MINTsC learns multi-way chromatin interactions from single cell high throughput chromatin conformation data

MINTsC is a novel computational method that leverages single-cell Hi-C data to infer multi-way chromatin interactions, offering well-calibrated statistical significance and revealing biologically relevant regulatory mechanisms while reducing the multiple-testing burden for association studies.

The Gist

Imagine your genome (your DNA) isn't just a long, straight string of instructions. Instead, think of it as a giant, tangled ball of yarn inside a tiny room (the cell nucleus). To make sense of the instructions, the cell folds this yarn into specific 3D shapes, bringing distant parts of the string close together so they can talk to each other.

For a long time, scientists could only see two parts of the yarn touching at a time. They knew that Point A might touch Point B, or Point C might touch Point D. But they couldn't easily see if three or more points were all touching each other at the exact same time to form a complex "meeting hub."

This is where a new tool called MINTsC comes in. Here is a simple breakdown of what the paper is about:

1. The Problem: The "Blurry Group Photo"

Scientists have a new camera technology called scHi-C that takes photos of the yarn ball inside individual cells.

• The Issue: These photos are very "noisy" and sparse. It's like trying to figure out who is standing in a circle at a party by looking at a thousand blurry, black-and-white snapshots where you can only see two people shaking hands at a time.
• The Old Way: Previous methods would just look at the "handshakes" (pairwise contacts) and try to guess if a group was meeting. But this often led to mistakes. It's like seeing Person A shake hands with Person B in one photo, and Person B shake hands with Person C in another, and wrongly concluding that A, B, and C were all in a group together.

2. The Solution: MINTsC (The "Detective")

The researchers built MINTsC (Multi-way INTeractions from single cell Hi-C). Think of MINTsC as a super-smart detective who looks at thousands of these blurry snapshots to find the truth.

• The "Clique" Hunt: In math, a "clique" is a group where everyone knows everyone. MINTsC looks for these cliques. It asks: "Did we see A touch B, B touch C, and A touch C in enough different cell photos to be sure they are actually a group?"
• Filtering the Noise: MINTsC is very careful. It ignores groups that only look like they are together because of random noise. It uses a special statistical "magnifying glass" to make sure the group is real before announcing it.

3. How It Works (The Analogy)

Imagine you are trying to figure out which students in a school are in the same study group.

• The Data: You have a log of who sat next to whom in the cafeteria every day for a month.
• The Challenge: You only see pairs sitting together. You never see the whole table at once.
• MINTsC's Job: It looks at the logs. If it sees Alice sitting with Bob, Bob with Charlie, and Alice with Charlie on many different days, it concludes: "Aha! Alice, Bob, and Charlie are a study group!"
• The Safety Check: If it sees Alice with Bob on Monday, and Bob with Charlie on Tuesday, but never Alice with Charlie, MINTsC says, "Nope, that's just a coincidence. They aren't a group."

4. Why Does This Matter? (The "Why")

Why do we care about these 3-way or 4-way meetings?

• The "Team Effort" Analogy: Think of a gene (a piece of DNA that makes a protein) as a lightbulb. Sometimes, one switch (an enhancer) turns it on. But often, multiple switches need to be flipped at the exact same time to turn the light on.
• The Discovery: MINTsC found that in the human brain, genes are often regulated by these "team efforts" where multiple enhancers gather around a gene to control it.
• The Disease Connection: The paper found that in the brains of people with Alzheimer's, these "team meetings" go wrong. Specifically, they found a gene called DKK3 (linked to Alzheimer's) that is controlled by two different genetic switches working together. If you only looked at the switches one by one, you wouldn't see the problem. But MINTsC saw the whole team failing to work together.

5. The Big Win

Before MINTsC, scientists had to guess which groups of DNA were working together, which meant testing millions of random combinations (a huge, expensive, and slow job).

• MINTsC's Superpower: It narrows down the list to the real groups. This is like going from searching for a needle in a haystack to being handed a map that points directly to the needle.
• The Result: It helps scientists find the "hidden rules" of how our genes work, potentially leading to better treatments for complex diseases like Alzheimer's, autism, and cancer.

In a nutshell: MINTsC is a new statistical tool that looks at single-cell DNA data to find groups of DNA strands that are physically touching each other at the same time. It cuts through the noise to reveal how multiple parts of our genome team up to control our health and disease.

Technical Summary

1. Problem Statement

Single-cell Hi-C (scHi-C) data provides high-resolution insights into the 3D organization of genomes within individual nuclei. While current analytical methods effectively identify cell types and pairwise chromatin interactions (loops), they largely overlook multi-way chromatin interactions (simultaneous contacts among three or more loci).

• The Gap: Disease-associated variants often involve complex regulatory mechanisms where multiple distal enhancers cooperate to regulate a single gene. Existing pairwise analyses fail to capture these higher-order cooperative structures.
• The Challenge: scHi-C data is inherently sparse and noisy. While some experimental platforms (e.g., SPRITE, scNanoHi-C) can directly detect multi-way contacts, scHi-C remains the most widely used for complex tissues. Inferring multi-way interactions from scHi-C requires distinguishing true simultaneous contacts from the spurious aggregation of pairwise contacts observed across different cells (heterogeneity).

