Cross-Cell-Type Folding Prediction

A transfer learning strategy where a source model pre-trained on a high-resolution Hi-C dataset (e.g., GM12878 lymphoblastoid cells) is partially retrained on a small amount of target cell-type data. This process updates the model's weights to learn the cell-type-specific chromatin interaction landscape while retaining generalizable features of genome folding, such as the mechanics of loop extrusion and compartmentalization, learned from the source domain.

An architectural approach that separates latent representations into cell-type-invariant and cell-type-specific features. The model learns to encode universal folding principles—like the biophysics of cohesin complex simulation—in a shared subspace, while mapping cell-type identity and differential CTCF binding site occupancy to a distinct subspace. This allows the prediction of 3D structure for an unseen cell type by combining universal folding rules with the new type's specific epigenomic signature.

Instead of relying on scarce Hi-C data, this method conditions the prediction model on widely available one-dimensional epigenomic tracks such as ATAC-seq, ChIP-seq for CTCF, and histone modification profiles. A graph neural network (GNN) or convolutional architecture ingests these tracks as input features alongside the DNA sequence, learning the direct mapping from the linear epigenomic state to the 3D Hi-C contact map without requiring any target cell-type interaction data.

The most extreme form of cross-cell-type prediction, where a model like Akita predicts a Hi-C contact map using only the raw DNA sequence as input. The model implicitly learns the grammar of CTCF binding site prediction and motif spacing from sequence alone. Cross-cell-type generalization is achieved because the model's predictions are driven by the static genomic sequence, providing a foundational folding landscape that can then be modulated by cell-type-specific factors in a downstream step.

A training paradigm where a single model is simultaneously trained on Hi-C data from multiple source cell types. By sharing a common backbone network, the model is forced to learn a robust, generalized representation of 3D genome folding. Auxiliary outputs for each cell type allow the network to capture shared structural features like A/B compartment identity and TAD boundaries, which are then more easily adapted to a new target cell type with minimal fine-tuning.

A technique using a domain discriminator network trained adversarially to ensure that the latent features extracted by the folding predictor are indistinguishable between source and target cell types. By backpropagating a reversed gradient, the main model learns to produce cell-type-agnostic representations of chromatin structure. This ensures that when the model processes epigenomic data from a new cell type, it maps it to a meaningful point in the 3D folding space learned from the source domain.

A machine learning paradigm where a model trained on a data-rich source cell type (e.g., GM12878) is fine-tuned on limited data from a target cell type (e.g., a rare neuronal subtype). This approach leverages shared architectural principles of genome folding, such as loop extrusion and compartmentalization, to bypass the prohibitive cost of generating new Hi-C data for every biological context.

A subfield of transfer learning that explicitly addresses the distribution shift between source and target Hi-C contact maps. Techniques include:

• Adversarial training to learn cell-type-invariant features
• Maximum Mean Discrepancy (MMD) minimization in latent space
• Gradient reversal layers to prevent the model from distinguishing source from target This ensures that predicted folding patterns are not biased by the source cell type's unique chromatin landscape.

The extreme case of cross-cell-type prediction where only a handful of experimental replicates (or even a single Hi-C library) are available for the target cell type. Models leverage meta-learning algorithms like MAML (Model-Agnostic Meta-Learning) to find an initialization that can rapidly adapt to new cell types after seeing only a few contact maps, mimicking the human ability to learn new concepts from sparse examples.

A learned, continuous vector representation that encodes the 3D genome folding identity of a specific cell type. By conditioning a folding predictor on this embedding, a single model can generate accurate Hi-C maps for multiple cell types. Similar cell types (e.g., different T-cell subtypes) cluster together in this space, allowing for zero-shot interpolation of folding patterns for uncharacterized intermediate states.

A strategy that bridges cell types by using shared epigenomic signals as an intermediate representation. A model is trained to predict 3D folding from DNase-seq, ChIP-seq, or ATAC-seq data in the source cell type. For the target, only the cheaper, easier-to-generate epigenomic data is required, and the model infers the corresponding 3D structure without ever seeing target Hi-C data.

The application of cross-cell-type methods to predict 3D genome organization across evolutionary distances. Models trained on mouse Hi-C data can predict human TAD boundaries and loops by leveraging synteny and conserved CTCF binding motifs. This reveals fundamental, sequence-encoded rules of chromosome folding that persist despite millions of years of divergence.