CoChat

ICAT: a novel algorithm to robustly identify cell states following perturbations in single-cell transcriptomes

Open access

Published version of a preprint

The DOI matches its Crossref record; 36 of the 37 references it cites were found in a registry or an index.

Abstract

Motivation: The detection of distinct cellular identities is central to the analysis of single-cell RNA sequencing (scRNA-seq) experiments. However, in perturbation experiments, current methods typically fail to correctly match cell states between conditions or erroneously remove population substructure. Here, we present the novel, unsupervised algorithm Identify Cell states Across Treatments (ICAT) that employs self-supervised feature weighting and control-guided clustering to accurately resolve cell states across heterogeneous conditions. Results: Using simulated and real datasets, we show ICAT is superior in identifying and resolving cell states compared with current integration workflows. While requiring no a priori knowledge of extant cell states or discriminatory marker genes, ICAT is robust to low signal strength, high perturbation severity, and disparate cell type proportions. We empirically validate ICAT in a developmental model and find that only ICAT identifies a perturbation-unique cellular response. Taken together, our results demonstrate that ICAT offers a significant improvement in defining cellular responses to perturbation in scRNA-seq data. Availability and implementation: https://github.com/BradhamLab/icat.

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 · 10.1093/bioinformatics/btad278

Cites

36 of 36 references verified0 citation errors · 0 not checkable

Statistics

Figures

Figure 1. Overview of the ICAT algorithm. (A) The schematic illustrates the ICATC implementation of ICAT. To identify cell states across treatments, ICAT first performs self-supervised feature weighting to find genes that discriminate cell identities among control cells alone, followed by semisupervised clustering using the newly transformed expression matrix. To learn feature weights, ICAT clusters control cells using Louvain community detection, then uses the newly generated cluster labels as input into NCFS to weight genes by their predictiveness. After applying the learned gene weights to the original gene expression matrix, ICAT clusters both treated and control cells together using a semisupervised implementation of Louvain community detection. During this process, ICAT holds the previously identified cluster labels for the control cells immutable. (B) The schematic illustrates the ICATC + T implementation, which expands feature weighting to treated cells to identify asymmetrical populations between treatments. Cells are split along treatments and independently clustered using the Louvain method; then, cluster labels are used to learn gene weights using NCFS in each treatment set independently. To retain asymmetrically informative genes, weights for each treatment are concatenated row wise and subsequently reduced to the maximum weight using a row-wise maxpool function. The reduced weight vector is then used to transform the original count matrix.
Figure 1. Overview of the ICAT algorithm. (A) The schematic illustrates the ICATC implementation of ICAT. To identify cell states across treatments, ICAT first performs self-supervised feature weighting to find genes that discriminate cell identities among control cells alone, followed by semisupervised clustering using the newly transformed expression matrix. To learn feature weights, ICAT clusters control cells using Louvain community detection, then uses the newly generated cluster labels as input into NCFS to weight genes by their predictiveness. After applying the learned gene weights to the original gene expression matrix, ICAT clusters both treated and control cells together using a semisupervised implementation of Louvain community detection. During this process, ICAT holds the previously identified cluster labels for the control cells immutable. (B) The schematic illustrates the ICATC + T implementation, which expands feature weighting to treated cells to identify asymmetrical populations between treatments. Cells are split along treatments and independently clustered using the Louvain method; then, cluster labels are used to learn gene weights using NCFS in each treatment set independently. To retain asymmetrically informative genes, weights for each treatment are concatenated row wise and subsequently reduced to the maximum weight using a row-wise maxpool function. The reduced weight vector is then used to transform the original count matrix.CC BY · PubMed Central
Figure 2. ICAT correctly identifies cell states in distinct experimental compositions. (A) UMAP projections of different cellular compositions in simulated datasets. Each dot represents a cell with circles representing control cells and crosses denoting treated cells. Dots are colored by ground truth identity (left column), cluster label produced by performing Louvain community detection on the raw count matrix (middle column), and clusters labeled produced by ICAT (right column). (B) Average agreement between ground truth label and cluster labels produced by clustering the raw data (blue) and ICAT (orange) as measured by the ARI. Error bars represent the 95% CIs for the mean ARI for each method. Five different cellular composition conditions were simulated: “All same,” both control and treated cells share the same two cell states; “Rx Unique,” treated cells contain a treatment-unique cell state; “Control Unique,” control cells contain a unique cell state; “Both Unique,” both treated and control cells contain treatment-specific cell states; and “None Same,” no shared cell states between treated and control cells. Each condition was simulated 15 times (n = 15). Simulations were evaluated using the ICATC implementation.
