FourC: identifying significant and differential contacts in 1D chromatin conformation data

The paper introduces FourC, an open-source Bayesian method that utilizes Gaussian processes to overcome the semi-quantitative limitations of 4C-seq data by identifying significant and differential chromatin contacts without requiring unique molecular identifiers (UMIs).

The Gist

The Big Picture: Mapping the Genome's "Social Network"

Imagine your DNA isn't just a long, straight string of beads. Inside the cell nucleus, it's actually a tangled ball of yarn, folded up so tightly that it fits in a space smaller than a grain of sand. But this isn't a messy tangle; it's a highly organized structure. Specific parts of the yarn touch each other to let genes talk to their "switches" (called enhancers).

Scientists use a technique called 4C-seq to take a snapshot of these touches. They pick one specific gene (the "Bait") and ask: "Who is this gene hugging right now?"

The Problem: The "Echo Chamber" Effect

The paper starts by identifying a major flaw in how scientists usually analyze these snapshots.

The Analogy: Imagine you are at a crowded party trying to count how many people are talking to the host.

• The Reality: You want to know how many unique people are talking to the host.
• The Flaw: The 4C-seq method is like a recording device that accidentally records the same person's voice over and over again because of a microphone glitch (called PCR duplication).
• The Result: If you just count the voices, you think 1,000 people are talking to the host, when really only 50 unique people are there. The extra 950 are just "echoes" or duplicates.

Because of this, previous methods often had to smooth out the data (like blurring a photo) to hide the noise. But blurring the photo also hides the fine details, making it hard to see exactly where the important conversations are happening.

The Solution: Enter "FourC"

The authors created a new tool called FourC. Instead of trying to count the exact number of "voices" (which is unreliable due to the echoes), FourC changes the question entirely.

The New Strategy:
Instead of asking, "How many times did we hear a voice?" FourC asks a simpler question: "Did we hear a voice at all?"

• The Binary Switch: FourC turns the data into a simple Yes/No (1 or 0).
  • Did we see a connection? Yes (1).
  • Did we not see it? No (0).
• Why this works: Even if the microphone creates 100 echoes of one person, the answer to "Did we hear them?" is still just Yes. By ignoring the count and focusing on the presence, FourC eliminates the noise caused by the echo chamber.

The Magic Tool: The "Smart Detective" (Gaussian Processes)

Now that FourC has a clean list of "Yes/No" connections, it needs to figure out which ones are real and which are just random noise.

The Analogy: Imagine you are a detective looking at a map of a city. You see a few dots where people were seen.

• Old Method: If you see a dot, you assume it's a crime scene. But sometimes dots appear randomly.
• FourC's Method: FourC uses a "Smart Detective" (a statistical model called a Bayesian Gaussian Process). This detective knows the rules of the city:
  • Rule 1: People usually hang out in clusters, not randomly scattered.
  • Rule 2: The further away you get from the center, the fewer people you see (this is called "distance decay").

The detective looks at the pattern of "Yes/No" dots. If there is a sudden, sharp cluster of "Yes" dots that breaks the normal pattern, the detective says, "Aha! This is a significant interaction!" It can spot these clusters even if they are tiny and were previously hidden by the noise.

What They Discovered: The "Priming" of Genes

The authors tested FourC on human stem cells turning into pancreatic cells. They looked at how genes interacted with their switches during this transformation.

The Discovery:
They found something surprising about how genes get "primed" to work.

• The Old Theory: We thought genes only reached out to their switches when they were ready to turn on.
• The FourC Finding: The genes actually reach out and touch their switches early on, even before the gene is turned on.
  • Analogy: It's like a student shaking hands with a professor in the hallway before the class starts. The student isn't asking a question yet, but they are establishing the connection so that when class starts, they can immediately ask for help.
• The Twist: When they used CRISPR (gene editing) to break the switch (the enhancer), the early handshake still happened! But when the gene finally tried to turn on, the connection failed. This proves that the structure (the handshake) can exist independently of the function (the conversation).

Why This Matters

1. Clearer Vision: FourC removes the "static" from the data, allowing scientists to see the genome's structure with much higher resolution.
2. Better Comparisons: It can accurately compare different cell types (like a healthy cell vs. a sick cell) to see exactly which connections are broken.
3. New Biology: It revealed that cells prepare their genetic "social network" in advance, long before they need to use it.

In a nutshell: The authors built a new filter that ignores the "echoes" of DNA data and focuses on the "presence" of connections. This allowed them to see the genome's hidden social network with crystal clarity, revealing that cells plan their future connections long before they actually need them.

Technical Summary

1. Problem Statement

4C-seq (Circularized Chromosome Conformation Capture followed by sequencing) is a cost-effective method for profiling interactions between a single genomic "bait" element and the rest of the genome. However, the data is inherently semi-quantitative due to two major technical limitations:

1. PCR Duplication: Unlike Hi-C, 4C-seq lacks Unique Molecular Identifiers (UMIs) and relies on a deterministic restriction digest step. This prevents positional deduplication, meaning distinct interaction counts cannot be distinguished from PCR amplification artifacts. High read counts often reflect PCR bias rather than true biological contact frequency.
2. 3C Sampling Biases: Read counts are confounded by fragment-level properties such as GC content, length, and distance from the bait.

Existing analysis methods (e.g., peakC) typically address these issues by averaging or smoothing over genomic windows. While this reduces noise, it sacrifices the high resolution of the assay and introduces arbitrary analysis parameters. Furthermore, standard count-based differential analysis tools (like DESeq2) struggle to distinguish true biological changes from PCR-induced noise, particularly in detecting subtle, localized chromatin topology changes.

