Functional annotation of human long noncoding RNAs via molecular phenotyping

1Over 50,000 loci in the human genome transcribe long noncoding RNAs (lncRNAs), which are defined as transcripts at least 200 nucleotides (nt) long with low or no protein-coding potential. Although lncRNA genes outnumber protein-coding genes in mammalian genomes, they are comparatively less conserved, lowly expressed, and more cell-type-specific. However, the evolutionary conservation of lncRNA promoters and the structural motifs of lncRNAs suggest that lncRNAs are fundamental biological regulators. To date, only a few hundred human lncRNAs have been extensively characterized, revealing their roles in regulating transcription, translation, and chromatin state.

2Our recent FANTOM5 computational analysis showed that 19,175 (out of 27,919) human lncRNA loci are functionally implicated. Yet, genomic screens are necessary to comprehensively characterize each lncRNA. One common approach of gene knockdown followed by a cellular phenotype assay typically characterizes a small percentage of lncRNAs for a single observable phenotype. For example, a recent large-scale screening using CRISPR interference (CRISPRi) found that ∼3.7% of targeted lncRNA loci are essential for cell growth or viability in a cell-type-specific manner. In addition, CRISPR-Cas9 experiments targeting splice sites identified ∼2.1% of lncRNAs that affect growth of K562, and a CRISPR activation study revealed ∼0.11% lncRNAs to be important for drug resistance in melanoma. However, many of these studies target the genomic DNA, potentially perturbing the chromatin architecture, or focus on a single cellular assay, possibly missing other relevant functions and underlying molecular pathways.

3As a part of the FANTOM6 pilot project, we established an automated high-throughput cell culture platform to suppress 285 lncRNAs expressed in human primary dermal fibroblasts (HDFs) using antisense LNA-modified GapmeR antisense oligonucleotide (ASO) technology. We then quantified the effect of each knockdown on cell growth and morphology using real-time imaging, followed by Cap Analysis Gene Expression (CAGE) deep sequencing to reveal molecular pathways associated with each lncRNA. In contrast to cellular phenotyping, molecular phenotyping provides a detailed assessment of the response to a lncRNA knockdown at the molecular level, allowing biological pathways to be associated to lncRNAs even in the absence of an observable cellular phenotype. All data and analysis results are publicly available (see Data access), and results can be interactively explored using our in-house portal (https://fantom.gsc.riken.jp/zenbu/reports/#FANTOM6).

4Human dermal fibroblasts are nontransformed primary cells that are commonly used for investigating cellular reprogramming, wound healing, fibrosis, and cancer. Here, an unbiased selection of lncRNAs expressed in HDFs was performed to choose 285 lncRNAs for functional interrogation (Methods; Supplemental Table S1; Fig. 1A–C). Using RNA-seq profiling of fractionated RNA, we annotated the lncRNA subcellular localization as the chromatin-bound (35%), nucleus-soluble (27%), or cytoplasmic (38%) (Fig. 1D). We then designed a minimum of five non-overlapping antisense oligonucleotides against each lncRNA (Supplemental Methods; Supplemental Table S2; Fig. 1E,F) and transfected them individually using an automated cell culture platform to minimize experimental variability (Fig. 1G). The overall knockdown efficiencies across 2021 ASOs resulted in median value of 45.4%, and we could successfully knockdown 879 out of 2021 (43.5%) ASOs (>40% knockdown efficiency in at least two primer pairs or >60% in one primer pair) (Supplemental Table S2). ASOs targeting exons or introns were equally effective, and knockdown efficiencies were independent of the genomic class, expression level, and subcellular localization of the lncRNA (Supplemental Fig. S1A–D).

5To evaluate the effect of each lncRNA knockdown on cell growth and morphology, we imaged ASO-transfected HDFs in duplicate every 3 h for a total of 48 h (Supplemental Table S3) and estimated their growth rate based on cell confluence measurements (Fig. 2A,B). First, we observed across all ASOs that changes in cell growth and morphological parameters were significantly correlated with knockdown efficiency (Supplemental Fig. S1E). Considering both successful knockdown and significant growth inhibition (Student's two-sided t-test FDR ≤ 0.05), 246 out of 879 ASOs (∼28%) showed cellular phenotype (Fig. 2C; Supplemental Table S3).

