Base-resolution prediction of transcription factor binding signals by a deep learning framework

Cross-Species Prediction of Transcription Factor Binding by Adversarial Training of a Novel Nucleotide-Level Deep Neural Network

Cross-species prediction of TF binding remains a major challenge due to the rapid evolutionary turnover of individual TF binding sites, resulting in cross-species predictive performance being consistently worse than within-species performance. In this study, a novel Nucleotide-Level Deep Neural Network (NLDNN) is first proposed to predict TF binding within or across species. NLDNN regards the task of TF binding prediction as a nucleotide-level regression task, which takes DNA sequences as input and directly predicts experimental coverage values. Beyond predictive performance, it also assesses model performance by locating potential TF binding regions, discriminating TF-specific single-nucleotide polymorphisms (SNPs), and identifying causal disease-associated SNPs. The experimental results show that NLDNN outperforms the competing methods in these tasks. Then, a dual-path framework is designed for adversarial training of NLDNN to further improve the cross-species prediction performance by pulling the domain space of human and mouse species closer. Through comparison and analysis, it finds that adversarial training not only can improve the cross-species prediction performance between humans and mice but also enhance the ability to locate TF binding regions and discriminate TF-specific SNPs. By visualizing the predictions, it is figured out that the framework corrects some mispredictions by amplifying the coverage values of incorrectly predicted peaks.

MulTFBS: A Spatial-Temporal Network with Multichannels for Predicting Transcription Factor Binding Sites

Revealing the mechanisms that influence transcription factor binding specificity is the key to understanding gene regulation. In previous studies, DNA double helix structure and one-hot embedding have been used successfully to design computational methods for predicting transcription factor binding sites (TFBSs). However, DNA sequence as a kind of biological language, the method of word embedding representation in natural language processing, has not been considered properly in TFBS prediction models. In our work, we integrate different types of features of DNA sequence to design a multichanneled deep learning framework, namely MulTFBS, in which independent one-hot encoding, word embedding encoding, which can incorporate contextual information and extract the global features of the sequences, and double helix three-dimensional structural features have been trained in different channels. To extract sequence high-level information effectively, in our deep learning framework, we select the spatial-temporal network by combining convolutional neural networks and bidirectional long short-term memory networks with attention mechanism. Compared with six state-of-the-art methods on 66 universal protein-binding microarray data sets of different transcription factors, MulTFBS performs best on all data sets in the regression tasks, with the average R2 of 0.698 and the average PCC of 0.833, which are 5.4% and 3.2% higher, respectively, than the suboptimal method CRPTS. In addition, we evaluate the classification performance of MulTFBS for distinguishing bound or unbound regions on TF ChIP-seq data. The results show that our framework also performs well in the TFBS classification tasks.

Comparative analysis of models in predicting the effects of SNPs on TF-DNA binding using large-scale in vitro and in vivo data

Non-coding variants associated with complex traits can alter the motifs of transcription factor (TF)-deoxyribonucleic acid binding. Although many computational models have been developed to predict the effects of non-coding variants on TF binding, their predictive power lacks systematic evaluation. Here we have evaluated 14 different models built on position weight matrices (PWMs), support vector machines, ordinary least squares and deep neural networks (DNNs), using large-scale in vitro (i.e. SNP-SELEX) and in vivo (i.e. allele-specific binding, ASB) TF binding data. Our results show that the accuracy of each model in predicting SNP effects in vitro significantly exceeds that achieved in vivo. For in vitro variant impact prediction, kmer/gkm-based machine learning methods (deltaSVM_HT-SELEX, QBiC-Pred) trained on in vitro datasets exhibit the best performance. For in vivo ASB variant prediction, DNN-based multitask models (DeepSEA, Sei, Enformer) trained on the ChIP-seq dataset exhibit relatively superior performance. Among the PWM-based methods, tRap demonstrates better performance in both in vitro and in vivo evaluations. In addition, we find that TF classes such as basic leucine zipper factors could be predicted more accurately, whereas those such as C2H2 zinc finger factors are predicted less accurately, aligning with the evolutionary conservation of these TF classes. We also underscore the significance of non-sequence factors such as cis-regulatory element type, TF expression, interactions and post-translational modifications in influencing the in vivo predictive performance of TFs. Our research provides valuable insights into selecting prioritization methods for non-coding variants and further optimizing such models.

