A deep generative model for estimating single-cell RNA splicing and degradation rates

Chikara Mizukoshi, Yasuhiro Kojima, Satoshi Nomura, Shuto Hayashi, Ko Abe, Teppei ShimamurabioRxiv 2023 · DOI

TL;DR

This paper introduces DeepKINET, a novel deep generative model that accurately estimates single-cell RNA splicing and degradation rates from standard scRNA-seq data. By overcoming limitations of previous methods that assumed uniform rates, DeepKINET reveals crucial cellular heterogeneity in post-transcriptional regulation, offering new insights into gene expression dynamics and disease mechanisms.

What Problem Does This Paper Address?

Gene expression is tightly regulated at multiple levels, with RNA splicing and degradation playing critical roles in determining the abundance of functional proteins. Aberrations in these processes are linked to various diseases, including cancer. While methods exist to estimate these "kinetic rates," they often struggle with the inherent variability within cell populations. Existing approaches either rely on specialized metabolic labeling techniques, which are not widely applicable, or make simplifying assumptions of uniform rates across all cells for a given gene. This uniformity assumption fails to capture the nuanced, cell-specific regulatory mechanisms that contribute to cellular heterogeneity and can lead to misinterpretations of biological processes. The authors aim to address this gap by developing a method that can estimate these kinetic rates at a single-cell resolution using readily available scRNA-seq data.

Methods

The core of DeepKINET is a deep generative model, specifically a variational autoencoder (VAE), that leverages RNA velocity principles. The model takes unspliced and spliced mRNA counts from scRNA-seq data as input. It first learns a latent representation of cell states and then uses this to predict the splicing and degradation rates for each gene in each cell. A key innovation is that these rates are modeled as cell-state-dependent, meaning they can vary from cell to cell, directly addressing the limitation of previous methods. The model is trained by optimizing its ability to reconstruct unspliced mRNA amounts, effectively learning the underlying kinetic parameters. DeepKINET also incorporates a conditional VAE framework to handle multiple experimental batches, allowing for more robust analysis across different datasets.

Datasets

The authors evaluated DeepKINET using several datasets:

  • Simulated data: Generated using SERGIO, with varying splicing and degradation rates, dropout rates, and cell numbers to rigorously test the model's accuracy and robustness.
  • Metabolic labeling data:
    • scEU-seq cell cycle dataset: Used to assess kinetic rate estimation against experimental data with temporal resolution.
  • scNT-seq hematopoietic dataset: Another metabolic labeling dataset used to evaluate transcription and degradation rate estimation.
  • Real-world biological datasets:
    • Forebrain dataset: Used to investigate RNA-binding protein (RBP) functions and their impact on gene clusters.
  • Breast cancer dataset: Applied to identify genes with altered kinetic rates in primary versus metastatic cells and to study RBP roles in cancer progression.
  • Erythroid lineage cells from MDS-RS patients: Used to analyze the effects of splicing factor mutations (SF3B1) on target gene kinetics.

Key Insights and Ideas

  • Cell-specific kinetics are crucial: The most significant insight is that RNA splicing and degradation rates are not uniform across cells for a given gene. DeepKINET's ability to capture this cell-to-cell variability is key to understanding complex biological processes and diseases.
  • Uncovering RBP functions: By clustering genes based on their kinetic rates, DeepKINET can identify groups of genes regulated by the same RNA-binding proteins, providing a powerful way to infer the functional roles of these proteins in splicing and degradation.
  • Splicing factor mutations have distinct kinetic impacts: The study demonstrates that mutations in splicing factors like SF3B1 lead to significantly altered splicing rates in target genes, offering a mechanistic link between genetic alterations and cellular dysfunction.
  • Kinetic changes in cancer metastasis: DeepKINET identified genes with altered splicing and degradation rates that are implicated in breast cancer metastasis, highlighting the role of post-transcriptional regulation in disease progression.

Key Results

  • Superior accuracy: DeepKINET consistently outperformed existing methods like cellDancer and DeepVelo in estimating splicing and degradation rates on simulated and metabolic labeling datasets, showing higher correlation coefficients with ground truth values.
  • Robustness to dropouts: The model demonstrated robustness to varying dropout rates, a common challenge in scRNA-seq data.
  • Identification of RBPs: In the forebrain dataset, DeepKINET successfully identified RNA-binding proteins (e.g., RBFOX1, RBFOX2) whose expression correlates with target gene splicing rates, validating known functions and suggesting new ones.
  • Cancer metastasis insights: In breast cancer, DeepKINET identified genes with kinetic rate changes associated with metastasis, including KDM6A and PRKAA1, and revealed how RBM47's influence on splicing rates differs between primary and metastatic cells.
  • Mutation effects: Analysis of SF3B1 mutations in erythroid cells showed significantly lower splicing rates in mutant cells, confirming DeepKINET's ability to capture the kinetic consequences of splicing factor dysregulation.

Limitations and Future Work

  • Correlation between splicing and degradation rates: The authors acknowledge that their unified model for estimating splicing and degradation rates can lead to some correlation between these estimates. While DeepKINET shows lower correlations than some other methods, suggesting better independence, this remains an area for improvement.
  • Fixed transcription rates: Currently, DeepKINET assumes fixed transcription rates. The authors suggest that extending the model to also estimate transcription rates dynamically could further enhance accuracy.
  • mRNA isoform complexity: The current RNA velocity analysis struggles to distinguish between mRNA isoforms, which is particularly relevant for diseases involving alternative splicing. Future work could address this to provide deeper biological insights.
  • Methodological enhancements: The authors mention the need for further methodological enhancements and additional constraints to improve the accuracy of kinetic rate estimation, especially given the challenges in simultaneously estimating RNA velocity and kinetic rates.

Summary updated

Abstract

A bstract Messenger RNA splicing and degradation are critical for gene expression regulation, the abnormality of which leads to diseases. Previous methods for estimating kinetic rates have limitations, assuming uniform rates across cells. We introduce DeepKINET, a deep generative model that estimates splicing and degradation rates at single-cell resolution from scRNA-seq data. DeepKINET outperformed existing methods on simulated and metabolic labeling datasets. Applied to forebrain and breast cancer data, it identified RNA-binding proteins responsible for kinetic rate diversity. DeepKINET also analyzed the effects of splicing factor mutations on target genes in erythroid lineage cells. DeepKINET effectively reveals cellular heterogeneity in post-transcriptional regulation.
BibTeX
@misc{mizukoshi2023deep,
  author = {Chikara Mizukoshi and Yasuhiro Kojima and Satoshi Nomura and Shuto Hayashi and Ko Abe and Teppei Shimamura},
  title = {{A deep generative model for estimating single-cell RNA splicing and degradation rates}},
  year = {2023},
  doi = {10.1101/2023.11.25.568659},
  howpublished = {bioRxiv},
}
Figure