DeepVelo: Single-cell Transcriptomic Deep Velocity Field Learning with Neural Ordinary Differential Equations

Zhanlin Chen, William C. King, Ahyeon Hwang, Mark Gerstein, Jing ZhangbioRxiv 2022 · DOI

TL;DR

DeepVelo is a novel framework that uses neural ordinary differential equations (ODEs) to model the complex, non-linear dynamics of gene expression in single cells over time. This approach allows for more accurate predictions of future cell states and the identification of key regulatory genes, overcoming limitations of previous methods.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) has revolutionized our ability to study cellular processes, but understanding the dynamic transitions between cell states remains a significant challenge. Traditional methods often struggle with the high dimensionality, sparsity, and non-linear nature of gene expression data. While RNA velocity offers a glimpse into immediate future states, it's limited to short timescales. Linear ODEs, though capable of longer-term predictions, fail to capture complex biological interactions like those involving transcription factors. This paper aims to bridge this gap by developing a method that can learn and model these intricate, non-linear dynamics over extended developmental periods.

Methods

DeepVelo combines a denoising variational autoencoder (VAE) with neural ODEs. The VAE is trained to learn the mapping between a cell's current gene expression state and its transcriptional velocity (the rate of change in gene expression). This learned mapping, represented by a neural network, effectively becomes a component of a neural ODE. By integrating this neural ODE with standard ODE solvers, DeepVelo can simulate the future (or past) gene expression states of individual cells. Key methodological contributions include:

  • Denoising VAE: The VAE is trained to not only predict RNA velocity but also to denoise the scRNA-seq data, improving the robustness of the learned dynamics.
  • Neural ODE Integration: The VAE's learned dynamics are embedded within an ODE framework, allowing for continuous modeling of gene expression changes over time.
  • Cell Criticality Index (CCI): A novel metric is introduced to quantify the instability of cell states, analogous to "kinetic energy" in dynamical systems, helping to identify critical transition points and fate commitment.
  • In Silico Perturbation Studies: The framework allows for computational experiments to identify "driver genes" that influence cell fate decisions.
  • Retrograde Trajectory Simulation: By reversing time, DeepVelo can reconstruct gene co-expression networks that are specific to developmental trajectories, revealing more coherent functional modules.

Datasets

The study utilized several publicly available scRNA-seq datasets:

  • Mouse pancreatic endocrinogenesis
  • Developing mouse dentate gyrus
  • Developing mouse neocortex (across multiple embryonic days)
  • Mouse gastrulation
  • Developing human forebrain

These datasets represent a variety of tissues, developmental timescales, and technical platforms, demonstrating the generalizability of DeepVelo. The paper does not introduce new benchmarks but compares DeepVelo against existing state-of-the-art methods like SparseVFC.

Key Insights and Ideas

  • Non-linear dynamics are crucial: The paper strongly emphasizes that linear models are insufficient for capturing the complexity of gene regulatory networks, particularly in developmental processes. DeepVelo's ability to learn non-linear relationships is a core strength.
  • Cellular dynamics can be chaotic: The study suggests that single-cell dynamical systems may exhibit chaotic properties, where small initial perturbations can lead to significantly different future states (the "butterfly effect"). This has implications for understanding developmental robustness and variability.
  • CCI as a measure of instability: The Cell Criticality Index provides a valuable new way to quantify how "unstable" or poised for change a cell state is, going beyond simple pseudotime ordering. It helps pinpoint critical junctures in development.
  • Retrograde analysis reveals trajectory-specific networks: Reversing developmental time allows for the discovery of gene co-expression networks that are specifically active during certain transitions, offering a more nuanced view than static network analysis.

Key Results

  • Improved Velocity Field Accuracy: DeepVelo demonstrated at least a 50% improvement in out-of-sample velocity field representation accuracy compared to state-of-the-art methods.
  • Accurate Trajectory Prediction: The framework successfully simulated developmental trajectories, accurately predicting cell states for out-of-sample cells and capturing complex transitions in various biological systems (pancreatic endocrinogenesis, dentate gyrus, neocortex).
  • Identification of Driver Genes: In silico perturbation studies successfully identified key genes that influence cell fate decisions, validated by existing biological knowledge and experimental results.
  • Enhanced Gene Co-expression Networks: Retrograde trajectory analysis using DeepVelo led to gene co-expression modules with significantly higher enrichment for cell-type-specific Gene Ontology (GO) terms (two orders of magnitude higher) compared to static cell clusters.

Limitations and Future Work

  • Single-cell dynamics only: Currently, DeepVelo models gene expression dynamics within individual cells. The authors suggest expanding the framework to incorporate interactions between spatially neighboring cells (using spatial transcriptomics) and integrating multi-modal data (protein, chromatin accessibility) for a more comprehensive representation of cell states.
  • Computational cost: While DeepVelo offers advantages, integrating neural ODEs can be computationally intensive. Future work could focus on optimizing computational efficiency.
  • Chaotic behavior implications: The observed chaotic behaviors in simulations warrant further investigation into their biological significance and potential role in emergent tissue-level phenomena. Synchronization models are proposed as a potential avenue for exploring this.

Summary updated

Abstract

Abstract Recent advances in single-cell RNA sequencing technology have provided unprecedented opportunities to simultaneously measure the gene expression profile and transcriptional velocity of individual cells, enabling us to sample gene regulatory network dynamics along developmental trajectories. However, traditional methods have faced challenges in modeling gene expression dynamics within individual cells due to sparse, non-linear (e.g., obligate heterodimer transcription factors), and high-dimensional measurements. Here, we present DeepVelo, a neural-network-based ordinary differential equation model that can learn non-linear, high-dimensional single-cell transcriptome dynamics and describe continuous gene expression changes within individual cells across time. We applied DeepVelo to multiple published datasets from different technical platforms and demonstrated its utility to 1) formulate transcriptome dynamics on different timescales, 2) measure the instability of cell states, and 3) identify developmental driver genes upstream of a signaling cascade. Benchmarking against state-of-the-art methods shows that DeepVelo can improve velocity field representation accuracy by at least 50% in out-of-sample cells. Further, perturbation studies revealed that single-cell dynamical systems may exhibit properties similar to those of chaotic systems. In summary, DeepVelo allows for the data-driven discovery of differential equations that delineate single-cell transcriptome dynamics. Teaser Embedding neural networks into ordinary differential equations to model gene expression changes within single cells across time.
BibTeX
@misc{chen2022deepvelo,
  author = {Zhanlin Chen and William C. King and Ahyeon Hwang and Mark Gerstein and Jing Zhang},
  title = {{DeepVelo: Single-cell Transcriptomic Deep Velocity Field Learning with Neural Ordinary Differential Equations}},
  year = {2022},
  doi = {10.1101/2022.02.15.480564},
  howpublished = {bioRxiv},
}
Figure