Interpretable single-cell transcription factor prediction based on deep learning with attention mechanism

Predicting the transcription factor binding site (TFBS) in the whole genome range is essential in exploring the rule of gene transcription control. Although many deep learning methods to predict TFBS have been proposed, predicting TFBS using single-cell ATAC-seq data and embedding attention mechanisms needs to be improved. To this end, we present IscPAM, an interpretable method based on deep learning with an attention mechanism to predict single-cell transcription factors. Our model adopts the convolution neural network to extract the data feature and optimize the pre-trained model. In particular, the model obtains faster training and prediction due to the embedded attention mechanism. For datasets, we take ATAC-seq, ChIP-seq, and DNA sequences data for the pre-trained model, and single-cell ATAC-seq data is used to predict the TF binding graph in the given cell. We verify the interpretability of the model through ablation experiments and sensitivity analysis. IscPAM can efficiently predict the combination of whole genome transcription factors in single cells and study cellular heterogeneity through chromatin accessibility of related diseases.

HEAP: a task adaptive-based explainable deep learning framework for enhancer activity prediction

Enhancers are crucial cis-regulatory elements that control gene expression in a cell-type-specific manner. Despite extensive genetic and computational studies, accurately predicting enhancer activity in different cell types remains a challenge, and the grammar of enhancers is still poorly understood. Here, we present HEAP (high-resolution enhancer activity prediction), an explainable deep learning framework for predicting enhancers and exploring enhancer grammar. The framework includes three modules that use grammar-based reasoning for enhancer prediction. The algorithm can incorporate DNA sequences and epigenetic modifications to obtain better accuracy. We use a novel two-step multi-task learning method, task adaptive parameter sharing (TAPS), to efficiently predict enhancers in different cell types. We first train a shared model with all cell-type datasets. Then we adapt to specific tasks by adding several task-specific subset layers. Experiments demonstrate that HEAP outperforms published methods and showcases the effectiveness of the TAPS, especially for those with limited training samples. Notably, the explainable framework HEAP utilizes post-hoc interpretation to provide insights into the prediction mechanisms from three perspectives: data, model architecture and algorithm, leading to a better understanding of model decisions and enhancer grammar. To the best of our knowledge, HEAP will be a valuable tool for insight into the complex mechanisms of enhancer activity.

GNet: An integrated context-aware neural framework for transcription factor binding signal at single nucleotide resolution prediction

Transcription factors (TFs) are important factors that regulate gene expression. Revealing the mechanism affecting the binding specificity of TFs is the key to understanding gene regulation. Most of the previous studies focus on TF-DNA binding sites at the sequence level, and they seldom utilize the contextual features of DNA sequences. In this paper, we develop an integrated spatiotemporal context-aware neural network framework, named GNet, for predicting TF-DNA binding signal at single nucleotide resolution by achieving three tasks: single nucleotide resolution signal prediction, identification of binding regions at the sequence level, and TF-DNA binding motif prediction. GNet extracts implicit spatial contextual information with a gated highway neural mechanism, which captures large context multi-level patterns using linear shortcut connections, and the idea of it permeates the encoder and decoder parts of GNet. The improved dual external attention mechanism, which learns implicit relationships both within and among samples, and improves the performance of the model. Experimental results on 53 human TF ChIP-seq datasets and 6 chromatin accessibility ATAC-seq datasets shows that GNet outperforms the state-of-the-art methods in the three tasks, and the results of cross-species studies on 15 human and 18 mouse TF datasets of the corresponding TF families indicate that GNet also shows the best performance in cross-species prediction over the competitive methods.

LangMoDHS: A deep learning language model for predicting DNase I hypersensitive sites in mouse genome

DNase I hypersensitive sites (DHSs) are a specific genomic region, which is critical to detect or understand cis-regulatory elements. Although there are many methods developed to detect DHSs, there is a big gap in practice. We presented a deep learning-based language model for predicting DHSs, named LangMoDHS. The LangMoDHS mainly comprised the convolutional neural network (CNN), the bi-directional long short-term memory (Bi-LSTM) and the feed-forward attention. The CNN and the Bi-LSTM were stacked in a parallel manner, which was helpful to accumulate multiple-view representations from primary DNA sequences. We conducted 5-fold cross-validations and independent tests over 14 tissues and 4 developmental stages. The empirical experiments showed that the LangMoDHS is competitive with or slightly better than the iDHS-Deep, which is the latest method for predicting DHSs. The empirical experiments also implied substantial contribution of the CNN, Bi-LSTM, and attention to DHSs prediction. We implemented the LangMoDHS as a user-friendly web server which is accessible at http:/www.biolscience.cn/LangMoDHS/. We used indices related to information entropy to explore the sequence motif of DHSs. The analysis provided a certain insight into the DHSs.

