The FANTOM5/CAT project is dedicated to exploring the regulatory landscape of human and mouse genomes by generating comprehensive transcriptomic data (Hon et al., 2017). Researchers may use this vast collection of transcriptome data from many biological contexts as a valuable tool to study the control of gene expression, transcriptional networks, and functional components of the mammalian genome. GENCODE is dedicated to maintaining protein-coding and non-coding genes (Frankish et al., 2019). Its main objective is to provide a thorough, current, and accurate annotation of these genomes by locating and describing the protein-coding genes, long non-coding RNAs (lncRNAs), and other functional non-coding RNAs. The annotation provides crucial details for the discovered genes and transcripts, such as gene loci, exon boundaries, coding sequences, untranslated regions (UTRs), and functional and structural annotations. There are now 62,703 genes reported by GENCODE, 19,393 of which are protein-coding and 19,928 long non-coding RNAs. In order to create these annotations, the research combines experimental and computational techniques, including high-throughput sequencing technologies, transcriptome data, protein-coding potential analysis, comparative genomics, and manual curation.

The ENCODE is a comprehensive repository of experimental data and metadata generated by the ENCODE project (Luo et al., 2020). The portal utilizes a standardized metadata architecture that facilitates understanding data in biological terms, enabling the representation of experiments and their analyses. The ENCODE database provides researchers access to a wide range of data, including genomic sequences, epigenetic marks, and transcription factor binding sites. LncExpDB is an expression database focused on human long non-coding RNA (lncRNA) genes (Li et al., 2021). Its main objective is to provide a wide-ranging collection of expression profiles for lncRNA genes, enabling the exploration of their expression characteristics, capacities, and potential functional significance. The database also aims to establish connections between lncRNAs and protein-coding genes across different biological contexts and conditions. LncBook, an extensive repository of lncRNAs) (Li et al., 2023), which is widely used in various studies. The updated version, LncBook 2.0, offers significant improvements and enhancements. This resource empowers users to unravel the functional significance of lncRNAs within different biological contexts.

The Genotype-Tissue Expression (GTEx) project is a valuable resource for studying human gene expression, regulation, and its association with genetic variation (Hon et al., 2017). This initiative involves collecting and analyzing gene expression data from tissues obtained from deceased human donors and their corresponding genotypic information. GTEx is an invaluable resource for researchers investigating the intricate relationship between genetic variation, gene expression, and the specific biology of different tissues. Furthermore, it establishes a valuable foundation for investigating the regulatory mechanisms governing gene expression. TANRIC is a database that maintains cancer-associated lncRNAs mainly extracted from TCGA. The Gene Expression Omnibus (GEO) is a globally accessible public repository that houses functional genomic data sets obtained through high-throughput microarray and next-generation sequencing technologies. These resources are indexed, crosslinked, and made searchable to facilitate easy access and exploration (Barrett et al., 2013).

LNCipedia is a database for collecting and annotating human long non-coding RNA (lncRNA) sequences (Volders et al., 2019). One of its main features is the consolidation of redundant transcripts from various sources, resulting in a consistent and reliable database with grouped transcripts. LNCipedia 5 has expanded its content by incorporating new lncRNAs from resources like FANTOM CAT. It also includes minor improvements, such as an enhanced filtering pipeline and support for official HGNC gene names. CANTATADb 2.0 is a database of LncRNAs that maintain lncRNA from plants and algae species (Szcześniak et al., 2019). Reads from hundreds of RNA-SEQ libraries were aligned with corresponding plant genomes. Gene prediction software was used to re-annotate plant genomes, and annotation data was used as a reference. NONCODE is a comprehensive database of non-coding RNAs, with a particular emphasis on lncRNAs (Zhao et al., 2021). The number of lncRNAs in NONCODEV6 has reached 644,510 including 173,112 human lncRNA transcript.

