Hi-Cformer enables multi-scale chromatin contact map modeling for single-cell Hi-C data analysis

Hi-Cformer is a transformer-based method that effectively models multi-scale chromatin contact maps in sparse single-cell Hi-C data to derive robust cell representations, accurately impute 3D genome features, and enable precise cell type annotation.

The Gist

Imagine your DNA isn't just a long, straight string of instructions, but a massive, tangled ball of yarn inside a tiny room (the cell nucleus). To make sense of this, the cell folds the yarn into specific shapes, bringing distant parts of the string close together so they can talk to each other. This 3D folding is crucial for deciding which genes turn on or off.

Scientists use a technique called Hi-C to take "photos" of this tangled yarn, showing which parts are touching. However, when they try to do this for single cells (looking at one cell at a time instead of a crowd), the photos come out incredibly blurry and full of holes. It's like trying to reconstruct a puzzle where 90% of the pieces are missing.

Enter Hi-Cformer, a new computer program designed to fix these blurry photos and understand the 3D shape of DNA in individual cells. Here is how it works, explained through simple analogies:

1. The Problem: The "Blurry Puzzle"

Think of single-cell Hi-C data as a broken mosaic. Because the technology is so sensitive, most of the tiles (data points) are missing.

• The Challenge: Existing tools try to fix the mosaic by looking at the whole picture at once, or by looking at just tiny, fixed-size squares. They miss the big picture and the fine details simultaneously. They can't see how a small local fold connects to the overall shape of the chromosome.

2. The Solution: The "Master Architect" (Hi-Cformer)

Hi-Cformer is like a Master Architect who uses a special type of AI (called a Transformer) to rebuild the mosaic. It doesn't just look at the whole room or just one corner; it looks at everything at multiple scales at the same time.

Here is the step-by-step process:

• Step 1: Breaking it into "Lego Bricks" (Multi-Scale Encoding)
  Instead of looking at the whole chromosome as one giant block, Hi-Cformer chops it up into different-sized Lego bricks.

  • It looks at tiny bricks (small local folds).
  • It looks at medium bricks (medium-sized loops).
  • It looks at giant bricks (the whole chromosome shape).
  • Analogy: Imagine reading a book. You don't just read the whole book at once; you read individual words, then sentences, then paragraphs, and finally the whole chapter. Hi-Cformer does this for DNA.

• Step 2: The "Smart Librarian" (The Transformer)
  Once the DNA is broken into these "bricks," Hi-Cformer uses a Transformer (the same technology behind chatbots like me) to organize them.

  • The Magic Trick: In a normal library, books on different shelves can't talk to each other. But Hi-Cformer has a special rule: Bricks on the same chromosome can talk to each other freely, but they can also listen to the "Head Librarian" (the whole chromosome summary) to understand the big picture.
  • This allows the AI to understand that a tiny fold in one part of the DNA is connected to a huge loop elsewhere, even if they are far apart.

• Step 3: Filling in the Blanks (Imputation)
  Because the original photos were so blurry (missing data), Hi-Cformer uses what it learned to predict and fill in the missing pieces.

  • Analogy: If you see a puzzle with a missing piece that looks like a blue sky, and you know the picture is of a beach, you can confidently guess the missing piece is a blue sky. Hi-Cformer does this for DNA, reconstructing the missing contacts so scientists can see the clear, 3D structure again.

3. What Can Hi-Cformer Do?

Once the "Master Architect" has rebuilt the DNA map, it can do three amazing things:

1. Sort the Cells (Cell Type Identification):
   Imagine a room full of people wearing different colored hats (different cell types). Because the DNA shapes are unique to each cell type, Hi-Cformer can look at the reconstructed DNA map and instantly say, "That's a brain cell," or "That's a skin cell," even if the cells look identical to the naked eye. It does this better than any previous tool.

2. Find the "Secret Doors" (Structural Features):
   DNA has "doors" (boundaries) that separate different neighborhoods. Hi-Cformer can spot these doors clearly, even in the noisy, blurry data, helping scientists understand how genes are regulated.

3. Teach the Computer to Label Cells (Annotation):
   Because it understands the DNA shapes so well, you can teach Hi-Cformer to act as a teacher. Once it learns what a "Liver Cell" DNA looks like, it can automatically label new, unknown cells as "Liver Cells" with high accuracy.

Why Is This a Big Deal?

Before Hi-Cformer, analyzing single-cell DNA was like trying to read a book written in a language you barely know, with half the pages torn out.

• Old Tools: Tried to guess the missing pages by looking at the whole book or just one sentence, often getting it wrong.
• Hi-Cformer: Reads the words, sentences, and chapters simultaneously, understands the context, and fills in the missing pages so accurately that you can read the story clearly again.

This allows scientists to finally see how individual cells in our bodies are organized, which is a huge step toward understanding diseases like cancer, where the DNA "folding" goes wrong.

Technical Summary

1. Problem Statement

Single-cell Hi-C (scHi-C) technology captures the 3D organization of chromatin in individual cells, offering insights into gene regulation and cellular heterogeneity. However, analyzing scHi-C data presents significant computational challenges:

• Extreme Sparsity and Noise: Due to limited sequencing depth, scHi-C data suffers from high dropout rates and uneven distribution of chromatin contacts.
• Multi-scale Complexity: Unlike bulk Hi-C, which reveals clear structures like A/B compartments and Topologically Associating Domains (TADs), single-cell maps exhibit strong diagonal signals and complex local patterns that vary across genomic distances.
• Limitations of Existing Methods: Current computational tools often model either entire chromosomal maps or pairwise interactions at a fixed resolution. They fail to explicitly capture the multi-scale structures inherent in chromatin contact matrices, hindering the simultaneous analysis of global context and fine-grained local features.

