PaperlayerView exact revision on arXiv

Source-linked AI summary

CellPhoneDB v5: inferring cell-cell communication from single-cell multiomics data

Kevin Troulé, Robert Petryszak, Martin Prete, James Cranley, Alicia Harasty, Zewen Kelvin Tuong, Sarah A Teichmann, Luz Garcia-Alonso, Roser Vento-Tormo

arXiv:2311.04567v2q-bio.CBq-bio.QM

TL;DR

Accurate cell-cell communication inference depends on the accuracy and scope of interaction priors. CellPhoneDB v5 adds tailored differential-expression queries, and the authors report that it more reliably approximates the nature of cell-cell interactions.

  • Problem

    Cell-cell communication inference depends on the accuracy and scope of its interaction prior.

    Source: Introduction · p. 2

  • Method

    CellPhoneDB v5 adds a differential-expression-based methodology for more tailored cell-cell interaction queries.

    Source: Introduction · p. 3

  • Results

    CellPhoneDB v5 more reliably approximates the nature of cell-cell interactions.

    Source: Introduction · p. 2

  • Takeaways & Limitations

    The tailored approach helps researchers address specific biological questions with greater precision.

    Source: Introduction · p. 3

    The authors note that cell-cell interaction databases must contend with context-dependent specificity.

    Source: Limitations · p. 10

Abstract

from arXiv · show

Cell-cell communication is essential for tissue development, regeneration and function, and its disruption can lead to diseases and developmental abnormalities. The revolution of single-cell genomics technologies offers unprecedented insights into cellular identities, opening new avenues to resolve the intricate cellular interactions present in tissue niches. CellPhoneDB is a bioinformatics toolkit designed to infer cell-cell communication by combining a curated repository of bona fide ligand-receptor interactions with a set of computational and statistical methods to integrate them with single-cell genomics data. Importantly, CellPhoneDB captures the multimeric nature of molecular complexes, thus representing cell-cell communication biology faithfully. Here we present CellPhoneDB v5, an updated version of the tool, which offers several new features. Firstly, the repository has been expanded by one-third with the addition of new interactions. These encompass interactions mediated by non-protein ligands such as endocrine hormones and GPCR ligands. Secondly, it includes a differentially expression-based methodology for more tailored interaction queries. Thirdly, it incorporates novel computational methods to prioritise specific cell-cell interactions, leveraging other single-cell modalities, such as spatial information or TF activities (i.e. CellSign module). Finally, we provide CellPhoneDBViz, a module to interactively visualise and share results amongst users. Altogether, CellPhoneDB v5 elevates the precision of cell-cell communication inference, ushering in new perspectives to comprehend tissue biology in both healthy and pathological states.

Introduction

This section presents CellPhoneDB v5 for inferring cell-cell communication from curated molecular interactions and single-cell expression data, describing its methods for targeted queries, prioritization, and visualization. It also covers applications and limitations of its inferences, which should be treated as hypotheses requiring corroboration.

Introduction

  • V5 adds differential-expression-based queries and prioritization using spatial co-occurrence, scATAC-seq-informed downstream TF activity through CellSign, and expression specificity for cell-type pairs.Spatial information restricts analyses to cell-type pairs coexisting in time and space.Source: Introduction · p. 3
  • CellPhoneDBViz and selective query and visualization strategies facilitate interpretation and sharing of results.The tool's broader purpose is to help researchers infer and prioritize context-specific cell-cell interactions across varied experiments.Source: Introduction · p. 3

Updates on the development of the protocol: the database

  • The expanded repository adds non-gene-encoded ligands such as steroid hormones, neurotransmitters and small molecules, whose abundance is approximated using the final biosynthetic enzyme and, when known, the transporter.It also adds manually curated interactions and findings from high-throughput surface-receptor screening that accounts for protein affinity and natural stoichiometry.Source: Updates on the development of the… · p. 3
  • Manual curation and representation of protein-complex subunits capture combinatorial specificity, which the authors say better reflects interaction biology and reduces false-positive inferences.The database excludes general protein–protein interactions from third-party repositories in favor of manually curated interactions with validated intercellular roles.Source: Updates on the development of the… · p. 4
  • The v5 repository contains 2,911 curated interactions—1,882 protein-mediated and 1,029 non-peptidic—plus 240 heteromeric proteins, 1,058 homomeric proteins and 120 non-peptidic ligands.Of the 120 non-peptidic ligands, 61 involve multiple proteins, including enzymes and transporters.Source: Updates on the development of the… · p. 4
  • The database classifies 2,508 interactions as directional ligand–receptor, 242 as adhesion and 161 as ligand–ligand, receptor–receptor or gap-junction interactions.Directionality supports tracing signaling cascades and connecting them to transcription-factor measurements.Source: Updates on the development of the… · p. 4
  • Because no single inference method suits every scenario, CellPhoneDB offers three transcriptomic integration methods for selection according to experimental design and research question.The simple method reports mean partner expression by cell-type pair, uses the least-expressed subunit for complexes and allows expression-fraction filtering; the statistical method extends it.Source: Updates on the development of the… · pp. 4–5
  • Differential-expression analysis enables flexible queries to identify interactions specific to cell states, experimental conditions, or differentiation processes.Users can tailor the null distribution, including by restricting which cell types it includes.Source: Updates on the development of the… · p. 5
  • The differential-expression query requires user-supplied cell types of interest and differentially expressed genes, then retrieves interactions meeting a user-specified expression threshold for the partner gene.Differential expression may be estimated across discrete cell types or along pseudotime using a method suited to the user's experimental design and research question.Source: Updates on the development of the… · p. 5
  • The expression-specificity score uses geometric means for complex subunits, sets complexes with absent subunits to zero, and combines scaled protein expression across interacting cell pairs.Scores range from 0 to 100, with higher scores indicating greater expression specificity.Source: Updates on the development of the… · pp. 5–6
  • Spatial transcriptomics restricts analyses to cell-type pairs coexisting in space, while temporal information can exclude cell types that do not coexist in time.These filters narrow candidate interactions to those between neighboring cells and can account for developmental processes, conditions, or cohorts.Source: Updates on the development of the… · p. 6
  • CellSign identifies activated receptors and prioritizes high-confidence interactions by leveraging downstream TF activity, using a manually revised literature resource restricted to highly specific TFs, with 212 highly specific direct receptor-to-TF relationships.Source: Updates on the development of the… · p. 6