6To assess globally whether the observed growth inhibition is lncRNA-specific, we used all 194 lncRNAs successfully targeted by at least two ASOs (Supplemental Fig. S2A) and found that ASOs targeting the same lncRNA were significantly more likely to have a concordant growth response than ASOs targeting different lncRNA (empirical P = 0.00037) (Supplemental Methods; Supplemental Fig. S2B). However, different ASOs targeting the same lncRNA typically showed different effects on growth, possibly due to variable knockdown efficiencies or differences in targeted lncRNA isoforms, as well as off-target effects. To reliably identify target-specific cellular phenotype, we applied conditional cutoffs based on the number of successful ASOs per each lncRNA (Supplemental Methods; Supplemental Fig. S2C) and identified 15/194 lncRNAs (7.7%) with growth phenotype (adjusted background <5%) (Supplemental Fig. S2D). We validated A1BG-AS1, which was previously implicated in cell growth, CATG00000089639, RP11-195F19.9, and ZNF213-AS1 by measuring the MKI67 proliferation protein marker upon knockdown with siRNAs and with selected ASOs (Fig. 2D; Supplemental Fig. S2E).

7In addition to cell growth, we also explored changes in cell morphology (Fig. 2E). Using a machine learning-assisted workflow (Methods), each cell was segmented and its morphological features representing various aspects of cell shapes and sizes were quantified (Fig. 2F; Supplemental Table S3; ). As an example, knockdown of 14/194 lncRNAs (7.2%) affected the spindle-like morphology of fibroblasts, as indicated by a consistent decrease in their observed eccentricity without reducing the cell number, suggesting possible cellular transformation toward epithelial-like states. Collectively, we observed 59/194 lncRNAs (∼30%) affecting cell growth and/or morphological parameters (Fig. 2G; Supplemental Table S3).

8Next, we selected 340 ASOs with high knockdown efficiencies (mostly >50%; median 71.4%) and sequenced 970 CAGE libraries to analyze 154 lncRNAs (Fig. 3A; Supplemental Table S4). To assess functional implications by individual ASOs, we performed differential gene expression, Motif Activity Response Analysis (MARA), and Gene Set Enrichment Analysis (GSEA) (Fig. 3B–F; ), and compared them with cellular phenotype.

9We globally observed significant knockdown-mediated transcriptomic changes (which generally correlated with KD efficiency) (Supplemental Fig. S3A), with ∼57% of ASOs showing at least 10 differentially expressed genes (FDR ≤ 0.05; abs[log2FC] > 0.5). For 84 divergent-antisense lncRNAs (targeted by 186 independent ASOs) (Supplemental Methods), we found their partner gene to be generally unchanged (median abs[log2FC] = ∼0.13), with an exception of two significantly down-regulated and three significantly up-regulated genes (FDR ≤ 0.05) (Supplemental Fig. S3B). We have, however, noticed a common response in a large number of ASOs (∼30%–35% of all responding ASOs), such as down-regulation of cell-cycle-related pathways, up-regulated stress genes and pathways, or altered cell metabolism and energetics (Supplemental Fig. S3C,D).

10When comparing knockdown-mediated molecular and cellular response, we found that transcription factor motifs that promote cell growth, including TFDP1, E2F1,2,3, and EP300, were positively correlated with the measured cell growth rate, whereas transcription factor motifs known to inhibit growth or induce apoptosis (e.g., PPARG, SREBPF, and STAT2,4,6) were negatively correlated (Fig. 3D; Supplemental Fig. S4A; Supplemental Table S6). Moreover, correlations of growth with GSEA pathways (Fig. 3F; Supplemental Fig. S4B; Supplemental Table S6) or with FANTOM5 coexpression clusters (Supplemental Fig. S4C) showed that cell growth and replication-related pathways were positively correlated with the measured growth rate, whereas those related to immunity, and cell stress and cell death were negatively correlated. We found that among 53 ASOs implicated in a growth-inhibition pathway based on the CAGE profiles, only 43% of them showed growth inhibition in the real-time imaging. This might suggest better sensitivity of transcriptomic profiling when detecting phenotypes as compared to live cell imaging methods, which are more prone to a delayed cellular response to the knockdown.

