TL;DR
This paper introduces DiCoLo, a novel computational framework for identifying genes that change their co-expression patterns within specific, localized cell neighborhoods across different biological conditions. DiCoLo achieves this without requiring cell clustering or cross-condition data integration, making it robust to batch effects and capable of uncovering subtle biological signals.
What Problem Does This Paper Address?
Single-cell RNA sequencing (scRNA-seq) allows us to study gene expression at an unprecedented resolution, revealing the heterogeneity of cells within a tissue or organism. A key challenge is understanding how genes coordinate their expression to drive specific cellular functions and how these coordination patterns change between different biological states (e.g., healthy vs. diseased, or different developmental stages). Existing methods often struggle with this because they either rely on pre-defined cell clusters, which might miss dynamic changes within subpopulations, or they try to align cells from different conditions. This alignment process can be problematic when there are strong technical variations (batch effects) or significant biological differences between the datasets, potentially distorting the true biological signals. The paper aims to fill this gap by providing a method that can detect localized changes in gene co-expression without these limitations.
Methods
DiCoLo employs a three-step workflow:
- Optimal Transport (OT) Distance for Gene Co-localization: For each condition, DiCoLo first constructs a cell graph based on cell similarity. It then uses Optimal Transport (OT) distances to quantify how similarly two genes' expression patterns are distributed across the cell manifold. Genes expressed in overlapping cell neighborhoods will have a small OT distance, indicating co-localization.
- Differential Graph Operator Construction: The OT distances are used to build a gene graph for each condition. DiCoLo then constructs "differential graph operators" that highlight differences in connectivity (co-localization) between the two conditions. These operators are designed to capture genes that are densely connected in one condition but sparsely connected in the other.
- Spectral Analysis for Differential Gene Identification: The differential graph operators are subjected to eigenvalue decomposition. The eigenvectors corresponding to the largest eigenvalues represent genes exhibiting differential co-localization. These genes can then be further grouped into modules for biological interpretation.
A key innovation is that DiCoLo operates independently on each condition's data, avoiding the need for data integration or batch correction, which can be a major source of error in other methods.
Datasets
The paper uses several datasets for evaluation:
- Human PBMC10k dataset: Used for an initial demonstration with simulated co-localized gene signals.
- Three real scRNA-seq datasets for benchmarking:
- Human pbmc10k: A baseline with no simulated batch effects.
- Human pbmcsca: Introduces inter-donor biological variation.
- Human pancreas: Introduces cross-platform technical variation (Drop-seq vs. inDrops).
- Mouse embryonic skin datasets: Used to study dermal condensate genesis in response to morphogen signaling mutants (SmoM2 and Wls mutants).
The paper also describes a simulation framework that introduces synthetic co-localized expression patterns into real scRNA-seq datasets to rigorously test the method's performance under various conditions.
Key Insights and Ideas
- Localized co-expression is crucial: The paper highlights that important biological processes often involve genes coordinating their expression within specific, small groups of cells, not just across broad cell types.
- Integration-free is a game-changer: By avoiding data integration and batch correction, DiCoLo sidesteps a major pitfall in scRNA-seq analysis, making it more robust to real-world data imperfections.
- Optimal Transport for spatial relationships: Using OT to measure gene co-localization across the cell manifold is a novel way to capture the "neighborhood" aspect of gene expression.
- Graph comparison for differential signals: Framing the problem as comparing gene graphs allows for a powerful spectral analysis to pinpoint genes with altered coordination.
Key Results
- Robust performance in simulations: DiCoLo consistently outperformed existing methods (DGCA for differential co-expression, and miloDE/LEMUR for cluster-free differential expression) in identifying simulated differential co-localization signals across various datasets and simulation parameters (neighborhood size, dropout rate, number of perturbed genes).
- Sensitivity to weak signals and batch effects: DiCoLo showed superior accuracy even with weak signals and strong batch effects, where other methods struggled.
- Biological discovery in mouse development: When applied to mouse hair follicle development data, DiCoLo identified condition-specific gene modules that revealed key regulatory programs and emerging cell populations driven by morphogen signaling perturbations, providing insights into dermal condensate differentiation. For example, it identified modules related to precocious DC formation, cell cycle exit, and disrupted Wnt/SHH pathway coordination.
Limitations and Future Work
- Computational cost: While optimized, the OT computation can still be demanding for very large datasets (many cells or genes). The authors suggest strategies like graph coarse-graining and sparsification for such cases.
- Multiple replicates: The current framework is designed for two conditions. Incorporating multiple biological replicates would improve statistical reliability and enable more rigorous cross-condition comparisons.
- Extension to spatial transcriptomics: The concept of co-localization naturally extends to spatial transcriptomics, and the authors suggest this as a promising future direction to link gene coordination with spatially organized biological processes.
- Expression thresholding: DiCoLo applies an expression threshold to filter out broadly expressed genes, which is a design choice that might exclude some biologically relevant signals in very rare or uniformly expressed genes.
Summary updated