Updates on the development of protocol: the visualisation tool, CellPhoneDBViz

  • CellPhoneDBViz [20] provides interactive web-based visualizations and searches for CellPhoneDB results, supporting result sharing and exploration of global patterns or selected cell pairs and microenvironments.Its visualizations include heatmaps, chord plots, and dot plots for CellPhoneDB inputs and outputs.Source: Updates on the development of the… · p. 7

Applications of the method

  • Method 2 identified 15 monocyte cytokine-mediated interactions specific to patients with severe COVID-19 [41] and differential ligand–receptor usage between LCAM-defined lung cancer groups [42].The LCAM analysis provided insight into immune-cell crosstalk in the tumors.Source: Applications of the method · p. 8
  • Spatial microenvironment modules shed light on how granulosa-cell-subtype compartmentalization influences oocyte differentiation during ovarian development [17] and prioritized invading trophoblast–maternal uterine interactions in placental development [18]; CellSign prioritized oocyte–granulosa interactions [17].Source: Applications of the method · p. 8

Limitations

  • CellPhoneDB estimates ligand and receptor abundance from gene expression, which may miss protein processing, modification, secretion, and diffusion; complex composition and cofactors also affect signalling specificity.Interactions can be context-dependent, and a protein may belong to different complexes with opposite effects.Source: Limitations · p. 10
  • No interaction database can capture every potential interaction; CellPhoneDB’s reliance on experimentally supported interactions may favor well-studied processes but reduces false positives [44].The authors caution that high-throughput inference can be challenging to interpret and recommend starting from a specific biological question.Source: Limitations · p. 10

Comparison with other approaches

  • CellPhoneDB v5 can prioritize interactions using spatial and chromatin-accessibility information, and combines user-provided TF activities with curated receptor-to-TF relationships.This strategy links ligand binding and receptor activation to downstream TF activation.Source: Comparison with other approaches · p. 11
  • A compared approach infers cell interactions directly from undeconvoluted spatial spots.Source: Comparison with other approaches · p. 11
  • Inference results should support or generate hypotheses and be corroborated with other in silico methods, complementary data modalities, or experimental validation.The passages recommend starting from a biological question or hypothesis and then seeking downstream corroboration.Source: Comparison with other approaches · p. 11

MATERIALS

CellPhoneDB requires expression and cell-type metadata files, with optional inputs for microenvironments and active transcription factors and a DEG file for method 3. Its materials also specify the software environment and database files supporting CellSign and gene-name synonym matching.

Input data files for running CellPhoneDB

  • Running CellPhoneDB requires normalized expression counts with gene identifiers and metadata assigning each cell to a user-defined cell type.Accepted gene identifier formats include Ensembl IDs, gene names, and hgnc_symbol; expression and metadata files can use .csv, .txt, .tsv, .tab, or .pickle formats.Source: Input data files for running CellPhoneDB · pp. 11–12
  • The optional microenvironments file groups cells by microenvironment, and the optional active-TFs file lists cell types with user-defined active TFs; a DEGs file is mandatory for method 3.Source: Input data files for running CellPhoneDB · p. 12

Software

  • CellPhoneDB requires Python 3.8 or higher, and input data can be reprocessed with Seurat, SCANPY, or other toolkits.Source: Software · p. 12

Input data files for building the database.

  • CellPhoneDB v5 keeps the previous database structure but adds transcription_factor_input.csv for receptor–TF relationships and uniprot_synonyms.tsv for matching gene-name synonyms to proteins.The receptor–TF file supports CellSign by linking receptors to user-provided TF activation status.Source: Input data files for building the… · pp. 12–13
  • The v5 database contains manually curated interactions and no longer collects general protein–protein interactions from third parties.Source: Input data files for building the… · p. 13
  • Plain-text outputs retain the previous structure and add files for method 3, CellSign, and specificity scoring.Source: Input data files for building the… · p. 15

PROCEDURE

CellPhoneDB v5 requires Python 3.8 or higher; tutorials provide extensive methods and examples for running the tool.

