Cell type composition drives patient stratification in single-cell RNA-seq cohorts

This study demonstrates that simple, interpretable cell-type composition metrics, particularly centered log-ratio-transformed proportions, outperform complex computational methods for unsupervised patient stratification in single-cell RNA-seq cohorts by capturing clinically relevant variation driven by cellular heterogeneity, and introduces the open-source R package scECODA to facilitate this approach.

The Gist

Imagine you are trying to understand a city by looking at a single, giant photograph of its entire population. For years, scientists did this with "bulk" biology: they took a tissue sample (like a piece of a tumor or a drop of blood), mashed it all together, and measured the average activity of every gene. It was like taking a photo of a crowd and saying, "The average person here is wearing a blue shirt." You missed the fact that half the crowd was wearing red, a few were wearing green, and the people in red were the ones actually running the show.

Then came Single-Cell RNA Sequencing (scRNA-seq). This technology is like having a camera that can take a high-definition photo of every single person in that crowd individually. Suddenly, we can see exactly who is there: the doctors, the construction workers, the artists, and the security guards.

But here's the new problem: Data Overload.
If you have a cohort of 100 patients, and each patient has 10,000 individual cells, you are looking at a million data points. Trying to find patterns in this massive, chaotic crowd using complex, super-smart AI computers is like trying to find a specific conversation in a stadium by listening to every single voice at once. It's slow, expensive, and often gets confused by the noise.

The Paper's Big Discovery: "The Crowd Composition"

The authors of this paper asked a simple question: Do we really need to listen to every single voice to understand the crowd?

They tested a bunch of fancy, complex computer methods against a much simpler idea: Just count the people.

They realized that in many diseases, the most important thing isn't what the individual cells are saying, but who is in the room and how many of them there are.

• In a healthy lung, you might have 50% Type A cells and 10% Type B cells.
• In a diseased lung, you might have 10% Type A and 50% Type B.

The "fancy" methods tried to analyze the complex gene conversations of every cell. The "simple" method just counted the heads.

The Result? The simple method won. Every time.
It was faster, cheaper, and actually better at separating sick patients from healthy ones than the super-complex AI models.

The Secret Sauce: "The Compositional Recipe"

The authors didn't just count heads; they used a specific mathematical trick called Centered Log-Ratio (CLR) transformation.

Think of it like baking a cake.

• If you have a recipe that says "1 cup of flour, 1 cup of sugar, 1 cup of eggs," and you accidentally add 2 cups of flour, you have to take something else away to keep the bowl full. The proportions change.
• In biology, if one type of cell multiplies, the percentage of all other cells automatically goes down, even if their actual numbers didn't change. This is called "compositional data."

Most computers get confused by this. They think, "Oh, the sugar went down, so the cake is ruined!" But the authors' method (which they call ECODA) understands the math of the recipe. It knows that if the flour went up, the sugar had to go down relatively, and it adjusts the math to see the real story.

Why This Matters (The "Aha!" Moments)

1. Simplicity is King: You don't need a supercomputer to find patient groups. A simple count of cell types, processed with the right math, works better than complex deep learning models. It's like realizing you can navigate a city with a simple map and a compass, rather than needing a GPS that tries to predict every traffic light.
2. The "Star Players": The study found that usually, only a tiny handful of cell types (maybe 5 or 10 out of 50) are responsible for the differences between patients. It's like realizing that in a soccer match, only the forwards and the goalkeeper really determine the score; the rest of the team is just doing their job.
3. It's Harder to Fake: Complex computer models often get tricked by "batch effects" (technical glitches, like taking photos with different cameras). The simple "count the heads" method is surprisingly tough to fool. It sees the biological truth even when the technical data is messy.
4. Real-World Translation: Because this method is so simple, it's easy to turn into a real-world medical test. Instead of needing a million-dollar machine to sequence every cell, a doctor might just need a simple test to count two specific types of cells (like a "Neutrophil-to-Lymphocyte Ratio") to predict if a patient will respond to cancer treatment.

The Takeaway

The paper introduces a new tool called scECODA. It's an open-source software package that lets researchers skip the complicated, slow, expensive AI models and go straight to the heart of the matter: Who is in the crowd, and in what numbers?

It turns out that for understanding disease and grouping patients, we don't need to overthink the details. Sometimes, the most powerful insight is just knowing who is showing up to the party and how many of them there are.

Technical Summary

1. Problem Statement

Single-cell RNA sequencing (scRNA-seq) offers high-resolution insights into cellular heterogeneity, yet translating this data into clinically meaningful patient stratification remains computationally challenging.

• The Gap: While bulk transcriptomics has long used unsupervised analysis to identify patient subgroups, scRNA-seq requires summarizing thousands of single-cell profiles into a sample-level representation.
• The Challenge: Existing state-of-the-art (SOTA) methods (e.g., deep generative models, tensor factorization, optimal transport) are computationally expensive and often fail to explicitly account for the compositional nature of cell-type data. Cell-type proportions are constrained to sum to one (simplex space), meaning a change in one cell type's abundance inherently alters the relative proportions of all others. Standard Euclidean distance metrics applied to raw proportions can distort biological relationships.
• The Question: Can simple, compositionally-aware representations of cell-type abundance outperform complex, high-dimensional embedding methods for unsupervised patient stratification?

2. Methodology

The authors conducted a comprehensive benchmark across 11 diverse scRNA-seq cohorts (697 samples) covering various biological conditions (cancer, autoimmune, infectious diseases, aging).