11Additionally, morphological changes were reflected in the molecular phenotype assessed by CAGE (Supplemental Fig. S4D). Cell radius and axis length were associated with GSEA categories related to actin arrangement and cilia, whereas cell compactness was negatively correlated with apoptosis. The extensive molecular phenotyping analysis also revealed pathways not explicitly associated with cell growth and cell morphology, such as transcription, translation, metabolism, development, and signaling (Fig. 3E).

12Next, to globally assess whether individual ASO knockdowns lead to lncRNA-specific effects, we scaled the expression change of each gene across the whole experiment and compared differentially expressed genes (Fig. 3B) of all possible ASO pairs targeting the same lncRNA target versus different lncRNAs (Supplemental Methods; Supplemental Table S5). We found that the concordance of the same target group was significantly greater than that of the different target group (comparing the Jaccard indices across 10,000 permutations) (Supplemental Fig. S5A), suggesting that ASO knockdowns are nonrandom and lead to more lncRNA specific effects than the nontargeting ASO pairs. Further, by requiring at least five common DEGs (FDR ≤ 0.05, abs[log2FC] > 0.5, abs[Z-score] > 1.645) and ASO-pairs significantly above the nontargeting ASO pairs background (P ≤ 0.05), we identified 16 ASO pairs, targeting 13 lncRNAs, exhibiting reproducible knockdown-mediated molecular responses in human dermal fibroblasts (Supplemental Fig. S5B). Corresponding GSEA pathways and MARA motifs of these 16 ASO pairs are shown in Supplemental Figure S5C.

13To evaluate whether the lncRNA-specific effects can be measured by other knockdown technologies, nine lncRNAs, with relatively mild growth phenotype, were subjected to siRNA knockdown. Measuring transcriptional response, we noted that higher concordance was observed for ASO modality alone (Supplemental Fig. S5D). The observed discrepancies in the transcriptional response between ASO- and siRNA-mediated knockdowns could be contributed by their mode of action and variable activities in different subcellular compartments. Next, a concordant response was found for (5/36) ASO-siRNA pairs targeting three lncRNAs (Supplemental Fig. S5E; Supplemental Table S5), enriched in the cytoplasm (MAPKAPK5-AS1), soluble nuclear fraction (LINC02454), and in the chromatin-bound fraction (A1BG-AS1). Although we cannot completely exclude the technical artifacts of each technology, concordant cellular response exhibited by using ASOs alone suggests that lncRNAs, in part, are essential regulatory elements in cells. Yet, our study generally warrants a careful assessment of specific findings from different knockdown technologies, including CRISPR-inhibition, and demonstrates a requirement of using multiple replicates in a given target per each modality.

14Extensive molecular and cellular phenotype data for each ASO knockdown can be explored using our portal https://fantom.gsc.riken.jp/zenbu/reports/#FANTOM6. As an example of an lncRNA associated with cell growth and morphology (Fig. 2G), we showcase ZNF213-AS1 (RP11-473M20.14). This lncRNA is highly conserved in placental mammals, moderately expressed (∼eight CAGE tags per million) in HDFs, and enriched in the chromatin-bound fraction. Four distinct ASOs (ASO_01, ASO_02, ASO_05, and ASO_06) strongly suppressed expression of ZNF213-AS1, whereas expression of the ZNF213 sense gene was not significantly affected in any of the knockdowns. The four ASOs caused varying degrees of cell growth inhibition (Fig. 4A). ASO_01 and ASO_06 showed a reduction in cell number, as well as an up-regulation of apoptosis and immune and defense pathways in GSEA, suggesting cell death. While cell growth inhibition observed for ASO_02 and ASO_05 was confirmed by MKI67 marker staining (Fig. 2D; Supplemental Table S7), the molecular phenotype revealed suppression of GSEA pathways related to cell growth, as well as to cell proliferation, motility, and extracellular structure organization (Fig. 4B). We also observed consistent down-regulation of motifs related to the observed cellular phenotype, for example, EGR1, EP300, SMAD1…7,9 (Fig. 4C).

