the precision-recall curve) 0.427, 0.253, 0.411 for the MFO, BPO, CCO ontology domains, respectively, much better than homology-based method BLAST (Fmax 0.117,0.121,0.207 and AUPRC 0.120, 0.120, 0.163). On the PDB-cdhit testset where the training and test proteins share higher sequence identity, GAT-GO obtains Fmax 0.637, 0.501, 0.542 for the MFO, BPO, CCO ontology domains, respectively, and AUPRC 0.662, 0.384, 0.481, significantly exceeding the just-published graph convolution method DeepFRI, which has Fmax 0.542, 0.425, 0.424 and AUPRC 0.313, 0.159, 0.193.

show abstract

“…In the field of transcription regulation it has for example been shown that predicting experimental transcription factor binding profiles can elucidate the joint binding syntax of the pluripotency TFs (Avsec et al, 2021b) and that predicting gene expression can reveal gene-regulatory features such as long-range enhancer-promoter interactions (Avsec et al, 2021a;Karbalayghareh et al, 2021).…”

Section: Discussionmentioning

confidence: 99%

“…The model that ultimately solves this defining challenge will have to be able to predict gene expression up to measurement errors from sequence alone in any cell type, possibly even across organisms. Recently the first papers have been published that among other features predict gene expression from sequence in a multitask learning objective (Avsec et al, 2021a;Karbalayghareh et al, 2021). Until now, these proposed methods predict gene expression only for specific cell types.…”

Section: Discussionmentioning

confidence: 99%

Statistical methods for biological sequence analysis for DNA binding motifs and protein contacts

Roth¹

View full text Add to dashboard Cite

Over the last decades a revolution in novel measurement techniques has permeated the biological sciences filling the databases with unprecedented amounts of data ranging from genomics, transcriptomics, proteomics and metabolomics to structural and ecological data. In order to extract insights from the vast quantity of data, computational and statistical methods are nowadays crucial tools in the toolbox of every biological researcher. In this thesis I summarize my contributions in two data-rich fields in biological sciences: transcription factor binding to DNA and protein structure prediction from protein sequences with shared evolutionary ancestry.In the first part of my thesis I introduce our work towards a web server for analysing transcription factor binding data with Bayesian Markov Models. In contrast to classical PWM or di-nucleotide models, Bayesian Markov models can capture complex inter-nucleotide dependencies that can arise from shape-readout and alternative binding modes. In addition to giving access to our methods in an easy-to-use, intuitive web-interface, we provide our users with novel tools and visualizations to better evaluate the biological relevance of the inferred binding motifs. We hope that our tools will prove useful for investigating weak and complex transcription factor binding motifs which cannot be predicted accurately with existing tools.The second part discusses a statistical attempt to correct out the phylogenetic bias arising in co-evolution methods applied to the contact prediction problem. Co-evolution methods have revolutionized the protein-structure prediction field more than 10 years ago, and, until very recently, have retained their importance as crucial input features to deep neural networks. As the co-evolution information is extracted from evolutionarily related sequences, we investigated whether the phylogenetic bias to the signal can be corrected out in a principled way using a variation of the Felsenstein's tree-pruning algorithm applied in combination with an independentpair assumption to derive pairwise amino counts that are corrected for the evolutionary history.Unfortunately, the contact prediction derived from our corrected pairwise amino acid counts did not yield a competitive performance. IIIPosnien for accompanying me on my journey as members of by TAC and Prof. Stephan Waack, Dr. Marieke Oudelaar and Prof. Burkhard Morgenstern for volunteering their time as examiners.I will be eternally grateful to those who offered help when I needed it the most, especially

show abstract

Chromatin interaction aware gene regulatory modeling with graph attention networks

Cited by 16 publications

References 29 publications

DeepSTARR predicts enhancer activity from DNA sequence and enables thede novodesign of enhancers

DeepSTARR predicts enhancer activity from DNA sequence and enables thede novodesign of enhancers

Accurate Protein Function Prediction via Graph Attention Networks with Predicted Structure Information

Statistical methods for biological sequence analysis for DNA binding motifs and protein contacts

Contact Info

Product

Resources

About