The human brain confers on us the abilities of perceptions, thoughts, emotions, actions, and memories. Over many years, the neuroscientists have strived to understand the molecular, cellular, circuit and behavior-level mechanisms that underlie these processes. Around a century ago, Santiago Ramon y Cajal proposed the neuron doctrine postulating that the relationship between nerve cells was not continuous, but contiguous. Cajal, in his Theory of Dynamic Polarization, described how information, in the form of electrical signals, travels within individual neurons, from their dendrites to their cell bodies and finally to their axons. We now know cognition, emotion, memory, and action are generated by circuits and networks of thousands to millions of interconnected neural cells, mostly neurons. Neural circuits are both anatomical and functional entities, composed of a series of interconnected neurons and glial cells with diverse properties and functions. However, it remains largely elusive how specific types of neural cells assembles the neural circuits in different brain regions and how specific neural circuits perform their signal processing functions during cognitive processes and behaviors. This requires detailed knowledge on the construction of neural circuits at the single-cell resolution and on the spatiotemporal pattern of neuronal activity (Poo et al., 2016). The United States BRAIN (Brain Research through Advancing Innovative Neurotechnologies) initiative was launched in 2013, which was intent to “accelerate the development and application of innovative technologies to construct a dynamic picture of brain function that integrates neuronal and circuit activity over time and space,” To achieve this goal “requires an integrated view of its component cell types, their local and long range synaptic connections, their electrical and chemical activity over time, and the functional consequences of that activity at the levels of circuits, the brain, and behavior” (NIH, 2014).

The process of neuronal specification, migration, and circuit formation is enormously complex in time and space during development, with multiple levels of regulation. Deciphering this process requires both a high-throughput neuronal subclass identification and an integrative approach that considers dynamic, multilayered transcriptional regulation (Molyneaux et al., 2015). However, current regulatory models are limited to a number of regulators, mostly transcription factors, which account for a limited subset of key nodes within a broader regulatory network that is believed to be far more complex (Greig et al., 2013). The transient expression, flexible structures, and dynamic localization of RNA molecules enable fine-tuning genome arrangement, scaffolding and transcription functions, thus precisely regulating gene expression in a time and site-specific manner. Recent work indeed points to the critical role of long non-coding RNAs (lncRNAs) in transcriptional and post-transcriptional regulation of gene expression, the formation of complexes with epigenetic regulatory machinery, and chromosomal architecture organization (Rinn and Chang, 2012; Quinn and Chang, 2016). Therefore, lncRNAs participate in numerous physiological and pathological processes including maintenance of pluripotency, lineage specification, organogenesis, tumorigenesis, and metabolism (Wang et al., 2011; Ramos et al., 2013; Li and Chang, 2014; Yang et al., 2014; Wu et al., 2015). Although recent reviews have covered many aspects of lncRNAs in the assembly and plasticity of neural circuits, this field is fast-growing with new evidence reinforcing the notion that lncRNAs are pivotal in cell fate determination and in modulating neural activity (Ng et al., 2013a; Aprea and Calegari, 2015; Briggs et al., 2015). In this review, we highlighted roles and mechanisms of lncRNAs in assembly, maintenance, plasticity and abnormality of neural circuitry. Given the cis- and trans- regulatory mechanisms by lncRNAs and/or their embedding DNA elements, along with far more uncharacterized lncRNAs than protein-coding genes, strategies and technologies in studying lncRNAs were also discussed. Finally, we speculate that findings in lncRNA studies would deepen our understanding on neural circuitry composition and functional dynamics in physiology and disease conditions.

