Local transcriptional covariation produces accurate estimates of cell phenotype

Sinan Ozbay, Aditya Parekh, Rohit SinghbioRxiv 2023 · DOI

TL;DR

This paper introduces Sceodesic, a novel algorithm that uses differential geometry to analyze gene co-expression patterns within specific cell states. By moving beyond global analysis, Sceodesic accurately captures cell phenotypes, leading to better predictions of protein expression and a deeper understanding of biological processes.

What Problem Does This Paper Address?

Single-cell RNA sequencing (scRNA-seq) is a powerful tool for understanding cellular heterogeneity, but its measurements are inherently noisy and sparse. While individual gene expression levels can be unreliable, researchers have turned to "gene programs" – sets of co-varying genes – to better represent cell states. However, existing methods for analyzing gene co-expression, like PCA and NMF, tend to focus on global patterns across the entire dataset. This can obscure subtle, cell-state-specific variations and lead to biologically irrelevant interpretations. The core challenge is the lack of a mathematically rigorous and biologically faithful framework for quantifying and comparing local gene co-expression patterns within distinct cell states.

Methods

The authors introduce Sceodesic, an algorithm that leverages differential geometry to analyze gene co-expression. The key innovation is the use of the log-Euclidean metric, which respects the inherent mathematical structure (the Riemannian manifold) of gene co-expression matrices. Unlike naive Euclidean distance, this metric accurately captures the "curvature" of these matrices, providing a more meaningful comparison of cell states.

Sceodesic works by:

  1. Cell Cohort Construction: Grouping cells into "cohorts" based on similar transcriptional profiles using clustering methods (like Gaussian Mixture Models).
  2. Local Covariance Estimation: Computing covariance matrices for each cell cohort, representing local gene co-expression. These matrices are regularized to ensure stability.
  3. Population Value Decomposition (PVD): Learning a shared set of "study-wide" gene programs (eigenvectors) from the local cohort programs. This allows for a common basis to compare across all cell states.
  4. Reconstruction and Program Expression: Reconstructing local co-expression matrices using the shared gene programs and estimating cell-specific "program expression" levels.

Datasets

The authors validate Sceodesic using several large-scale scRNA-seq datasets, including:

  • Human Lung Atlas (2.3 million cells)
  • One-million cell survey of embryonic development (Cao et al.)
  • Human Breast Cell Atlas (800,000 cells)
  • DCM/ACM Heart Cell Atlas (881,000 cells)

They also utilize:

  • CITE-seq data: For predicting protein expression levels from RNA expression.
  • Perturb-seq data: To assess how well Sceodesic distinguishes between wild-type and perturbed cell states.
  • Simulated data: Generated using the SERGIO simulator to test robustness against noise and varying regulatory network structures.

Key Insights and Ideas

  • Local is More Informative: Focusing on gene co-expression within specific cell states, rather than globally, is crucial for accurately capturing cell phenotypes.
  • Geometry Matters: Gene co-expression matrices live on a curved space (a Riemannian manifold), and naive Euclidean distance fails to respect this geometry, leading to inaccurate comparisons. The log-Euclidean metric provides a mathematically sound way to navigate this space.
  • Gene Programs as Phenotype Signatures: Sceodesic's interpretable gene programs act as powerful, low-dimensional signatures that can predict other biological modalities (like protein expression) and accurately reflect experimental perturbations.
  • Bridging the Transcriptome-Proteome Gap: Sceodesic's programs show significant promise in bridging the gap between RNA expression and protein abundance, a long-standing challenge.

Key Results

  • Superior Prediction of Protein Expression: Sceodesic outperformed PCA and NMF in predicting surface protein expression levels from scRNA-seq data, achieving the best performance for 60% of proteins tested.
  • Accurate Perturbation Analysis: Sceodesic effectively distinguished between wild-type and perturbed cell states and grouped functionally similar perturbations, outperforming PCA and NMF in case-control separation.
  • Meaningful Developmental Signatures: Sceodesic identified gene programs that strongly correlated with developmental time across various cell types, capturing biologically relevant processes like neuron projection.
  • Robustness to Noise: The log-Euclidean metric demonstrated robustness to simulated dropout and outlier noise, a common issue in scRNA-seq data.

Limitations and Future Work

  • Hyperparameter Sensitivity: The performance of Sceodesic is influenced by hyperparameters like the number of cell cohorts and variable genes. The authors provide recommendations but acknowledge that optimal choices may vary.
  • Computational Cost: While efficient, Sceodesic's runtime scales linearly with dataset size. The authors note that replacing sparse PCA with regular PCA can speed up computation if interpretability is less critical.
  • Future Directions: The authors suggest extending Sceodesic to incorporate both first and second moments of gene expression, creating multivariate normal representations. They also envision its application to other fields where co-expression is important, such as analyzing protein abundance across cohorts.

Summary updated

Abstract

Abstract The utility of single-cell RNA sequencing (scRNA-seq) is premised on the notion that transcriptional state can faithfully reflect cell phenotype. However, scRNA-seq measurements are noisy and sparse, with individual transcript counts showing limited correlation with cell phenotype markers such as protein expression. To better characterize cell states from scRNA-seq data, researchers analyze gene programs–—sets of covarying genes–—rather than individual transcripts. We hypothesized that more accurate estimation of gene covariation, especially at a local (i.e., cell-state) rather than global (i.e., experimental) scale, could better capture cell phenotypes. However, the field lacks appropriate mathematical frameworks for analyzing gene covariation: coexpression is quantified as a symmetric positive-definite matrix, where even basic operations like arithmetic differences lack biological interpretability. Here we introduce Sceodesic, which exploits the Riemannian manifold structure of gene coexpression matrices to quantify cell state-specific coexpression patterns using the log-Euclidean metric from differential geometry. Unlike principal components analysis and non-negative matrix factorization, which infer only global covariation, Sceodesic efficiently discovers local covariation patterns and organizes them into interpretable, linear gene programs. Sceodesic outperforms existing approaches in predicting protein expression levels, distinguishing transcriptional responses to gene perturbations, and identifying biologically meaningful programs in fetal development. By respecting the mathematical structure of gene coexpression, Sceodesic bridges the gap between biological variability and statistical analysis of scRNA-seq data, enabling more accurate characterization of cell phenotypes. Software availability https://singhlab.net/Sceodesic
BibTeX
@misc{ozbay2023local,
  author = {Sinan Ozbay and Aditya Parekh and Rohit Singh},
  title = {{Local transcriptional covariation produces accurate estimates of cell phenotype}},
  year = {2023},
  doi = {10.1101/2023.11.09.566448},
  howpublished = {bioRxiv},
}
Figure