Figure 2. ICAT correctly identifies cell states in distinct experimental compositions. (A) UMAP projections of different cellular compositions in simulated datasets. Each dot represents a cell with circles representing control cells and crosses denoting treated cells. Dots are colored by ground truth identity (left column), cluster label produced by performing Louvain community detection on the raw count matrix (middle column), and clusters labeled produced by ICAT (right column). (B) Average agreement between ground truth label and cluster labels produced by clustering the raw data (blue) and ICAT (orange) as measured by the ARI. Error bars represent the 95% CIs for the mean ARI for each method. Five different cellular composition conditions were simulated: “All same,” both control and treated cells share the same two cell states; “Rx Unique,” treated cells contain a treatment-unique cell state; “Control Unique,” control cells contain a unique cell state; “Both Unique,” both treated and control cells contain treatment-specific cell states; and “None Same,” no shared cell states between treated and control cells. Each condition was simulated 15 times (n = 15). Simulations were evaluated using the ICATC implementation.CC BY · PubMed Central
Figure 3. ICAT outperforms current methods for cell state identification and is robust to experimental conditions. (A) UMAP projections of raw (left), ICATC processed (middle), and ICATC+T processed count matrices for the simulated data. Projections show ICATC and ICATC + T correctly mix shared populations (red and green), whereas only ICATC + T isolates asymmetrical populations (purple, yellow). ICAT performance for simulated data was further evaluated using the ICATC + T implementation only. (B) Select marker and perturbed gene expression patterns are displayed as violin plots for three simulated control cell types (1–3) under normal (C) and perturbed (P) conditions. P(C1)+ is a stimulated and perturbed version of cell type 1; perturbation-specific cell types P4 and P5 express distinct marker genes. (C) The percentage of perturbed genes used to assess robustness to perturbation severity is shown; values range from 1% to 25%. (D) The average number of marker genes per cell identity used to assess robustness to signal strength is shown; values range from 10 to 105 (0.7%–7% of total genes). (E) The set of cell identity proportions used to test the ability to identify rare cell states is shown. The number of cells per treatment-label pair ranges from 50 to 175. (F) Method performance is compared as the fraction of perturbed genes increases (F1) and as the average number of marker genes per population increases (F2). Left panels display results for stand-alone methods: ICAT, No Int. Seurat, Scanorama, and Harmony; while right panels show results for ICAT-extended workflows: ICATSeurat, ICATScan, and ICATHarm. Results are depicted as averages with 95% CIs shown by shading. (G) Method performance is compared as the proportion of cell types is varied. The Gini coefficient reflects the degree of population size disproportion among cell states. Left panel: ICAT, Seurat, Scanorama, and Harmony. Right panel: ICATSeurat, ICATScan, and ICATHarm.
Figure 3. ICAT outperforms current methods for cell state identification and is robust to experimental conditions. (A) UMAP projections of raw (left), ICATC processed (middle), and ICATC+T processed count matrices for the simulated data. Projections show ICATC and ICATC + T correctly mix shared populations (red and green), whereas only ICATC + T isolates asymmetrical populations (purple, yellow). ICAT performance for simulated data was further evaluated using the ICATC + T implementation only. (B) Select marker and perturbed gene expression patterns are displayed as violin plots for three simulated control cell types (1–3) under normal (C) and perturbed (P) conditions. P(C1)+ is a stimulated and perturbed version of cell type 1; perturbation-specific cell types P4 and P5 express distinct marker genes. (C) The percentage of perturbed genes used to assess robustness to perturbation severity is shown; values range from 1% to 25%. (D) The average number of marker genes per cell identity used to assess robustness to signal strength is shown; values range from 10 to 105 (0.7%–7% of total genes). (E) The set of cell identity proportions used to test the ability to identify rare cell states is shown. The number of cells per treatment-label pair ranges from 50 to 175. (F) Method performance is compared as the fraction of perturbed genes increases (F1) and as the average number of marker genes per population increases (F2). Left panels display results for stand-alone methods: ICAT, No Int. Seurat, Scanorama, and Harmony; while right panels show results for ICAT-extended workflows: ICATSeurat, ICATScan, and ICATHarm. Results are depicted as averages with 95% CIs shown by shading. (G) Method performance is compared as the proportion of cell types is varied. The Gini coefficient reflects the degree of population size disproportion among cell states. Left panel: ICAT, Seurat, Scanorama, and Harmony. Right panel: ICATSeurat, ICATScan, and ICATHarm.CC BY · PubMed Central
