Integration of single-cell multi-omic data with graph-based topic modelling

The paper introduces bionSBM, a graph-based topic modelling method that integrates single-cell multi-omic data to achieve superior clustering performance and biological interpretability compared to state-of-the-art approaches.

The Gist

Imagine you are trying to understand a massive, chaotic library. But instead of just books, this library contains millions of tiny, living "cells," and each cell has three different types of notebooks:

1. The Transcriptome: A list of instructions the cell is currently reading (genes).
2. The Epigenome: A list of which instructions are allowed to be read (chromatin accessibility).
3. The Proteome: A list of tools the cell is wearing on its surface (proteins).

For a long time, scientists could only read one notebook at a time. But new technology now lets us read all three notebooks for the same cell simultaneously. The problem? It's like trying to organize a library where every book is written in a different language, some pages are missing, and the ink is smudged. It's incredibly messy.

This paper introduces a new tool called bionSBM to help organize this mess. Here is how it works, using simple analogies:

1. The Old Way: The "One-Size-Fits-All" Translator

Previous methods tried to force all these different notebooks into a single, flat list. Imagine trying to translate a novel, a recipe, and a phone book into one giant paragraph. It's hard to tell what belongs to what.

• The Problem: These old methods often had to "smooth out" the data first, like trying to make a jagged mountain look like a flat plain so it fits on a map. In doing so, they sometimes lost the unique, jagged details that make a cell special. They also required scientists to guess how many groups of cells existed beforehand (like guessing there are exactly 12 genres of books before you start sorting).

2. The New Way: bionSBM (The "Smart Network" Organizer)

The authors created bionSBM, which treats the data like a social network rather than a list.

• The Party Analogy: Imagine a huge party where guests (cells) are talking to each other. But instead of people, the "guests" are also the topics they are discussing (genes, proteins, DNA peaks).
• The Graph: bionSBM draws a map where lines connect a "Cell" to the "Gene" it is using. If a cell is using a specific gene heavily, the line is thick. If it's not using it, there is no line.
• The Magic: Instead of forcing everything into one list, bionSBM looks at this web of connections and asks: "Who naturally hangs out together?"
  • It finds groups of cells that act like a "clique" at the party.
  • It finds groups of genes that act like a "conversation topic."
  • Crucially, it does this separately for each type of notebook (genes, DNA, proteins) but links them together. It doesn't force a gene to look like a protein; it respects their differences.

3. Why It's Better: The "Detective" Advantage

The paper tested bionSBM against other top tools (ShareTopic and Mowgli) using real biological data from human and mouse cells.

• Better Sorting: bionSBM was better at correctly identifying what type of cell it was looking at (e.g., "This is a B-cell," "This is a T-cell") without needing to be told how many types to look for. It figured out the number of groups automatically, like a detective who finds the clues rather than being told how many suspects to expect.
• Clearer Stories: Because it keeps the different "notebooks" separate, it can tell a clearer story.
  • Example: In a group of blood cells, bionSBM found a specific "topic" (a group of genes) that was unique to B-cells. It then looked at the DNA notebook and found a specific "switch" (a DNA peak) that turned those genes on. It even found the "master switch" (a transcription factor) that controls the whole process.
  • It's like finding a specific recipe, seeing exactly which ingredients were used, and identifying the chef who wrote it, all at once.

4. The Bottom Line

Think of bionSBM as a super-smart librarian who doesn't just stack books on a shelf. Instead, they build a complex web of connections, noticing that certain books always get borrowed by the same people, and certain people always read the same themes.

• It handles the mess: It works with noisy, incomplete data without needing to "clean" it too much first.
• It finds the truth: It groups cells more accurately than current methods.
• It explains the "Why": It doesn't just say "these cells are similar"; it explains why by showing exactly which genes and DNA switches are driving that similarity.

In short, bionSBM helps scientists make sense of the incredibly complex "multi-omic" data, turning a chaotic library of life into an organized, understandable story about how our cells work.

Technical Summary

1. Problem Statement

Single-cell multi-omics technologies (e.g., 10X Multiome, SHARE-seq, CITE-seq) enable the simultaneous profiling of multiple molecular layers (transcriptome, epigenome, surface proteins) within individual cells. However, analyzing these data presents significant challenges:

• Complexity and Heterogeneity: Data is high-dimensional, sparse, and noisy, with features exhibiting different statistical characteristics (counts, continuous values, binary).
• Integration Difficulties: Existing methods often struggle to integrate paired data without aggressive preprocessing (scaling, harmonization, or batch correction), which can distort biological signals.
• Limitations of Current Algorithms:
  • Deep Learning (Autoencoders): Effective for dimensionality reduction but sensitive to input scale differences and often act as "black boxes" with limited interpretability.
  • Latent Dirichlet Allocation (LDA): Assumes a unimodal Dirichlet prior, imposing a strong homogeneity assumption that conflicts with the heterogeneity of biological systems.
  • Non-Negative Matrix Factorization (NMF): Limited by linearity and tends to create steep cluster boundaries, which may not reflect the continuous nature of biological differentiation.
• Interpretability: There is a need for methods that not only cluster cells but also provide probabilistic, biologically interpretable "topics" (groups of coordinated features) for each modality.