15As cell motility pathways were affected by the knockdown, we tested whether ZNF213-AS1 could influence cell migration. Based on the wound-closure assay after transient cell growth inhibition (mitomycin C and serum starvation) (Supplemental Fig. S2F,G), we observed a substantial reduction of wound closure rate (∼40% over a 24-h period) in the ZNF213-AS1-depleted HDFs (Fig. 4D,E). The reduced wound healing rate should thus mainly reflect reduced cell motility, further confirming affected motility pathways predicted by the molecular phenotype.

16As these results indicated a potential role of ZNF213-AS1 in cell growth and migration, we used FANTOM CAT Recount 2 atlas, which incorporates The Cancer Genome Atlas (TCGA) data set, and found relatively higher expression of ZNF213-AS1 in acute myeloid leukemia (LAML) and in low-grade gliomas (LGG) as compared to other cancers (Supplemental Fig. S6A). In LAML, the highest expression levels were associated with mostly undifferentiated states, whereas in LGG, elevated expression levels were found in oligodendrogliomas, astrocytomas, and in IDH1 mutated tumors, suggesting that ZNF213-AS1 is involved in modulating differentiation and proliferation of tumors (Supplemental Fig. S6B–E). Further, univariate Cox proportional hazard analysis as well as Kaplan-Meier curves for LGG were significant and consistent with our findings (HR = 0.61, BH FDR = 0.0079). The same survival analysis on LAML showed a weak association with poor prognostic outcome, but the results were not significant (Supplemental Fig. S6F,G).

17Next, we investigated in detail RP11-398K22.12 (ENSG00000229852), where the knockdowns by two independent ASOs (ASO_03, ASO_05) successfully reduced the expression of the target lncRNA (67%–82% knockdown efficiency, respectively) and further down-regulated its neighboring genes, KCNQ5 and its divergent partner novel lncRNA CATG00000088862.1 (Fig. 5A). Although the two genomic loci occupy Chromosome 6 and are 650 kb away, Hi-C analysis (Supplemental Methods; Supplemental Fig. S7; Supplemental Table S8) showed that they are located within the same topologically associated domain (TAD) and spatially colocalized (Fig. 5B). Moreover, chromatin-enrichment and single molecule RNA-FISH of RP11-398K22.12 (Fig. 5C; Supplemental Table S9) suggested its highly localized cis-regulatory role.

18In FANTOM5, expression levels of RP11-398K22.12, KCNQ5, and CATG00000088862.1 were enriched in brain and nervous system samples, whereas GTEx showed their highly specific expression in the brain, particularly in the cerebellum and the cerebellar hemisphere (Fig. 5D). GTEx data also showed that expression of RP11-398K22.12 was highly correlated with the expression of KCNQ5 and CATG00000088862.1 across neuronal tissues (Fig. 5E,F), with the exception of cerebellum and the cerebellar hemisphere, potentially due to relatively lower levels of KCNQ5 and CATG00000088862.1, whereas levels of RP11-398K22.12 remained relatively higher. Additionally, we found an eQTL SNP (rs14526472) overlapping with RP11-398K22.12 and regulating expression of KCNQ5 in brain caudate (P = 4.2 × 10−6; normalized effect size −0.58). All these findings indicate that RP11-398K22.12 is implicated in the nervous system by maintaining the expression of KCNQ5 and CATG00000088862.1 in a cis-acting manner.

19This study systematically annotates lncRNAs through molecular and cellular phenotyping by selecting 285 lncRNAs from human dermal fibroblasts across a wide spectrum of expression, conservation levels and subcellular localization enrichments. Using ASO technology allowed observed phenotypes to be associated to the lncRNA transcripts, whereas, in contrast, CRISPR-based approaches may synchronically influence the transcription machinery at the site of the divergent promoter or affect regulatory elements of the targeted DNA site. Knockdown efficiencies obtained with ASOs were observed to be independent of lncRNA expression levels, subcellular localization, and of their genomic annotation, allowing us to apply the same knockdown technology to various classes of lncRNAs.

