AnnoLnc2: the one-stop portal to systematically annotate novel lncRNAs for human and mouse.

With the abundant mammalian lncRNAs identified recently, a comprehensive annotation resource for these novel lncRNAs is an urgent need. Since its first release in November 2016, AnnoLnc has been the only online server for comprehensively annotating novel human lncRNAs on-the-fly. Here, with significant updates to multiple annotation modules, backend datasets and the code base, AnnoLnc2 continues the effort to provide the scientific community with a one-stop online portal for systematically annotating novel human and mouse lncRNAs with a comprehensive functional spectrum covering sequences, structure, expression, regulation, genetic association and evolution. In response to numerous requests from multiple users, a standalone package is also provided for large-scale offline analysis. We believe that updated AnnoLnc2 (http://annolnc.gao-lab.org/) will help both computational and bench biologists identify lncRNA functions and investigate underlying mechanisms.

INTRODUCTION

Long noncoding RNAs (lncRNAs) have been demonstrated to play many important regulatory roles via various mechanisms (1,2), and many characteristics have been used to infer their functions. For example, analyzing the natural selection pressures on lncRNAs can reveal whether they are functional (3), and their subcellular localization indicates where they perform functions (4). Functional enrichment of protein-coding genes whose expression profiles are highly correlated with lncRNAs provides important hints for potential lncRNA functionality (5). In addition, it has been reported that many lncRNAs perform their functions by interacting with other cellular factors; thus, studying interacting partners of lncRNAs, e.g. proteins and miRNAs, provides insights into their possible molecular mechanisms (6,7). Multiple tools have been developed to annotate lncRNA functions via particular mechanisms (e.g., LncTar (8) for predicting interacting RNA partners, SFPEL-LPI (9) for predicting lncRNA–protein interactions, lncFunTK (10) for annotating GO functions and regulatory networks based on coexpressed genes along with transcription factors and miRNA binding profiles, and lncLocator (11) and iLoc-lncRNA (12) for predicting lncRNA subcellular localization), along with several databases available for known lncRNAs (e.g. LncBook (13) and NONCODE (14)) (see Supplementary Table S1 for a detailed comparison).

As part of our continual attempts to provide the community with intuitive online tools for annotating novel lncRNAs effectively and efficiently, we present AnnoLnc2, the new version of the AnnoLnc web server (15). AnnoLnc2 supports on-the-fly annotation for any human or mouse lncRNA. In addition to supporting new species and new annotation modules, AnnoLnc2 has been fully updated, with many new datasets incorporated (Table 1). To enable computational biologists to run high-throughput analysis efficiently, we also provide a fully functional standalone package for large-scale offline analysis. All these resources are freely available at http://annolnc.gao-lab.org/ and open to all users without login requirements.

Table 1.

Summary of annotation data sources for each module

Module	Annotation data
Expression	Human: 52 RNA-seq samples for 26 tissues (58), 39 CCLE cancer cell lines (59), and 4 RNA-seq samples for H1 and GM12878 from ENCODE (60);
	Mouse: 56 RNA-seq samples for 28 tissues and 13 RNA-seq samples for 7 cell lines from ENCODE (60).
Transcriptional regulation	Human: 13 515 ChIP-seq samples for 1339 TFs (25);
	Mouse: 10 728 ChIP-seq samples for 738 TFs (25).
Subcellular localization	Human: 40 RNA-seq samples covering 10 cell lines (29).
miRNA regulation	Human: 170 AGO CLIP-seq samples in various human tissues (e.g. brain cortex) and cell lines (e.g. H1, HK-2 and osteoblast);
	Mouse: 153 AGO CLIP-seq samples in various mouse tissues (e.g., cortex and liver) and cell lines (e.g. mESC, CD4+ T cell and keratinocytes).
Protein interaction	Human: 385 CLIP-seq samples for 188 RBPs in various human tissues (e.g. brain, adrenal gland and hippocampus) and cell lines (e.g. HEK293T, HeLa and K562);
	Mouse: 238 CLIP-seq samples for 62 RBPs in various mouse tissues (e.g. brain, testis, and cortex) and cell lines (e.g. B cell, mESC, 220-8 cell and N2A cell).
Genetic association	Human: 96 455 trait-associated SNPs from the GWAS catalog (42), 298 590 SNPs in linkage disequilibrium with GWAS SNPs from dbSNP (61), and 4 365 369 eQTLs from GTEx (41);
	Mouse: 4013 phenotypic alleles and 997 QTL alleles from MGI (43).
Evolution	Human: phyloP and phastCons scores for the primate, mammal, and vertebrate clades were computed by a local implemented UCSC MultiZ-Phast pipeline based on 100-way multiple alignments from UCSC Genome Browser (62); derived allele frequency (DAF) is based on the 1000 Genomes Project (47);
	Mouse: phyloP and phastCons scores for the glire, euarchontoglire, placental and vertebrate clades were based on 60-way multiple alignments from UCSC Genome Browser (62).

