Take ACTION to characterize the functional identity of single cells

Shahin Mohammadi, Vikram Ravindra, David F. Gleich, Ananth GramabioRxiv 2016 · DOI

TL;DR

This paper introduces ACTION, a novel computational framework that goes beyond simply classifying cell types to characterizing their functional identity. ACTION infers primary cellular functions, reconstructs regulatory networks, and identifies novel melanoma subtypes with distinct survival rates and therapeutic responses, offering a more nuanced understanding of cellular heterogeneity.

What Problem Does This Paper Address?

Traditional methods for identifying cell types from single-cell RNA sequencing (scRNA-seq) data often rely on distance metrics that struggle to capture weakly expressed but crucial genes. These methods also frequently require users to specify the number of cell types beforehand and don't fully elucidate the functional differences between identified cell types. The authors aim to address these limitations by developing a method that can automatically characterize the functional identity of cells, uncovering their specialized roles and the regulatory mechanisms driving them, which is crucial for understanding complex biological systems and developing personalized therapies.

Methods

ACTION is built upon three main components:

  1. Biologically-inspired cell similarity metric: This metric is designed to suppress universally expressed genes (which are less informative for cell identity) and enhance cell type-specific genes. It involves projecting gene expression profiles onto an orthogonal subspace to remove the effect of universally expressed genes and then boosting the contribution of cell type-specific genes using an information-theoretic approach.
  2. Geometric approach for identifying primary functions: ACTION models cellular functions as points in a high-dimensional space. It identifies "archetypes" (extreme corners of a polytope) that represent distinct primary functions. Each cell's functional identity is then represented as a convex combination of these archetypes. A modified successive projection algorithm (SPA) is used to find candidate archetypes, and a novel monitoring technique helps determine the optimal number of archetypes without overfitting. Archetypal analysis (AA) is then used to relax these archetypes and allow for contributions from multiple cells.
  3. Statistical framework for constructing cell-type specific transcriptional regulatory networks (TRNs): This component identifies key marker genes for each archetype and then infers the transcription factors (TFs) that significantly regulate these markers. The minimum hypergeometric (mHG) method is used to assess the significance of TFs and their target genes, ultimately constructing TRNs that explain the functional identity of each cell type.

Datasets

The authors utilized the following publicly available single-cell gene expression datasets:

  • Brain (GEO: GSE67835): 466 cells from various human brain cell types.
  • CellLines (GEO: GSE81861): 561 cells from seven different human cell lines. This dataset was specifically used for benchmarking and evaluating robustness to dropout.
  • Melanoma (GEO: GSE72056): 4,645 malignant, immune, and stromal cells from human melanoma tumors, classified into 7 major types. This dataset was central to identifying novel melanoma subtypes.
  • MouseBrain (GEO: GSE60361): 3005 cells from the mouse cortex and hippocampus.

Additionally, they used the Genomics of Drug Sensitivity in Cancer Project website for pharmacogenomic profiling of cell lines to assess therapeutic indications.

Key Insights and Ideas

  • Functional identity is more informative than just cell type: The paper argues that understanding what a cell does (its functional identity) is more critical than just knowing what it is (its cell type), especially for complex tissues and diseases like cancer.
  • Universally expressed genes can obscure cell identity: By developing a metric that down-weights these genes, ACTION can better highlight the subtle but crucial differences driven by cell-specific genes.
  • Geometric interpretation of cellular functions: Representing functions as corners of a geometric shape (polytope) provides an intuitive and powerful way to model cellular specialization.
  • Identifying novel melanoma subtypes: ACTION successfully identified previously unknown melanoma subtypes with distinct survival rates and therapeutic responses, highlighting its clinical relevance.
  • Reconstructing context-specific regulatory networks: The ability to build TRNs for specific cell states or subtypes allows for a deeper understanding of the molecular mechanisms driving their function and behavior.

Key Results

  • Improved cell type identification: ACTION outperformed or tied with state-of-the-art methods in cell type identification across multiple datasets, as measured by Adjusted Rand Index (ARI), F-score, and Normalized Mutual Information (NMI).
  • Robustness to noise: The ACTION metric demonstrated stable performance even with varying degrees of gene expression dropout.
  • Identification of novel melanoma subtypes: ACTION identified two novel, phenotypically distinct subclasses of MITF-high melanoma patients with differential survival rates and therapeutic responses.
  • Construction of TRNs: The method successfully constructed TRNs for identified archetypes, revealing key regulatory factors and potential therapeutic targets. For example, it identified MYC, LEF1, and E2F1 as critical factors in subclass C melanoma, associated with worse outcomes.
  • Scalability: ACTION was shown to scale more gracefully than other methods for larger datasets.

Limitations and Future Work

  • Computational cost: While ACTION scales well, the preconditioning step using the minimum volume ellipsoid method can be time-consuming, depending on the solver used.
  • Parameter tuning for other methods: The authors note that other methods often require user-specified parameters (e.g., number of cell types for SIMLR, subspace dimension for MDS/Isomap), whereas ACTION is largely non-parametric.
  • Rare cell type identification: While ACTION shows sensitivity to rare cell types, the paper notes that distinguishing between noise and true rare cell types can be challenging, with PreSPA+PCHA identifying percentages less than 2% as noise/outlier.
  • Future work: The authors suggest further analysis of the identified regulatory networks to understand their contribution to patient survival and to explore therapeutic indications for the identified melanoma subclasses. The developed framework is also applicable to other complex biological systems beyond cancer.

Summary updated

Abstract

Single-cell transcriptomic data has the potential to radically redefine our view of cell type identity. Cells that were previously believed to be homogeneous are now clearly distinguishable in terms of their expression phenotype. Methods for automatically characterizing the functional identity of cells, and their associated properties, can be used to uncover processes involved in lineage differentiation as well as sub-typing cancer cells. They can also be used to suggest personalized therapies based on molecular signatures associated with pathology. We develop a new method, called ACTION , to infer the functional identity of cells from their transcriptional profile, classify them based on their dominant function, and reconstruct regulatory networks that are responsible for mediating their identity. Using ACTION , we identify novel Melanoma sub-types with differential survival rates and therapeutic responses, for which we provide biomarkers along with their underlying regulatory networks.
BibTeX
@misc{mohammadi2016take,
  author = {Shahin Mohammadi and Vikram Ravindra and David F. Gleich and Ananth Grama},
  title = {{Take ACTION to characterize the functional identity of single cells}},
  year = {2016},
  doi = {10.1101/081273},
  howpublished = {bioRxiv},
}
Figure