snoHUB2

Data processing, coexpression analysis, and functional prediction workflow used by snoHUB2.

Overview

snoHUB2 generates predicted functional associations for human small nucleolar RNAs (snoRNAs) by combining gene expression data from ARCHS4 [1] with biological gene sets from Enrichr [2, 3, 4]. The computational workflow consists of data extraction, sample and feature filtering, expression transformation, quantile normalization, pairwise snoRNA to protein-coding gene correlation, averging of snoRNA to protein-coding genes by term, and ranking of these gene-set terms for each retained snoRNA.

The primary functional association score is the mean Pearson correlation between a snoRNA and the protein-coding genes assigned to a biological term. The resulting scores are used to order terms within individual gene-set libraries and across all included libraries.

ARCHS4 expression data and gene annotation

Human gene-level expression data was obtained from ARCHS4 [1]. The expression matrix contained genes as rows and RNA sequencing samples as columns. Sample metadata distributed within ARCHS4 was used to distinguish bulk RNA sequencing samples from samples predicted to represent single-cell experiments. Samples were classified as bulk RNA sequencing when their single-cell probability was below 0.05.

Gene annotations were used to split the matrix according to gene biotype. Rows annotated as snoRNAs were assigned to the snoRNA expression matrix. Rows annotated as protein-coding genes were assigned to the protein-coding expression matrix. Both matrices retained the same sample columns but contained different gene rows.

Gene symbols, Ensembl gene identifiers, biotype assignments, and selected sample indices were preserved. These identifiers were retained in a consistent order across the raw-expression, normalized-expression, correlation, pathway-ranking, and overlapping-gene files.

Bulk-sample selection and expression filtering

Sample selection was performed after excluding samples that did not satisfy the bulk RNA sequencing criterion (Single-cell probability > 0.05). For every retained bulk sample, the mean snoRNA expression was calculated across all annotated snoRNA rows. The median of the resulting sample-level mean snoRNA expression values was used as the sample-selection threshold. Samples with mean snoRNA expression above this median were included in the eligible high-snoRNA-expression sample pool.

A random subset of 10,000 samples was selected without replacement from the eligible sample pool. A fixed random seed was used so that the same sample indices could be reproduced in repeated executions of the workflow. The selected sample indices were used for both the snoRNA and protein-coding matrices.

Feature filtering was performed within the selected sample subset. A snoRNA or protein-coding gene was retained only when its raw expression value was greater than zero in at least 50 selected samples. This filter was applied separately to the snoRNA and protein-coding matrices.

ARCHS4 labels genes with the Ensembl biotype, so its protein-coding rows include entries that are not separate named genes. A protein-coding gene filter therefore keeps one row per named gene, leaving 19,299 of the 22,805 retained rows. It removes readthrough fusions, predicted genes without an HGNC symbol, genes that HGNC classifies as pseudogenes or non-coding RNAs, and extra copies on alternate scaffolds. Correlations are calculated for all rows, and the filter is applied before genes are ranked and before gene-set terms are scored, so every downstream result uses the same genes.

The original nonzero-expression status of each retained feature in every selected sample was preserved as a Boolean expression mask. These masks are later used to define the samples included in each pairwise snoRNA-gene correlation calculation.

Expression transformation and normalization

Retained expression values were transformed using a base-2 logarithm after adding one to each value. Adding one allowed zero-valued observations to remain defined after the logarithmic transformation. The same transformation was applied to both the snoRNA and protein-coding expression matrices.

Quantile normalization was applied after the logarithmic transformation. The snoRNA and protein-coding matrices were combined and normalized together as one matrix, so both share one expression scale. Values in every selected sample were sorted by rank. The mean value for each rank was calculated across samples. Each observation was then replaced by the mean value corresponding to its rank, and tied values received the average of their ranks' means. Values were then returned to their original gene ordering and separated back into the two matrices.