RNAcentral is a state-of-the-art database that consolidates the non-coding sequences of 44 RNA resources, including more than 18 million ncRNA sequences from various organisms (RNAcentral Consortium, 2021). The new version of RNAcentral also contains the secondary structure of 13 million sequences, making it the world’s most extensive 2D structure database. CHESS is an extensive collection of human genes derived from approximately 10,000 RNA sequencing experiments (Pertea et al., 2018). The database comprises a comprehensive set of genes, encompassing 19,838 protein-coding genes and 17,624 lncRNA) genes. RNA Atlas is an experimentally derived database that advances our understanding of human non-coding RNAs (ncRNAs). By analyzing a diverse set of 300 human samples, including 45 tissues, 162 cell types, and 93 cell lines, researchers have expanded the existing catalog of ncRNAs, identifying a total of 44,428 long non-coding RNAs (lncRNAs). This comprehensive dataset, accessible through the R2 webtool, provides a foundation for further exploration of RNA biology and function (Lorenzi et al., 2021).

LncATLAS is a subcellular localization database of lncRNA, where information has been obtained from RNA-sequencing data of human cell-lines (Mas-Ponte et al., 2017). The RNA-seq datasets were obtained from the ENCODE database. In order to quantify subcellular localization, a measure called relative concentration index (RCI) was introduced. LncSLdb is a specialized database designed to collate and manage qualitative and quantitative information on the subcellular localization of lncRNAs (Wen et al., 2018). The latest release of LncSLdb encompasses data on more than 11,000 lncRNA transcripts derived from three species (human, mouse, and fruit fly). RNALocate v2.0 is a comprehensive repository of information on RNA subcellular localization. It is compiled from scientific literature, public databases, and RNA sequencing datasets (Cui et al., 2022). It contains around 200 thousand entries, and each entry provides detailed information about RNA, including their subcellular localization. Additionally, it provides three prediction tools to cater to different user requirements.

The LncRNADisease2 database is dedicated to maintaining ncRNA-associated diseases, including cancer, cardiovascular diseases, and neurological disorders (Bao et al., 2019). It contains a collection of experimentally validated lncRNAs linked to specific diseases, detailed annotations on their molecular functions, subcellular localizations, interacting proteins, and experimental evidence. The LncRNADisease database documents 205,959 lncRNA-disease associations and 2,297 lncRNA causative associations. The Lnc2Cancer v3.0 database is a comprehensive resource focusing on experimentally validated associations between lncRNAs and cancers (Gao et al., 2021). The updated version of Lnc2Cancer includes new features, such as an increased number of cancer-associated lncRNA entries. The database also offers details on microRNAs, transcription factors, genetic variations, methylation, enhancers, and other regulatory mechanisms.

Fluorescence In Situ Hybridization (FISH) is a molecular cytogenetic technique commonly used to identify the location of biological molecules like DNA and entire chromosomes in a cell. One of the significant features of FISH is that it allows the detection of many biological molecules simultaneously. Single-molecule FISH (smFISH) is a modified version of FISH for single-cell resolution. It allows for visualizing the localization and abundance of individual RNA molecules within cells or tissues (Femino et al., 1998; Raj and Tyagi, 2010; Lubeck and Cai, 2012). In smFISH, RNA molecules are hybridized with fluorescently labeled probes to generate a fluorescent signal that can be detected using fluorescence microscopy. Single Molecule Inexpensive FISH (smiFISH) is a user-friendly and versatile approach for visualizing and quantifying RNA. It uses fluorescently tagged secondary detector oligonucleotides and unlabeled primary probes (Tsanov et al., 2016). The primary probes, which are gene-specific, are unlabeled, making them cost-effective to synthesize. This cost advantage allows for using more probes per mRNA, significantly improving detection efficiency. Single-nucleotide variant FISH (SNV-FISH) is a highly sophisticated method that detects single-nucleotide variations in an RNA transcript (Levesque et al., 2013; Symmons et al., 2019). One more variety of FISH is inosineFISH or inoFISH, which capture image of adenosine-to-inosine RNA altering occasions with single-particle resolution (Mellis et al., 2017).