A. Proposed Method: ECODA

The authors introduced Exploratory COmpositional Data Analysis (ECODA), a baseline approach based on:

1. Cell-Type Annotation: Aggregating single cells into known cell types.
2. Compositional Transformation: Applying a Centered Log-Ratio (CLR) transformation to cell-type counts. This converts data from the simplex space to Euclidean space, handling the constant-sum constraint and zero-inflation (via pseudocounts).
   • Formula: $clr(x) = \ln(x_i / g(x))$, where $g(x)$ is the geometric mean of all components.
3. Dimensionality Reduction: Using Principal Component Analysis (PCA) on the CLR-transformed matrix to generate sample embeddings.

B. Benchmarking Framework

The study compared ECODA against:

• Baselines: Pseudobulk gene expression (average gene expression per sample) and raw cell-type frequencies (without log-ratio transformation).
• SOTA Methods:
  • Deep Generative Models: MrVI, scPoli.
  • Distributional/Density Methods: PILOT, GloScope, GloProp.
  • Factor Decomposition: MOFA+, scITD.
• Evaluation Metrics:
  • Unsupervised Separation: Adjusted Rand Index (ARI), Graph Modularity, and Analysis of Similarities (ANOSIM) to measure how well methods recover known biological ground truths (e.g., disease status, treatment response).
  • Robustness: Sensitivity to batch effects and different cell-type annotation strategies (expert vs. automated vs. unsupervised clustering).
  • Efficiency: Computational runtime and memory usage.

C. Software Tool

The authors developed scECODA, an open-source R package to facilitate this workflow, handling data transformation, distance calculation, and visualization.

3. Key Results

A. Performance Superiority of Simple Baselines

• ECODA Outperforms SOTA: ECODA consistently achieved the highest separation scores (ARI, Modularity, ANOSIM) across the majority of datasets and biological conditions.
• Pseudobulk vs. Composition: Whole-sample pseudobulk (average gene expression) also performed surprisingly well, ranking second or third. However, when pseudobulk was calculated per cell type (isolating transcriptional changes from compositional shifts), performance dropped significantly. This suggests that compositional shifts are the primary driver of inter-sample variation in these cohorts, rather than cell-type-specific transcriptional reprogramming.
• Importance of CLR: Methods using raw frequencies or arcsine transformations performed significantly worse than CLR-transformed data, confirming the necessity of proper compositional data analysis.

B. Computational Efficiency

• Speed: ECODA and Pseudobulk embeddings were generated in seconds on standard hardware.
• Cost: SOTA methods (MrVI, scPoli, GloScope) required hours of computation on GPU-enabled hardware and often necessitated downsampling large cohorts to avoid out-of-memory errors. ECODA scales linearly and is feasible for cohorts of hundreds of samples without specialized hardware.

C. Robustness and Interpretability

• Batch Effects: ECODA demonstrated superior robustness to technical batch effects (e.g., 3' vs. 5' sequencing chemistries). In contrast, pseudobulk gene expression representations were heavily confounded by batch effects, obscuring biological signals.
• Annotation Granularity: ECODA performance was robust across different annotation strategies. While high-resolution expert labels yielded the best results, unsupervised Leiden clustering and automated tools (HiTME, scATOMIC) achieved comparable performance. Crucially, performance dropped only when granularity was too coarse (e.g., broad cell lineages).
• Signal Concentration: The stratification signal was often driven by a small subset of Highly Variable Cell types (HVCs). In many datasets, retaining only the top 12–29% of cell types (by variance) was sufficient to maintain high stratification performance.

D. Specific Examples

• Adams (Lung Fibrosis): Separation of IPF, COPD, and Control was driven almost entirely by two cell types: Alveolar Type 2 (ATII) and Peribronchial Vascular Endothelial cells.
• Kfoury (Prostate Metastasis): Immature B cells and Tumor-Inflammatory Monocytes (TIMs) effectively separated tumor locations.
• Gong & Sharma (Aging/CMV): CLR-transformed proportions clearly separated individuals by age and CMV serostatus, driven by specific T-cell subsets (e.g., $\gamma\delta$ T cells, CD4+ effector memory).

4. Key Contributions

1. Paradigm Shift: Demonstrates that for unsupervised patient stratification in scRNA-seq, simple cell-type compositional representations often outperform complex deep learning or tensor-based methods.
2. Methodological Rigor: Establishes that treating cell-type proportions as compositional data (via CLR transformation) is critical for accurate biological inference, correcting a common oversight in current pipelines.
3. Efficiency: Provides a scalable solution (ECODA) that reduces computational costs by orders of magnitude, making large-cohort analysis accessible without GPU clusters.
4. Interpretability: Unlike "black box" embeddings, ECODA directly links cohort structure to specific cell populations, facilitating mechanistic interpretation and clinical translation (e.g., identifying specific cell ratios as biomarkers).
5. Tooling: Release of scECODA, an R package that standardizes this workflow for the community.

5. Significance and Implications

• Clinical Translation: The findings suggest that clinically relevant patient subgroups are largely defined by the relative abundance of specific cell populations rather than subtle gene expression changes within those populations. This supports the use of lower-plex, cost-effective clinical assays (like flow cytometry or IHC) to measure specific cell ratios as biomarkers, rather than requiring full transcriptomic profiling for every patient.
• Quality Control: The robustness of ECODA to batch effects makes it a powerful tool for early quality control in large cohort studies to detect confounders.
• Future Directions: The authors propose that future sample representation methods should prioritize compositional awareness and interpretability over sheer model complexity. The approach is modality-agnostic and can be extended to spatial omics and flow cytometry.

In conclusion, the paper argues that less is more: by respecting the mathematical constraints of compositional data, simple baseline methods can achieve superior, faster, and more interpretable patient stratification than current state-of-the-art complex models.