RESULTS

Effectively annotating novel lncRNAs in human and mouse

To systematically annotate novel lncRNAs, AnnoLnc2 is equipped with multiple annotation modules, covering sequence and structure, expression and regulation, function and interaction, as well as genetic association and evolution (Figure 1).

Figure 1.

Framework of AnnoLnc2. The AnnoLnc2 web server provides on-the-fly services for annotating novel lncRNAs from 10 perspectives, covering from sequence and structure, expression and regulation, function and interaction, to evolution and association, for human and mouse.

Sequence and structure

For inputted lncRNAs, AnnoLnc2 will first try to map them against the latest human (hg38) and mouse (mm10) genome builds. During mapping, AnnoLnc2 tries to associate the inputted lncRNAs with known lncRNAs by searching GENCODE gene models based on genomic location and intron/exon structure, and matched hits will be reported as part of the genomic mapping results. In addition, AnnoLnc2 also reports known repeats (16) in the inputted lncRNAs given their potential importance for lncRNAs’ functionality (17,18).

LncRNAs tend to acquire complex structures that are strongly linked to their biological functions (19,20). AnnoLnc2 performs de novo secondary structure prediction with the ViennaRNA/RNAfold program (21). To help identify potential structural/functional domains, the result is further visualized as an interactive graph, allowing flexible base coloring on the basis of various scores, including conservation and energy entropy.

Expression and regulation

LncRNAs exhibit temporal-spatial expression with sophisticated regulation (22,23). Based on the LocExpress (24) algorithm, AnnoLnc2 implements on-the-fly expression profiling for hundreds of samples. Specifically, the AnnoLnc2 web server supports simultaneous expression calling in 95 human samples (52 normal tissue samples, 39 cancer cell lines from Cancer Cell Line Encyclopedia (CCLE) and 4 ENCODE cell line samples) or 69 mouse samples (56 normal tissue samples and 13 ENCODE cell line samples) for any lncRNA (Supplementary Table S2) in nearly real time. Compared to AnnoLnc, AnnoLnc2 offers a more abundant and reliable representation of lncRNA expression with a novel species (mouse) and a ∼50% increased human sample pool. Likewise, AnnoLnc2 also presents an eight-fold increase in its transcriptional regulation module, with 91.7 million ChIP-seq-based binding sites for 1339 human transcription factors (TFs) as well as 80.8 million sites for 738 mouse TFs (25). For miRNA-based regulation, AnnoLnc2 implements a dedicated AGO-CLIP-based module to identify putative functional miRNA-binding events for inputted lncRNAs based on 170 human datasets and 153 mouse datasets and allows users to run TargetScan (26) to predict lncRNA-miRNA interactions. Furthermore, for each predicted miRNA binding site, AnnoLnc2 provides a structural view, with this site highlighted for linking the binding site and structural context.