2. Methodology: The FourC Model

The authors developed FourC, an open-source statistical framework based on a Bayesian Bernoulli Spatial Generalized Linear Mixed Model (SGLMM). The core innovation is the transformation of fragment count data into binary presence/absence indicators to bypass PCR duplication effects.

A. Binarization Strategy

• Theoretical Basis: The authors model the initial fragment count $y_i$ as a Poisson distribution ($Pois(\lambda_i)$), where $\lambda_i$ is the intrinsic capture rate. PCR amplification is modeled as a branching process.
• Transformation: Instead of modeling the final read count $r_i$ (which is inflated by PCR), they define a binary indicator $o_i = 1[r_i > 0]$.
• Justification: Assuming libraries are sufficiently deep, the probability of observing a fragment is tied to its intrinsic capture rate: $P(o_i=1) = 1 - e^{-\lambda_i}$. By applying a complementary log-log (cloglog) link function, the model can perform inference directly on $\lambda_i$ without needing to model the PCR duplication process explicitly.
  $$\text{cloglog}(p_i) = \log(-\log(1-p_i)) = \log(\lambda_i)$$

B. Statistical Model Structure

The model uses a Latent Gaussian Model (LGM) estimated via Integrated Nested Laplace Approximation (INLA) for computational efficiency. The linear predictor for the log-capture rate includes:

1. Fixed Effects: Technical covariates including fragment GC content, length, fragment type (blind vs. non-blind), and sample/replicate indicators. These are modeled using natural splines.
2. Spatial Random Effects (Gaussian Processes): To capture spatial patterns in chromatin contacts, the model employs two distinct Gaussian Process (GP) terms:
   • Coarse-scale GP ($f_{Coarse}$): Models large-scale patterns (e.g., ~100 kb distance decay).
   • Fine-scale GP ($f_{Fine}$): Models residual local variations (e.g., ~10 kb) to identify specific enhancer-promoter loops.
   • Implementation: The GPs are constructed using a Matern covariance function approximated via a stochastic partial differential equation (SPDE) on a lattice (Extended Lattice Kriging), allowing for efficient computation in 1D.

C. Inference and Testing

• Significant Interactions: Identified by finding regions where the posterior probability of the fine-scale term being positive exceeds a threshold (e.g., $P(f_{Fine}(s) > 0 | y) > 0.9$). This isolates local enrichments from the global distance decay.
• Differential Interactions: Calculated as the difference in total spatial trends between conditions ($\Delta = (f_{Fine}^B + f_{Coarse}^B) - (f_{Fine}^A + f_{Coarse}^A)$). Significance is determined if the credible interval for $\Delta$ excludes zero.

3. Key Contributions

1. Novel Statistical Framework: First method to explicitly model 4C-seq data at the fragment level using a binarized Bernoulli likelihood, effectively removing PCR duplication bias without losing resolution.
2. Spatial Modeling: Integration of multi-scale Gaussian processes to separate global distance decay from local regulatory interactions, avoiding the need for arbitrary window smoothing.
3. Software Package: Release of the FourC R package, utilizing inlabru for fast Bayesian inference, making the method accessible to the community.
4. Biological Insight: Application of the method to pancreatic differentiation and CRISPR perturbation data to reveal new biological mechanisms regarding enhancer-promoter dynamics.

4. Results

The authors validated FourC using 4C-seq data from a human pancreatic stem cell differentiation system (ESC $\to$ DE $\to$ GT $\to$ PP) and CRISPR perturbations of GATA6 and ONECUT1 enhancers.

• Superior Peak Detection: Compared to peakC, FourC successfully identified functional enhancers (e.g., ONECUT1+e12, SOX17+e10) that were missed by peakC. FourC's results showed higher concordance with known functional marks (CTCF binding, H3K27ac) and reduced false positives caused by PCR artifacts.
• Validation with Hi-C: Differential contact estimates from FourC showed high correlation with deduplicated Hi-C data. In contrast, standard count-based methods (DESeq2 applied to 4C counts) produced discordant results, particularly near the bait where PCR duplication is highest.
• Detection of Subtle Changes: FourC successfully detected localized chromatin topology changes induced by CRISPRi (enhancer silencing), which were too subtle for binning-based methods like DESeq2 to detect.
• CRISPR Deletion vs. Silencing:
  • CRISPRi (Silencing): Produced subtle, localized reductions in looping.
  • CRISPR-Cas9 (Deletion): Caused broader structural changes, including decreased interactions in the target TAD and increased interactions in adjacent TADs.
• Enhancer Rewiring Mechanism: Analysis of ONECUT1 revealed a "priming" mechanism. Contacts between the promoter and the distal enhancer (+e664) were established early in differentiation (ESC to GT stages) independent of the enhancer's functional status (observed even in CRISPR knockouts). However, the accumulation and strengthening of these contacts at the PP stage (gene activation) were dependent on a functional enhancer.

5. Significance

• Resolution Preservation: FourC allows researchers to utilize the full resolution of 4C-seq data without the resolution loss inherent in window-based smoothing.
• Robustness to Artifacts: By decoupling biological signal from PCR noise, FourC provides more reliable estimates of contact frequencies, enabling the detection of subtle regulatory changes that were previously obscured.
• Biological Discovery: The method revealed that chromatin looping can be established in an "enhancer-function independent" manner (priming) before gene activation, a nuance that required high-resolution, noise-free data to detect.
• Generalizability: While designed for 4C-seq, the framework is adaptable to other 1D 3C-based assays by adjusting the likelihood function (e.g., switching to Poisson or Negative Binomial if UMIs are available).

In summary, FourC represents a significant methodological advance in 3D genomics, transforming semi-quantitative 4C-seq data into a robust, high-resolution tool for studying the dynamics of enhancer-promoter interactions during development and disease.