Epigenomic Feature Engineering

Converts continuous epigenomic signal tracks (e.g., BigWig files from ATAC-seq or ChIP-seq) into discrete, model-ready numerical matrices. This involves binning the genome into fixed-width windows (e.g., 25bp, 200bp) and aggregating signal intensity within each bin.

• Coverage normalization: Adjusts for sequencing depth using methods like RPKM or TPM.
• Smoothing: Applies Gaussian kernels to reduce stochastic noise inherent in sparse assays.
• Log transformation: Stabilizes variance and compresses the dynamic range of highly skewed signal distributions.

Represents the underlying DNA sequence as a numerical tensor for model consumption. The simplest method is one-hot encoding, creating a 4-channel binary matrix (A, C, G, T) of shape [4, sequence_length].

• K-mer frequency vectors: Counts occurrences of short subsequences (3-mers to 6-mers) to capture local compositional biases.
• Learned embeddings: Uses pre-trained genomic foundation models like the Nucleotide Transformer to generate dense, context-aware vector representations that encode evolutionary and functional semantics.
• Positional encoding: Injects information about relative or absolute position within the chromosome, crucial for transformer architectures.

Integrates heterogeneous epigenomic data types into a unified feature vector for each genomic locus. A single region's representation might concatenate signals from DNase-seq, multiple histone modifications (H3K4me3, H3K27ac), and DNA methylation beta values.

• Channel stacking: Treats each assay as a separate channel in a multi-dimensional input tensor, analogous to RGB channels in computer vision.
• Feature fusion: Employs late-fusion architectures where separate encoders process each assay type before concatenation in a shared latent space.
• Missing data imputation: Uses models like Avocado to computationally infer missing assay tracks for cell types not experimentally profiled.

Engineers spatial relationship features that capture the regulatory grammar of the linear genome. These features explicitly model the distance between a candidate regulatory element and its target gene's transcription start site (TSS).

• Genomic distance bins: Log-transforms and buckets the base-pair distance to the nearest TSS or exon boundary.
• CTCF loop anchors: Encodes binary features indicating whether a locus is anchored within a CTCF-mediated chromatin loop identified by Hi-C.
• Topologically Associating Domain (TAD) membership: Assigns a categorical feature for the TAD boundary status, as regulatory interactions are largely constrained within these insulated neighborhoods.

Incorporates metrics of negative selection as a prior on functional importance. PhyloP and PhastCons scores quantify nucleotide-level evolutionary constraint across multiple sequence alignments of dozens of vertebrate species.

• Base-wise priors: Provides a static, pre-computed feature indicating purifying selection, independent of cell-type-specific assays.
• Grantham distance: Encodes the biochemical severity of amino acid substitutions for coding variants.
• CADD integration: Uses features from the Combined Annotation Dependent Depletion framework, which integrates conservation with epigenomic context to score variant deleteriousness.

Generates features by systematically perturbing the input sequence and measuring the predicted change in a model's output. This transforms a static sequence into a mutation sensitivity map.

• Saturation mutagenesis: Performs all possible single-nucleotide substitutions and records the delta in predicted chromatin accessibility.
• Allelic imbalance features: Encodes the difference in predicted transcription factor binding affinity between reference and alternate alleles.
• Gradient-based saliency: Uses Integrated Gradients or DeepLIFT to compute nucleotide-resolution importance scores, which serve directly as feature vectors for downstream variant effect predictors.

A transformer-based architecture developed by DeepMind that predicts gene expression and epigenomic tracks from DNA sequence. Enformer's key innovation is its use of dilated convolutions and multi-head attention to capture regulatory interactions up to 200 kilobases away, dramatically improving enhancer-gene linkage prediction. The model outputs 5,313 genomic tracks covering CAGE expression, ChIP-seq, and DNase-seq across multiple human cell types.