SABER, an acronym for signal amplification by exchange reaction, is a technique that enhances the signals obtained from oligonucleotide-based FISH probes by attaching long, single-stranded DNA chains (Kishi et al., 2019). It can amplify RNA/DNA signals in cells and tissues. Click-amplifying FISH is a precise method that achieves remarkable signal amplification with a high gain (Rouhanifard et al., 2018). It allows the detection of RNA species using low-magnification microscopy and RNA-based flow cytometry. Sequential FISH (seqFISH) is an advanced technology that identifies thousands of molecules within single cells while preserving their spatial context (Eng et al., 2019). SeqFISH allows researchers to obtain super-resolved images that provide insights into genomic-level processes with exceptionally high efficiency and accuracy. Multiplexed error-robust FISH (MERFISH) is a powerful single-molecule imaging technique developed for determining copy numbers of thousands of RNA molecules (Chen et al., 2015; Xia et al., 2019). It revolutionizes transcriptome-scale RNA imaging at the single-cell level by utilizing error-robust barcodes to encode individual RNA species. Notably, the MERFISH protocol has undergone updates to enhance detection efficiency, allowing the transcriptome-scale imaging of approximately 1,000 RNA species in single cells.

STARmap is a novel approach that allow to measure gene expression at single cell compartment in intact tissue, providing insights into the activity of over a thousand genes (Wang et al., 2018). In the future, integrating this intact-system gene expression measurement with complementary methodologies will offer cellular-resolution analysis. One of the powerful technologies in this category is in situ sequencing (ISS), which utilizes padlock probes and rolling circle amplification. An enhanced version of ISS, known as HybISS (hybridization-based in situ sequencing), has been developed to improve the spatial detection of RNA transcripts (Gyllborg et al., 2020). HybISS incorporates modifications in probe design, enabling a new barcoding system through sequence-by-hybridization chemistry. The HybISS design offers increased flexibility, allowing for enhanced multiplexing and improved signal-to-noise ratio, all while maintaining the efficiency required for imaging large fields of view. Fluorescent In Situ Sequencing (FISSEQ) is a molecular technique that allows the simultaneous visualization of RNA molecules and their sequences in intact biological samples (Lee et al., 2015). FISSEQ uses barcoded oligonucleotide probes to hybridize targeted RNA molecules. These probes are amplified and labeled with fluorescent dyes, allowing the visualization of the RNA molecules. In the widely adopted RNA labeling method, query RNA molecule is labeled with a specific stem-loop. This stem-loop sequence binds to the bacteriophage coat protein and fused to a fluorescent protein that enables the visualization of the labeled RNA using fluorescence microscopy (Heinrich et al., 2017). Using CRISPR-Cas based RNA imaging, which offers target programmability without genetically altering the target, it is possible to track a variety of endogenous long-chain RNAs (like mRNA) in live cells in real-time (Abudayyeh et al., 2017; Chen et al., 2020). The CRISPR-mediated RNA imaging technique involves the use of a modified CRISPR-Cas system that targets specific RNA molecules and recruit fluorescent proteins.

In the last decade, RNA sequencing (RNA-seq) has been heavily used for a wide range of omics, including a comprehensive analysis of gene expression and sequencing of the transcriptome. It now encompasses various methods that enable the investigation of diverse aspects of RNA biology. Furthermore, the advent of long-read sequencing and direct RNA-seq technologies, coupled with improved computational tools for data analysis, has revolutionized RNA-seq (Wang et al., 2009). In contrast to classical sequencing technologies, RNA-Seq offers superior coverage and enhanced resolution. Microarray RNA profiling is an older technique that allows the simultaneous detection and quantification of thousands of RNA transcripts in a single experiment (Schena et al., 1995). This procedure has limitations, including its reliance on pre-designed probes, which may not detect all RNA transcripts. A tiling array consists of short nucleic acid fragments fixed onto a substrate (Bertone et al., 2004). These arrays are specifically designed to cover the entire genome of the target species. Tiling arrays are widely utilized in various applications, such as ChIP assays, to determine genome binding or to identify transcribed regions. By employing tiling arrays, researchers can obtain valuable insights into the binding patterns of the genome in ChIP assays.