The resulting sample columns had a common empirical expression distribution. The normalized snoRNA matrix, normalized protein-coding matrix, expression masks, gene identifiers, gene symbols, and selected sample indices were stored in an HDF5 file for downstream analysis.

Pairwise snoRNA to protein-coding gene coexpression

Pearson correlation coefficients were calculated between every retained snoRNA and every retained protein-coding gene. Each calculation was restricted to samples in which both features had nonzero raw expression.

Samples in which either member of the pair had zero raw expression were excluded from the calculation for that specific pair. The overlapping sample set therefore differed between snoRNA-gene pairs according to their expression patterns.

For each valid pair, Pearson correlation was calculated using the normalized expression values from the shared nonzero samples.

The Pearson correlation coefficient, corresponding p-value, and number of shared nonzero samples were recorded for every valid pair. The resulting HDF5 correlation file contained protein-coding genes as rows and snoRNAs as columns. After the protein-coding gene filter leaves one row per gene symbol, the 500 most positively correlated and 500 most negatively correlated genes for each snoRNA were written to top_correlated_genes_snoRNA.csv, labeled by direction.

Gene-set libraries and term matching

Functional terms were obtained from 6 gene set libraries from Enrichr [2, 3, 4]. All 6 were scored, and the reports and set analysis show the first 5:

  • GWAS Catalog 2025
  • MGI Mammalian Phenotype Level 4 2024
  • KEGG 2026
  • WikiPathways 2024 Human
  • Gene Ontology Biological Process 2025
  • ChEA 2022, used only as a feature of the similarity map (see Similarity views). Its terms are transcription factor target sets, so the reports and set analysis show transcription factors from ChEA3 and ChEA-KG instead (see Transcription factor analysis) rather than a second ChEA result.

Each library consisted of named terms and the protein-coding gene symbols assigned to those terms. Gene symbols were standardized between the Enrichr libraries and the correlation matrix before matching. For every term, the original gene list was intersected with the protein-coding genes represented in the snoRNA-gene correlation matrix.

The original gene-set size and the number of matched protein-coding genes were recorded separately. The original gene-set size represents the complete number of genes supplied for the term by the source library. The overlap size represents the number of genes from that term that were present in the processed protein-coding correlation matrix.

Terms with fewer than 5 overlapping protein-coding genes were excluded from functional scoring. This minimum-overlap criterion was applied independently to every term in every library.

Functional association scoring and ranking

For each retained snoRNA and each eligible gene-set term, correlations for the matched protein-coding genes were extracted from the correlation matrix.

The snoRNA-term association score was calculated as the unweighted arithmetic mean of the available snoRNA-gene Pearson correlations for those matched genes.

Every matched gene-level correlation contributed equally to the term score. No gene-specific weighting, enrichment p-value, pathway-size correction, or variance-based adjustment was applied to the mean.

The pathway score table contained the snoRNA symbol, snoRNA Ensembl identifier, library name, term name, overlap size, original gene-set size, matched matrix-row count, overlapping gene symbols, and mean Pearson correlation.

Terms were ranked separately for each snoRNA and gene-set library. Descending sorting produced the highest-correlated ranking. Ascending sorting produced the lowest-correlated ranking. Direction and rank were stored with every selected term in the final ranking file.

The current web report displays the 10 highest-correlated terms across all included libraries and the 10 highest-correlated terms within each individual library.

The overview table also reports general annotation for the queried snoRNA. Chromosome location and length are retrieved from the Ensembl REST API [5]. The publication count is the PubMed [6] hit count for the snoRNA's parent symbol: a numbered copy suffix is dropped, so SNORD115-12 is counted as SNORD115 and SNORD116-1 as SNORD116. PubMed splits hyphenated symbols into words, so this search also finds papers that name only one copy. Letter suffixes such as SNORD3A mark separate genes and are kept. Counts are fetched from PubMed when a snoRNA is opened and re-fetched once they are more than a week old, so they follow new publications. Expression across tissues, cell types, and cell lines is described in its own section.

