Systematic clustering alignment and feature characterization for single-cell omics using ACE-OF-Clust

The paper introduces ACE-OF-Clust, a scalable four-step workflow that addresses the clustering alignment problem in single-cell omics by integrating multiple clustering solutions, comparing models against annotations, and identifying informative features to enhance the interpretability and robustness of cellular heterogeneity analysis.

The Gist

Imagine you are a detective trying to sort a massive, chaotic crowd of people into different groups based on what they are wearing, how they talk, and where they are standing. In the world of biology, this "crowd" is a collection of cells, and the "clothing and talk" are their genetic instructions (genes). Scientists use computer programs to sort these cells into types (like "T-cell," "cancer cell," or "muscle cell") to understand how our bodies work or how diseases like cancer develop.

However, there's a big problem: The sorting machines are unreliable.

If you run the same sorting program twice, or use two different programs, you might get completely different groupings. One time, a specific group of people might be labeled "Group A," and the next time, they are labeled "Group C." Sometimes the machine splits a big group into two tiny ones, and other times it smashes two groups together. This is called the "Clustering Alignment Problem." It's like trying to compare two maps of the same city where one calls a street "Main St." and the other calls it "First Ave," or where one map shows a park as a single blob and the other splits it into five tiny squares.

Enter ACE-OF-Clust. Think of this tool as a super-smart translator and map-maker that fixes these messy maps so scientists can finally compare them fairly.

Here is how it works, using simple analogies:

1. The "Multiple Guesses" Strategy (Multiple Clustering)

Instead of trusting just one run of a sorting program (which might be a fluke), ACE-OF-Clust tells the computer to run the sorting process many times, like asking 10 different detectives to sort the same crowd.

• The Problem: Detective A says "Group 1" is the red shirts. Detective B says "Group 3" is the red shirts.
• The ACE-OF-Clust Solution: It uses a clever algorithm (called Clumppling) to look at all 10 detective reports and say, "Okay, even though the labels are different, these three reports are actually describing the same group of red shirts." It aligns them so everyone is speaking the same language.

2. The "Fuzzy" vs. "Hard" Sorting (Mixed-Membership)

Traditional sorting is like a hard cut: You are either in the "Red Shirt Club" or the "Blue Shirt Club." You can't be in both.

• The Reality: In biology, cells are often "fuzzy." A cell might be 70% "Red Shirt" and 30% "Blue Shirt" because it's in the middle of changing from one type to another (like a caterpillar turning into a butterfly).
• The ACE-OF-Clust Solution: It handles this "fuzziness" perfectly. It doesn't force a cell into a single box. Instead, it tracks how much of each "club" a cell belongs to. This helps scientists see the gradual transitions between cell types, which hard sorting misses.

3. Finding the "Star Players" (Feature Characterization)

Once the groups are aligned, the tool asks: "Which genes are actually doing the work to separate these groups?"

• The Analogy: Imagine you are sorting a crowd of musicians. You want to know: Is it the drummers that separate the rock band from the jazz band? Or is it the saxophone players?
• The Innovation: Most tools just look for genes that are "loud" (highly variable). ACE-OF-Clust looks for genes that are strategically important. It calculates a "separation score." If a gene is the only thing that makes a specific group of cells unique, ACE-OF-Clust highlights it as a "clustering-informative feature." It's like finding the one specific detail that proves a suspect is guilty, rather than just listing everything they own.

4. The "Multi-Omic" Detective (Cross-Modal Comparison)

Sometimes scientists have two different types of clues for the same cells:

1. RNA-seq: What genes are the cells reading? (The script).
2. ATAC-seq: What parts of the DNA are open and ready to be read? (The open pages).

• The Problem: The script might say "Rock Band," but the open pages say "Jazz Band." Which one is right?
• The ACE-OF-Clust Solution: It aligns the sorting results from both types of data. If a specific gene (from the script) and a specific open DNA region (from the pages) both point to the same group of cells, ACE-OF-Clust flags them as a regulatory link. It's like finding a fingerprint and a DNA sample that both match the same suspect, giving you much stronger evidence that they are connected.

Why Does This Matter?

Before ACE-OF-Clust, scientists were often guessing which sorting result was "real" or just picking one and hoping for the best. This tool:

• Reduces Guesswork: It shows you where the sorting is stable and where it's shaky.
• Finds Hidden Patterns: It catches the "fuzzy" cells that are in transition, which are often the most interesting ones in disease.
• Connects the Dots: It helps link genetic switches (DNA) to the actual genes they control, even if they are far apart in the genome.

In short: ACE-OF-Clust is the ultimate referee for single-cell biology. It takes the chaotic, conflicting results from different computer programs, aligns them into a single, clear picture, and points out exactly which genetic clues are the most important for understanding how our cells work and how diseases like cancer evolve.

Technical Summary

1. Problem Statement

Single-cell omics analyses (scRNA-seq, spatial transcriptomics, and multi-omics) rely heavily on clustering to identify cell types and states. However, current workflows face three critical challenges:

• The Clustering Alignment Problem: Stochasticity (random initialization), local optima, and varying parameter settings (e.g., number of clusters $K$) lead to "label switching" and distinct clustering modes across runs. Comparing results from different runs or different algorithms is difficult because cluster labels are arbitrary and inconsistent.
• Limitations of Hard Clustering: Standard "hard" clustering (assigning one label per cell) fails to capture continuous biological variation, such as transient states, spatial gradients, or hybrid cell types.
• Lack of Systematic Feature Characterization: While "marker genes" are often identified post-hoc via differential expression, there is a lack of systematic methods to quantify how specific features (genes) drive the clustering hierarchy itself, particularly in mixed-membership (soft) clustering frameworks.
• Multi-omic Integration Gaps: Comparing clustering results across different omic modalities (e.g., RNA vs. ATAC-seq) is rarely done systematically, missing opportunities to identify cross-omic regulatory links.