20We investigated the cis-regulation of nearby divergent promoters, which has been reported as one of the functional roles of lncRNA. However, in agreement with previous studies, we did not observe general patterns in the expression response of divergent promoters (Supplemental Fig. S3B). Recent studies suggest that transcription of lncRNA loci that do not overlap with other transcription units may influence RNA polymerase II occupancy on neighboring promoters and gene bodies. Thus, it is plausible that transcription of targeted lncRNA was maintained, despite suppression of mature or nascent transcripts using ASOs. This further suggests that the functional responses described in this study are due to interference of processed transcripts present either in the nucleus, the cytoplasm, or both. Although it is arguable that ASOs may interfere with general transcription by targeting the 5′-end of nascent transcripts and thus releasing RNA polymerase II, followed by exonuclease-mediated decay and transcription termination (aka “torpedo model”), most of the ASOs were designed across the entire length of the transcript. Since we did not broadly observe dysregulation in nearby genes, interference of transcription or splicing activity is less likely to occur.

21We observed a reduction in cell growth for ∼7.7% of our target lncRNA genes, which is in line with previous experiments using CRISPRi-pooled screening, which reported 5.9% (in iPS cells) of lncRNAs exhibiting a cell growth phenotype. Although these rates are much lower than for protein-coding genes, recurrent observations of cell growth phenotypes (including cell death) strongly suggest that a substantial fraction of lncRNAs play an essential role in cellular physiology and viability. Further, when applying image-based analysis, we found that lncRNAs affect cell morphologies (Fig. 2G), which has not been so far thoroughly explored.

22Several lncRNAs such as MALAT1, NEAT1, and FIRRE have been reported to orchestrate transcription, RNA processing, and gene expression but are not essential for mouse development or viability. These observations advocate for assays that can comprehensively profile the molecular changes inside perturbed cells. Therefore, in contrast to cell-based assays, functional elucidation via molecular phenotyping provides comprehensive information that cannot be captured by a single phenotypic assay. Herein, the number of overlapping differentially expressed genes between two ASOs of the same lncRNA targets indicated that 10.9% of lncRNAs exert a reproducible regulatory function in HDF.

23Although the features of selected lncRNAs are generally similar to those of other lncRNAs expressed in HDFs (Fig. 1B–D), the cell-type-specific nature of lncRNAs and the relatively small sampling size (119 lncRNAs with knockdown transcriptome profiles) used in our study may not fully represent the whole extent of lncRNA in other cell types. However, lncRNA targets that did not exhibit a molecular phenotype may be biologically relevant in other cell types or cell states. At the same time, our results showed that particular lncRNAs expressed broadly in other tissues (e.g., in the human brain) were functional in HDFs (such as RP11-398K22.12). Although the exact molecular mechanisms of RP11-398K22.12 are not yet fully understood, its potential role in HDFs suggests that lncRNAs may be functionally relevant across multiple tissues in spite of the cell-type-specific expression of lncRNAs.

24Further, we used siRNA technology to knockdown lncRNA targets as a method for independent validation. When comparing the transcriptomes perturbed by ASOs and siRNAs, concordance was observed only for three out of nine lncRNAs. This discrepancy is likely due to different modes of actions of the two technologies. Whereas ASOs invoke RNase H-mediated cleavage, primarily active in the nucleus, the siRNAs use the RNA-inducing silencing complex (RISC) mainly active in the cytoplasm. LncRNAs are known to function in specific subcellular compartments and their maturity, secondary structures, isoforms, and functions could be vastly different across compartments. Since the majority of functional lncRNAs are reported to be inside the nucleus, ASO-mediated knockdowns, which mainly target nuclear RNAs, are generally more suitable for functional screenings of our lncRNA (62% found in the nuclear compartment). Besides, the dynamics of secondary effects mediated by different levels of knockdown from different technologies are likely to be observed as discordance when considering the whole transcriptome, where this kind of discordance has been reported previously. In contrast, in the MKI67 assay, where only a single feature such as growth phenotype is assayed, siRNA knockdown revealed higher reproducibility with ASO knockdown. This suggested that the growth phenotype might be triggered by different specific pathways in ASO- and siRNA-knockdowns.