Computational prediction and characterization of cell-type-specific and shared binding sites

MOTIVATION

Cell-type-specific gene expression is maintained in large part by transcription factors (TFs) selectively binding to distinct sets of sites in different cell types. Recent research works have provided evidence that such cell-type-specific binding is determined by TF's intrinsic sequence preferences, cooperative interactions with co-factors, cell-type-specific chromatin landscapes and 3D chromatin interactions. However, computational prediction and characterization of cell-type-specific and shared binding sites is rarely studied.

RESULTS

In this article, we propose two computational approaches for predicting and characterizing cell-type-specific and shared binding sites by integrating multiple types of features, in which one is based on XGBoost and another is based on convolutional neural network (CNN). To validate the performance of our proposed approaches, ChIP-seq datasets of 10 binding factors were collected from the GM12878 (lymphoblastoid) and K562 (erythroleukemic) human hematopoietic cell lines, each of which was further categorized into cell-type-specific (GM12878- and K562-specific) and shared binding sites. Then, multiple types of features for these binding sites were integrated to train the XGBoost- and CNN-based models. Experimental results show that our proposed approaches significantly outperform other competing methods on three classification tasks. Moreover, we identified independent feature contributions for cell-type-specific and shared sites through SHAP values and explored the ability of the CNN-based model to predict cell-type-specific and shared binding sites by excluding or including DNase signals. Furthermore, we investigated the generalization ability of our proposed approaches to different binding factors in the same cellular environment.

AVAILABILITY AND IMPLEMENTATION

The source code is available at: https://github.com/turningpoint1988/CSSBS.

SUPPLEMENTARY INFORMATION

Supplementary data are available at Bioinformatics online.

DLoopCaller: A deep learning approach for predicting genome-wide chromatin loops by integrating accessible chromatin landscapes

In recent years, major advances have been made in various chromosome conformation capture technologies to further satisfy the needs of researchers for high-quality, high-resolution contact interactions. Discriminating the loops from genome-wide contact interactions is crucial for dissecting three-dimensional(3D) genome structure and function. Here, we present a deep learning method to predict genome-wide chromatin loops, called DLoopCaller, by combining accessible chromatin landscapes and raw Hi-C contact maps. Some available orthogonal data ChIA-PET/HiChIP and Capture Hi-C were used to generate positive samples with a wider contact matrix which provides the possibility to find more potential genome-wide chromatin loops. The experimental results demonstrate that DLoopCaller effectively improves the accuracy of predicting genome-wide chromatin loops compared to the state-of-the-art method Peakachu. Moreover, compared to two of most popular loop callers, such as HiCCUPS and Fit-Hi-C, DLoopCaller identifies some unique interactions. We conclude that a combination of chromatin landscapes on the one-dimensional genome contributes to understanding the 3D genome organization, and the identified chromatin loops reveal cell-type specificity and transcription factor motif co-enrichment across different cell lines and species.

Prediction of Back-splicing sites for CircRNA formation based on convolutional neural networks

BACKGROUND

Circular RNAs (CircRNAs) play critical roles in gene expression regulation and disease development. Understanding the regulation mechanism of CircRNAs formation can help reveal the role of CircRNAs in various biological processes mentioned above. Back-splicing is important for CircRNAs formation. Back-splicing sites prediction helps uncover the mysteries of CircRNAs formation. Several methods were proposed for back-splicing sites prediction or circRNA-realted prediction tasks. Model performance was constrained by poor feature learning and using ability.

RESULTS

In this study, CircCNN was proposed to predict pre-mRNA back-splicing sites. Convolution neural network and batch normalization are the main parts of CircCNN. Experimental results on three datasets show that CircCNN outperforms other baseline models. Moreover, PPM (Position Probability Matrix) features extract by CircCNN were converted as motifs. Further analysis reveals that some of motifs found by CircCNN match known motifs involved in gene expression regulation, the distribution of motif and special short sequence is important for pre-mRNA back-splicing.

CONCLUSIONS

In general, the findings in this study provide a new direction for exploring CircRNA-related gene expression regulatory mechanism and identifying potential targets for complex malignant diseases. The datasets and source code of this study are freely available at: https://github.com/szhh521/CircCNN .