Current data from the ENCODE consortium suggest that up to 75% of the human genome may be transcribed (Djebali et al., 2012), but only about 1–2% of the human genome seems to encode protein (Birney et al., 2007; Church et al., 2009). Thus, most transcripts are non-protein-coding RNA (ncRNA) transcripts (Chodroff et al., 2010). LncRNAs are usually defined as non-protein coding transcripts longer than 200 nucleotides (nt) to exclude small regulatory RNAs such as short interfering RNAs (siRNAs), microRNAs (miRNAs), small nucleolar RNAs (snoRNAs), Piwi-interacting RNAs (piRNAs), and other short RNAs. Occasionally, functional short peptides can be derived from lncRNAs (Matsumoto et al., 2017). Until now, the NONCODE database have annotated 101,700 lncRNA genes in the human genome (Zhao et al., 2016). Remarkably, 40% of lncRNAs are expressed predominantly in the brain (Derrien et al., 2012). While many lncRNA genes overlap protein-coding genes in sense or antisense directions, others resides in genomic regions previously termed “gene deserts,” between protein-coding genes (intergenic) (Carninci et al., 2005; Cheng et al., 2005; Katayama et al., 2005; Kapranov et al., 2007a,b; Qureshi et al., 2010). In a recent study, using FANTOM5 (Functional Annotation of Mammalian cDNA) cap analysis of gene expression (CAGE) data, Hon et al. generated an atlas of nearly 30,000 human lncRNA genes with typical 5′ ends and expression profiles across 1,829 human primary cell types and tissues. Interestingly, most intergenic lncRNAs (lincRNAs) originate from enhancers rather than from promoters. Incorporating genetic and expression data implicates around 20,000 potentially functional lncRNAs in multiple diseases and in transcriptional regulation (Hon et al., 2017).

Some lncRNAs have distinct molecular biogenesis features compared to protein-coding transcripts. Using native elongating transcript sequencing (mNET-seq), Schlackow et al. found human lincRNAs and protein-coding pre-mRNAs are transcribed by different Pol II phospho-CTD (the C-terminal domain) isoforms. LincRNAs are rarely spliced, mainly non-polyadenylated, and are stabilized in the nucleoplasm (Schlackow et al., 2017). LncRNA conservation includes four dimensions: the sequence, structure, function, and expression from syntenic genome loci (Diederichs, 2014). In fact, lncRNA exons are significantly more conserved than neutrally evolving sequences, albeit at lower levels than protein-coding genes (Derrien et al., 2012). Interestingly, lncRNA promoters are more conserved than their exons, and nearly as conserved as promoters of protein-coding gene (Guttman et al., 2009).

LncRNAs are generally expressed at lower levels than protein-coding transcripts, and exhibit more cell- and tissue-specific expression patterns. Moreover, lncRNA expression is vigorously regulated during neural development (Mercer et al., 2010; Belgard et al., 2011; Aprea et al., 2013; Molyneaux et al., 2015) and upon neural activity (Lipovich et al., 2012; Barry et al., 2014), suggesting their specific functional roles. Analyzing in situ hybridization data from ABA (the Allen Brain Atlas), numerous lncRNAs are found to be expressed in the adult mouse brain and most of them were present in specific neuroanatomical structures or cell types such as particular cortical regions or the hippocampus (Mercer et al., 2008). Similarly to the expression of fate-determining protein-coding genes, these region-specific and dynamic expression patterns of lncRNAs could be orchestrated by cis-regulatory elements (enhancers), chromatin status, and cell-type-specific or activity-dependent transcription factors (TFs) (Ramos et al., 2013).

Production of neurons and glial cells during neural development is an intricate but highly stereotyped process that necessitates accurate spatiotemporal controlling of neural stem/progenitor cells (NSPCs) self-renewal and differentiation (Zhou, 2012). The mature mammalian neocortex, for example, is a multi-layered structure and the layers-specific projection neurons are generated sequentially by cortical neural stem/precursor cells (NSPCs) lying in the ventricular zone/subventricular zone (VZ/SVZ) over developmental time. Intriguingly, cortical NPCs acquire restrictions in fate potential progressively over developmental time, which are largely cell-intrinsic (Mcconnell, 1995; Desai and Mcconnell, 2000; Shen et al., 2006; Gaspard et al., 2008). In contrast, cortical interneurons, which usually make local and inhibitory connections, are produced mostly from precursors in the ventral telencephalon and cortical hem and undergo tangential migration into the cortex (Anderson et al., 2002; Wonders and Anderson, 2006). Notably, most neurons are not directly derived from bipolar radial glial neural precursor cells (RGCs) but are from multipolar intermediate progenitor cells, which are direct progenies of RGCs and may undergo a few rounds of replication prior to differentiation. This so-called indirect cortical neurogenesis is more prevalent in primates than in rodents (Qian et al., 2000; Franco and Muller, 2013; Greig et al., 2013). Cortical neurogenesis is followed by gliogenesis, which occurs mostly after birth in mice.

