CoChat

Exploiting single-cell expression to characterize co-expression replicability

Open access

The DOI matches its Crossref record; 66 of the 67 references it cites were found in a registry or an index, and 1 are not checked yet.

Abstract

BACKGROUND: Co-expression networks have been a useful tool for functional genomics, providing important clues about the cellular and biochemical mechanisms that are active in normal and disease processes. However, co-expression analysis is often treated as a black box with results being hard to trace to their basis in the data. Here, we use both published and novel single-cell RNA sequencing (RNA-seq) data to understand fundamental drivers of gene-gene connectivity and replicability in co-expression networks. RESULTS: We perform the first major analysis of single-cell co-expression, sampling from 31 individual studies. Using neighbor voting in cross-validation, we find that single-cell network connectivity is less likely to overlap with known functions than co-expression derived from bulk data, with functional variation within cell types strongly resembling that also occurring across cell types. To identify features and analysis practices that contribute to this connectivity, we perform our own single-cell RNA-seq experiment of 126 cortical interneurons in an experimental design targeted to co-expression. By assessing network replicability, semantic similarity and overall functional connectivity, we identify technical factors influencing co-expression and suggest how they can be controlled for. Many of the technical effects we identify are expression-level dependent, making expression level itself highly predictive of network topology. We show this occurs generally through re-analysis of the BrainSpan RNA-seq data. CONCLUSIONS: Technical properties of single-cell RNA-seq data create confounds in co-expression networks which can be identified and explicitly controlled for in any supervised analysis. This is useful both in improving co-expression performance and in characterizing single-cell data in generally applicable terms, permitting cross-laboratory comparison within a common framework.

