TL;DR
This paper introduces GeneTrajectory, a novel method for inferring gene dynamics in single-cell RNA sequencing data. By focusing on gene trajectories rather than cell trajectories, GeneTrajectory can disentangle complex, concurrent biological processes that often confound traditional approaches, leading to a more accurate understanding of gene regulation during development and differentiation.
What Problem Does This Paper Address?
Single-cell RNA sequencing (scRNA-seq) has revolutionized our ability to study cellular processes like differentiation and development. A common approach is to infer a "pseudotime" for each cell, representing its progression through a biological process. However, when cells are undergoing multiple, independent processes simultaneously (e.g., cell cycle and differentiation), this single pseudotime can become ambiguous, mixing signals and obscuring the true dynamics of individual gene programs. This makes it difficult to pinpoint the specific genes driving each distinct process.
Methods
GeneTrajectory tackles this challenge by shifting the focus from cells to genes. Its core innovation lies in using optimal transport (OT) metrics, specifically the Wasserstein distance, to quantify the dissimilarity between the expression patterns of different genes across a cell-cell graph.
Here's a breakdown of the key steps:
- Cell Graph Construction: A k-nearest neighbor (kNN) graph is built from a low-dimensional embedding of the cells, preserving the cell manifold structure. This graph defines the "road map" for gene expression flow.
- Gene-Gene Wasserstein Distance: The expression of each gene is treated as a distribution across the cells. The Wasserstein distance is then computed between these gene distributions. This metric is robust to noise and captures how gene expression patterns co-vary across the cell graph, accounting for the cell graph's geometry.
- Gene Trajectory Inference: Pairwise gene-gene distances are converted into affinities, forming a gene graph. Diffusion maps are then used to identify distinct gene trajectories, representing coordinated gene programs.
- Gene Ordering: Genes within each identified trajectory are ordered based on their spectral embedding, revealing their pseudotemporal sequence within that specific biological process.
The method is designed to be computationally efficient, employing strategies like cell graph coarse-graining and gene graph sparsification.
Datasets
The authors evaluated GeneTrajectory using:
- Simulated scRNA-seq datasets: These datasets were generated to mimic various biological scenarios, including cyclic processes, differentiation with branching, and combinations of linear/cyclic and multi-level differentiation/cyclic processes. These simulations allowed for ground truth comparisons.
- Human myeloid lineage dataset: Publicly available 10x Genomics PBMC data was used to study monocyte maturation and dendritic cell differentiation.
- Mouse embryonic skin dataset: Data from wildtype and Wntless (Wls) knockout mouse embryos was used to investigate dermal condensate (DC) genesis and the role of Wnt signaling.
Key Insights and Ideas
- Genes, not cells, define trajectories: The fundamental insight is that distinct biological processes are driven by coordinated gene programs, and tracking these gene programs directly is more robust than relying on a single cell pseudotime when multiple processes are active.
- Optimal transport for gene similarity: Using Wasserstein distance to compare gene expression distributions across a cell graph is a powerful way to capture gene co-regulation, robustly handling noise and the underlying cell manifold structure.
- Deconvolving concurrent processes: GeneTrajectory excels at separating intertwined biological processes (like cell cycle and differentiation) that confound traditional cell trajectory methods. This is crucial for understanding complex developmental pathways.
- Uncovering hidden gene programs: The method can identify gene programs that might be obscured by noise or the presence of other concurrent processes, leading to new discoveries about biological regulation.
Key Results
- GeneTrajectory accurately identified gene trajectories and their pseudotemporal order in simulations, outperforming traditional cell trajectory methods in gene ordering tasks.
- In the myeloid lineage differentiation dataset, GeneTrajectory successfully resolved three distinct gene trajectories corresponding to early monocyte maturation, later CD16+ monocyte differentiation, and type-2 dendritic cell differentiation.
- For the mouse skin development data, GeneTrajectory deconvolved cell cycle and dermal condensate differentiation processes, revealing that DC progenitors actively proliferate and exit the cell cycle at the end of differentiation.
- The method identified specific biological defects in Wls mutant embryos by showing how attenuated Wnt signaling disrupts the DC differentiation gene program.
- GeneTrajectory demonstrated robustness to variations in cell numbers and sparsity levels in simulations.
Limitations and Future Work
- Branch identification: The current branch identification procedure requires interactive optimization and might be unstable if branches vary significantly in length and size.
- Directionality: GeneTrajectory does not automatically infer the direction of progression along each trajectory; this must be determined by inspecting endpoint genes.
- Uninformative genes: Genes that are homogeneously expressed across many cells ("uninformative genes") can sometimes be colocalized with genes of interest in the gene embedding, potentially causing confusion.
- Future Work: The authors suggest exploring the application of their optimal transport framework to other single-cell modalities (e.g., ATAC-seq, spatial transcriptomics) and for multi-modal data integration. They also propose investigating the use of gene-gene graphs to improve cell organization.
Summary updated