25Previous studies suggest that lncRNAs regulate gene expression in trans epigenetically, via direct or indirect interaction with regulators such as DNMT1 or by directly binding to DNA (triplex) or other RNA-binding proteins. Analysis of cellular localization by fractionation followed by RNA-seq and in situ hybridization can indicate whether a given lncRNA may act in trans by quantifying its abundance in the nuclear soluble fraction as compared to cytoplasm. Although most lncRNAs in the nuclear soluble fraction may affect pathways associated with chromatin modification, additional experiments to globally understand their interaction partners will elucidate the molecular mechanism behind trans-acting lncRNAs.

26In summary, our study highlights the functional importance of lncRNAs regardless of their expression, localization, and conservation levels. Molecular phenotyping is a powerful and generally more sensitive to knockdown-mediated changes platform to reveal the functional relevance of lncRNAs that cannot be observed based on the cellular phenotypes alone. With additional molecular profiling techniques, such as RNA duplex maps in living cells to decode common structural motifs, and Oxford Nanopore Technology (ONT) to annotate the full-length variant isoforms of lncRNAs, the structure-to-functional relationship of lncRNAs may be elucidated further in the future.

27The gene models used in this study were primarily based on the FANTOM CAGE-associated transcriptome (CAT) at permissive level as defined previously. From this merged assembly, there were ∼2000 lncRNAs robustly expressed in HDFs (TPM ≥ 1). However, we selected lncRNA knockdown targets in an unbiased manner to broadly cover various types of lncRNAs (TPM ≥ 0.2). Briefly, we first identified a list of the lncRNA genes expressed in HDFs, with RNA-seq expression at least 0.5 fragments per kilobase per million and CAGE expression at least 1 tag per million. Then, we manually inspected each lncRNA locus in the ZENBU genome browser for (1) its independence from neighboring genes on the same strand (if any), (2) support from RNA-seq (for exons and splicing junctions) and CAGE data (for TSSs) of its transcript models, and (3) support from histone marks at TSSs for transcription initiation (H3K27ac) and along the gene body for elongation (H3K36me3), from the Roadmap Epigenomics Consortium. A representative transcript model, which best represents the RNA-seq signal, was manually chosen from each locus for design of antisense oligonucleotides. In total, 285 lncRNA loci were chosen for ASO suppression. Additional controls (NEAT1, protein coding genes) (Supplemental Table S1) were added, including MALAT1 as an experimental control. For details, please refer to the Supplemental Methods.

28ASOs were designed as RNase H-recruiting locked nucleic acid (LNA) phosphorothioate gapmers with a central DNA gap flanked by 2–4 LNA nucleotides at the 5′ and 3′ ends of the ASOs. For details, please refer to the Supplemental Methods.

29Robotic automation (Hamilton) was established to provide a stable environment and accurate procedural timing control for cell culturing and transfection. In brief, trypsin-EDTA detachment, cell number and viability quantification, cell seeding, transfection, and cell harvesting were performed with automation. All transfections were divided into 28 runs on a weekly basis. ASO transfection was performed with duplication. In each run, there were 16 independent transfections with ASO negative control A (NC_A, Exiqon) and 16 wells transfected with an ASO targeting MALAT-1 (Exiqon).

30The HDF cells were seeded in 12-well plates with 80,000 cells in each well 24 h prior to the transfection. A final concentration of 20 nM ASO and 2 µL Lipofectamine RNAiMAX (Thermo Fisher Scientific) were mixed in 200 µL Opti-MEM (Thermo Fisher Scientific). The mixture was incubated at room temperature for 5 min and added to the cells, which were maintained in 1 mL complete medium. The cells were harvested 48 h posttransfection by adding 200 µL RLT buffer from the RNeasy 96 kit (Qiagen) after PBS washing. The harvested lysates were kept at −80°C. RNA was extracted from the lysate for real-time quantitative RT-PCR (Supplemental Methods).