After neural cells were generated in appropriate numbers, at right times, and in the correct places, they establish functional connections required for normal brain function. To form connections, neurons extend long processes, axons and dendrites, which allow synapse formation (synaptogenesis). Neurite outgrowth and synaptogenesis involve complex regulations on gene expression and signal transduction. Neurons can alter their synaptic connections and the relative strength of individual connections in response to increases or decreases in their activity. This so-called neural plasticity accounts for memory, learning, and cognition, as well as the brain's capability to recover from damage. Compared to studies on lncRNAs' roles in fate specifications of neural cells, little is known regarding lncRNAs' functions in modulating nerite growth, synaptogenesis and neural plasticity, which is partly due to their dynamic features, scarcity of research material and hurdles in functional validation (Puthanveettil et al., 2013).

Nonetheless, emerging evidence indicates both nuclear and synaptic lncRNAs are actively involved in these processes. Comparative sequence analysis of genomic regions covering 150 presynaptic genes discovered highly conserved elements in non-protein coding regions in eight vertebrate species. Many of these “non-exon-associated and non-protein-coding” elements can transcribe and were predicted to form a highly stable stem-loop RNA structure, whereas some conserved noncoding elements correlate with specific gene expression profiles (Hadley et al., 2006). This early work implied that non-coding transcripts are prevalent in genomic regions of presynaptic genes and may have regulatory roles in transcriptional regulation.

Since lncRNAs regulate neural development and function, it's not surprising that mutation or dysregulation of lncRNAs has implications in pathogenesis of mental illness and neurodegenerative diseases such as autism spectrum disorder (ASD), depression, schizophrenia, amyotrophic lateral sclerosis (ALS), Alzheimer's disease and neuropathic pain. Some neural-specific lncRNAs have been emerged as potential therapeutic targets.

LncRNAs related to cognitive functions or synaptic plasticity or other psychiatry diseases, including GOMAFU, BDNF-AS, and DISC2, may potentially contribute to major depressive disorder (MDD) (Huang et al., 2017). In a microarray-based study, about two thousand lncRNAs were found to be differentially expressed in peripheral blood samples from major depression disorder (MDD) patients (Liu Z. et al., 2014), but their diagnostic and therapeutic implications remain to be elucidated. A recent genome-wide study characterized thousands of lncRNAs to be differentially expressed in ASD peripheral leukocytes. Gene ontology (GO) analysis of these lncRNA gene loci predicted neurological regulations of the synaptic vesicle cycling, long-term depression and potentiation to be mainly involved, including SHANK2-AS and BDNF-AS (Wang et al., 2015). Similarly, a large-scale study applied RNA sequencing (RNA-seq) of 251 post-mortem samples of frontal and temporal cortex and cerebellum from 48 individuals with ASD and 49 control human subjects, and identified 60 differentially expressed lncRNAs (Parikshak et al., 2016). Twenty of these lncRNAs were previously shown to interact with microRNA (miRNA)–protein complexes, and 9 with the fragile X mental retardation protein (FMRP), whose mRNA targets are enriched in ASD risk genes (Parikshak et al., 2013; Iossifov et al., 2014). These data show that dysregulation of lncRNAs is an integral component of the transcriptomic signature of ASD (Parikshak et al., 2016). LncRNA GOMAFU/MIAT is downregulated in post-mortem cortical gray matter from schizophrenia (SZ) patients. GOMAFU associates with splicing factors SRSF1 (serine/arginine-rich splicing factor 1) and QKI and dysregulation of GOMAFU results in aberrant splicing of DISC1 and ERRB4, two SZ-associated genes (Barry et al., 2014). Another study discovered that GOMAFU mediates mouse anxiety-like behavior probably via association with BMI1, a key member of the polycomb repressive complex 1 (PRC1), to repress the expression of beta crystallin (Crybb1), one of the SZ-related genes (Spadaro et al., 2015).