2. Methodology: ACE-OF-Clust

The authors introduce ACE-OF-Clust (Alignment, Comparison, and Evaluation of Omics Features in Clustering), a four-step framework built upon the existing tool Clumppling.

Step 0: Multiple Clustering Generation

Users generate multiple clustering results from the same data using various models (e.g., Seurat, Scanpy, FastTopics) and parameter settings (different $K$ values). This step is user-defined and not implemented within the tool itself.

Step 1: Clustering Alignment

Using Clumppling, the framework aligns multiple clustering runs:

• Mode Identification: Runs with the same $K$ are grouped into representative "modes" (distinct clustering solutions).
• Cross-$K$ Alignment: Modes across different $K$ values are aligned to capture cluster splitting and merging patterns.
• Cross-Model Alignment: A two-level procedure aligns modes across different software tools/models into a common reference frame.
• Output: A unified set of aligned membership matrices ($Q$) where columns (clusters) are permuted to match the optimal alignment pattern.

Step 2: Quantitative Comparison and Evaluation

The framework quantifies differences between aligned solutions:

• Metrics: It calculates the Normalized Hamming Distance (NHD) for hard clustering and the Average Total Membership Difference ($\Delta$) for mixed-membership clustering.
• Annotation Comparison: It compares clustering results against ground-truth annotations (e.g., cell types or morphotypes) to assess consistency and identify unstable cell groups.
• Visualization: Uses structure plots (stacked bar charts) and UMAP projections to visualize consensus and disagreement.

Step 3: Feature Characterization (Clustering Profiles)

For mixed-membership clustering (which outputs a feature-level matrix $P$ representing relative feature expression per cluster), ACE-OF-Clust constructs Clustering Profiles for each feature:

• Sorting: Feature values in $P$ are sorted in increasing order ($p_{j\ell_1} \le \dots \le p_{j\ell_K}$).
• Log Fold Change (LFC) Vector: Computes log-fold changes between adjacent sorted values ($L_j^k = \log_2(p_{j\ell_{k+1}}/p_{j\ell_k})$).
• Key Metrics:
  • Weighted $P$ Sum ($\tilde{p}_j$): Measures the total relative expression of a feature across all clusters, weighted by cluster size.
  • sepLFC (Separation Log Fold Change): The maximum LFC gap between adjacent clusters. High sepLFC indicates a feature strongly separates specific subsets of clusters.
• Goal: Identify "clustering-informative features" that have high weighted sums and high separation gaps, distinguishing them from traditional marker genes.

Step 4: Multi-omic Integration

The framework aligns clustering results across modalities (e.g., RNA and ATAC) to:

• Quantify variability within annotated cell groups across omics.
• Identify cross-omic regulatory links by finding gene-peak pairs that separate the same aligned clusters, even if they are not genomically proximal.

3. Key Contributions

1. Systematic Alignment Framework: Extends Clumppling to single-cell transcriptomics and spatial data, solving the label-switching and multi-modality issues to enable robust comparison of clustering solutions.
2. Feature-Level Characterization: Introduces a novel "clustering profile" and metrics (Weighted $P$ sum, sepLFC) to quantify how genes drive clustering structure, moving beyond simple differential expression.
3. Multi-omic Regulatory Discovery: Provides a method to infer regulatory relationships by correlating clustering-informative features across modalities (e.g., linking distal chromatin peaks to genes based on shared clustering behavior).
4. Open-Source Tool: Implementation as a Python package (ace-of-clust) compatible with standard pipelines (Scanpy, Seurat, FastTopics).

4. Results

The authors validated ACE-OF-Clust on three datasets:

• PBMC3k (scRNA-seq Benchmark):

  • Demonstrated that even with identical settings, different tools (Seurat vs. Scanpy) and random seeds produce substantially different clustering modes.
  • Showed that CD4+ T cells are highly unstable across models, while CD14+ Monocytes are consistent.
  • Proved that running clustering once is insufficient; alignment reveals the "major mode" and quantifies uncertainty.

• Human Breast Cancer (Spatial Transcriptomics):

  • Applied to 10X Visium data with histological annotations.
  • Hard clustering struggled to resolve tumor-edge regions, while mixed-membership clustering captured continuous transitions between morphotypes (e.g., IDC vs. DCIS).
  • Identified non-HVG (Highly Variable Gene) genes like COX6C as critical for clustering, challenging the standard practice of filtering to HVGs.

• PBMC10k (Multi-omic: RNA + ATAC):

  • Revealed that ATAC-seq clustering is more variable than RNA-seq for specific T-cell subsets.
  • Successfully identified candidate regulatory links (gene-peak pairs) that separate the same cell clusters across modalities.
  • Found that many high-confidence regulatory links are non-proximal (not adjacent in the genome) but share clustering-informative signals, suggesting distal regulation.

5. Significance and Recommendations

• Robustness: The paper argues that single-run clustering is risky. Researchers should run multiple iterations and align them to identify stable cell groups and uncertain assignments.
• Feature Selection: The authors recommend using all genes (or a broader set) for mixed-membership clustering rather than restricting to HVGs, as non-HVGs can be highly informative for cluster separation.
• Interpretability: By quantifying feature contributions via clustering profiles, the tool helps prioritize genes for downstream biological validation and gene regulatory network inference.
• Generalizability: While focused on single-cell omics, the alignment logic applies to any unsupervised clustering task requiring the comparison of multiple solutions.

Limitations: The method assumes the same set of cells across runs (unsuitable for pseudotime trajectories with distinct cells) and relies on the stability of the feature-level matrix $P$ in mixed-membership models.