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1186/s13059-016-0964-6

Cites

66 of 67 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 What lies beneath: co-expression can reflect different combinations of cell-state or compositional variation. Each panel shows a different scenario in which cell state and composition affect the expression of two genes (A and B), yielding different types of co-expression. Two cell types are colored in red and blue. In the top panel, both cell types have state-dependent variation that causes co-expression within each (r ~ 0.75). In addition, there is co-expression due to compositional variation (r ~ 0.75). In the bottom left panel only compositional variation is apparent (r ~ 0.65), there is no relationship between gene A and gene B within the cell types (r ~ 0). This is the opposite in the bottom right panel. Here, there is only variation within the cell types (r ~ 0.95) but no compositional effect across cell types (r ~ 0). The exact value the compositional correlations take would vary in real data since combinations of the underlying cell types would fill in intermediate points, but the three cases would still occur as described; other possibilities due to noise or other complex scenarios (e.g. Yule-Simpson effect) are also possible
Fig. 1 What lies beneath: co-expression can reflect different combinations of cell-state or compositional variation. Each panel shows a different scenario in which cell state and composition affect the expression of two genes (A and B), yielding different types of co-expression. Two cell types are colored in red and blue. In the top panel, both cell types have state-dependent variation that causes co-expression within each (r ~ 0.75). In addition, there is co-expression due to compositional variation (r ~ 0.75). In the bottom left panel only compositional variation is apparent (r ~ 0.65), there is no relationship between gene A and gene B within the cell types (r ~ 0). This is the opposite in the bottom right panel. Here, there is only variation within the cell types (r ~ 0.95) but no compositional effect across cell types (r ~ 0). The exact value the compositional correlations take would vary in real data since combinations of the underlying cell types would fill in intermediate points, but the three cases would still occur as described; other possibilities due to noise or other complex scenarios (e.g. Yule-Simpson effect) are also possibleCC BY · PubMed Central
Fig. 2 scRNA-seq networks have lower functional connectivity than bulk RNA-seq networks and state-driven performance is highly correlated with compositional performance. a Individual bulk and scRNA-seq networks were aggregated ten times in random order by averaging rank co-expression values, then re-ranking the resulting network. Receiver operating characteristic (ROC) scores for GO slim categories were calculated as each additional network was added. Individual runs are shown in gray (bulk) and pink (sc), and means are shown in black and red. Performance rises with aggregation for both bulk and sc networks, but sc networks have much lower baseline functional connectivity. Inset: Aggregate single-cell network (163 cell types) GO slim AUROCs are plotted against those of the aggregate bulk network (163 experiments). Performance is strongly correlated (r = 0.89) but is consistently lower in the single-cell aggregate. b Aggregate cell-type network GO slim AUROCs are plotted against those of the pseudo-compositional aggregate network. Performance is strongly correlated
Fig. 2 scRNA-seq networks have lower functional connectivity than bulk RNA-seq networks and state-driven performance is highly correlated with compositional performance. a Individual bulk and scRNA-seq networks were aggregated ten times in random order by averaging rank co-expression values, then re-ranking the resulting network. Receiver operating characteristic (ROC) scores for GO slim categories were calculated as each additional network was added. Individual runs are shown in gray (bulk) and pink (sc), and means are shown in black and red. Performance rises with aggregation for both bulk and sc networks, but sc networks have much lower baseline functional connectivity. Inset: Aggregate single-cell network (163 cell types) GO slim AUROCs are plotted against those of the aggregate bulk network (163 experiments). Performance is strongly correlated (r = 0.89) but is consistently lower in the single-cell aggregate. b Aggregate cell-type network GO slim AUROCs are plotted against those of the pseudo-compositional aggregate network. Performance is strongly correlatedCC BY · PubMed Central
Fig. 3 Comparative network analysis shows higher functional connectivity, semantic similarity, and convergent co-expression of UMI-based aggregates. a For each batch network, functional connectivity was benchmarked against 108 GO slim categories then networks were randomly selected and aggregated ten times. Networks built from raw UMI data are shown in black and count-per-million (CPM) standardized data are shown in red. As in our meta-analysis of single-cell networks, performance rises with aggregation, indicating an overlap in functional signal among networks. CPM networks have significantly lower functional connectivity than UMI networks (mean UMI = 0.54 +/– 0.002, mean CPM + 0.54 +/– 0.001, p <0.05 Wilcoxon rank sum, n = 8). Inset: Boxplot of synaptic gene performance for UMI and CPM networks. Though mean GO slim performance is modest, connectivity of this functionally relevant gene set is high (mean UMI = 0.73 +/– 0.004, mean CPM = 0.69 +/– 0005). b Semantic similarity of top 1 % network connections assessed by the number of shared GO functions. The red line indicates mean semantic similarity of all genes. Lower semantic similarity is observed for CPM, removal of unwanted variation (RUV), and binary expression aggregates compared to UMI-based networks. c Plot shows pairwise comparisons of top 1 % network connections based on the Jaccard index. UMI-based networks are more similar to one another than to CPM, RUV, and binary expression networks. d Standard deviation of aggregate co-expression values, red line marks the amount of variance expected by chance. All aggregates are more variable than random, indicating the presence of replicable co-expression among individual networks
Fig. 3 Comparative network analysis shows higher functional connectivity, semantic similarity, and convergent co-expression of UMI-based aggregates. a For each batch network, functional connectivity was benchmarked against 108 GO slim categories then networks were randomly selected and aggregated ten times. Networks built from raw UMI data are shown in black and count-per-million (CPM) standardized data are shown in red. As in our meta-analysis of single-cell networks, performance rises with aggregation, indicating an overlap in functional signal among networks. CPM networks have significantly lower functional connectivity than UMI networks (mean UMI = 0.54 +/– 0.002, mean CPM + 0.54 +/– 0.001, p <0.05 Wilcoxon rank sum, n = 8). Inset: Boxplot of synaptic gene performance for UMI and CPM networks. Though mean GO slim performance is modest, connectivity of this functionally relevant gene set is high (mean UMI = 0.73 +/– 0.004, mean CPM = 0.69 +/– 0005). b Semantic similarity of top 1 % network connections assessed by the number of shared GO functions. The red line indicates mean semantic similarity of all genes. Lower semantic similarity is observed for CPM, removal of unwanted variation (RUV), and binary expression aggregates compared to UMI-based networks. c Plot shows pairwise comparisons of top 1 % network connections based on the Jaccard index. UMI-based networks are more similar to one another than to CPM, RUV, and binary expression networks. d Standard deviation of aggregate co-expression values, red line marks the amount of variance expected by chance. All aggregates are more variable than random, indicating the presence of replicable co-expression among individual networksCC BY · PubMed Central
Fig. 4 Functional connectivity is dependent on expression level in UMI networks. a A UMI-based network is plotted as a function of gene expression level, with co-expression values for highly expressed genes in the bottom left corner and lowly expressed genes in the top right corner. A characteristic sunset pattern is observed in A, with the highest co-expression values occurring between highly expressed genes. b AUROC for synaptic set performance in n-fold cross-validation is plotted against median expression level for each held-out gene (plot shows the results from one network). A strong positive relationship was observed between expression level and performance (mean rs = 0.72 +/– 0.038). Inset: Box-and-whisker plot of synaptic set performance in networks made from all genes compared to networks made from high expressing genes only (median > =16 in each batch). A significant reduction in performance was observed in filtered networks (mean AUROC in all gene networks = 0.73 +/– 0.004, mean AUROC in high expressing networks = 0.49 +/– 0.03, p <0.0002 Wilcoxon rank sum test, n = 8)
Fig. 4 Functional connectivity is dependent on expression level in UMI networks. a A UMI-based network is plotted as a function of gene expression level, with co-expression values for highly expressed genes in the bottom left corner and lowly expressed genes in the top right corner. A characteristic sunset pattern is observed in A, with the highest co-expression values occurring between highly expressed genes. b AUROC for synaptic set performance in n-fold cross-validation is plotted against median expression level for each held-out gene (plot shows the results from one network). A strong positive relationship was observed between expression level and performance (mean rs = 0.72 +/– 0.038). Inset: Box-and-whisker plot of synaptic set performance in networks made from all genes compared to networks made from high expressing genes only (median > =16 in each batch). A significant reduction in performance was observed in filtered networks (mean AUROC in all gene networks = 0.73 +/– 0.004, mean AUROC in high expressing networks = 0.49 +/– 0.03, p <0.0002 Wilcoxon rank sum test, n = 8)CC BY · PubMed Central
Fig. 5 Node degree and network performance are predicted by expression level in the single-cell aggregate network. Top: GO slim AUROCs and predicted AUROCs based on node degree are plotted for single-cell and bulk aggregates (163 networks in each). Functional connectivity in both aggregates is dependent on node degree. Bottom: GO slim AUROCs and predicted AUROCs based on median gene expression are plotted. Single-cell aggregate performance is predicted by expression; however, there is no relationship between expression and bulk aggregate performance
Fig. 5 Node degree and network performance are predicted by expression level in the single-cell aggregate network. Top: GO slim AUROCs and predicted AUROCs based on node degree are plotted for single-cell and bulk aggregates (163 networks in each). Functional connectivity in both aggregates is dependent on node degree. Bottom: GO slim AUROCs and predicted AUROCs based on median gene expression are plotted. Single-cell aggregate performance is predicted by expression; however, there is no relationship between expression and bulk aggregate performanceCC BY · PubMed Central
Fig. 6 Differential connectivity of autism candidate genes in BrainSpan networks is predicted by expression level. Networks were built using BrainSpan RNA-seq data for individuals with ten or more samples (16 pre-natal, 18 post-natal networks). Connectivity of autism candidate genes is plotted against mean expression, with linear regression lines for pre-natal and post-natal networks colored in black and red (adjusted R2 = –0.05, = 0.23, respectively)
Fig. 6 Differential connectivity of autism candidate genes in BrainSpan networks is predicted by expression level. Networks were built using BrainSpan RNA-seq data for individuals with ten or more samples (16 pre-natal, 18 post-natal networks). Connectivity of autism candidate genes is plotted against mean expression, with linear regression lines for pre-natal and post-natal networks colored in black and red (adjusted R2 = –0.05, = 0.23, respectively)CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 67 references

