TL;DR
This paper introduces Velorama, a novel computational framework that leverages RNA velocity data to infer causal gene regulatory networks (GRNs) from single-cell data. Velorama can handle complex, branching cell differentiation trajectories and, crucially, estimates the speed of transcription factor (TF) interactions, revealing a link between TF speed and their regulatory function in development and disease.
What Problem Does This Paper Address?
Understanding how genes interact to control cellular function is a fundamental goal in biology, often represented by gene regulatory networks (GRNs). While single-cell RNA sequencing (scRNA-seq) has revolutionized our ability to study these dynamics, inferring causal GRNs from this data remains challenging. Existing methods often rely on ordering cells along a linear "pseudotemporal" axis, which fails to capture the complex, branching nature of cell differentiation. Furthermore, many methods don't fully utilize the rich temporal information encoded in RNA velocity, a powerful technique that predicts future cell states. This gap limits our ability to uncover the precise causal relationships and the temporal dynamics of gene regulation, especially in the context of development and disease.
Methods
Velorama's core innovation lies in its ability to model cell differentiation as a directed acyclic graph (DAG) of cells, rather than a linear sequence. This DAG can be constructed from either pseudotime or, more powerfully, from RNA velocity data. Velorama directly incorporates RNA velocity-based cell-to-cell transition probabilities, allowing it to capture complex trajectories.
The inference of causal GRNs is performed using a generalized Granger causal (GC) inference framework adapted for DAGs. This approach looks for lagged interactions between transcription factors (TFs) and their target genes. A key feature of Velorama is its ability to estimate the speed of these TF-target interactions, based on the lag between a TF's expression change and its target's expression change. This is achieved through a multilayer neural network architecture that learns these relationships.
Datasets
The authors evaluated Velorama on:
- Four synthetic datasets generated by SERGIO, designed to simulate various differentiation landscapes (linear, bifurcation, trifurcation, and tree-shaped). These datasets provided ground-truth GRNs for benchmarking.
- Six real single-cell RNA-seq datasets, including:
- Mouse endocrinogenesis
- Mouse dentate gyrus neurogenesis
- Human hematopoiesis
- Cardiac hypertrophy in cardiomyocytes
- A549 cells responding to pracinostat
- Human corticogenesis (using both gene expression and chromatin accessibility data)
- Two acute myeloid leukemia (AML) datasets for studying disease progression and treatment response.
Key Insights and Ideas
- DAGs are better than linear models for complex differentiation: Representing cell states as a DAG, rather than forcing them into a linear pseudotime, is crucial for accurately capturing branching trajectories in cell differentiation.
- RNA velocity unlocks finer-grained causal inference: Incorporating RNA velocity data significantly improves GRN inference accuracy compared to using pseudotime alone, highlighting its power in capturing dynamic cellular processes.
- TF interaction speed matters: Velorama's ability to estimate TF interaction speeds is not just a technical feat; it reveals biologically meaningful distinctions. Slow TFs are linked to fundamental regulatory processes like chromatin remodeling and are associated with diseases like gliomas, while fast TFs are more involved in specific functions and linked to neuropsychiatric disorders.
- Cooperative regulation across different speeds: TFs with different interaction speeds often work together. Slow TFs (like chromatin remodelers) may cooperate with faster TFs to drive cell-type-specific functions and differentiation.
Key Results
- Substantial performance gains: Velorama outperformed existing GRN inference methods by 3.7–4.8x in terms of area under the precision-recall curve (AUPRC) on synthetic datasets.
- RNA velocity boosts accuracy: Using RNA velocity data improved AUPRC by 1.8–3x compared to pseudotime-based inference.
- Disease associations:
- Slow TFs were significantly linked to gliomas and were less cell-type specific.
- Fast TFs were implicated in neuropsychiatric diseases and were more cell-type specific.
- Pioneer factors are slow: Slow TFs were significantly enriched for pioneer factors, which are known to initiate chromatin accessibility.
- Corticogenesis insights: In human corticogenesis, slow TFs were associated with gliomas, while fast TFs were linked to neuropsychiatric disorders.
Limitations and Future Work
- Scalability: While Velorama is efficient, very large datasets (over 100,000 cells) might benefit from data sketching techniques.
- Integration of modalities: The authors suggest that integrating Velorama with other data modalities, such as chromatin accessibility and TF binding data, could further improve performance.
- Ensembling approaches: Combining pseudotime and RNA velocity predictions through ensembling could harness the unique strengths of both.
- Broader application of speed estimation: The ability to measure TF interaction speeds is a promising new direction for understanding the role of temporal control in gene regulation across various tissues and diseases.
Summary updated