BARTsc identifies key transcriptional regulators from single-cell omics data

BARTsc is a novel computational method that leverages public ChIP-seq profiles to accurately infer cell-type-specific transcriptional regulators from diverse single-cell omics data, outperforming existing tools and successfully identifying and experimentally validating key regulators in complex biological systems like pancreatic cancer.

The Gist

The Big Picture: Finding the "Bosses" of the Cell

Imagine your body is a massive, bustling city. Inside this city, there are billions of tiny workers (cells). Some workers are construction crews (muscle cells), some are security guards (immune cells), and some are messengers (nerve cells).

Even though all these workers have the same blueprints (DNA), they do very different jobs. How do they know what to do? They have managers (Transcriptional Regulators or TRs). These managers don't do the heavy lifting themselves; instead, they flip switches on the blueprints to tell the workers which instructions to follow.

The Problem: Scientists have a hard time figuring out who the managers are.

• The Old Way: Scientists used to look at the workers' names (gene expression) or look for specific "lock shapes" (motifs) on the blueprints. But this is like trying to guess who the CEO of a company is just by looking at the office furniture or the company logo. It often leads to wrong guesses because a manager might be very quiet (low expression) but still in charge, or the "lock" might be there but no one is actually using it.
• The New Challenge: Now, we have high-tech microscopes (single-cell sequencing) that let us see every single worker in the city individually. But this creates a massive amount of data that is hard to organize.

The Solution: BARTsc (The Detective Tool)

The authors created a new computer program called BARTsc. Think of BARTsc as a super-smart detective that solves the mystery of "Who is the boss?" for every type of cell.

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

1. Grouping the Workers (Clustering)

Instead of looking at one lonely cell at a time (which is like trying to understand a whole orchestra by listening to one violinist), BARTsc groups similar cells together into "neighborhoods" or clusters. It asks: "What makes this group of nerve cells different from this group of immune cells?"

2. Gathering Evidence (The ChIP-seq Library)

BARTsc has a massive library of "wanted posters." These are real, physical photos taken in labs showing exactly where different managers (TRs) have been standing and flipping switches in the past. This is called ChIP-seq data.

• Analogy: Imagine you want to know who runs a specific bakery. Instead of guessing, you look at a giant archive of security camera footage showing exactly where every known baker has been standing in bakeries across the country.

3. Connecting the Dots (The "Signature" Match)

BARTsc looks at the unique "signature" of a cell group (the genes that are turned on). It then asks the library: "Which manager's security footage matches this signature the best?"

• If a group of cells is acting like a "firefighter," BARTsc checks the library and sees that the manager NF-kB has a history of standing right next to firefighter switches.
• It doesn't just look at the name of the manager; it looks at the activity and the location of the switches.

4. The "Double-Check" (Multi-omics)

The coolest part is that BARTsc can look at two types of evidence at once:

1. The Blueprint (RNA): What instructions are being read?
2. The Open Doors (ATAC): Which parts of the blueprint are unlocked and ready to be read?

• Analogy: Imagine you are trying to figure out who is running a party.
  • RNA only is like listening to the music playing (what's happening).
  • ATAC only is like looking at which doors are open (what could happen).
  • BARTsc (Multi-omics) listens to the music and checks the open doors simultaneously. If the music is loud and the doors to the DJ booth are wide open, you can be 100% sure the DJ is in charge. This makes BARTsc much more accurate than tools that only look at one or the other.

The Big Discovery: A New Suspect in Cancer

To prove it works, the scientists used BARTsc to study Pancreatic Cancer (PDAC).

• They found a specific group of cancer cells that were growing super fast and were very aggressive.
• BARTsc pointed a finger at a manager named NELFA.
• Before this, scientists didn't think NELFA was a big deal in pancreatic cancer.
• The Proof: The scientists went into the lab and "fired" NELFA (turned it off) in cancer cells. The result? The cancer cells stopped growing and slowed down.
• Analogy: It's like a detective finding a new suspect in a crime, and then catching them in the act, proving they were the one pulling the trigger.

Why This Matters

• It's Smarter: It finds the real bosses, even if they are quiet or if the clues are tricky.
• It's Faster: It can process huge amounts of data quickly.
• It's Versatile: It works on different types of cells, from the brain to the blood to tumors.

In a nutshell: BARTsc is a powerful new tool that uses a massive library of past evidence to figure out exactly which genetic managers are running the show in every type of cell. This helps scientists understand how healthy cells work and, more importantly, how to stop cancer cells from taking over.

Technical Summary

1. Problem Statement

Inferring functional transcriptional regulators (TRs), such as transcription factors (TFs) and chromatin regulators, from single-cell (sc) omics data (scRNA-seq, scATAC-seq, and scMultiome) is a critical challenge in functional genomics. Existing computational methods suffer from several inherent limitations:

• Reliance on Co-expression: Many methods rely on TF-target co-expression, which often reflects correlation rather than causation and can be masked by confounding factors or shared regulatory programs.
• Motif Enrichment Limitations: Methods based on TF binding motif (TFBM) enrichment in gene-proximal regions yield high false-positive rates (due to unbound motifs) and false negatives (due to motif-less binding). They also often overlook distal enhancers.
• Bulk vs. Single-Cell: Previous tools like the original BART were designed for bulk sequencing data and do not account for cellular heterogeneity or the ability to compare distinct cell clusters within a single-cell dataset.
• Data Sparsity: Single-cell data is often too sparse at the individual cell level to distinguish biological signals, necessitating analysis at the cluster level without introducing artifacts from imputation.