Several recent studies have well demonstrated that, in addition to expression level, the subcellular localization of lncRNAs is also under intricate regulation (27,28). In an attempt to map the lncRNA subcellular localization landscape at fine-scale, we designed and implemented a novel module to empirically characterize lncRNA cytoplasmic/nuclear localization preference in ten human cell lines (29). Briefly, the inputted lncRNAs are compared with a precompiled localization catalog (Supplementary Figure S1). AnnoLnc2 then returns the localization preference, measured by the fold change in nuclear/cytosolic expression in all ten cell lines, along with the corresponding FDR-adjusted P-value (Figure 2C, see Supplementary Figure S1 for the detailed pipeline). Moreover, many studies have shown that multiple sequence motifs contribute to lncRNA localization (30). AnnoLnc2 will report known localization-related motifs in the query sequence based on a curated list compiled from published researches (Supplementary Table S3) (31–35). Meanwhile, we also implemented a dedicated analysis module for running motif identification (64) over a set of selected lncRNAs in the standalone version.

Figure 2.

AnnoLnc2 web server. Users can run AnnoLnc2 through a three-step operation (A) and view detailed annotation results, as well as download all results by one click (B). A well-studied lncRNA, NEAT1, is an important component of nuclear paraspeckles (63). (C) Annotation results from the subcellular localization module show that NEAT1 is strongly located in the nucleus in multiple cell lines. Each bar represents logarithm scaled nuclear/cytoplasmic localization ratio, with corresponding FDR marked upon it.

Function and interaction

LncRNAs play functional roles via various interactions (36). AnnoLnc2 reports lncRNA–protein interactions based on both experimentally validated data and sequence-oriented prediction (37). Briefly, we first manually curated and processed recently published human and mouse CLIP-seq datasets for RNA-binding proteins (RBPs) from GEO (see Supplementary Table S4 for a detailed list as well as Supplementary Figure S2 for details on data processing) and then merged these newly identified binding sites with published sites (38) to obtain the final catalog. By integrating 385 human CLIP-seq datasets (for 188 RBPs) and 238 mouse CLIP-seq datasets (for 62 RBPs), AnnoLnc2 nearly quadruples the number of experimentally validated RNA-binding proteins compared with that of AnnoLnc and provides the most comprehensive online resource for CLIP-seq-based lncRNA–protein interaction annotation so far. Similar to that of miRNA binding sites, each protein binding site can be viewed in a structural context. For RNA-binding proteins out of the list, we utilized lncPro (37) to predict lncRNA–protein interaction ab initio. Moreover, benefiting from its own effective expression profiling module, AnnoLnc2 also systematically computes all coexpressed genes of the inputted lncRNA and annotates the lncRNA with enriched GO terms for these genes at the biological process and molecular function levels using the ‘Guilt-by-Association’ strategy (5,22,39).

Evolution and association

Association studies such as genome-wide association studies (GWASs) and expression quantitative trait loci (eQTLs) offer unique opportunities to characterize lncRNAs associated with diseases and other traits (40,41). In addition to annotating the inputted lncRNA with variants from the NHGRI GWAS Catalog (42) (with ∼25% more candidate variants compared to that in AnnoLnc), AnnoLnc2 also integrates all human eQTLs from GTEx (41) and mouse phenotypic and QTL alleles from MGI (43). When reporting human GWAS SNPs, all SNPs linked with the tag SNP are also reported, along with their position relative to the aligned lncRNA (e.g. ‘promoter’, ‘exon’ or ‘intron’) and all relevant information of the tag SNP (i.e. the associated trait, the P-value and the PubMed ID of the paper reporting this SNP). When reporting mouse alleles, the allele's type, ID, symbol and name, its associated marker's ID and name as well as the PubMed ID of the paper reporting the allele are listed.