2. Methodology: bionSBM

The authors propose bionSBM, a graph-based topic modelling method rooted in Hierarchical Stochastic Block Models (hSBM) extended to multipartite graphs (nSBM).

• Graph Construction:
  • Input count matrices (e.g., scRNA-seq, scATAC-seq, ADTs) are transformed into a weighted multipartite network.
  • Nodes represent distinct entities: cells and features (genes, peaks, proteins).
  • Edges connect features to cells based on expression levels or openness, with weights proportional to the measurement intensity. Features are not directly connected to each other.
• Inference Mechanism:
  • Bayesian Framework: The algorithm maximizes the posterior probability $P(\text{model}|\text{data})$ by minimizing the Description Length (DL), a principle of Minimum Description Length (MDL).
  • Agnostic Priors: Unlike LDA, bionSBM does not impose a unimodal prior. It uses a flat (non-informative) prior, allowing the data to dictate the distribution of topics and clusters.
  • MCMC Simulation: A Markov Chain Monte Carlo (MCMC) procedure iteratively adjusts node assignments (cells to clusters, features to topics) to minimize the description length.
• Key Technical Features:
  • Modality Independence: It treats each omic layer separately in the output. A "topic" is specific to a modality (e.g., an mRNA-topic vs. a peak-topic), avoiding the mixing of features found in other methods.
  • No Preprocessing Scaling: Because it operates on graph structures, it does not require cross-modality scaling or batch correction, preserving the native statistical properties of each layer.
  • Automatic Parameter Selection: The model automatically infers the optimal number of clusters and topics via the MDL principle, removing the need for users to specify hyperparameters.
  • Implementation: Integrated into the Python single-cell ecosystem (AnnData and Muon objects) for scalability.

3. Key Contributions

1. Novel Algorithm: Introduction of bionSBM, the first application of n-partite Stochastic Block Models specifically designed for integrating and interpreting single-cell multi-omic data.
2. Theoretical Advancement: Moving away from unimodal priors (LDA) and linear factorizations (NMF) to a graph-based, agnostic prior approach that better fits biological heterogeneity.
3. Modality-Specific Interpretability: Unlike other topic models that produce mixed-feature topics, bionSBM generates distinct, modality-specific topic assignments (e.g., specific gene sets for mRNA, specific motifs for ATAC), facilitating clearer biological interpretation.
4. Robust Benchmarking: Comprehensive evaluation against state-of-the-art tools (ShareTopic and Mowgli) across six diverse datasets (human and mouse, varying cell types and technologies).

4. Results

The authors evaluated bionSBM on six datasets (PBMC, BMMCMultiOme, HSPC, MouseSkin, Spleen, BMMCCite) covering scRNA-seq, scATAC-seq, and CITE-seq.

• Cell Type Identification (Clustering Performance):
  • Measured using Normalized Mutual Information (NMI/NMI)* against ground-truth annotations.
  • bionSBM consistently outperformed ShareTopic and Mowgli, particularly in complex datasets with high cell type diversity (e.g., Spleen with 35 types, BMMCCite with 37 types).
  • It achieved superior performance in retrieving ground-truth labels with high specificity.
• Topic Specificity and Distinctiveness:
  • Specificity: bionSBM demonstrated significantly higher topic specificity (how cluster-specific the dominant topic is) compared to competitors. This indicates that the identified topics are strongly enriched for specific cell types.
  • Distinctiveness: All methods showed similar distinctiveness (separation between in-group and out-of-group probabilities), but bionSBM's higher specificity makes its signatures more biologically actionable.
• Biological Interpretability:
  • The method successfully recovered known regulatory programs by linking peak-topics (chromatin accessibility) to mRNA-topics (gene expression).
  • Case Studies:
    • B Cells: Identified PAX5 motif enrichment in peak-topics and PAX5 expression in mRNA-topics.
    • GMPs: Linked IRF8 motif enrichment to GMP differentiation.
    • Erythroblasts: Connected KLF1 motif and expression to erythropoiesis.
  • These findings confirmed that bionSBM captures cell-type-specific regulatory mechanisms across modalities.
• Scalability: The algorithm handles large sparse graphs efficiently, with running times and memory usage comparable to or better than competitors, as shown in supplementary analyses.

5. Significance

• Biological Discovery: bionSBM bridges the gap between computational clustering and biological insight. By providing modality-specific, probabilistic topics, it allows researchers to explain why cells are grouped together (e.g., "These cells cluster because they share high expression of Gene X and open chromatin at Motif Y").
• Generalizability: The framework is not limited to current technologies; it can be applied to any new omic measurement summarizable as a count matrix.
• Practical Utility: The tool is released as an open-source Python package, integrated with standard single-cell data structures (AnnData/Muon), ensuring reproducibility and ease of adoption for the research community.
• Paradigm Shift: It challenges the reliance on deep learning "black boxes" and rigid statistical priors, offering a statistically rigorous, interpretable, and flexible alternative for multi-omic integration.

In summary, bionSBM represents a significant step forward in single-cell analysis by leveraging graph theory to integrate multi-omic data without losing biological nuance, offering superior clustering accuracy and deep mechanistic interpretability.