Cited by

See all 85

Metadata

Authors
Crow, M., Paul, A., Ballouz, S., Huang, Z. J., & Gillis, J.
Affiliations
Cold Spring Harbor Laboratory (US)
Corresponding author
Gillis, Jesse
Venue
Genome Biology
Published
2016-05-06
Volume
17
Issue
1
Pages
101-101
ISSN
  • 1465-6906 (ISSN-L)
  • 1474-760X
OpenAlex
W2361797231
Cited by
76
References
66
Citations by year
2026: 5 · 2025: 7 · 2024: 6 · 2023: 7 · 2022: 13 · 2021: 6 · 2020: 9 · 2019: 5 · 2018: 10 · 2017: 7 · 2016: 1
FWCI
3.98
Type
journal article
Language
en
Open access
Licence
Locations
  • Genome biology (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • Cold Spring Harbor Laboratory Institutional Repository (Cold Spring Harbor Laboratory) (acceptedVersion)
  • UNSWorks (University of New South Wales, Sydney, Australia) (submittedVersion, cc-by-nc-nd)
Funders
  • National Institutes of Health (1R01MH102616-01)
  • T. and V. Stanley (Private Donation)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Bioinformatics and Genomic Networks
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Single-cell and spatial transcriptomics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Gene expression and cancer classification
Keywords
co-expression network, single-cell RNA sequencing, cortical interneurons, BrainSpan, semantic similarity, interlaboratory comparison
MeSH terms
Animals, Cell Separation, Reproducibility of Results, Sequence Analysis, RNA, Gene Expression Profiling, Mice, Gene Regulatory Networks, Single-Cell Analysis

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Exploiting single-cell expression to characterize co-expression replicability. https://app.cochat.ai/papers/10.1186/s13059-016-0964-6
The paper
Crow, M., Paul, A., Ballouz, S., Huang, Z. J., & Gillis, J. (2016). Exploiting single-cell expression to characterize co-expression replicability. Genome Biology. https://doi.org/10.1186/s13059-016-0964-6

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.