Serial analysis of gene expression (SAGE) is a well-established technique developed to study the transcriptome by identifying and quantifying transcripts, including non-coding RNAs (ncRNAs) (Velculescu et al., 1995). This method relies on restriction enzymes to generate short, unbiased cDNA sequences known as SAGE tags. It involves several steps, including isolating and converting mRNA into cDNA, cutting the cDNA into smaller fragments, and adding tags to each fragment. The fragments are then combined and sequenced, providing information about which genes are present and how many copies of each gene are being expressed. The limited length of the sequence tags in SAGE may need to provide more information to accurately determine the identity of a new sequence. To address this limitation, a modified version of SAGE that utilizes more extended tags has been developed, demonstrating increased usefulness in transcript identification. Cap analysis gene expression (CAGE) provides information about transcript abundance and promoter identification (Shiraki et al., 2003). It is a high throughput method that can identify and quantify 5′ ends of capped RNAs. It has been used for study of RNA subcellular localization (Djebali et al., 2012) and identification of transcription start sites (ENCODE Project Consortium, 2012). The PARE is another technology used to identify and quantify small RNA molecules in a sample (German et al., 2009). PARE works by capturing the 5′ ends of RNA fragments using a technique called ligation-mediated PCR. It is often used to study the targets and functions of miRNAs.

GRO-seq was developed to study the transcriptional activity of the genome at a global scale. It works by labeling nascent RNA transcripts using BrUTP (bromouridine triphosphate) and isolating and sequencing the labeled RNA molecules (Gardini, 2017). This method allows researchers to identify the specific genomic regions where transcription occurs and the direction and rate of transcription. GRO-seq is often used to study gene expression regulation and identify novel non-coding RNAs.

RIP-Chip is used for isolating and sequencing RNA transcripts that bind to RNA-binding proteins (Keene et al., 2006). Various downstream methods, such as high-throughput sequencing, characterize the RNA molecules associated with the RNA binding proteins. RIP is often used to study post-transcriptional regulation of gene expression, particularly the regulation of mRNA stability and translation. Transcript IsoForm Sequencing (TIF-Seq) is a powerful technology that has been used to study transcript isoforms at the genome scale (Pelechano et al., 2013). In TIF-Seq, both the 5′ and 3′ ends are simultaneously sequenced and it allows the user to accurately determine the stand and end sites of individual RNA molecules. Selective 2′-hydroxyl acylation analyzed by primer extension (SHAPE) was developed to study the structure and folding of RNA molecules at single nucleotide resolution (Smola et al., 2015). SHAPE experiments use the reactivity of the RNA ribose 2′-OH towards hydroxyl-selective electrophilic reagents to model the secondary and tertiary structure of RNA molecules. PARS is another method for studying RNA structure and folding (Wan et al., 2013).

APEX-RIP allows the identification of RNA-protein interactions by combining two techniques: APEX (Ascorbate Peroxidase-mediated Protein Localization and Protein-Protein Interaction Profiling) and RIP (RNA Immunoprecipitation) (Kaewsapsak et al., 2017). APEX-RIP has been used to identify RNA-binding proteins and their associated RNAs in a variety of systems, including cancer cells, neuronal cells, and embryonic stem cells. Biochemical fractionation is a technique used to isolate specific cellular components or organelles from complex mixtures such as whole cells or tissues. The distribution and expression of RNA molecules in different subcellular fractions is studied by combining biochemical fractionation and RNA-seq (Tilgner et al., 2012). By fractionating cells or tissues into different organelles or subcellular compartments, researchers can isolate RNAs that are localized or enriched in specific cell regions. This approach has been used to study various aspects of RNA biology, including RNA localization.

RNAlocate is a comprehensive resource that provides information on the subcellular localization of different types of RNA molecules (Zhang et al., 2017). The current version of RNALocate contains a vast collection of over 37,700 manually curated entries, each supported by experimental evidence of RNA-associated subcellular localization. The database encompasses more than 21,800 coding and non-coding RNAs across 65 species. It covers a wide range of 42 subcellular locations, focusing on Homo sapiens and Mus musculus. RNALocate is a valuable repository for researchers seeking knowledge about RNA localization and its functional implications. The wealth of data available in RNALocate has been utilized in the development of several popular datasets. Notably, it has contributed to creating datasets used in lncLocator (Cao et al., 2018) and iLoc-lncRNA (Su et al., 2018). These resources leverage the comprehensive information provided by RNALocate to enhance the study of subcellular localization patterns in long non-coding RNAs (lncRNAs) and other RNA molecules.