Overall term selection and overlapping-gene reports

Overall highest-correlated terms were selected by collecting the highest-direction term rankings across all included libraries for a queried snoRNA. The collected terms were then sorted by mean Pearson correlation in descending order. Duplicate library-term combinations were removed, and the first 10 terms were retained.

Library-specific displays were generated independently. For each library, terms in the highest direction were sorted by mean Pearson correlation in descending order. Duplicate term names were removed, and up to 10 terms were retained.

Gene-level results were generated for each selected overall term. Protein-coding genes assigned to the term were matched to the correlation matrix, and their pairwise correlations with the selected snoRNA were retrieved.

The protein-coding gene filter leaves one row per gene symbol, so each gene in a term has one correlation with the selected snoRNA. Correlation distance was calculated by subtracting the Pearson correlation from one.

Genes were ordered by ascending correlation distance. This ordering is equivalent to sorting by descending Pearson correlation. Duplicate gene symbols were removed after gene-level aggregation, and up to 25 genes were retained for each term.

The downloadable overlapping-gene files contain gene rank, gene symbol, Pearson correlation, and correlation distance. The report lists the top 10 of these genes for each library-specific term, in one table per library. The overall and library-specific term tables contain term rank, library or term name, mean Pearson correlation, overlap size, and original gene-set size.

Tables and figures are are built in the browser. Table downloads are written as UTF-8 CSV files. Figure downloads are rendered as PNG files with a white background.

Sequence and predicted structure

Each snoRNA's sequence was retrieved from the Ensembl REST API [5] by Ensembl gene ID and written as RNA. The sequence was matched to RNAcentral [10] by exact sequence, and the stored R2DT [9] diagram for that entry was used when one was available. R2DT draws a sequence on a curated family template, which for snoRNAs comes from Rfam [11]. Sequences with no stored diagram were folded with RNAfold from the ViennaRNA Package [12], using default parameters, to give the minimum free energy structure.

These diagrams are predictions. The method is stated under each diagram. RNAfold models the RNA alone, while snoRNAs function bound to snoRNP proteins. Rfam templates for C/D box snoRNAs can contain few base pairs, so an open ring in the diagram does not show that those nucleotides are unpaired. Sequences follow the current Ensembl annotation, recorded in the downloadable table, and can differ from Ensembl release 107, which was used for gene filtering.

Transcription factor analysis

The same up to 500 positively correlated protein-coding genes per snoRNA that are submitted to Perturb-Seqr were submitted to the ChEA3 [22] API. ChEA3 ranks transcription factors by the overlap of the genes with their targets in each of its libraries, built from ChIP-seq experiments, co-expression, co-occurrence, and the literature. The report keeps the top 25 transcription factors of its integrated mean-rank ranking, which averages a factor's rank across these libraries, so a lower score is a stronger result. For each factor, the report lists the libraries that ranked it and how many of the genes are among its targets.

ChEA-KG [23] builds a transcription factor regulatory network from ChEA3 and draws, for a gene set, the network among the top 10 factors of this same ranking. The Open the network in ChEA-KG button sends the genes to ChEA-KG when it is clicked; the network is drawn on ChEA-KG, not on snoHUB2. The ChEA3 results were computed once for all snoRNAs and are served with the reports.

Perturb-Seqr analysis

Positively correlated protein-coding genes for each snoRNA were submitted to the Perturb-Seqr [21] API. Up to 500 positively correlated genes were used when at least that many were available; otherwise, all available positively correlated genes were submitted.

Each supported perturbation library was queried separately so that a snoRNA could retain results from libraries that returned matches even when other libraries returned no results. Unrecognized submitted identifiers were retained in the query gene list rather than removed before submission.

The report groups results by perturbation type and library, then separates up-regulated and down-regulated perturbations. Each library panel reports query status, query-gene count, total enrichment-result count, perturbation rank, nominal and adjusted p-values, odds ratio, overlap, and significance status.