Amyloid precursor protein (APP) is sequentially cleaved by beta-site APP-cleaving enzyme 1 (BACE1), β-secretase, and γ-secretase to generate the toxic Aβ₄₂ peptide. Defective Aβ₄₂ clearance and elevated BACE1 expression contribute to Aβ₄₂ accumulation and AD progression. BACE1-AS, the antisense transcript of BACE1, can bind to and stabilize BACE1 transcripts, thus increasing BACE1 protein levels. Interestingly, BACE1-AS is induced by Aβ₄₂ peptide, leading to increased BACE1 mRNA stability and amyloid accumulation via a positive feedback loop. Consistently, expression of BACE1 and BACE1-AS is elevated in brains of AD, and knockdown of BACE1-AS reduced BACE1 levels in vivo (Faghihi et al., 2008; Liu T. et al., 2014). The neuromuscular disorder spinal muscular atrophy (SMA) is caused by insufficient expression of SMN (survival motor neuron) protein, and the primary goal of SMA therapeutics is to increase SMN levels (Lefebvre et al., 1997). LncRNA SMN-AS1 is enriched in neurons and suppresses SMN expression by recruiting the polycomb repressive complex-2 (PRC2) to SMN promoter. Targeting SMN-AS1 with antisense oligonucleotides (ASOs) increases SMN expression both in cultured neurons and in mice, indicating SMN-AS1 has potential to be a novel therapeutic target for treating SMA (D'ydewalle et al., 2017).

Large-scale RNA dysregulations are essential molecular hallmarks in neurodegenerative diseases including ALS and FTLD (Frontotemporal lobe dementia; Polymenidou et al., 2012). This is mostly due to the presence of aberrant protein states (proteinopathy) of two essential RNA/DNA binding proteins TDP-43 and FUS (Fused in sarcoma) in affected neurons, including cytosolic translocation, truncation, phosphorylation, ubiquitination, and aggregates formation (Lagier-Tourenne et al., 2010; Da Cruz and Cleveland, 2011). Although TDP-43 and FUS regulate the processing of an array of RNA molecules including non-coding RNAs, no specific RNA was yet identified as major causal factor of ALS and FTLD (Tollervey et al., 2011). The association with TDP-43 or FUS/TLS could allow lncRNAs to carry out their cellular function. On the other hand, the dynamics of association/dissociation of RNAs with TDP-43 or FUS might contribute to TDP-43 and FUS proteinopathies (Yang et al., 2015).

The above findings implicate the correlation between dysregulation of lncRNAs and neurological diseases. However, many were in vitro studies with very few mechanistic hints. Thus, we are still far from understanding the extent and mechanisms that lncRNAs are involved in disease brains.

Although a number of lncRNAs have been found to be involved in most, if not all, aspects of neural circuitry assembly and plasticity, many essential questions remain to be answered. First, in contrast to the abundancy of lncRNAs characterized, very few of them are essential for embryonic development, cell fate choices or circuitry functions. So, as many may ask, are lncRNAs largely transcriptional noise or non-functional? It's quite possible that most lncRNAs only play subtly regulatory roles, and that certain lncRNAs are not normally required but become essential upon neuronal activation or injury. Second, compared to proteins, most lncRNAs have low sequence conservation across species or among homologs, though evolutionary conservation of RNA secondary structures may exist across species. It's, therefore, hard to identify parallel or redundant pathways and related molecular mechanisms. Third, lncRNAs may exert functions in cis (transcripts dependent or independent), and/or in trans (chromatin remodeling, histone modification, DNA methylation, transcription and splicing regulation etc.). Moreover, the embedding DNA elements that transcribe lncRNAs may have cis-regulatory roles. These conditions greatly confound experiment design and data interpretation for functional studies of lncRNAs. Thus, it's not surprising loss-of-function studies in vivo or in vitro using different approaches (e.g., RNAi, antisense oligonucleotides, genomic deletion, polyadenylation insertion, promoter deletion/inversion and CRISPR-Cas9 mediated gene inactivation, etc.) may lead to distinct phenotypes (Bernard et al., 2010; Schorderet and Duboule, 2011; Eissmann et al., 2012; Li et al., 2013). Since each technique has advantages and limitations, researchers are required to exhaustedly apply necessary approaches and develop new technologies to elucidate lncRNAs' roles and mechanisms.