Figure 4. ICAT outperforms current integration methods in identifying cell states across treatments in real datasets. scRNA-seq data from Tian, Kagohara, and Kang studies is compared. (A) Spider plots compare the performance for each algorithm within each dataset for ARI, LISI, and DB quality metrics. (B) Lollipop plots highlight the differences in the metrics for each method across the same datasets. (C) Spider plots comparing Seurat to ICATSeurat, Scanorama to ICATScan, and Harmony to ICATHarm, respectively. All metrics are scaled from 0 to 1, where 1 is best (closest to the apex of each corner). DB, Davies–Bouldin metric, and otherwise as in Fig. 2. ICAT performance for the Tian dataset was evaluated using the ICATC + T implementation, whereas Kagohara and Kang datasets were evaluated using the ICATC implementation.
Figure 4. ICAT outperforms current integration methods in identifying cell states across treatments in real datasets. scRNA-seq data from Tian, Kagohara, and Kang studies is compared. (A) Spider plots compare the performance for each algorithm within each dataset for ARI, LISI, and DB quality metrics. (B) Lollipop plots highlight the differences in the metrics for each method across the same datasets. (C) Spider plots comparing Seurat to ICATSeurat, Scanorama to ICATScan, and Harmony to ICATHarm, respectively. All metrics are scaled from 0 to 1, where 1 is best (closest to the apex of each corner). DB, Davies–Bouldin metric, and otherwise as in Fig. 2. ICAT performance for the Tian dataset was evaluated using the ICATC + T implementation, whereas Kagohara and Kang datasets were evaluated using the ICATC implementation.CC BY · PubMed Central
Figure 5. ICAT most accurately defines subpopulation response to perturbation. SMART-seq2 was performed on cells isolated from controls or from embryos treated with the perturbants chlorate or MK-886. (A) ICAT predicts five clusters from the combined data (A1), with treatment-dependent subpopulation compositions (G-test likelihood test, FDR < 0.05; ad-hoc pairwise Fisher’s Exact test, FDR < 0.05) (A2). MK-induced differences are defined by reciprocal expression patterns of ICAT-identified genes, sm50 and pks2, in subpopulations 3 (sm50+/pks2−) and 4 (sm50−/pks2+) (A3). (B) ICAT predictions for control (B1, B3) and MK-866 (B2, B4) are validated by HCR FISH analysis (B5, B6). Automated detection and segmentation of PMCs in vivo from 3D image data are shown as various colors; each color represents an individual cell (1–2). The expression of sm50 (red) and pks2 (cyan) transcripts are shown in the same embryos; arrowheads indicate PMCs that express only sm50 or only pks2 (3–4). MK-866 treated embryos exhibit statistically significantly fewer sm50+/pks2− PMCs (B5) and more sm50−/pks2+ PMCs (B6) than controls, consistent with the predictions from ICAT (A). Dot plot error bars denote 95% CIs of the mean with each dot representing an individual embryo [nDMSO = 26, nMK886 = 37; B5 (Binomial GLM, μDMSO = 3.59%, μMK886 = 1.91%; βMK886 = −0.71, P < 10−3; B6 Binomial GLM, μDMSO = 0.19%, μMK886 = 2.71%; βMK886 = 2.60, P < 10−4)]. (C) Seurat (C1) fails to find any treatment effect on PMC subpopulation composition, while Scanorama (C2), and Harmony (C3) predict one and two disrupted populations, respectively (G-test likelihood test, FDR < 0.05; ad hoc pairwise Fisher’s Exact test, FDR < 0.05).
Figure 5. ICAT most accurately defines subpopulation response to perturbation. SMART-seq2 was performed on cells isolated from controls or from embryos treated with the perturbants chlorate or MK-886. (A) ICAT predicts five clusters from the combined data (A1), with treatment-dependent subpopulation compositions (G-test likelihood test, FDR < 0.05; ad-hoc pairwise Fisher’s Exact test, FDR < 0.05) (A2). MK-induced differences are defined by reciprocal expression patterns of ICAT-identified genes, sm50 and pks2, in subpopulations 3 (sm50+/pks2−) and 4 (sm50−/pks2+) (A3). (B) ICAT predictions for control (B1, B3) and MK-866 (B2, B4) are validated by HCR FISH analysis (B5, B6). Automated detection and segmentation of PMCs in vivo from 3D image data are shown as various colors; each color represents an individual cell (1–2). The expression of sm50 (red) and pks2 (cyan) transcripts are shown in the same embryos; arrowheads indicate PMCs that express only sm50 or only pks2 (3–4). MK-866 treated embryos exhibit statistically significantly fewer sm50+/pks2− PMCs (B5) and more sm50−/pks2+ PMCs (B6) than controls, consistent with the predictions from ICAT (A). Dot plot error bars denote 95% CIs of the mean with each dot representing an individual embryo [nDMSO = 26, nMK886 = 37; B5 (Binomial GLM, μDMSO = 3.59%, μMK886 = 1.91%; βMK886 = −0.71, P < 10−3; B6 Binomial GLM, μDMSO = 0.19%, μMK886 = 2.71%; βMK886 = 2.60, P < 10−4)]. (C) Seurat (C1) fails to find any treatment effect on PMC subpopulation composition, while Scanorama (C2), and Harmony (C3) predict one and two disrupted populations, respectively (G-test likelihood test, FDR < 0.05; ad hoc pairwise Fisher’s Exact test, FDR < 0.05).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 36 references

