MiCBuS: Marker Gene Mining for Unknown Cell Types Using Bulk and Single Cell RNA-Seq Data

MiCBuS is a computational method that identifies marker genes for unknown cell types by generating Dirichlet-pseudo-bulk RNA-seq data to overcome the limitations of heterogeneous bulk and incomplete single-cell RNA-seq datasets where traditional differential analysis fails.

The Gist

The Big Problem: The "Missing Piece" Puzzle

Imagine you are a detective trying to solve a mystery in a crowded city (a biological tissue). You have two main tools to figure out who lives there:

1. The "City Census" (Bulk RNA-seq): This is a report that tells you the total population of the city. It says, "There are 10,000 people here." But it's a bit blurry; it mixes everyone together. You can't tell exactly how many bakers, doctors, or artists are there, just the total noise of the crowd.
2. The "Street Survey" (Single-cell RNA-seq): This is a high-tech survey where you interview people one by one. It's very detailed! You can see exactly who the bakers and doctors are. However, this survey has a flaw: some people are too shy, too small, or hiding in hard-to-reach places. The survey misses them completely.

The Dilemma:
When you look at the "City Census" (Bulk), you see a total population. When you look at the "Street Survey" (Single-cell), you see a list of known people. But when you compare them, the Census says there are more people than the Survey found.

Who are these missing people? What are they doing? Traditional methods can't tell you. They can only analyze the people they saw in the survey. The "missing" people (the Unknown Cell Types) remain a mystery, and we don't know what genes make them special.

The Solution: MiCBuS (The "Ghost Hunter" Algorithm)

The authors created a new tool called MiCBuS (Marker Gene Mining for Unknown Cell Types Using Bulk and Single Cell RNA-Seq Data). Think of MiCBuS as a clever detective who uses a bit of magic to find the missing people.

Here is how it works, step-by-step:

Step 1: The "Best Guess" Estimate

MiCBuS looks at the detailed "Street Survey" (Single-cell data) and the blurry "City Census" (Bulk data). It tries to guess the proportions of the people it does know.

• Analogy: It looks at the survey and says, "Okay, I see 20% Bakers and 30% Doctors." It assumes these are the only people in the city for a moment.

Step 2: Creating "Ghost Towns" (Dirichlet-pseudo-bulk)

This is the magic trick. MiCBuS knows the "Street Survey" is incomplete. So, it creates fake versions of the City Census using a mathematical recipe (called a Dirichlet distribution).

• Analogy: Imagine MiCBuS creates 20 different "Ghost Towns." In these towns, the number of Bakers and Doctors changes slightly every time (sometimes 25% Bakers, sometimes 18%), but crucially, these Ghost Towns only contain the people the Survey found. They do not contain the missing "Unknown" people.

Step 3: The "Shadow Comparison"

Now, MiCBuS compares the Real City Census (which has everyone, including the missing ones) against the Ghost Towns (which only have the known people).

• The Logic: If the Real Census has a lot of "Baker noise" that the Ghost Towns don't have, MiCBuS knows: "Aha! The difference must be caused by the missing people!"
• It looks for genes that are loud in the Real Census but quiet in the Ghost Towns. These loud genes must belong to the missing, unknown cell types.

Why This Matters

Before MiCBuS, if a cell type was missed by the survey, scientists were blind to it. They couldn't study it.

• Old Way: "We found 5 types of cells. Let's study them." (Ignoring the fact that the Census says there are 6 types).
• MiCBuS Way: "We found 5 types, but the Census says there's a 6th. Let's use this math trick to figure out what the 6th type looks like, even though we never saw it directly."

The Results: It Works!

The authors tested MiCBuS in two ways:

1. Simulation (The Practice Run): They took real data, hid some cell types on purpose (like playing hide-and-seek), and asked MiCBuS to find them. MiCBuS successfully identified the "hiding" cells and found their unique genetic signatures (Marker Genes).
2. Real Data (The Real Crime Scene): They used real human tissue data (like pancreas and lung cancer). Even when the data was messy or incomplete, MiCBuS managed to find the genes belonging to the "unknown" cells.

The Takeaway

MiCBuS is like a detective who can identify a suspect they have never seen, just by noticing what is missing from the witness list.

By combining a blurry group photo (Bulk RNA-seq) with a detailed but incomplete list of names (Single-cell RNA-seq), MiCBuS can mathematically reconstruct the "ghosts" in the room and tell us exactly what makes them unique. This helps scientists understand diseases better, find new drug targets, and map the complex ecosystems inside our bodies more accurately.

Technical Summary

1. Problem Statement

The identification of cell type-specific marker genes is fundamental to understanding tissue heterogeneity, disease mechanisms, and therapeutic targets. However, current methodologies face significant limitations in real-world scenarios:

• Heterogeneous Bulk RNA-seq: Traditional bulk RNA-seq amalgamates gene expression from multiple cell types, resulting in low resolution. While deconvolution methods exist, they rely on reference profiles that often miss specific cell types present in the tissue, leading to "unknown" cell types.
• Incomplete scRNA-seq: Single-cell RNA sequencing (scRNA-seq) offers high resolution but often fails to capture certain cell types due to technical limitations (e.g., cell size, fragility, low abundance) or biological differences (e.g., sample source mismatch). This results in "incomplete" datasets where specific cell populations are entirely missing.
• The Gap: Traditional differential expression analysis (DEA) applied to either bulk or incomplete scRNA-seq data can only identify markers for known or captured cell types. No existing statistical method can reliably identify marker genes for unknown cell types (those present in bulk data but absent in the scRNA-seq reference).