2. Methodology: MINTsC

The authors introduce MINTsC (Multi-way INTeractions from single cell Hi-C), an empirical Bayes framework designed to infer multi-way interactions by modeling scHi-C data as a multilayer network where cells are layers and genomic loci are nodes.

A. Statistical Framework

1. Null Model (Dirichlet-Multinomial Spline):

   • MINTsC models pairwise contact counts using a Dirichlet-Multinomial distribution.
   • To account for the well-known genomic distance bias (contact probability decreases with distance), the model employs a natural cubic spline at the "band" level (groups of locus pairs with similar genomic distances).
   • A Dirichlet prior is placed on locus-pair probabilities within bands to distribute the band-level probability, allowing for the estimation of bias-adjusted pairwise test statistics (z-scores and p-values).

2. Clique Detection and Aggregation:

   • Multi-way interactions are defined as cliques (fully connected subgraphs) in the contact network.
   • Pre-filtering: To prevent spurious cliques formed by merging pairwise contacts from different cell groups, MINTsC requires candidate cliques to be fully observed (all edges present) in at least a minimum number of cells ($c$).
   • Order Statistics: For a candidate clique of size $N$ (containing $M = \binom{N}{2}$ edges), MINTsC aggregates pairwise evidence across cells. It calculates a clique p-score based on the $r$-th smallest p-value among the edges (order statistics).
   • Null Distribution: Under the assumption of independence (or weak dependence), the distribution of the $r$-th smallest p-value follows a Beta distribution ($Beta(r, M-r+1)$). This allows for the analytical calculation of well-calibrated p-values for the entire clique.

3. Multiple Testing Control:

   • The method applies the Benjamini-Hochberg (BH) procedure to the clique-level p-values to control the False Discovery Rate (FDR).
   • Post-filtering: Optional statistical co-occurrence tests are applied to further eliminate artifacts.

B. Computational Strategy

• The method uses a two-stage modeling approach: first fitting a Poisson spline model at the band level to estimate global distance bias, then allocating probabilities to specific locus pairs.
• This strategy significantly reduces computational burden compared to fitting a spline for every individual locus pair.

3. Key Contributions

• First Method for scHi-C Multi-way Inference: MINTsC is the first computational tool specifically designed to learn multi-way chromatin interactions from standard scHi-C data.
• Robust Statistical Framework: It introduces a principled empirical Bayes approach that corrects for genomic distance bias and provides analytically tractable, well-calibrated p-values for multi-way structures.
• Reduction of Multiple Testing Burden: By aggregating pairwise signals into clique-level tests, MINTsC drastically reduces the search space for molecular QTL and epistatic SNP studies, making it feasible to test for complex genetic interactions.

4. Results and Validation

The authors validated MINTsC using diverse datasets and orthogonal technologies:

• 3D Spatial Co-localization (scMicro-C & Dip-C):
  • Applied to GM12878 scMicro-C and mouse brain Dip-C data.
  • MINTsC-inferred cliques showed significantly smaller maximum pairwise 3D spatial distances compared to non-significant cliques.
  • In GM12878, MINTsC cliques co-localized in a significantly higher proportion of cells, with an empirical FDR of ~8% (at 10% threshold).
• Cross-Technology Validation (SPRITE & scNanoHi-C):
  • SPRITE: MINTsC-inferred cliques in GM12878 and mESCs exhibited significantly higher SPRITE enrichment scores than non-significant cliques.
  • scNanoHi-C: In mESCs, MINTsC cliques correlated with higher counts of long-read concatemers (which physically link multiple loci), confirming true multi-way contacts.
• Biological Validation (Methylation & Gene Regulation):
  • Methylation: In human prefrontal cortex data, MINTsC cliques showed stronger conditional dependence in DNA methylation profiles (partial correlations), supporting the biological reality of the interactions.
  • Gene Regulation: Cliques involving genes and multiple enhancers ($g \leftrightarrow e_1 \leftrightarrow e_2$) were strongly supported by Activity-By-Contact (ABC) scores and showed higher gene expression levels compared to genes without multi-way interactions.
• Baseline Comparison:
  • Compared against a baseline derived from SnapHi-C (pairwise loop caller) and pseudo-bulk O/E ratios.
  • MINTsC identified more multi-way interactions with markedly better FDR control and comparable power in synthetic and real data.

5. Biological Significance and Applications

• Epistatic SNP Discovery: The authors applied MINTsC to human prefrontal cortex data to investigate epistatic effects in Alzheimer's disease (ROS/MAP study).
  • They identified significant SNP-SNP interaction effects for genes like DKK3 (linked to amyloid pathology) and CPLX2 (linked to synaptic plasticity).
  • While individual SNPs had weak effects, their interaction within a multi-way chromatin context significantly explained gene expression variance.
• Molecular QTLs: The method enables the probing of molecular QTLs by reducing the multiple-testing burden, allowing researchers to test specific candidate interactions rather than exhaustively scanning all SNP combinations.

6. Conclusion

MINTsC successfully bridges the gap between sparse single-cell Hi-C data and the biological reality of multi-way chromatin interactions. By leveraging order statistics and a robust null model, it provides a statistically sound method to uncover higher-order genome architecture. This capability is crucial for understanding complex gene regulation, epistasis, and the genetic basis of complex diseases where multiple enhancers cooperate. The tool is publicly available at https://github.com/keleslab/mintsc.