Chromatin State Annotation

Annotation relies on the combinatorial presence or absence of specific histone marks rather than single modifications. For example, H3K4me3 at a promoter combined with H3K27ac at an adjacent region distinguishes an active promoter-enhancer pair from a poised one marked by H3K4me1 and H3K27me3. Hidden Markov Models (HMMs) and Dynamic Bayesian Networks are used to segment the genome by learning these emission probabilities from ChIP-seq data.

Standardized ontologies classify segments into biologically meaningful categories:

• Active TSS: High H3K4me3, H3K9ac; marks transcription start sites
• Flanking TSS: Enrichment immediately upstream/downstream of active promoters
• Strong/Weak Enhancers: Distinguished by H3K27ac and H3K4me1 intensity
• Bivalent/Poised Promoter: Dual H3K4me3 and H3K27me3 marks, keeping developmental genes silent but ready
• Heterochromatin: H3K9me3 enrichment, constitutive repression
• Quiescent: No significant mark enrichment, often intergenic

Chromatin state maps are critical for non-coding variant prioritization. Genome-wide association study (GWAS) SNPs are systematically overlapped with state annotations to test for enrichment in specific functional categories. A variant falling within an active enhancer state in a disease-relevant tissue provides a mechanistic hypothesis for regulatory disruption, guiding downstream functional validation via reporter assays or CRISPR interference.

Modern approaches bypass ChIP-seq experiments entirely by predicting chromatin states directly from DNA sequence alone. Models like DeepSEA, Basenji2, and Enformer use convolutional and transformer architectures to predict epigenomic tracks and derived state annotations from raw nucleotide input. This enables in silico annotation of any genome, including personal genomes or species without experimental data, and powers in-silico mutagenesis to assess variant impact.

A deep learning paradigm where a model predicts genome-wide epigenomic tracks, such as chromatin accessibility or histone marks, solely from raw DNA sequence input. This approach bypasses the need for experimental assays in unseen cell types, enabling in silico annotation of regulatory elements. Key architectures include convolutional neural networks with dilated convolutions and transformers to capture long-range interactions up to 200 kilobases away.

A computational perturbation technique used to interpret chromatin state predictions. It systematically introduces virtual mutations into a DNA sequence to quantify their predicted impact on a model's output. Key applications include:

• Saturation mutagenesis: Testing every possible single-nucleotide variant in a regulatory element.
• Motif disruption: Assessing the functional importance of specific transcription factor binding sites.
• Variant effect prediction: Prioritizing non-coding variants that may alter chromatin state.

A neural network training strategy where a single model simultaneously predicts multiple epigenomic assays across different cell types. By sharing latent representations, the model leverages common regulatory grammars to improve generalization. This is essential for chromatin state annotation because states are defined by combinatorial patterns of histone modifications, DNase hypersensitivity, and transcription factor binding, not single assays in isolation.

The compressed, high-dimensional vector representation learned by an autoencoder or foundation model that captures the underlying structure of complex epigenomic data. In chromatin state annotation, navigating this space allows for:

• Cross-cell-type comparison of regulatory landscapes.
• Clustering of genomic segments with similar combinatorial mark patterns.
• Interpolation to predict states in uncharacterized cell types or conditions.

A model interpretability method that attributes the prediction of a genomic neural network to its input sequence. It works by accumulating gradients along a path from a neutral baseline (e.g., a shuffled sequence) to the actual input. For chromatin state annotation, this reveals which specific nucleotides and motif syntaxes drive the classification of a segment as a promoter, enhancer, or repressed region.