Skip to content

ENVI / COVET

Verdict: Conceptually the most sophisticated niche definition — niche as covariance tensor, not just mean. Captures regulatory structure invisible to simpler methods, at the cost of computational complexity.

Citation: Haviv D, et al. "The covariance environment defines cellular niches for spatial inference." Nature Biotechnology 42:1830-1842, 2024. DOI

Problem Setup

Can we define cell niches using the gene-gene covariance structure of their spatial neighbors, rather than just neighbor composition or mean expression? And can this richer niche representation improve imputation of whole transcriptomes from spatial data?

Niche Definition

Covariance-based: Niche = the gene-gene covariance tensor computed over a cell's spatial neighbors. Two neighborhoods with the same mean expression but different correlation structure represent different niches.

This is a fundamentally richer definition than composition-based (who is nearby) or expression-based (what is expressed nearby). It captures how genes co-vary with each other in the local environment.

Architecture

COVET (COVariance Environment Tensor):

  1. For each cell, identify its spatial neighbors.
  2. Compute the gene-gene covariance matrix across those neighbors for a selected set of genes.
  3. This covariance matrix is the COVET niche representation.

ENVI (ENVironment Imputation):

  1. Train on paired spatial + scRNA-seq data from the same tissue.
  2. Use COVET niche representations as conditioning variables in a conditional VAE.
  3. The VAE learns to impute full transcriptomes from spatial data conditioned on the local niche covariance.

Evaluation

  • Demonstrated on mouse brain (MERFISH + scRNA-seq), mouse intestine (Slide-seq + scRNA-seq), and other tissues.
  • Compared against imputation methods that use simpler niche definitions (mean expression, composition).
  • COVET-conditioned imputation outperforms alternatives, demonstrating that covariance carries biological information beyond means.

Honest Assessment

Strengths:

  • Theoretically compelling: two neighborhoods with the same average gene expression but different regulatory structure (different covariances) are biologically different. COVET captures this.
  • ENVI's imputation is a practical killer application — enables whole-transcriptome analysis from gene-panel spatial data.
  • From Dana Pe'er's lab (MSKCC), with strong theoretical foundations.

Limitations:

  • The covariance tensor grows quadratically with the number of genes. Practical only with gene subsets (typically ~50-200 genes) or dimensionality reduction.
  • Covariance estimation requires enough cells in each neighborhood for stable estimates — sparse regions may produce noisy tensors.
  • Harder to interpret biologically than composition or mean expression. What does it mean that genes A and B are correlated in neighborhood X but not Y?
  • Computationally more demanding than simpler niche methods.

Design Decision: COVET bets that regulatory structure (covariance) is more biologically meaningful than snapshot statistics (means, proportions). This is likely true — gene regulation is fundamentally about co-expression patterns. But the practical cost of computing and interpreting covariance tensors limits adoption.