The human genome contains billions of base pairs, but only a tiny fraction directly codes for proteins. The remaining non-coding space is a vast, intricate regulatory landscape filled with promoters, enhancers, and repressors that dictate when and where genes are turned on. To understand cellular identity, development, and disease, we must map these functional elements across the genome.
Historically, researchers tracked individual biochemical modifications on histone proteins—the molecular spools around which DNA winds—using technologies like ChIP-seq. However, looking at a single modification in isolation provides an incomplete and noisy picture. The true functional identity of a genomic region is written in a combinatorial language of multiple overlapping epigenetic marks.
To decode this language, computational biologists treat the genome not as a collection of independent data tracks, but as a unified sequence of discrete functional states.
The Combinatorial Language of Chromatin
Instead of treating each histone modification as an independent variable, modern genomics views them as co-occurring signals. As shown by Rye et al., chromatin marks organize into highly reproducible, distinct combinatorial patterns that correlate with specific genomic contexts.
For instance, the simultaneous presence of the marks H3K4me3 and H3K27ac typically denotes an active promoter, while H3K4me1 combined with H3K27ac marks an active distal enhancer. Conversely, marks like H3K9me3 consistently signal deeply repressed heterochromatin.
To systematically decode this language across the entire genome, researchers needed a framework that could compress multi-dimensional, continuous data tracks into a single, interpretable map of discrete functional units. This process is known as chromatin state segmentation.
Why Hidden Markov Models Work
The breakthrough in solving this segmentation problem came from applying probabilistic sequence models, specifically multivariate Hidden Markov Models (HMMs). Pioneered in this domain by the ENCODE pilot project (Birney et al.) and expanded into widely used frameworks by Ernst, HMMs are uniquely suited to genomic data for two fundamental reasons: emission patterns and spatial transitions.
Hidden States: [ Active Promoter ] ---> [ Transcribed Region ]
| |
Probabilistic Emission Probabilistic Emission
| |
Observed Marks: (H3K4me3, H3K27ac) (H3K36me3)
Emission Probabilities: Capturing Combinations
At any given position (or window) along a chromosome, the model observes a vector of marks—for example, a binary representation of which histone modifications are present or absent. The HMM treats the underlying functional identity of that window (e.g., "Active Enhancer" or "Poised Promoter") as a hidden state that probabilistically "emits" a specific combination of observed marks. By analyzing the co-occurrence of these marks across the entire genome, the model learns the signature of each state automatically, without requiring manual, ad-hoc rules.
Transition Probabilities: Leveraging Spatial Context
Biology operates with spatial continuity; genomic features do not change randomly from one base pair to the next. An active promoter state is highly likely to be adjacent to a transcribed region, but it is highly unlikely to immediately transition into a deeply repressed heterochromatin domain.
The HMM captures this linear syntax through transition probabilities, which model the likelihood of moving from one hidden state to another along the chromosome. This spatial awareness effectively filters out experimental noise and smooths the segmentation, as demonstrated by Wu et al..
From Epigenetic Maps to Disease Biology
By training these models on massive datasets, large-scale consortia have generated standardized atlases of the human regulatory landscape. For example, Kundaje et al. integrated data across 111 reference human epigenomes to map functional states across diverse tissue types, providing a systematic annotation of the human genome.
These discrete maps have fundamentally transformed how we interpret the genetic basis of complex human diseases. The vast majority of disease-associated variants identified in Genome-Wide Association Studies (GWAS) fall within non-coding regions, making their mechanisms difficult to discern. Overlapping these variants with chromatin state maps reveals exactly which cell types and regulatory elements are disrupted.
Using this approach, Disanto et al. demonstrated that genetic variants associated with Multiple Sclerosis are significantly and specifically active within the enhancer states of B cells. Similarly, Gjoneska et al. leveraged conserved chromatin states between mice and humans to show that Alzheimer's disease risk variants are highly enriched in immune-related enhancer states rather than strictly neuronal ones. By converting raw, continuous biochemical signals into discrete, functional states, chromatin state modeling provides the necessary map to navigate the non-coding genome.
Key Foundational Work
Assessing the performance of different high-density tiling microarray strategies for mapping transcribed regions of the human genome (2006) — contributed 1 concept elements to this direction.
Kernels for gene regulatory regions (2005) — contributed 1 concept elements to this direction.