Cross-Cell-Type Generalization

The model relies exclusively on raw DNA sequence as input, avoiding cell-type-specific features that would prevent generalization. By learning directly from nucleotide patterns, the model captures universal regulatory grammars—such as transcription factor binding motifs—that are shared across cell types. This sequence-only approach ensures the model does not overfit to experimental artifacts or cell-type-specific signal tracks present in the training data.

The model is trained to simultaneously predict epigenomic profiles for multiple source cell types using shared hidden representations. This forces the network to learn a common regulatory code rather than memorizing cell-type-specific patterns. Key benefits include:

• Shared representations capture universal motif syntax
• Auxiliary tasks act as a regularizer, reducing overfitting
• Zero-shot transfer emerges when the latent space encodes cell-type-invariant features

The architecture incorporates a learned cell-type embedding vector that modulates the network's predictions. During training, the model associates each cell type with a dense vector in a continuous latent space. For unseen target cell types, a new embedding can be inferred or interpolated from related types, allowing the model to condition its predictions on cell-type identity without requiring explicit training data for that type.

The model processes extended genomic windows of 100–200 kilobases, capturing distal regulatory elements like enhancers that act over large genomic distances. Architectures such as dilated convolutions or transformer attention mechanisms enable the model to integrate information across these long ranges. This is critical for generalization because enhancer-promoter interactions are often conserved across cell types even when the specific regulatory activity differs.

Before fine-tuning on specific epigenomic assays, the model undergoes self-supervised pre-training on massive unlabeled genomic sequences. Pretext tasks such as masked nucleotide prediction force the model to learn fundamental DNA grammar, evolutionary constraints, and motif syntax. This pre-training phase builds a rich internal representation of genomic sequence function that transfers robustly to unseen cell types during downstream epigenomic prediction tasks.

The model quantifies its epistemic uncertainty when predicting on held-out cell types, distinguishing between confident generalizations and unreliable extrapolations. Techniques include:

• Monte Carlo dropout during inference
• Deep ensembles with multiple model initializations
• Prediction intervals that flag low-confidence regulatory calls This allows downstream analyses to filter predictions based on confidence thresholds, ensuring only reliable cross-cell-type transfers are used.

A neural network training strategy where a single model simultaneously predicts multiple epigenomic assays across different cell types. By learning a shared biological representation from diverse training signals, the model is forced to capture universal regulatory grammar rather than memorizing cell-type-specific noise. This shared latent space is the primary mechanism enabling cross-cell-type generalization to held-out assays.

The process of adapting a model pre-trained on a large, general epigenomic corpus to a specific, data-scarce target task. A model first learns universal regulatory syntax from abundant cell types, then undergoes fine-tuning on limited data from a rare or difficult-to-assay target cell type. This paradigm directly addresses the cross-cell-type generalization challenge by decoupling general feature learning from task-specific adaptation.

Large-scale neural networks pre-trained on massive, diverse epigenomic datasets to learn universal regulatory grammars. These models, such as the Nucleotide Transformer or Enformer, ingest data from hundreds of cell types and assays. Their scale enables emergent zero-shot or few-shot cross-cell-type generalization, where the model predicts regulatory activity in a completely unseen cell type without any additional training.

The computational prediction of missing epigenomic assay data for an unmeasured cell type using a model trained on available reference epigenomes. This is a direct application of cross-cell-type generalization, where the model leverages correlations learned from a reference panel of profiled cell types to infer the epigenomic landscape of a query cell type. Techniques often use matrix factorization or deep generative models.

The compressed, high-dimensional vector representation learned by an autoencoder or foundation model that captures the underlying structure of complex epigenomic data. A well-structured latent space organizes cell types by their true biological similarity. Cross-cell-type generalization succeeds when the latent representation of an unseen cell type can be interpolated from its neighbors, allowing the decoder to generate accurate regulatory tracks.

The statistical assessment of a model's confidence in its predictions, critical for evaluating cross-cell-type generalization. When predicting on an unseen cell type, a model should exhibit high epistemic uncertainty in regions where the target biology diverges from its training distribution. Distinguishing this model ignorance from inherent data noise is essential for safely deploying generalized predictions in clinical or research settings.