Cited by

See all 10

Metadata

Authors
Hawkins, D. Y., Zuch, D. T., Huth, J., Rodriguez-Sastre, N., McCutcheon, K. R., Glick, A., Lion, A. T., Thomas, C. F., Descoteaux, A. E., Johnson, W. E., & Bradham, C. A.
Affiliations
  • Boston University (US)
  • Biology Department, Boston University , 24 Cummington Mall , Boston, MA 02215, United States
Corresponding author
Bradham, Cynthia A
Venue
Bioinformatics
Published
2023-04-22
Volume
39
Issue
5
ISSN
  • 1367-4803 (ISSN-L)
  • 1367-4811
OpenAlex
W4366742301
Cited by
10
References
37
Citations by year
2026: 2 · 2025: 3 · 2024: 4 · 2023: 1
FWCI
0.89
Type
journal article
Language
en
Open access
Locations
  • Bioinformatics (publishedVersion, cc-by)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by-nc-nd)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • OpenBU (Boston University) (publishedVersion, cc-by)
Funders
National Science Foundation Integrative Organismal Systems (1656752)
Topics
  • 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
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Gene Regulatory Network Analysis
Keywords
ICAT, cell state identification, single-cell transcriptomics, perturbation experiments, single-cell RNA sequencing, unsupervised clustering, population substructure, developmental model
MeSH terms
Algorithms, Cluster Analysis, Sequence Analysis, RNA, Gene Expression Profiling, Single-Cell Analysis, Transcriptome

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

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.