TL;DR
This paper introduces cell2fate, a novel Bayesian model for RNA velocity that improves cell fate prediction by using a "module-based" approach. By decomposing complex transcriptional dynamics into interpretable modules, cell2fate offers a more accurate and biologically insightful way to understand cellular differentiation, even in challenging scenarios like rare cell types or subtle transcriptional changes.
What Problem Does This Paper Address?
Understanding how cells differentiate and transition to new fates is fundamental to biology. RNA velocity is a powerful technique that uses the ratio of spliced to unspliced mRNA molecules in single cells to infer the future transcriptional state of a cell, essentially predicting its "velocity." However, existing RNA velocity models often rely on simplified biophysical assumptions or complex numerical methods to solve the underlying equations. These simplifications can lead to inaccuracies, especially when dealing with complex or subtle changes in gene expression, or in rare cell populations. The challenge is to develop a method that is both biophysically accurate and computationally tractable, allowing for a deeper and more reliable understanding of cellular dynamics.
Methods
The core innovation of cell2fate lies in its approach to solving the ordinary differential equations (ODEs) that govern RNA velocity. Instead of relying on coarse approximations or extensive numerical simulations, cell2fate linearizes the ODEs. This linearization allows the model to decompose the complex transcriptional dynamics into a set of interpretable "modules." Each module represents a simplified, independent component of transcriptional regulation with its own dynamics.
This modular approach offers several advantages:
- Biophysical Accuracy: It allows for a more accurate representation of transcriptional dynamics.
- Interpretability: The modules provide a direct link between RNA velocity and statistical dimensionality reduction techniques (like factor analysis), making the inferred dynamics more understandable.
- Computational Efficiency: The linearized ODEs can be solved analytically, making the model computationally efficient.
cell2fate is implemented as a fully Bayesian model, incorporating refinements to account for technical variations in scRNA-seq data (e.g., overdispersion, detection sensitivity, ambient RNA). It uses a hierarchical prior structure for regularization and is fitted using stochastic variational inference.
Datasets
The authors benchmarked cell2fate on five publicly available scRNA-seq datasets:
- Developing mouse dentate gyrus
- Developing mouse pancreas
- Mouse erythroid maturation
- Human bone marrow
- Mouse bone marrow (with low UMI counts)
They also applied cell2fate to a newly generated spatial transcriptomics dataset of the developing human brain.
Key Insights and Ideas
- RNA Velocity Modules as Building Blocks: The concept of decomposing complex transcriptional dynamics into a smaller set of interpretable modules is a significant conceptual advance. These modules act like "building blocks" of cellular differentiation.
- Bridging RNA Velocity and Dimensionality Reduction: cell2fate provides a direct biophysical link between RNA velocity and statistical dimensionality reduction methods, offering a more mechanistic interpretation of these techniques.
- Uncovering Subtle Dynamics: The modular approach allows cell2fate to capture weak dynamical signals and complex transcriptional changes that are often missed by other methods, particularly in rare or mature cell types.
- Spatial-Temporal Integration: The ability to map these RNA velocity modules onto spatial transcriptomics data opens up new avenues for understanding how temporal processes are organized within tissues.
Key Results
- Superior Performance in Benchmarks: cell2fate consistently outperformed ten other RNA velocity methods across five scRNA-seq datasets, as measured by the CBDir metric (cross-boundary directional correctness). It correctly inferred cell fate trajectories in all datasets, whereas most other methods failed in at least one setting.
- Resolving Challenging Trajectories: cell2fate successfully resolved the late maturation trajectory of granule neurons in the mouse dentate gyrus, a feat that eluded other methods. It also accurately captured complex "multi-rate kinetic" gene dynamics in mouse erythroid maturation.
- Granular Dissection of Maturation: In the developing human brain, cell2fate identified seven sequentially activated modules that finely dissected the maturation of excitatory neurons, revealing distinct temporal windows of gene expression.
- Spatial Mapping of Differentiation: By integrating with cell2location, cell2fate successfully mapped these neuronal differentiation modules to specific spatial locations within the developing human cortex, aligning with known patterns of progenitor localization and neuronal migration.
Limitations and Future Work
- Simplifying Assumptions: The authors acknowledge that cell2fate still makes simplifying assumptions, such as constant degradation rates and the absence of stochastic bifurcations.
- Future Extensions: They suggest future work could incorporate cell-specific splicing and degradation rates, stochastic rates at lineage branching points, and causal connections between transcription rates and dynamic gene regulatory networks.
- Cell-Cell Interactions: A key future direction is to integrate RNA velocity module mapping with spatial cell-cell interaction tools to identify putative interactions driving specific differentiation steps.
Summary updated