Inferring extrinsic factor-dependent single-cell transcriptome dynamics using a deep generative model

Yasuhiro Kojima, Yuko Arioka, Haruka Hirose, Shuto Hayashi, Yusuke Mizuno, Keiki Nagaharu, Hiroki Okumura, Masato Ishikawa, Kohshi Ohishi, Yutaka Suzuki, Norio Ozaki, Teppei ShimamurabioRxiv 2024 · DOI

TL;DR

This paper introduces ExDyn, a novel deep generative model that infers cell state dynamics from single-cell transcriptomes while explicitly accounting for extrinsic factors like experimental conditions and cell-cell interactions. ExDyn not only predicts these dynamics more accurately than existing methods but also identifies the specific extrinsic factors driving them, offering a powerful new tool for understanding cellular heterogeneity.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) has revolutionized our ability to study cellular heterogeneity and development. A key technique, RNA velocity, uses splicing kinetics to infer the future state of cells, providing insights into dynamic processes. However, existing RNA velocity methods largely ignore the crucial influence of extrinsic factors – external signals and environmental cues like experimental conditions, neighboring cells, or chromatin accessibility – on cell fate decisions. This oversight limits our understanding of how these external influences shape cellular trajectories, especially in complex biological systems where multiple factors interact. ExDyn aims to bridge this gap by developing a model that can disentangle and quantify the impact of these extrinsic factors on cell state dynamics.

Methods

ExDyn is built upon a deep generative model, specifically a conditional variational autoencoder (cVAE), which is a standard approach for scRNA-seq data analysis. The novelty lies in its integration of:

  • Splicing Kinetics: ExDyn explicitly models the relationship between spliced and unspliced RNA transcripts, a core principle of RNA velocity.
  • Extrinsic Factor Conditioning: The model is designed to incorporate and learn from various extrinsic factors (e.g., experimental conditions, colocalization scores, TF activity) as inputs.
  • Counterfactual Inference: ExDyn can predict how cell states would evolve under different hypothetical conditions, allowing for a deeper understanding of causal relationships.
  • Principal Mode Analysis: Using Singular Value Decomposition (SVD) on the Jacobian matrix of cell state dynamics with respect to extrinsic factors, ExDyn can identify the dominant modes of dynamic change driven by these factors and pinpoint the key contributing extrinsic features.

Datasets

The authors evaluated ExDyn using several datasets:

  • Simulated scRNA-seq data: Used to validate the model's ability to accurately estimate differential dynamics between conditions.
  • PERK-KO and WT neurosphere scRNA-seq data: To study the effect of gene knockout on cell state dynamics.
  • Human hematopoiesis multi-omics data: Combining transcriptome and chromatin accessibility data to investigate epigenetic regulation of cell state dynamics.
  • Human squamous cell carcinoma spatial transcriptome data: To analyze the influence of cell-cell communication and colocalization on tumor cell heterogeneity.

Key Insights and Ideas

  • Extrinsic factors are critical drivers: The study strongly suggests that extrinsic factors are not just noise but actively shape cell state dynamics, often more so than subtle differences in initial cell states.
  • Disentangling complex interactions: ExDyn's ability to identify specific extrinsic factors (e.g., a particular ligand, a gene knockout, or a specific TF activity) that drive distinct cellular behaviors is a major breakthrough.
  • Counterfactual predictions reveal causality: The model's capacity to predict "what if" scenarios allows researchers to move beyond correlation and infer causal relationships between environmental cues and cell fate.
  • Beyond simple RNA velocity: ExDyn extends the power of RNA velocity by providing a framework to understand why cells are moving in certain directions, not just where they are going.

Key Results

  • Superior Accuracy: ExDyn demonstrated higher accuracy in RNA velocity estimation compared to existing methods like scVelo, VeloVI, and CellDancer, particularly at bifurcation points where cell states are similar but dynamics diverge.
  • PERK KO Elucidation: ExDyn successfully identified how PERK knockout alters neurosphere differentiation dynamics, highlighting specific cell populations and their abundance changes.
  • Chromatin Influence: The model revealed that chromatin activity significantly influences hematopoietic stem cell differentiation dynamics, identifying key transcription factors involved.
  • Tumor Heterogeneity Drivers: In squamous cell carcinoma, ExDyn pinpointed fibroblasts and specific secreted ligands as key drivers of tumor cell heterogeneity and invasion.

Limitations and Future Work

  • Micro-duration focus: The current study focuses on short-term dynamics inherent to RNA velocity. The authors suggest extending ExDyn to infer growth and death rates, and to analyze longer time-course data for developmental or disease states.
  • Computational complexity: While not explicitly detailed, deep generative models can be computationally intensive. Future work might focus on optimizing efficiency.
  • Experimental validation: While the paper demonstrates biological plausibility and identifies candidates for experimental validation, further direct experimental confirmation of all findings would strengthen the conclusions.

Summary updated

Abstract

1 Abstract RNA velocity estimation helps elucidate temporal changes in the single-cell transcriptome. However, current methodologies for inferring single-cell transcriptome dynamics ignore extrinsic factors, such as experimental conditions and neighboring cell. Here, we propose ExDyn—a deep generative model integrated with splicing kinetics for estimating cell state dynamics dependent on extrinsic factors. ExDyn enables the counterfactual inference of cell state dynamics under different conditions. Among the extrinsic factors, ExDyn can extract key features which have large effects on cell state dynamics. ExDyn correctly estimated the difference in dynamics between two conditions and showed better accuracy over existing RNA velocity methods. ExDyn were utilized for unveiling the effect of PERK-knockout on neurosphere differentiation, hematopoietic stem cell differentiation driven by chromatin activity and the dynamics of squamous cell carcinoma cells dependent on colocalized neighboring cells. These results demonstrated that ExDyn is useful for analyzing key features in the dynamic generation of heterogeneous cell populations.
BibTeX
@misc{kojima2024inferring,
  author = {Yasuhiro Kojima and Yuko Arioka and Haruka Hirose and Shuto Hayashi and Yusuke Mizuno and Keiki Nagaharu and Hiroki Okumura and Masato Ishikawa and Kohshi Ohishi and Yutaka Suzuki and Norio Ozaki and Teppei Shimamura},
  title = {{Inferring extrinsic factor-dependent single-cell transcriptome dynamics using a deep generative model}},
  year = {2024},
  doi = {10.1101/2024.04.01.587302},
  howpublished = {bioRxiv},
}

Loading related content...

Figure