DNA Methylation State Inference

Models must explicitly encode CpG dinucleotide positions as distinct tokens or positional features. Unlike standard genomic models that treat all cytosines equally, methylation predictors use specialized tokenization that distinguishes CpG-context cytosines from non-CpG cytosines. This often involves:

• Dual-channel inputs separating CpG and non-CpG positions
• One-hot encoding with an additional channel for CpG status
• Learned embeddings that capture the unique mutational and chemical properties of methylatable sites

These models are trained to predict methylation levels without requiring bisulfite conversion data as input. Instead, they learn the sequence determinants of methylation from paired reference data (e.g., Whole Genome Bisulfite Sequencing or Illumina 450K/EPIC arrays). During inference, only the reference genome sequence is required, enabling:

• Methylation prediction for any sequenced genome
• Retrospective analysis of existing sequencing data
• Avoidance of bisulfite-induced DNA degradation and GC bias

Methylation state at a given CpG is influenced by distal regulatory elements and local chromatin context. Effective models employ receptive fields spanning 1-10 kilobases around each target site. Architectures achieve this through:

• Dilated convolutions that exponentially expand the receptive field without parameter explosion
• Transformer attention mechanisms that directly model interactions between distant CpGs
• Hybrid CNN-transformer stacks that capture both local motif syntax and long-range dependencies

A single model typically predicts methylation levels across multiple cell types, tissues, or conditions simultaneously. This multi-task design leverages shared sequence features while learning tissue-specific regulatory logic. Output heads produce:

• Beta values (0-1) representing methylation fraction per CpG
• Tissue-specific predictions from a shared sequence encoder
• Differential methylation signals between conditions This approach improves generalization to underrepresented cell types through cross-tissue knowledge transfer.

DNA methylation is typically symmetric across CpG dyads due to maintenance methyltransferase activity during replication. Models enforce or learn this biological constraint through:

• Siamese network architectures that process forward and reverse strands identically
• Loss functions that penalize strand-asymmetric predictions
• Data augmentation with reverse-complement sequences during training This inductive bias reduces overfitting and ensures predictions respect the underlying biochemistry of DNMT1-mediated methylation maintenance.

Production methylation models quantify prediction confidence to distinguish high-confidence calls from ambiguous regions. Techniques include:

• Monte Carlo dropout at inference time to estimate posterior variance
• Ensemble disagreement as a proxy for epistemic uncertainty
• Quantile regression to produce prediction intervals rather than point estimates
• Per-CpG confidence scores that flag regions requiring experimental validation This is critical for clinical applications where false methylation calls could misdirect diagnostic decisions.

The gold-standard experimental technique that treats DNA with sodium bisulfite to convert unmethylated cytosines to uracils while leaving 5-methylcytosines intact. Subsequent PCR amplification replaces uracils with thymines, creating a detectable sequence difference between methylated and unmethylated states.

• Whole-genome bisulfite sequencing (WGBS) provides single-base resolution
• Reduced representation bisulfite sequencing (RRBS) enriches for CpG-rich regions
• Deep learning models trained on WGBS data aim to bypass this labor-intensive process entirely by predicting methylation states in silico

A deep learning paradigm where a model predicts genome-wide epigenomic tracks—including DNA methylation, chromatin accessibility, and histone modifications—solely from raw DNA sequence input. This approach treats the genome as a deterministic regulatory code that can be decoded by sufficiently expressive neural architectures.

• Enformer and Basenji2 are landmark architectures in this category
• Methylation prediction is a subset task within multi-task epigenomic models
• The central hypothesis is that sequence context contains sufficient information to predict cell-type-specific methylation patterns

A computational perturbation technique that systematically introduces virtual mutations into a DNA sequence and measures the predicted change in methylation state output. This reveals which nucleotides are causally responsible for establishing or maintaining methylation patterns.

• Saturation mutagenesis tests all possible single-nucleotide variants at each position
• The difference between reference and mutated predictions quantifies variant effect size
• Used to identify methylation quantitative trait loci (mQTLs) and interpret non-coding variants in genome-wide association studies

The statistical assessment of a model's confidence in its methylation predictions, distinguishing between epistemic uncertainty (model ignorance due to limited training data) and aleatoric uncertainty (inherent biological noise or stochastic methylation events).

• Monte Carlo dropout and deep ensembles are common techniques
• High-uncertainty regions often correspond to partially methylated domains or cell-type heterogeneity
• Production methylation inference systems must flag low-confidence predictions to prevent downstream analytical errors