TL;DR
This paper introduces Genes2Genes (G2G), a novel computational framework for aligning single-cell RNA sequencing (scRNA-seq) trajectories. G2G accurately captures both matches and mismatches at the gene level, overcoming limitations of existing methods and enabling deeper insights into cellular dynamics, disease progression, and optimization of in vitro culture conditions.
What Problem Does This Paper Address?
Single-cell RNA sequencing allows us to track dynamic changes in cell populations over time, space, or in response to perturbations. To compare these dynamic processes between different conditions (e.g., in vitro vs. in vivo differentiation, or healthy vs. diseased states), researchers often align their "trajectories" – essentially, the inferred developmental paths of cells. Traditional methods, like Dynamic Time Warping (DTW), are good at finding overall similarities but struggle with:
- Assumption of definite matches: They often assume every point in one trajectory must have a corresponding match in the other.
- Handling mismatches: They can't effectively identify or quantify situations where parts of a trajectory are missing or significantly different (like insertions or deletions in biological sequences).
- Focus on average expression: They typically rely on the mean expression of genes, ignoring the full distribution, which can be noisy in single-cell data.
This gap limits our ability to precisely pinpoint where and why biological processes diverge between conditions, hindering efforts to understand disease or optimize experimental systems.
Methods
Genes2Genes (G2G) tackles these challenges with a novel dynamic programming (DP) framework that unifies aspects of both Dynamic Time Warping (DTW) and biological sequence alignment (specifically, Gotoh's algorithm). Key innovations include:
- Five-state alignment: G2G models five states: match (M), compression warp (W), expansion warp (V), insertion (I), and deletion (D). This allows for nuanced representation of similarities and differences.
- Bayesian information-theoretic cost function: Instead of simple distance metrics, G2G uses a Minimum Message Length (MML) inference approach to quantify the "cost" of a match. This considers the full gene expression distribution (mean and variance) for each time point, making it more robust to noise and capturing distributional differences.
- Gene-level alignment: G2G performs alignment independently for each gene, providing a detailed view of how individual gene expression dynamics align across trajectories.
- Distributional interpolation: Before alignment, G2G interpolates gene expression data to create smooth, time-aligned trajectories, accounting for cell density and estimating Gaussian distributions at each interpolated time point.
Datasets
G2G was evaluated on:
- Simulated datasets: A benchmark dataset was generated using Gaussian Processes to simulate three main patterns: Matching, Divergence, and Convergence, with variations in bifurcation points. This allowed for controlled testing of G2G's accuracy against existing methods.
- Published real-world datasets:
- Murine bone marrow-derived dendritic cells stimulated with PAM3CSK or LPS.
- Idiopathic Pulmonary Fibrosis (IPF) dataset comparing healthy lung epithelial cells to diseased cells.
- Human T cell development data, comparing in vitro differentiation in an artificial thymic organoid (ATO) system with in vivo T cell development.
Key Insights and Ideas
- Mismatches are informative: G2G's ability to explicitly model mismatches (insertions/deletions) is crucial. These aren't just noise; they can indicate missing cell states, differential gene expression, or significant shifts in biological processes.
- Gene-level resolution is powerful: Aligning each gene individually reveals heterogeneity in how different genes contribute to trajectory differences, which would be masked by a single, global alignment.
- In vitro vs. in vivo T cell development: G2G identified that TNFɑ signaling, critical for in vivo T cell maturation, is not fully recapitulated in vitro. This suggests specific targets for optimizing in vitro culture conditions to better mimic in vivo development.
- Disease insights: In IPF, G2G highlighted differences in epithelial-mesenchymal transition (EMT) pathways, pinpointing potential therapeutic targets to prevent pathological cell state differentiation.
Key Results
- High accuracy: G2G demonstrated superior accuracy in identifying expected alignment patterns on simulated datasets compared to CellAlign and TrAGEDy, achieving accuracy rates of 98.6-100% for various patterns.
- Robustness to noise: The MML-based cost function, considering full gene expression distributions, makes G2G more robust to the inherent noise in scRNA-seq data.
- Identified biological differences:
- In T cell development, G2G revealed that in vitro differentiated T cells matched an immature in vivo state and lacked TNFɑ signaling expression, guiding potential optimization strategies.
- In IPF, G2G identified EMT as a key pathway with differences occurring at early/mid differentiation stages, suggesting therapeutic targets.
- Improved clustering: G2G's gene-level alignments enabled more accurate hierarchical clustering of genes with similar alignment patterns.
Limitations and Future Work
- Pseudotime reliability: The accuracy of G2G's alignment is dependent on the quality of the input pseudotime estimates. The authors recommend using reliable pseudotime inference methods.
- Linear trajectories: The current G2G version is designed for comparing two linear trajectories. Future work will explore extending it to handle branched trajectories, which are common in developmental processes.
- Parameter optimization: While G2G provides default parameters, further research could automate the inference of parameters like the window size for interpolation and transition probabilities for the five-state machine.
- Adaptive window size: An adaptive window size for Gaussian kernel-based interpolation could potentially improve sensitivity to variance in cell expression.
Summary updated