31The HDF cells were transfected manually in 96-well plates to facilitate high-throughput real-time imaging. The cells were seeded 24 h before transfection at a density of 5200 cells per well. A final concentration of 20 nM ASO and 2 µL Lipofectamine RNAiMAX (Thermo Fisher Scientific) were mixed in 200 µL Opti-MEM (Thermo Fisher Scientific). After incubating at room temperature for 5 min, 18 µL of the transfection mix was added to 90 µL complete medium in each well. The ASOs were divided into 14 runs and transfected in duplicate. Each plate accommodated six wells of NC_A control, two wells of MALAT1 ASO control, and two wells of mock-transfection (Lipofectamine alone) control.

32Phase-contrast images of transfected cells were captured every 3 h for 2 d with three fields per well by the Incucyte live-cell imaging system (Essen Bioscience). The confluence in each field was analyzed by the Incucyte software. The mean confluence of each well was taken along the timeline until the mean confluence of the NC_A control in the same plate reached 90%. The growth rate in each well was calculated as the slope of a linear regression. A normalized growth rate of each replicate was calculated as the growth rate divided by the mean growth rate of the six NC_A controls from the same plate. Negative growth rate was derived when cells shrink and/or detach. As these rates of cell depletion could not be normalized by the rate of growth, negative values were maintained to indicate severe growth inhibition. Student's t-test was performed between the growth rate of the duplicated samples and the six NC_A controls, assuming equal variance.

33For each transfection, a representative phase-contrast image at a single time point was exported from the Incucyte time-series. These raw images were first transformed to probability maps of cells by pixel classification using ilastik (1.3.2). The trained model was then applied to all images where the predicted probability maps of cells (grayscale, 16 bits tiff format) were subsequently used for morphology quantification in CellProfiler (3.1.5). For details, please refer to the Supplemental Methods.

34For the selected four lncRNA targets showing >25% growth inhibition, we used two siRNAs and two ASOs with independent sequences. The transfected cells were fixed by adding prechilled 70% ethanol and incubated at −20°C. The cells were washed with FACS buffer (2% FBS in PBS, 0.05% NaN3) twice. FITC-conjugated MKI67 (20Raj1, eBioscience) was applied to the cells and subjected to flow cytometric analysis. Knockdown efficiency by siRNA was determined by real-time quantitative RT-PCR using the same three primer pairs as for ASO knockdown efficiency. For details, please refer to the Supplemental Methods.

35The HDF cells were transfected with 20 nM ASO as described earlier in 12-well plates. The cells were replated at 24 h posttransfection into a 96-well ImageLock plate (Essen BioScience) at a density of 20,000 cells per well. At 24 h after seeding, cells form a spatially uniform monolayer with 95%–100% cell confluence. The cells were incubated with 5 µg/mL mitomycin C for 2 h to inhibit cell division. Then, medium was refreshed and a uniform scratch was created in each well by the WoundMaker (Essen BioScience). The closure of the wound was monitored by Incucyte live-cell imaging system (Essen Bioscience) every 2 h for 24 h. The RNA was harvested after the assay for real-time quantitative RT-PCR. For details, please refer to the Supplemental Methods.

36Four micrograms of purified RNA were used to generate libraries according to the nAnT-iCAGE protocol. For details, please refer to the Supplemental Methods.

37Hi-C libraries were prepared essentially as described previously with minor changes to improve the DNA yield of Hi-C products. For details, please refer to the Supplemental Methods.

38All raw and processed sequencing data generated in this study have been submitted to the DNA Data Bank of Japan (DDBJ; https://www.ddbj.nig.ac.jp/) under accession numbers DRA008311, DRA008312, DRA008436, and DRA008511 or can be accessed through the FANTOM6 project portal https://fantom.gsc.riken.jp/6/datafiles. The analysis results can be downloaded from https://fantom.gsc.riken.jp/6/suppl/Ramilowski_et_al_2020/data/ and interactively explored using our in-house portal https://fantom.gsc.riken.jp/zenbu/reports/#FANTOM6.

39The authors declare no competing interests.