Latest breakthroughs in genome engineering technology utilizing CRISPR (clustered regularly interspaced short palindromic repeats) and Cas9 system have dramatically accelerated biomedical researches (Doudna and Charpentier, 2014; Hsu et al., 2014). It has been widely used for generating genetic-modified cells, plants and animals (Cong et al., 2013; Niu et al., 2014; Peng et al., 2014); for disease modeling and genetic corrections (Platt et al., 2014; Cox et al., 2015); as well as for repressing (CRISPRi) or inducing (CRISPRa) gene expressions without altering genomic sequences (Gilbert et al., 2013, 2014; Konermann et al., 2014). In a few proof-of-principle studies, CRISPR-Cas9 has been successfully applied to lncRNA studies in cells and in animals (Ho et al., 2015; Ghosh et al., 2016; Zhu et al., 2016). A genome-scale deletion screening for functional lncRNAs were carried out using a lentiviral paired-guide RNA (gRNA) CRISPR-Cas9 library targeting hundreds of lncRNAs. This screen identified fifty-one lncRNAs that can enhance or slow down human cancer cell growth. Next, nine lncRNA candidates were validated utilizing CRISPR–Cas9-mediated genomic deletion, CRISPRa or CRISPRi, functional rescue and transcriptome profiling. This study indicates high-throughput genome deletion method mediated by CRISPR–Cas9 enables rapid identification of functional non-coding elements (Zhu et al., 2016). Using the minimal CRISPRi (dCas9) system targeting the roX locus in the Drosophila cells leads to an efficient and specific knockdown of roX1 and roX2 lncRNAs. Moreover, this minimal CRISPRi system inhibits roX expression efficiently in vivo and leads to loss-of-function phenotype, thus validating the method in a multicellular model organism (Ghosh et al., 2016). To explore if certain RNA molecule can exert transactivation or adapter roles, Schechner et al. developed a targeted localization method called CRISPR-Display utilizing Cas9 to deploy RNA cargos to specific DNA loci. A distinct feature of CRISPR-Display is that it makes possible for multiplexing of different functions at multiple loci in the same cell (Shechner et al., 2015). The ever-growing innovation of CRIPSR-Cas9 technique would also enable detection and editing of DNA and RNA with high specificity and sensitivity (Abudayyeh et al., 2016; Gootenberg et al., 2017; Qin et al., 2017). Nonetheless, caution must be taken when applying CRISPR-Cas9 techniques in lncRNA studies: First, deletion of lncRNA genes overlapping protein-coding genes' promoter/enhancer or intron should be avoided; Second, effects of CRISPRa and CRISPRi on promoter/enhancer elements that shared by both lncRNA and protein-coding genes should be taken into account when analyzing phenotypes. Finally, possible off-target effects can be addressed by applying multiple gRNAs and performing rescue experiments.

LncRNAs exert roles by associating with cellular macro-molecules including chromatin, DNA, RNA or proteins. Current biochemical means using RNA-centric or protein-centric strategies can identified these molecules and been extensively reviewed elsewhere (Mchugh et al., 2014). Technology breakthroughs in physics, chemistry, molecular biology and neuroscience would allow researcher to carry on high-throughput investigations of lncRNAs at single-cell, circuitry and animal levels.

One of the primary aim of the BRAIN initiative is to “identify and provide experimental access to the different brain cell types to determine their roles in health and disease” (NIH, 2014). However, we are still far from identifying and characterizing the component cells comprising the neural circuits, especially for glial cells. This is partially due to lack of defined biomarkers and dynamic changes of cell properties upon stimulation or depression. Latest advances in cell labeling using genetic and viral means, high throughput purification, e.g., FACS (fluorescence-activated cell sorting) and microfluidic devices, enable researchers to isolate neural cells from embryonic and adult brain under a variety of conditions. Moreover, recent development in profiling transcriptomes and epigenomes from as few as single cells markedly advanced molecular census of neural cell in embryonic and adult brains (Usoskin et al., 2015; Zeisel et al., 2015; Poulin et al., 2016; Tasic et al., 2016; Telley et al., 2016). Tasic et al. established a cellular classification of primary visual cortex in adult mice based on single-cell transcriptome analysis. A total of 49 transcriptomic cortical cell types, including 19 glutamatergic, 23 GABAergic, and 7 non-neuronal types were identified from around 1,600 cells labeled by Cre reporters. Interestingly, many transcriptomic cell types showed discrete anatomical and physiological characteristics, thus validating that the single-cell transcriptomic profiles can reflect specific properties of neural cells (Tasic et al., 2016).

