Super-enhancers are characterized by high levels of Mediator binding and are major contributors to the expression of their associated genes. They exhibit high levels of local chromatin interactions and a higher order of local chromatin organization. On the other hand, lncRNAs can localize to specific DNA sites by forming a RNA:DNA:DNA triplex, which in turn can contribute to local chromatin organization. In this paper, we characterize a new class of lncRNAs called super-lncRNAs that target super-enhancers and which can contribute to the local chromatin organization of the super-enhancers. Using a logistic regression model based on the number of RNA:DNA:DNA triplex sites a lncRNA forms within the super-enhancer, we identify 442 unique super-lncRNA transcripts in 27 different human cell and tissue types; 70% of these super-lncRNAs were tissue restricted. They primarily harbor a single triplex-forming repeat domain, which forms an RNA:DNA:DNA triplex with multiple anchor DNA sites (originating from transposable elements) within the super-enhancers. Super-lncRNAs can be grouped into 17 different clusters based on the tissue or cell lines they target. Super-lncRNAs in a particular cluster share common short structural motifs and their corresponding super-enhancer targets are associated with gene ontology terms pertaining to the tissue or cell line. Super-lncRNAs may use these structural motifs to recruit and transport necessary regulators (such as transcription factors and Mediator complexes) to super-enhancers, influence chromatin organization, and act as spatial amplifiers for key tissue-specific genes associated with super-enhancers.

The Smyd1 gene encodes a lysine methyltransferase specifically expressed in striated muscle. Because Smyd1-null mouse embryos die from heart malformation prior to formation of skeletal muscle, we developed a Smyd1 conditional-knockout allele to determine the consequence of SMYD1 loss in mammalian skeletal muscle. Ablation of SMYD1 specifically in skeletal myocytes after myofiber differentiation using Myf6cre produced a non-degenerative myopathy. Mutant mice exhibited weakness, myofiber hypotrophy, prevalence of oxidative myofibers, reduction in triad numbers, regional myofibrillar disorganization/breakdown and a high percentage of myofibers with centralized nuclei. Notably, we found broad upregulation of muscle development genes in the absence of regenerating or degenerating myofibers. These data suggest that the afflicted fibers are in a continual state of repair in an attempt to restore damaged myofibrils. Disease severity was greater for males than females. Despite equivalent expression in all fiber types, loss of SMYD1 primarily affected fast-twitch muscle, illustrating fiber-type-specific functions for SMYD1. This work illustrates a crucial role for SMYD1 in skeletal muscle physiology and myofibril integrity.

Long noncoding RNAs (lncRNAs) can exert their function by interacting with the DNA via triplex structure formation. Even though this has been validated with a handful of experiments, a genome-wide analysis of lncRNA-DNA binding is needed. In this paper, we develop and interpret deep learning models that predict the genome-wide binding sites deciphered by ChIRP-Seq experiments of 12 different lncRNAs. Among the several deep learning architectures tested, a simple architecture consisting of two convolutional neural network layers performed the best suggesting local sequence patterns as determinants of the interaction. Further interpretation of the kernels in the model revealed that these local sequence patterns form triplex structures with the corresponding lncRNAs. We uncovered several novel triplexes forming domains (TFDs) of these 12 lncRNAs and previously experimentally verified TFDs of lncRNAs HOTAIR and MEG3. We experimentally verified such two novel TFDs of lncRNAs HOTAIR and TUG1 predicted by our method (but previously unreported) using Electrophoretic mobility shift assays. In conclusion, we show that simple deep learning architecture can accurately predict genome-wide binding sites of lncRNAs and interpretation of the models suggest RNA:DNA:DNA triplex formation as a viable mechanism underlying lncRNA-DNA interactions at genome-wide level.

Background: Chromosomes are organized into units called topologically associated domains (TADs). TADs dictate regulatory landscapes and other DNA-dependent processes. Even though various factors that contribute to the specification of TADs have been proposed, the mechanism is not fully understood. Understanding the process for specification and maintenance of these units is essential in dissecting cellular processes and disease mechanisms. Results: In this study, we report a genome-wide study that considers the idea of long noncoding RNAs (lncRNAs) mediating chromatin organization using lncRNA:DNA triplex-forming sites (TFSs). By analyzing the TFSs of expressed lncRNAs in multiple cell lines, we find that they are enriched in TADs, their boundaries, and loop anchors. However, they are evenly distributed across different regions of a TAD showing no preference for any specific portions within TADs. No relationship is observed between the locations of these TFSs and CTCF binding sites. However, TFSs are located not just in promoter regions but also in intronic, intergenic, and 3’UTR regions. We also show these triplex-forming sites can be used as predictors in machine learning models to discriminate TADs from other genomic regions. Finally, we compile a list of important “TAD-lncRNAs” which are top predictors for TADs identification. Conclusions: Our observations advocate the idea that lncRNA:DNA TFSs are positioned at specific areas of the genome organization and are important predictors for TADs. LncRNA:DNA triplex formation most likely is a general mechanism of action exhibited by some lncRNAs, not just for direct gene regulation but also to mediate 3D chromatin organization.

Genomes are organized into self-interacting chromatin regions called topologically associated domains (TADs). A significant number of TAD boundaries are shared across multiple cell types and conserved across species. Disruption of TAD boundaries may affect the expression of nearby genes and could lead to several diseases. Even though detection of TAD boundaries is important and useful, there are experimental challenges in obtaining high resolution TAD locations. Here, we present computational prediction of TAD boundaries from high resolution Hi-C data in fruit flies. By extensive exploration and testing of several deep learning model architectures with hyperparameter optimization, we show that a unique deep learning model consisting of three convolution layers followed by a long short-term-memory layer achieves an accuracy of 96%. This outperforms feature-based models’ accuracy of 91% and an existing method's accuracy of 73–78% based on motif TRAP scores. Our method also detects previously reported motifs such as Beaf-32 that are enriched in TAD boundaries in fruit flies and also several unreported motifs.

Comprehensive understanding of intratumor heterogeneity requires identification of molecular markers, which are capable of differentiating different subpopulations and which also have clinical significance. One important tool that has been addressing this issue is single cell RNA-Sequencing (scRNASeq) that allows the quantification of expression profiles of transcripts in individual cells in a population of cancer cells. Using the expression profiles from scRNASeq, current studies conduct analysis to group cells into different subpopulations using clustering algorithms. In this study, we explore scRNASeq cancer data from a different perspective. We focus on scRNASeq data originating from cancer cells pertaining to a particular cancer type, where the cell type or the subpopulation to which each cell belongs is known. We investigate if the “cell type” of a cancer cell can be predicted based on the expression profiles of a small set of transcripts. We outline a predictive analytics pipeline to accurately predict 6 breast cancer cell types using single cell gene expression profiles. Instead of building predictive models using the complete human transcripts, the pipeline first eliminates predictors with low expression and low variance. A multinomial penalized logistic regression further reduces the size of the predictors to only 308, out of which 34 are long non-coding RNAs. Tuning of predictive models shows support vector machines and neural networks as the most accurate models achieving close to 98% prediction accuracies. We also find that mixture of protein coding genes and long non-coding RNAs are better predictors compared to when the two sets of transcripts are treated separately. A signature risk score originating from 65 protein coding genes and 5 lncRNA predictors is associated with prognostic survival of TCGA breast cancer patients. This association was maintained when the risk scores were generated using 65 PCGs and 5 lncRNA separately. We further show that predictors restricted to a particular cell type serve as better prognostic markers for the respective patient subtype. Our results show that in general, the breast cancer cell type predictors are also associated with patient survivability and hence have clinical significance.