Perturb-Seqr returns signatures with an adjusted p-value of at most 0.05; libraries with more than 1,000 such signatures are retrieved in pages, so every significant signature is kept. The per-library charts show 10 up-regulated and 10 down-regulated signatures per snoRNA-library query. When a direction has fewer than 10 significant signatures, it was queried again without the cutoff and the next-best signatures by p-value were added. These are drawn in grey, did not reach an adjusted p-value of 0.05, and are used only in the charts: they do not enter the consensus tables, abstracts, or generated papers. Libraries with no significant signature for a snoRNA have no chart.

Four additional tables combine the libraries: drugs and chemicals (9 libraries) and gene perturbations (7 libraries), each split into perturbations that up-regulate and down-regulate the submitted genes. A perturbation is listed once per table and ranked by consistency across resources: the number of libraries with at least one significant signature in that direction, then its best adjusted p-value. P-values are not combined across libraries: the best adjusted p-value is the smallest Perturb-Seqr adjusted p-value among the perturbation's signatures, reported with its library, and it only breaks ties between rows supported by the same number of libraries. Several signatures from one library, such as different doses, time points, or cell lines, count as one library, because they are not independent evidence. Drug names are matched across libraries ignoring case and salt forms (for example, "ralimetinib dimesylate" and "ralimetinib"). Gene perturbations are matched by uppercase gene symbol, which maps most mouse genes to their human ortholog, and the knockdown, knockout, or overexpression modes are listed; a table flags genes where loss- and gain-of-function perturbations point the same way. Conflicts are reported and never change the ranking. Each row counts the libraries significant only in the opposite direction and the libraries significant in both directions, for example at different doses or in different cell lines. A knockdown or knockout and an overexpression that point in opposite directions agree, so that pair is not counted as a conflict; signatures with no stated mode are compared with all others. Each row can be expanded to show, per library and direction, the rank of the best signature among that library's significant signatures (the same rank as the per-library charts), its adjusted p-value, and the number of signatures. The tables keep every perturbation supported by two or more libraries and the 25 most significant perturbations found in one library.

Report abstracts

The abstract at the top of each report is built in Pipeline/09_snoRNA_description.py as one paragraph in a fixed order: approach, the snoRNA profiled, positively correlated themes, inversely correlated themes, the top co-expressed protein-coding genes and nearest snoRNAs, Perturb-Seqr evidence, and a statement that the predictions need experimental validation. The same abstract is used in the downloadable report papers.

Only the two theme sentences are written by a large language model, because grouping many terms into a few themes needs interpretation. The model receives only the overall highest-correlated terms and the highest- and lowest-correlated terms in each library for that snoRNA. It is told to describe co-expression only and to add no outside facts. An answer that uses wording implying function (for example "involved in" or "suggests"), or that falls outside 35 to 130 words, is rejected and requested again.

The other sentences are filled in from the data, not by the model. The snoRNA class comes from the symbol prefix (SNORD, SNORA, or SCARNA). The top three positively correlated protein-coding genes and the two nearest snoRNAs come from the report tables. The Perturb-Seqr sentence names the most significant drug signature and the most significant gene perturbation signature exactly as they appear in the Perturb-Seqr table, with their direction and adjusted p-value.

snoRNA similarity views

The snoRNA report includes an interactive similarity map. It is not part of the overview, term tables, or Perturb-Seqr panels. Those sections do not show UMAP or t-SNE coordinates.

Each snoRNA is represented by its mean Pearson correlation with gene-set terms from the ranked term tables, which hold the 50 highest and 50 lowest terms for each snoRNA in each of the 6 gene-set libraries, including ChEA 2022, which is not shown in the reports. Terms ranked for fewer than 25% of snoRNAs are dropped, leaving 741 terms, all of which are used. A term that was not ranked for a given snoRNA is filled with the mean score of that term rather than with zero, so an unranked term is not read as an absent correlation.