Current annotations of brain lncRNAs are unfinished, partly because of the selection of polyadenylated (polyA) transcripts in most studies and RNA-seq libraries not preserving strand information (Miller et al., 2014; Darmanis et al., 2015). Single-cell RNA sequencing is even harder to detect lncRNAs because there's much less starting material and lncRNAs are generally less abundant than protein-coding transcripts. Nonetheless, latest single-cell transcriptome studies indeed correlate lncRNA profiles with developmental stages and cell identities. Using single cell RNA-seq to analyze roughly 100 individual cells from human embryonic stem cells (hESCs) and human preimplantation embryos, Yan et al. identified 2,733 previously uncharacterized lncRNAs, many of which are specifically expressed in developmental stages (Yan et al., 2013). In another single-cell study, more than 600 novel multi-exonic lncRNAs were discovered using micro-dissected adult subventricular zone (SVZ) tissues (Luo et al., 2015). Single-cell sequencing of hundreds of human cortical cells revealed that many lncRNAs are enriched in individual cells, and are cortical layer and cell type-specific (Liu et al., 2016), which coincides with previous studies showing lncRNAs provide more cell identity information during the development of mammalian cortex than protein-coding transcripts (Molyneaux et al., 2015). We speculate that lncRNA profiling at sing-cell level, along with high-throughput single molecule fluorescent in situ hybridization (smFISH), would greatly advance the census of neurons and glial cell, especially astrocytes, in the context of neural development and plasticity (Femino et al., 1998; Raj et al., 2008). Moreover, transcriptome dissection of cells in cerebral organoid derived from human pluripotent cells or NSPCs would advance our understanding on lncRNAs in brain evolution, development and diseases (Lancaster et al., 2013; Fatehullah et al., 2016).

N⁶-Methyladenosine (m⁶A) is a widespread, reversible chemical modification of polyadenylated mRNAs and lncRNAs in eukaryotes, implicated in many aspects of RNA metabolism including regulations of stability, transport and translation (Fu et al., 2014). Antibody-based N⁶-methyladenosine (m⁶A) RNA immunoprecipitation followed by high-throughput sequencing (MeRIP-seq) has been developed to profile the transcriptome-wide distribution of m⁶A, revealing m⁶A is distributed in more than 7,000 mRNA and 250 lncRNA transcripts in human cells (Dominissini et al., 2012). In mouse brain, m⁶A is present in mRNA at low levels throughout embryogenesis but increases dramatically by adulthood, suggesting that upregulation of m⁶A levels accompanies neuronal maturation. Moreover, lncRNAs transcribed by RNA polymerase II are also subject to m⁶A methylation, and long intergenic noncoding RNAs (lincRNAs) had significantly higher m⁶A levels than mRNAs or pseudogenes, but its biological implication is largely unknown and awaits future explorations (Meyer et al., 2012; Molinie et al., 2016). Interesting, a long non-coding RNA antisense to FOXM1 (FOXM1-AS) promotes the interaction of m⁶A demethylase ALKBH5 with FOXM1 nascent transcripts, which facilitates m⁶A demethylation of FOXM1 pre-mRNA at its 3′UTR. Demethylated FOXM1 pre-mRNAs have higher affinity with HuR, a RNA binding protein, which stabilizes FOXM1 to promote glioblastoma stem-like cells self-renewal and tumorigenesis (Zhang et al., 2017).

In summary, the knowledge of lncRNAs in neural circuitry assembly has been greatly expanded in recent years, whereas how lncRNAs exert roles in circuitry function in physiologic and pathologic conditions are much less known. Future studies would use modern genetic labeling, live-imaging, electrophysiology, behavioral and high-throughput means to explore how lncRNA-expressing neural cells are spatiotemporally participated in circuitry assembly and function, which can provide solid evidence that lncRNAs are essential fate and activity markers/determinants of neural cells.

AW and JW collected references and wrote the review. YL and YZ wrote the review.