2. Methodology: Hi-Cformer

Hi-Cformer is a transformer-based deep learning framework designed to model multi-scale chromatin contact maps. It draws inspiration from language models (like BERT), treating chromatin contact maps as sequences of tokens.

Core Architecture

The model consists of three primary modules:

1. Multi-scale Encoder:

   • Chromosomal Map Embeddings: Uses Principal Component Analysis (PCA) to encode entire intra-chromosomal contact maps into fixed-dimensional vectors, capturing global chromosomal architecture.
   • Block Embeddings: Extracts diagonal blocks of varying sizes (e.g., 8, 16, 32, 64, 128 bins) from contact maps. These blocks correspond to local genomic regions (often TAD-like) and are encoded via scale-specific Multi-Layer Perceptrons (MLPs).
   • Sequence Construction: All embeddings are concatenated into an ordered sequence based on chromosome index, block size, and position.

2. Transformer Module (Chromosome-Aware):

   • Contextual Embeddings: Adds learnable embeddings for chromosome identity, block size, and genomic position to the input sequence.
   • Biologically Constrained Attention: A key innovation is the structured attention mask.
     • Block-level: Block embeddings can only attend to other blocks and the map embedding of the same chromosome (modeling intra-chromosomal dependencies).
     • Global-level: Chromosomal map embeddings can attend to all other chromosomes (modeling inter-chromosomal context).
   • This design allows the model to integrate local high-density signals with long-range interactions without violating biological constraints.

3. Multi-scale Decoder:

   • Reconstructs the input data at three hierarchical levels:
     • Cell-level: Reconstructs a unified whole-cell contact map (used for imputation).
     • Chromosome-level: Reconstructs individual chromosome maps.
     • Block-level: Reconstructs local sub-maps at multiple scales.

Training Strategy

• Self-Supervised Learning: Trained using a Masked Language Modeling (MLM) task where random blocks of embeddings are masked (replaced with [MASK]), forcing the model to learn long-range dependencies to reconstruct them.
• Preheating: A pre-training phase where the transformer module is temporarily removed to initialize the encoder and decoder weights before the full model training begins.
• Loss Function: A weighted combination of reconstruction losses at the cell, chromosome, and block levels.

3. Key Contributions

• Multi-scale Modeling: First method to explicitly model chromatin contact maps as sequences of multi-scale blocks within a transformer architecture, capturing both fine-grained local patterns and global structural context.
• Chromosome-Aware Attention: Introduces a novel attention mechanism that respects the biological hierarchy of the genome, preventing biologically implausible cross-chromosome block interactions while allowing global context integration.
• Versatility: The framework is flexible enough to handle different resolutions (e.g., 50-kb for specific loci) and can be adapted for both unsupervised tasks (embedding, imputation) and supervised tasks (cell type annotation).

4. Results

The authors evaluated Hi-Cformer on five public scHi-C datasets (Ramani2017, Lee2019, Tan2021A/B, Wu2024) against six baselines (Higashi, scDEC-Hi-C, scHiCluster, PCA, LDA, HiCRep/MDS).

• Cell Embedding & Clustering:

  • Hi-Cformer generated low-dimensional cell embeddings that significantly outperformed baselines in separating cell types.
  • Metrics: Achieved average improvements of 6.41% in NMI and 29.35% in ARI over the second-best method (Higashi).
  • Robustness: Demonstrated superior performance in identifying rare cell types (e.g., GM12878 in Ramani2017) and distinguishing subtle neuronal subtypes (Lee2019) even under high dropout rates.

• Data Imputation:

  • Hi-Cformer outperformed Higashi and scHiCluster in reconstructing sparse contact maps.
  • Metrics: Showed median improvements of ~69% in Pearson Correlation and ~66% in Cosine Similarity against bulk Hi-C references.
  • Structural Preservation: High SSIM scores indicated that Hi-Cformer preserved spatial structural organization (e.g., TAD boundaries) better than baselines, which tended to over-smooth data.

• Biological Feature Identification:

  • TAD-like Boundaries: Imputed data revealed clearer TAD-like boundaries and cell-type-specific insulation scores.
  • A/B Compartments: Imputed data yielded compartment scores more consistent with bulk data and H3K27ac ChIP-seq signals.
  • Case Study: Successfully resolved structural differences in the ABL1 locus (chr9) between K562 and GM12878 cells, correlating with known oncogenic fusion events.

• Supervised Cell Type Annotation:

  • When adapted for supervised learning, Hi-Cformer achieved higher accuracy, Cohen's kappa, and macro-F1 scores than scHiClassifier, Logistic Regression, and Random Forest.
  • It demonstrated strong cross-dataset generalization, maintaining performance when training on one dataset (e.g., Tan2021A) and testing on another (Tan2021B).

5. Significance

Hi-Cformer represents a paradigm shift in single-cell 3D genome analysis by leveraging the power of transformers to handle the unique sparsity and multi-scale nature of scHi-C data.

• Scientific Impact: It enables the reliable detection of cell-type-specific 3D genome features (TADs, compartments) that were previously obscured by noise, facilitating deeper insights into how chromatin organization drives cellular identity and disease (e.g., leukemia).
• Methodological Impact: By treating genomic contact maps as sequences, it opens the door for applying advanced NLP techniques to genomics, including transfer learning and multi-omics integration (e.g., combining with gene expression or methylation data).
• Practical Utility: As a robust tool for imputation and annotation, it lowers the barrier for analyzing sparse scHi-C datasets, making high-quality 3D genome analysis accessible for diverse biological studies.