The matrix is standardized and reduced with principal component analysis in scikit-learn [16] (up to 50 components). UMAP [14] and t-SNE [15] are then fit on those components with cosine distance and a fixed random seed. The coordinates and the ten nearest snoRNAs for each snoRNA are precomputed by the analysis pipeline, so the web server only reads a table.

The neighbor table is ranked by cosine similarity of the principal components, not by distance on the plot, so it is identical in the UMAP and t-SNE views. Both plots show similar co-expression profiles, not expression-space distance. Closeness does not establish shared function, and it can also reflect shared host genes, expression level, or how RNA-seq libraries were prepared. The static gene-space UMAP figures written by the analysis pipeline are not displayed in the web report.

Set analysis

The Set analysis page predicts shared functions for a set of 2 to 100 snoRNAs entered by the user. It finds the protein-coding genes most related to the set the way ARCHS4 and Geneshot predict genes for a gene set [1, 24]: each gene's Pearson correlation with the snoRNAs, from the same correlation matrix and gene filter as the reports, is averaged across the set. A missing correlation, where the gene and snoRNA share no samples in which both are detected, counts as 0. The 200 genes with the highest positive mean are the related genes. As in lncHUB2, the 200 genes with the lowest negative mean, the negatively correlated genes, can be analyzed instead. For each gene, the page also shows how many snoRNAs in the set have it among their own top 500 correlated genes of the same sign, and a heatmap shows the correlations of the top 30 genes with each snoRNA. Correlations are stored at half precision, which changes them by less than 0.001.

The related genes are tested against the 5 gene-set libraries shown in the reports with a hypergeometric test, using the protein-coding genes kept by the gene filter as the background. P-values are adjusted with the Benjamini-Hochberg procedure, and a term is significant at an adjusted p-value below 0.05.

The Open in Enrichr button sends the related genes to Enrichr [2, 3, 4] through its public API and opens the results, where they are tested against all of Enrichr's libraries. Only the genes are sent, and only when the button is clicked.

The Open in Perturb-Seqr button sends the related genes to Perturb-Seqr [21], where they can be searched against drug and gene perturbation signatures. The genes are sent only when the button is clicked, the snoRNA names are not sent, and the search runs on Perturb-Seqr, not on snoHUB2.

The Open in ChEA-KG button sends the related genes to ChEA-KG [23], which draws the regulatory network among the top transcription factors. As with Perturb-Seqr, the genes are sent only when the button is clicked.

The related genes are also analyzed with Perturb-Seqr and ChEA3 [22] when a set is submitted. Perturb-Seqr results use the same 9 drug and 7 gene perturbation libraries as the snoRNA reports, keep signatures with an adjusted p-value up to 0.05, and rank perturbations by how many libraries support them, then by best adjusted p-value. When a table has fewer than 10 significant perturbations, the strongest non-significant ones fill it to 10 and are marked as such. Transcription factors are ranked as in the reports and in ChEA-KG, by ChEA3's integrated mean rank across libraries, and the top 25 are shown.

Set analysis results are predictions from co-expression. They do not show that the snoRNAs act on these functions. A set of near-identical copies, such as the SNORD115 example, can share genes because the copies are hard to tell apart in RNA-seq. Submitted lists are not stored.

Genome-wide top correlated genes

The snoRNA report includes a table of the protein-coding genes with the largest positive Pearson correlations to the queried snoRNA, and a second view of the genes with the largest negative correlations. This ranking uses the genome-wide correlation matrix rather than the genes assigned to an Enrichr term, after the protein-coding gene filter leaves one row per gene symbol. The positive list excludes correlations of zero or below, and the negative list excludes correlations of zero or above. Each list is ranked separately, starting at 1.

The website reads top_correlated_genes_snoRNA.csv and lets the user display the top 10, 25, 50, 100, 250, or 500 genes when that many were stored for the snoRNA. The default display is 25 genes. The table is numbered as Table 3 in the report, after the most similar snoRNAs and before the overall highest-correlated terms.