2. Methodology: MiCBuS Framework

MiCBuS (Marker Gene Mining for Unknown Cell Types Using Bulk and Single Cell RNA-Seq) is a novel computational framework designed to bridge this gap by integrating bulk RNA-seq and incomplete scRNA-seq data. The method operates in three primary steps:

Step 1: Cell Type Proportion Estimation

• Input: Bulk RNA-seq data and an incomplete scRNA-seq reference (containing only known cell types).
• Process: The method assumes the scRNA-seq reference contains all cell types initially. It uses a reference-based deconvolution algorithm (defaulting to SECRET) to estimate the proportions of the known cell types within the bulk samples.
• Outcome: An initial vector of cell type proportions ($\mathbf{p}_{initial}$). Note that this step ignores the existence of unknown cell types, effectively attributing their signal to the known types or background.

Step 2: Dirichlet-Pseudo-Bulk RNA-seq Generation

• Concept: To create a control group that mimics the bulk data excluding the unknown cell types, MiCBuS generates synthetic "Dirichlet-pseudo-bulk" samples.
• Mechanism:
  • It models the cell type proportions using a Dirichlet distribution: $\mathbf{p} \sim \text{Dirichlet}(\boldsymbol{\alpha})$.
  • The parameter $\boldsymbol{\alpha}$ is defined as $\mathbf{p}_{initial} \times s$, where $s$ is a scaling parameter controlling the variance (spread) of the distribution.
  • A smaller $s$ increases variability to ensure the generated proportions cover the true (unknown) proportions, while a larger $s$ creates a more concentrated distribution.
• Synthesis: Using the SimBu R package, random cell type proportions are drawn from this Dirichlet distribution. These proportions are then used to generate synthetic bulk RNA-seq samples based only on the gene expression profiles of the known cell types from the incomplete scRNA-seq data.
• Result: A set of "Dirichlet-pseudo-bulk" samples that represent what the bulk data would look like if the unknown cell types were absent.

Step 3: Differential Analysis and Marker Identification

• Comparison: The original heterogeneous Bulk RNA-seq data (containing unknown cells) is compared against the generated Dirichlet-pseudo-bulk data (containing only known cells).
• Algorithm: DESeq2 is employed to perform differential expression analysis.
  • Contrast: Bulk vs. Dirichlet-pseudo-bulk.
  • Model: Negative binomial distribution with Wald test.
  • Filtering: Genes significantly upregulated in the original Bulk data (FDR < 0.05, log2FoldChange > 0) are identified as psMarkers (pseudo-marker genes) specific to the unknown cell types.

3. Key Contributions

• Novel Statistical Framework: MiCBuS is the first method specifically designed to identify marker genes for cell types that are missing from scRNA-seq references but present in bulk data.
• Dirichlet-Pseudo-Bulk Generation: The introduction of generating pseudo-bulk samples via a Dirichlet distribution allows for the simulation of realistic cellular composition variability without needing prior knowledge of the unknown cell types.
• Dual-Input Integration: It effectively leverages the high-throughput nature of bulk RNA-seq and the cell-type resolution of scRNA-seq, even when the latter is incomplete.
• Open Source Implementation: The method is implemented in R and freely available on GitHub.

4. Results and Validation

The authors evaluated MiCBuS through extensive simulation studies and real data analyses.

Simulation Studies

• Human Pancreas Dataset:
  • Setup: Simulated bulk data from 6 cell types, with 1 to 3 cell types masked (treated as unknown).
  • Performance: MiCBuS successfully identified hundreds of upregulated genes for the masked cell types (e.g., Beta and Acinar cells).
  • Validation:
    • Database Check: Identified known markers (e.g., NKX6.1 for Beta cells) from the CellMarker2.0 database.
    • Jaccard Index: Compared psMarkers against ground-truth markers derived from complete scRNA-seq. High overlap was observed for the specific unknown cell types, with near-zero overlap for other known cell types, confirming specificity.
  • Robustness: The method remained stable across 20 simulation runs and under varying levels of added noise.
• Metastatic Lung Adenocarcinoma:
  • Setup: Used primary tumor scRNA-seq as reference and metastatic tumor data (containing an extra cell type: oligodendrocytes) to generate synthetic bulk.
  • Result: Successfully identified markers for the oligodendrocytes, which were absent in the primary tumor reference.

Real Data Analysis

• Cell Line Mixtures (Cobos et al., 2023):
  • Setup: Used mixtures of 6 cell lines (3 breast cancer, monocytes, lymphocytes, stem cells). Masked specific lines (e.g., THP1, Jurkat) from the reference.
  • Performance:
    • Bulk Reference: High accuracy in identifying markers for masked cell lines using individual cell-type bulk RNA-seq as the reference.
    • scRNA-seq Reference: Successfully identified markers even when using incomplete scRNA-seq as the reference, though performance was slightly lower due to sequencing platform differences.
  • Validation: Confirmed that psMarkers were upregulated in the masked cell lines and overlapped significantly with manually curated marker genes.

5. Significance and Implications

• Unlocking "Dark Matter" in Tissues: MiCBuS enables researchers to characterize elusive cellular components that are consistently missed by standard scRNA-seq protocols, providing a more complete picture of tissue composition.
• Improved Deconvolution: The marker genes identified by MiCBuS can serve as high-quality references for cellular deconvolution algorithms, improving the accuracy of estimating cell type proportions in heterogeneous bulk samples.
• Downstream Analysis: The identified markers facilitate pathway enrichment and Gene Ontology (GO) analysis for unknown cell types, linking them to specific biological processes and disease states.
• Practical Utility: The method is robust to noise and requires only standard bulk and scRNA-seq data, making it applicable to a wide range of existing datasets where cell type coverage is incomplete.

In conclusion, MiCBuS provides a robust, statistically sound solution to a critical bottleneck in transcriptomics: the inability to define the molecular identity of cell types that are present in a tissue but invisible to single-cell sequencing.