The dataset for lncLocator was extracted from RNALocate version 1 (Zhang et al., 2017). A total of 1,361 lncRNA entries with curated subcellular localization were retrieved from the RNAlocate database, and multiple entries for the same lncRNA were combined. Only those lncRNA were retained, with their sequence information in NCBI (Sayers et al., 2022) and Ensembl (Cunningham et al., 2022). Seven subcellular locations were used for classification, and there were 19 combinations of these locations in the dataset. Redundancy was removed using CD-HIT, which was used to remove sequences with more than 80% similarity. Multi-locational lncRNAs were removed, and lncRNAs associated with only one location were allowed. Two locations (Endoplasmic reticulum and synapse) were removed as they had very few samples. Finally, a dataset of 612 lncRNA was created, covering five subcellular locations.

The dataset used in iLoc-lncRNA (Su et al., 2018) was created from RNALocate version 1 (Zhang et al., 2017). 923 lncRNA sequences with annotated subcellular localizations were obtained, and CD-HIT reduced redundancy at 80% similarity. After removing similar sequences, 655 non-redundant lncRNA sequences were obtained. The dataset comprises 156 lncRNAs from the nucleus, 426 lncRNAs from the cytoplasm, 43 lncRNAs from the ribosome, and S4 contains 30 lncRNAs from the exosome. This dataset was also used in lncLocation, Locate-R, and lncLocPred.

The training dataset used in KD-KLNMF (Zhang and Qiao, 2020) is obtained from Yang et al. (2020), where they have used 923 sequences for the training dataset. CD-HIT was used at a similarity index threshold of 80% to remove redundant sequences from the dataset. The final training dataset contains 644 lncRNA sequences - 154 sequences from the nucleus, 417 sequences from the cytoplasm, 43 from the ribosome, and 30 from the exosome. In order to balance the dataset, SMOTE (Synthetic Minority Oversampling Technique) was applied to the dataset multiple times post-feature extraction, such that the number of samples in the minority classes equaled that of the majority class. In DeepLncLoc, LncRNAs located in only one location were selected for model construction, the dataset was created from RNALocate v1 (Zeng et al., 2022). GM-lncLoc (Cai et al., 2023) obtained datasets from iLoc-lncRNA and lncLocator, which were originally extracted from RNALocate. Both datasets were subjected to oversampling using SMOTE, and all the classes were equally represented. Dataset1 had 292 lncRNA samples, while dataset2 had 417 samples. Also, an independent dataset was created using the test dataset from DeepLncLoc. GraphLncLoc (Li et al., 2023) uses the benchmark dataset of DeepLncLoc for training their graph convolution network-based model.

Apart from RNALocate, some methods have used RNA-seq data to generate localization data by quantifying the gene expression in subcellular locations.

DeepLncRNA (ENCODE Project Consortium, 2012) used 93 RNA-seq samples from 14 human immortalized cell lines from the ENCODE database (ENCODE Project Consortium, 2012), of which 45 were from the cytosol, and 48 were from the nucleus. Differential transcript expression was used to quantify the differences in lncRNA transcript abundances between nuclear and cytosolic cellular fractions for each cell type. Log2 fold change was used to determine enrichment in the nucleus or cytosol. A sequence was assigned to the cytosol if the log2 fold change <0, and it will be assigned to the nucleus if the log2 fold change >2.8. Applying this log2 fold change threshold resulted in a dataset containing 4,380 cytosolic lncRNAs and 4,298 nuclear lncRNAs.

LncLocator 2.0 (Lin et al., 2021) sourced data from two databases—nucleotide sequences from the GENCODE (Frankish et al., 2019) project and localization information from lncATLAS (Mas-Ponte et al., 2017), combined by using common gene IDs. lncATLAS uses relative concentration index (RCI) for quantifying subcellular localization. RCI is defined as the log ratio between concentrations measured by Fragments Per Kilobase Million (FPKM) in two samples (Cytoplasm/Nucleus). Different numbers of lncRNA samples are available based on the type of cell line as they have filtered out the lncRNA with Cytoplasm/Nucleus RCI in the range [−1, 1]. The dataset used in lncLocator 2.0 was also used to develop the model in TACOS (Jeon et al., 2022), which is also a cell-line-specific predictor. However, TACOS utilizes only ten cell lines to develop their classifier.