Graph Neural Network (GNN) for Chromatin

The foundational step where linear DNA is abstracted into a graph structure. Nodes represent genomic loci (bins at specific resolutions like 1kb or 10kb), initialized with multi-scale features including DNA sequence embeddings, epigenomic signals, and chromatin accessibility. Edges are defined by a combination of linear adjacency, CTCF binding site co-occurrence, and learnable attention mechanisms, creating a sparse connectivity matrix that captures both local and long-range dependencies.

The core computational mechanism where nodes iteratively aggregate information from their neighbors to update their representations. A typical layer performs:

• Message Computation: A neural network transforms the features of a source node and its edge attributes.
• Aggregation: A permutation-invariant function (sum, mean, or max) pools messages from all neighbors.
• Update: A recurrent or feed-forward network combines the aggregated message with the node's current state. This process allows the model to learn that two distal loci are spatially proximal based on shared protein-binding profiles.

After message passing, the final node embeddings are used to predict chromatin interactions. A bilinear decoder or a multi-layer perceptron takes the embeddings of a pair of nodes and outputs a score representing the contact probability. This is trained against experimental Hi-C contact maps using a loss function that accounts for the power-law decay of contacts with genomic distance, often a Poisson negative log-likelihood or a mean squared error on distance-normalized signals.

To ensure predictions are physically plausible, GNNs are regularized with principles from polymer physics. This includes enforcing the excluded volume effect (no two loci can occupy the same space) and modeling the contact probability decay $P(s) \sim s^{-1}$ characteristic of a fractal globule. These constraints can be integrated via a physics-informed loss term that penalizes structurally impossible conformations, guiding the GNN to learn representations that respect the thermodynamic principles of chromatin folding.

Standard message passing struggles with very long-range interactions due to over-smoothing. To address this, architectures integrate graph attention mechanisms (GAT) . Instead of aggregating from all neighbors uniformly, attention layers compute dynamic weighting coefficients. A node representing an enhancer can learn to assign high attention weights to a distal promoter node if their feature profiles suggest a regulatory interaction, effectively creating a shortcut for information flow across the graph and enabling the capture of specific chromatin loops.

Chromatin organization is hierarchical (loops, TADs, compartments). GNNs use differentiable pooling layers like DiffPool or Top-K pooling to learn this structure. The model progressively coarsens the graph, merging nodes into super-nodes that represent TADs, and then further into A/B compartments. This hierarchical representation allows the decoder to predict contact maps at multiple resolutions simultaneously, ensuring that both fine-scale loops and broad compartmental domains are accurately reconstructed from the same learned embeddings.

A graph attention network that predicts chromatin interactions from ChIA-PET data by representing genomic bins as nodes and their interactions as edges. Key innovations:

• Uses multi-head attention to weigh the importance of different epigenomic features
• Integrates CTCF binding strength and cohesin loading signals as node attributes
• Outperforms convolutional baselines on loop prediction tasks
• Demonstrates that relational inductive biases are critical for capturing long-range dependencies

A generative adversarial framework where the generator produces plausible chromatin interaction graphs and the discriminator distinguishes them from real Hi-C-derived structures. Architecture details:

• Generator uses graph convolutional layers to iteratively refine node positions in 3D space
• Discriminator employs GraphSAGE operators to classify real vs. generated structures
• Enforces polymer physics constraints through a custom loss term penalizing steric clashes
• Enables sampling of diverse structural ensembles rather than a single consensus structure

A hierarchical GNN that learns multi-scale chromatin organization by applying differentiable pooling to coarsen the genomic graph from fine-scale bins to TAD-level representations. Design choices:

• Base layer: GATv2 convolutions with edge features encoding linear distance
• Pooling: DiffPool assigns bins to TAD clusters in an end-to-end learnable manner
• Readout: Global pooling produces a fixed-length embedding for whole-chromosome structure prediction
• Achieves state-of-the-art on insulation score prediction across multiple cell types

A model that predicts 3D genome structures directly from DNA sequence by constructing a graph where nodes are 2kb bins and edges represent predicted interaction probabilities. Core components:

• Node features: DNA language model embeddings from a pre-trained genomic transformer
• Edge prediction: Message-passing neural network with edge gates that learn distance-dependent interaction rules
• Structure decoding: Variational graph autoencoder reconstructs pairwise Euclidean distance matrices
• Trained end-to-end on Hi-C data across multiple species, demonstrating cross-species transfer

A GNN specifically designed for predicting enhancer-promoter interactions by modeling the 3D genome as a bipartite graph. Architecture:

• Two node types: enhancer nodes and promoter nodes with distinct feature encoders
• Edges represent candidate interactions, initialized with CTCF ChIP-seq signal and ATAC-seq accessibility
• Uses relational graph convolutions to propagate information along the bipartite structure
• Loss function incorporates Hi-C contact frequency as soft supervision for edge weights
• Identifies disease-relevant enhancer-gene links missed by linear distance-based methods

A hybrid architecture that combines graph attention networks with transformer-style positional encodings to capture both local neighborhood structure and global chromatin context. Key features:

• Node positional encodings: Laplacian eigenvectors of the genomic distance graph
• Attention mechanism: Graphormer-style spatial bias terms that modulate attention weights based on genomic distance
• Edge features: Stratum-adjusted contact probabilities pre-computed from training Hi-C maps
• Achieves strong performance on A/B compartment prediction and TAD boundary calling simultaneously
• Demonstrates that explicit spatial encoding improves generalization to held-out chromosomes