Selection is a direct indicator of biological functions (3,44). To effectively identify selection signals across various evolutionary clades, we reimplemented the UCSC MultiZ-Phast pipeline (45,46) and recomputed conservation scores for the primate, mammal and vertebrate clades. Briefly, we downloaded the 100-way human multiple alignment data from UCSC Genome Browser, constructed a phylogeny for each clade with phyloFit (45), and computed phastCons and phyloP scores by PHAST and phyloP, respectively. In addition, the server will also list the derived allele frequency (DAF) of all the SNPs from the 1000 Genomes Project (47) that fall within the lncRNA locus to analyze potential population-scale selection.

Powerful and user-friendly interfaces

AnnoLnc2 takes RNA sequences in fasta format as input, and users can paste or upload sequences on the AnnoLnc2 homepage and choose corresponding species from a drop-down option list. After that, users simply need to click the ‘GO’ button, and AnnoLnc2 will analyze all inputted sequences (up to 100 sequences) simultaneously (Figure 2A, B). After a job is submitted, AnnoLnc2 will return a link to users for retrieving results in the future. AnnoLnc2 provides an item for each best alignment of all sequences, and users can view detailed annotation results from the link in the ‘Status’ column. All of the results are shown in interactive figures and tables and allow users to view detailed numeric information and search for items of interest, such as TFs, RBPs, traits, or functions. Of note, AnnoLnc2 offers a ‘Summary’ text section for briefly describing the annotation results for each module. To make using AnnoLnc2 more convenient, AnnoLnc2 supports batch uploading and downloading to avoid laboring operations for fetching results one by one. The AnnoLnc2 web server is now hosted on an elastic cloud server from the Huawei Cloud running an Ubuntu Linux system (18.04 LTS) with 16 CPU and 64 GB memory.

Moreover, in response to user requests, we also provide a standalone version of AnnoLnc2 for large-scale offline analysis (available from http://annolnc.gao-lab.org/download.php). In addition to providing all functionalities of the online server, the standalone AnnoLnc2 package is fully customized, and users are able to not only specify their own datasets for modules such as expression profiling and lncRNA-protein interaction but also fine tune/change the code for their specific requirements (or even port it to an entirely new species!). To facilitate customization, we also provide the online AnnoLnc2 Developer Kit as a set of scripts and guidelines for particular conversion and maintenance tasks. Currently, the standalone package requires a reasonable Linux server with a minimal requirement of 8 GB memory (see http://annolnc.gao-lab.org/download.php for the installation guide).

DISCUSSION

As the successor of AnnoLnc, AnnoLnc2 supports annotating both human and mouse novel lncRNAs with wider and more comprehensive data resources and provides a standalone version for large-scale customized annotation, along with a more responsive and user-friendly Web interface. The major improvements are listed below (also see the full list at http://annolnc.gao-lab.org/help.php#link-update).

AnnoLnc2 is now available for both Mus Musculus (mm10) and Homo sapiens (hg38), compared to Homo sapiens (hg19) only in AnnoLnc.

A novel module has been added for assessing lncRNAs’ subcellular localization based on nucleus/cytosol-separated profiling data (29).

All annotations have been updated with more datasets incorporated. For example, the ChIP-seq dataset now covers 1339 human TFs and 738 mouse TFs (versus 159 human TFs in AnnoLnc), and the updated CLIP-seq dataset contains 188 human RBPs and 62 mouse RBPs (versus 51 human RBPs in AnnoLnc).

In response to numerous requests from multiple users, a standalone version is available for large-scale offline analysis.

To the best of our knowledge, AnnoLnc2 is currently the most comprehensive annotation tool for analyzing novel lncRNAs in both human and mouse. These updates significantly enhance AnnoLnc2 as an effective and efficient tool for annotating lncRNAs and inspiring hypotheses for further investigation, which we would like to demonstrate below with two cases.

The polymorphisms in human lncRNA NR_046105.1 were reported to influence schizophrenia (48). Based on the results from the genetic association module, we found 72 SNPs (71 intronic) enriched in the NR_046105.1 gene locus. Among them, 69 SNPs are in linkage disequilibrium with schizophrenia-associated tag SNPs, e.g. rs1625579 (48,49) (Supplementary Table S5). Consistently, the expression profile (Figure 3A) indicates that NR_046105.1 is exclusively highly expressed in the brain, which supports the important roles of NR_046105.1 in the brain.