Expression across tissues, cell types, and cell lines

The report ranks the normal tissues, primary cell types, and cell lines in which each snoRNA has the highest average expression in ARCHS4 [1]. Groups come from a curated list rather than from free-text GEO labels: about 2,800 human cell lines from Cellosaurus [17] (lines in DepMap or ENCODE, or cited at least 10 times) with CLO [18] identifiers, 50 normal organs with UBERON [19] identifiers, and 48 primary cell types with Cell Ontology [20] identifiers. Each entry carries the ontology's synonyms, plus synonyms found in curated ARCHS4 sample metadata and reviewed against expression data.

Each bulk sample (single-cell probability below 0.05) is assigned to at most one group from its GEO characteristics_ch1 fields, source name, and title. A named cell line takes priority, then a primary cell type, then a tissue. Derived cell lines count toward their parent line, and finer anatomy counts toward its organ, so hippocampus counts as brain. Tissue and cell-type samples from diseased donors are dropped, as are organoids. Cells differentiated from induced pluripotent or embryonic stem cells form separate groups labeled "(iPSC/ESC-derived)".

Expression is TMM-normalized counts per million (CPM). Dividing by library size adjusts each sample for how deeply it was sequenced, and the trimmed mean of M-values (TMM) factor [13] corrects for a few abundant transcripts taking up the library. Factors were estimated from all genes against the average of the samples used. Samples with a factor outside 0.1 to 10 are left out, because a single inflated sample could otherwise decide a group's mean. A group's value is the mean CPM over its samples, shown with the standard deviation (SD) across those samples. The mean is used instead of the median because most samples have no reads for any single snoRNA, so the median would be zero for most snoRNAs. Samples with no snoRNA reads at all are left out.

Library preparation strongly affects snoRNA reads. Libraries made by removing ribosomal RNA keep about 80 times more snoRNA reads per million than libraries made by selecting poly(A) RNA, because most snoRNAs lack a poly(A) tail. A group's average therefore also reflects how its studies prepared their libraries, and the SD is often larger than the mean because a few samples carry most of the reads.

A group is kept when it has at least 50 GEO samples (GSMs) from at least 3 GEO studies (GSEs). Its value uses up to 100 samples, taken in turn from each study so no single study dominates. With each value, the report gives how many of the samples had at least one read (for example, 20 of 100 GSMs) and how many of the studies had at least one sample with a read (for example, 11 of 41 GSEs). Values supported by few samples or a single study should be read with caution. Rankings near the top of a list can change between independent sets of studies, so the list is a guide to where a snoRNA is most expressed rather than an exact order. The report shows the 15 highest groups per category as Table 1, linked to their ontology terms. If the precomputed table is absent, the report lists names from the ARCHS4 [1] website atlas, then GTEx [7] median TPM values greater than zero and Harmonizome [8] expression associations with positive standardized scores, without values.

Software and data storage

Data extraction and numerical processing were implemented in Python. HDF5 files were read and written using h5py. Matrix operations were performed using NumPy. Tabular processing, identifier matching, grouping, ranking, and CSV generation were performed using pandas.

Pearson statistics were calculated using SciPy. Similarity coordinates are calculated in the pipeline with scikit-learn and umap-learn.

Large expression and correlation matrices were stored in HDF5 format to permit incremental access without loading the complete source dataset into memory. Final snoRNA-term rankings and ranked overlapping-gene results were stored as CSV files.