2. Methodology: BARTsc

BARTsc is a computational framework designed to predict functional TRs by leveraging large collections of public ChIP-seq profiles and differential genomic features from clustered single-cell data. It operates in three main stages:

A. Data Input and Feature Extraction

• Input: Supports unimodal (scRNA-seq, scATAC-seq) and bimodal (scMultiome) data.
• Clustering: Cells are grouped into clusters based on similarity or prior annotations. Analysis is performed at the cluster level to overcome data sparsity.
• Feature Sets: Two types of differential feature sets are generated:
  1. Cell-Cluster Signature Features: Genes or regions specifically upregulated in a target cluster compared to all others.
  2. Pairwise Differential Features: Features differentially expressed/accessible between every pair of distinct cell clusters.

B. Cis-Regulatory Profile Inference

BARTsc infers a genome-wide cis-regulatory profile for each feature set, mapping them to a unified set of Union DNaseI Hypersensitive Sites (UDHS):

• scRNA-seq (Gene Sets): Uses adaptive lasso regression on a compendium of >1,000 public H3K27ac ChIP-seq profiles to select and weight a subset that best explains the input gene set's expression pattern.
• scATAC-seq (Region Sets): Directly maps normalized peak signals to UDHS to generate a profile scored by chromatin accessibility.
• scMultiome (Bimodal): Infers profiles from RNA and ATAC modalities separately, then generates a consensus cis-regulatory profile using a rank-aggregation method (geometric mean) that balances contributions from both modalities.

C. Association Scoring and Analysis

• Association Score: Calculates the Area Under the Receiver Operating Characteristic Curve (AUROC) between the inferred cis-regulatory profile and known TR binding profiles (ChIP-seq).
• Two-Pronged Analysis:
  1. Signature Analysis (Cell-Centric): Identifies TRs that best explain the signature features of a specific cell cluster.
  2. Cross-Cell-Cluster Analysis (TR-Centric): Quantifies the relative activity of TRs across clusters using a Deviation Ratio (DR). This involves paired statistical tests (Wilcoxon signed-rank) between up-regulated and down-regulated feature sets of cluster pairs. The Mean Deviation Ratio (MDR) summarizes a TR's overall relative activity in a specific cluster.

D. Key Regulator Identification

BARTsc integrates three metrics to rank key regulators for each cell cluster:

1. Signature Score: Contribution to cluster-specific expression.
2. MDR: Relative activity compared to other clusters.
3. Uniqueness (dMDR): How unique the activity is to the target cluster.
   These ranks are aggregated using the Irwin-Hall distribution to generate a final consensus ranking and P-value.

3. Key Contributions

• Novel Framework: First method to integrate public ChIP-seq data directly with single-cell differential features to infer TR activity, moving beyond motif and co-expression limitations.
• Multi-Modal Integration: A robust bimodal mode for scMultiome data that significantly improves prediction accuracy by prioritizing regulatory elements supported by both transcriptomic and epigenomic signals.
• Relative Activity Quantification: Introduces the MDR metric to quantify TR activity differences across cell types, addressing the limitation that high expression does not always equal high activity.
• AI-Assisted Benchmarking: Developed a generative-AI-assisted pipeline to curate a ground-truth database of TR-cell type relationships from literature, enabling rigorous benchmarking.

4. Results

• Benchmarking Performance:
  • Tested on mouse cortex scRNA-seq and human PBMC scMultiome datasets.
  • BARTsc consistently outperformed state-of-the-art tools (SCENIC+, MAESTRO, BITFAM, chromVAR) in identifying known functional TRs, achieving the highest F1 scores, sensitivity, and specificity.
  • In the bimodal mode, BARTsc identified key regulators (e.g., EOMES, TBX21 in NK cells) that were missed or ranked lower by unimodal (RNA-only or ATAC-only) analyses.
• Robustness: Successfully identified known lineage-specific regulators (e.g., SOX9 in astrocytes, PAX5 in B cells, SPI1 in myeloid cells) and correctly distinguished functional TRs from non-functional or non-expressed ones.
• Novel Discovery in Pancreatic Cancer (PDAC):
  • Applied to a human PDAC scMultiome dataset, BARTsc identified established regulators and discovered NELFA as a novel key regulator driving proliferation in a specific high-proliferative ductal malignant sub-cluster (c5).
  • Experimental Validation: RNAi-mediated knockdown of NELFA in PANC-1 cells significantly downregulated cell division genes (e.g., CENPU, KIF2C) and retarded tumor cell growth, confirming NELFA's role as a proliferation driver.

5. Significance

• Biological Insight: BARTsc provides a deeper understanding of cell-type-specific regulatory programs and the dynamic activity of TRs across complex tissues, overcoming the limitations of relying solely on gene expression or motif presence.
• Therapeutic Potential: By identifying novel drivers of disease states (like NELFA in PDAC), the method facilitates the discovery of new therapeutic targets.
• Versatility: The tool is applicable to diverse biological systems (immune cells, brain, cancer) and data types (RNA, ATAC, Multiome), making it a versatile resource for the single-cell genomics community.
• Open Source: Implemented as an R package, BARTsc is freely available to accelerate research in transcriptional regulation.

In summary, BARTsc represents a significant methodological advance by bridging the gap between single-cell omics heterogeneity and comprehensive ChIP-seq reference data, offering a more accurate and biologically relevant approach to deciphering transcriptional regulatory networks.