Analysis methods

  • Tutorials provide an extensive description of CellPhoneDB v5 methods and examples for running the tool.Available at https://github.com/ventolab/CellphoneDB.Source: Analysis methods · p. 15

Installation. Timing 10 min

1 Create conda environment and install CellPhoneDB

For CellPhoneDB use in Jupyter hubs, add the environment to the Jupyter kernel.

1 Create conda environment and install CellPhoneDB

  • Register the CellPhoneDB environment as a Jupyter kernel with `python -m ipykernel install --user --name 'cpdb'`.Source: §1 · p. 15

1 List database versions available

This section shows how to display the remote CellPhoneDB database versions using the `db_releases_utils` module.

  • The snippet imports `HTML` and `display` from `IPython.display`.Source: §1 · p. 15
  • The snippet imports `db_releases_utils` from `cellphonedb.utils`.Source: §1 · p. 15
  • The code calls `get_remote_database_versions_html()` to retrieve remote database version information for display.Source: §1 · p. 15

2 Download database

The section identifies a CellPhoneDB database utility import and specifies the format and directory-structure requirements for database updates.

2 Download database

  • The code imports the db_utils utility from cellphonedb.utils.Source: §2 · p. 16

Update database. Timing 1 min

1 Create database from input files

The section introduces the CellPhoneDB database utility by importing `db_utils` from `cellphonedb.utils`.

1 Create database from input files

  • The code imports `db_utils` from `cellphonedb.utils`.Source: §1 · p. 16

1 Define input files … 2 Running method 2

The documentation distinguishes a simple analysis from a statistical analysis and describes their callable methods and input options. Method 2 reuses method 1’s inputs and adds active-TF and subsampling options.

2 Running method 1 (simple analysis)

  • Method 1 imports `cpdb_analysis_method` from `cellphonedb.src.core.methods`.Source: §2 · p. 16
  • Method 1 accepts a TSV microenvironment-definition file and a Boolean option to score interactions.Source: §2 · pp. 16–17

1 Define input files

  • Method 2 uses the same input files as method 1, plus additional inputs.Source: §1 · p. 17

2 Running method 2

  • Method 2 imports `cpdb_statistical_analysis_method` from `cellphonedb.src.core.methods`.Source: §2 · p. 17
  • Method 2 accepts a TSV microenvironment-definition file and a Boolean option to score interactions.Source: §2 · p. 17
  • Method 2 also accepts an active-TF-per-cell TSV and subsampling options for whether to subsample cells, log1p-transformed counts, the number of PCs, and the number of cells.Source: §2 · p. 17

1 Define input files … 2 Steps executed during scoring

CellPhoneDB scoring filters and summarizes gene expression by cell type, aggregates heteromer subunits, scales expression, and multiplies interaction-protein values; scoring requires log-normalised expression data.

1 Define input files

  • Method 3 uses the same input files as method 1, plus additional files.Source: §1 · p. 17

2 Running method 3

  • Method 3 accepts a microenvironment-definition TSV and an active-TFs-per-cell TSV; score_interactions toggles interaction scoring.Source: §2 · p. 18

Run method 2 and 3 with scoring.

2 Steps executed during scoring

  • Scoring filters non-interaction and low-prevalence genes, computes cell-type means, geometrically aggregates heteromer subunits, then scales gene and heteromer means to 0–10.Source: §2 · p. 19
  • It multiplies the scaled mean expression of interacting proteins as a proxy for interaction relevance.Source: §2 · p. 19

TROUBLESHOOTING · Anticipated results

The troubleshooting section directs users to Table 2 for common errors. The anticipated-results section describes CellPhoneDB v5 applications, its expanded inference and prioritisation capabilities, and hypotheses derived from its findings that were experimentally validated.

TROUBLESHOOTING

Anticipated results

  • CellPhoneDB v5 was used to examine epithelial–fibroblast interactions in the human endometrium in response to hormones [19].Source: Anticipated results · p. 20
  • The methods were applied to examine how granulosa cells compartmentalise and drive oocyte differentiation [17].Source: Anticipated results · p. 20
  • The methods were used to investigate interactions between invading trophoblast cells and maternal uterine cells during early human development [18].Source: Anticipated results · p. 20
  • The methods were used to identify interactions between trans-synaptic pacemaker cells and glial cells in the heart [16].Source: Anticipated results · p. 20
  • CellPhoneDBViz provides a platform to view and share cell-cell communication findings during regeneration and differentiation.Source: Anticipated results · p. 20
  • The updates leverage other omics, add GPCR and non-peptidic ligand modalities, enable differential-expression queries, score interaction specificity, and add visualisation and sharing tools.Source: Anticipated results · p. 20
  • CellPhoneDB v5 findings produced biologically relevant hypotheses, some of which were subsequently experimentally validated.Source: Anticipated results · p. 20

Competing interests

S.A.T. disclosed recent consulting and advisory roles, an equity holding, and a co-founder role.

Loading 2311.04567v2…