References

[1] Lachmann A, et al. Massive mining of publicly available RNA-seq data from human and mouse. Nature Communications. 2018;9(1):1366. PMID: 29636450

[2] Xie Z, et al. Gene Set Knowledge Discovery with Enrichr. Current Protocols. 2021;1(3):e90. PMID: 33780170

[3] Chen EY, et al. Enrichr: interactive and collaborative HTML5 gene list enrichment analysis tool. BMC Bioinformatics. 2013;14:128. PMID: 23586463

[4] Kuleshov MV, et al. Enrichr: a comprehensive gene set enrichment analysis web server 2016 update. Nucleic Acids Research. 2016;44(W1):W90-W97. PMID: 27141961

[5] Dyer SC, et al. Ensembl 2025. Nucleic Acids Research. 2025;53(D1):D948-D957. PMID: 39656687

[6] Sayers EW, et al. Database resources of the National Center for Biotechnology Information in 2026. Nucleic Acids Research. 2026;54(D1):D20-D27. PMID: 41385079

[7] GTEx Consortium. The GTEx Consortium atlas of genetic regulatory effects across human tissues. Science. 2020;369(6509):1318-1330. PMID: 32913098

[8] Rouillard AD, et al. The harmonizome: a collection of processed datasets gathered to serve and mine knowledge about genes and proteins. Database (Oxford). 2016;2016:baw100. PMID: 27374120

[9] McCann H, et al. R2DT: a comprehensive platform for visualizing RNA secondary structure. Nucleic Acids Research. 2025;53(4). PMID: 39921562

[10] RNAcentral Consortium. RNAcentral 2021: secondary structure integration, improved sequence search and new member databases. Nucleic Acids Research. 2021;49(D1):D212-D220. PMID: 33106848

[11] Kalvari I, et al. Rfam 14: expanded coverage of metagenomic, viral and microRNA families. Nucleic Acids Research. 2021;49(D1):D192-D200. PMID: 33211869

[12] Lorenz R, et al. ViennaRNA Package 2.0. Algorithms for Molecular Biology. 2011;6:26. PMID: 22115189

[13] Robinson MD, Oshlack A. A scaling normalization method for differential expression analysis of RNA-seq data. Genome Biology. 2010;11(3):R25. PMID: 20196867

[14] McInnes L, Healy J, Saul N, Großberger L. UMAP: Uniform Manifold Approximation and Projection. Journal of Open Source Software. 2018;3(29):861. DOI: 10.21105/joss.00861

[15] van der Maaten L, Hinton G. Visualizing Data using t-SNE. Journal of Machine Learning Research. 2008;9:2579-2605. Link

[16] Pedregosa F, et al. Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research. 2011;12:2825-2830. Link

[17] Bairoch A. The Cellosaurus, a Cell-Line Knowledge Resource. Journal of Biomolecular Techniques. 2018;29(2):25-38. PMID: 29805321

[18] Sarntivijai S, et al. CLO: The cell line ontology. Journal of Biomedical Semantics. 2014;5:37. PMID: 25852852

[19] Mungall CJ, et al. Uberon, an integrative multi-species anatomy ontology. Genome Biology. 2012;13(1):R5. PMID: 22293552

[20] Diehl AD, et al. The Cell Ontology 2016: enhanced content, modularization, and ontology interoperability. Journal of Biomedical Semantics. 2016;7(1):44. PMID: 27377652

[21] Taub LD, Gardner JK, Diamant I, Clarke DJB, Ma'ayan A. Perturb-Seqr: Signature Search Engine Integrating Connectivity Mapping Resources. bioRxiv. 2026. DOI: 10.64898/2026.10.06.757137 Tool: https://perturbseqr.maayanlab.cloud/

[22] Keenan AB, et al. ChEA3: transcription factor enrichment analysis by orthogonal omics integration. Nucleic Acids Research. 2019;47(W1):W212-W224. PMID: 31114921

[23] Byrd AI, et al. ChEA-KG and ChEA-KG-TS: a network-based transcription factor enrichment analysis tool with an accompanying time-series workflow. Nucleic Acids Research. 2026;54(W1):W343-W358. PMID: 42163770

[24] Lachmann A, et al. Geneshot: search engine for ranking genes from arbitrary text queries. Nucleic Acids Research. 2019;47(W1):W571-W577. PMID: 31114885