Figure 3.

Case study of the human lncRNA NR_046105.1. (A) Expression profile of NR_046105.1 in normal samples, where it is exclusively highly expressed in the brain. (B-C) Functions of NR_046105.1 predicted by the coexpression-based functional annotation module at the biological process (B) or molecular function (C) level. (D) Annotation results from the genetic association module. The SNP rs11804556 is not only a tag SNP but also an eQTL. (E) KLF12 is predicted to interact with NR_046105.1 by the protein interaction module.

We then explored how NR_046105.1 performs its function. The coexpression-based functional annotation module found that NR_046105.1 is involved in ‘chemical synaptic transmission’ and ‘synaptic signaling’ from the ‘biological process’ category (Figure 3B), which is consistent with the finding of a previous study (50). In addition, its enriched ‘molecular function’ terms suggested that NR_046105.1 regulates the activity of neurotransmitter reporters and ion channels (Figure 3C), which may partially explain how it affects synaptic transmission and leads to brain disease. Further inspection of GWAS SNP and eQTL annotations showed that the SNP rs11804556 is associated with ‘cognitive performance’ and the expression of NR_046105.1 (Figure 3D), suggesting that this variant affects cognition by regulating gene expression. Based on the results from the genetic association module, we also found that most eQTLs (98/101) in NR_046105.1 influence DPYD expression. Considering that DPYD was reported to be involved in pancreatic cancer (51), as is miR-137 (product of NR_046105.1) (52,53), these eQTLs may provide an explanation for their relationship. Moreover, He et al. (53) found that miR-137 plays a role in pancreatic cancer by targeting KLF12, and the protein interaction module of AnnoLnc2 also discovered this interaction (Figure 3E).

Similarly, the well-studied mouse lncRNA Pvt1 (Ensembl ID: ENSMUST00000180432.8) was reported to regulate atrial fibrosis by acting as a sponge of miR-128-3p (54), which was also identified by the miRNA regulation module of AnnoLnc2 (Figure 4A). Moreover, Pvt1 was predicted to play roles in the regulation of immune system processes based on enriched functions at the biological process level in cell lines (Figure 4B), in agreement with the finding of a previous study (55). According to the results from the functional annotation module, this process may be achieved by regulating ‘immunoglobulin receptor activity’ (Figure 4C). In addition, we found that Pvt1 might regulate translation (Figure 4D), which was further supported by the annotated multiple binding sites of LIN28A and TIA1 on Pvt1 (Figure 4E), both of which have been reported to act as regulators of mRNA translation (56,57). Overall, AnnoLnc2 leads to the new hypothesis that Pvt1 is involved in translation regulation by acting as a sponge of the RNA binding proteins LIN28A and TIA1.

Figure 4.

Case study of the mouse lncRNA Pvt1. (A) miRNA annotation results revealed that miR-128-3p might interact with Pvt1. (B–D) Functions of Pvt1 predicted by the coexpression-based functional annotation module at the biological process (B) or molecular function (C) level in cell line samples and at the biological process level in tissue samples (D). (E) Integrated view in UCSC Genome Browser displaying the binding spectrum of LIN28A and TIA1 on Pvt1.

During past years, the AnnoLnc web server has been widely used by the community, serving 50 000+ users around the world, with millions of sequences analyzed. In the future, we will continue our efforts to improve the services provided by AnnoLnc to research communities. More specifically, we will keep updating backend data sources and support more species, including fruit fly and nematode, as well as plants, such as Arabidopsis. Beyond the current modules, other types of helpful annotations, such as RNA modification and machine learning-based function or localization prediction, could also be incorporated to better understand lncRNA functionalities and potential mechanisms. Finally, we recognize the urgent need for a user-friendly and interactive interface for a wider research community and will continue improving the server interface based on users’ feedback.

Click here for additional data file.