Enhancer Prediction

Models ingest heterogeneous data types to recognize the chromatin state signatures characteristic of active enhancers. Key inputs include:

• Histone modification ChIP-seq (e.g., H3K27ac, H3K4me1) marking active and primed enhancers
• ATAC-seq or DNase-seq for open chromatin accessibility
• DNA sequence motifs for transcription factor binding site grammar
• DNA methylation status for cell-type-specific silencing patterns Multi-task architectures learn shared representations across these modalities, improving generalization to unseen cell types.

Enhancers regulate genes across distances exceeding 100 kilobases. Effective models capture this distal dependency through:

• Dilated convolutions that exponentially expand the receptive field without parameter explosion
• Self-attention mechanisms that compute pairwise interactions between all positions
• Squeeze-and-excitation blocks that recalibrate channel-wise feature responses Architectures like Enformer use a combination of convolutional towers and transformer blocks to integrate regulatory context across 200 kb input windows.

Enhancer prediction is fundamentally a binary classification task on genomic bins, with severe class imbalance—true enhancers constitute less than 1% of the non-coding genome. Mitigation strategies include:

• Focal loss to down-weight easy negative examples during training
• Stratified sampling ensuring balanced minibatches
• Precision-recall curves as primary evaluation metrics over accuracy
• Area under the precision-recall curve (auPRC) for threshold-independent comparison Models are typically validated against orthogonal experimental data such as STARR-seq or MPRA.

Enhancer activity is highly context-dependent, varying dramatically across tissues. Modern approaches leverage transfer learning to adapt models across cell types:

• Pre-training on large compendia like ENCODE and Roadmap Epigenomics
• Fine-tuning on target cell types with limited assay data
• Chromatin accessibility priors used as input features to condition predictions
• Multi-task heads that simultaneously predict activity across dozens of tissues, sharing sequence-level representations This strategy reduces the experimental burden of generating cell-type-specific training data.

Regulatory models must be interpretable to extract biological insight. In silico saturation mutagenesis systematically mutates every nucleotide in an input sequence and measures the predicted change in enhancer probability. This reveals:

• Transcription factor binding motifs driving enhancer activity
• Nucleotide-resolution importance scores visualized as saliency maps
• Motif syntax and spacing constraints between cooperative factors
• Variant effect predictions for non-coding genetic variants associated with disease Integrated gradients and DeepLIFT provide complementary attribution methods.

Computational predictions are validated against high-throughput experimental measurements of enhancer activity:

• Massively Parallel Reporter Assays (MPRA) test thousands of synthesized sequences simultaneously
• STARR-seq measures self-transcribing enhancer activity genome-wide
• CRISPRi perturbation confirms endogenous enhancer function
• eQTL colocalization links predicted enhancers to target gene expression Leading models achieve auPRC values exceeding 0.5 on held-out chromosomes, substantially outperforming conservation-based methods.

The foundational epigenomic assay for enhancer discovery. ATAC-seq uses a hyperactive Tn5 transposase to cleave and tag open DNA regions, directly revealing active regulatory elements.

• Mechanism: Tn5 inserts sequencing adapters into nucleosome-depleted regions
• Signal: Peaks of accessibility correlate with active enhancers and promoters
• Integration: ATAC-seq data serves as a primary training label for deep learning models like Enformer and Basenji
• Resolution: Single-nucleotide footprinting reveals transcription factor binding sites within accessible regions

Post-translational modifications on histone tails that define enhancer state and activity. ChIP-seq against specific marks provides the ground truth for supervised enhancer prediction models.

• H3K27ac: Marks active enhancers and promoters; distinguishes active from poised elements
• H3K4me1: Enriched at both active and poised enhancers; a primary discovery mark
• H3K4me3: Dominant at active promoters; used to differentiate enhancers from promoters
• H3K27me3: Polycomb-repressed chromatin; marks silenced or poised developmental enhancers
• Combinatorial logic: Models learn the interplay of multiple marks to classify regulatory state

A computational perturbation method that systematically mutates every nucleotide in an input sequence to measure the predicted change in enhancer activity. This reveals the regulatory grammar encoded in DNA.

• Saturation mutagenesis: Each position is mutated to all three alternative bases
• Output: A mutation effect map highlighting critical nucleotides for enhancer function
• Motif discovery: Clusters of high-effect positions correspond to transcription factor binding sites
• Variant interpretation: Prioritizes non-coding variants that disrupt enhancer activity
• Tools: Integrated into Selene, DeepSEA, and custom PyTorch pipelines

The gold-standard experimental validation method for enhancer predictions. MPRA simultaneously tests the regulatory activity of thousands of synthesized DNA sequences by measuring their transcribed barcodes.

• Design: Oligonucleotide pool with candidate enhancers cloned upstream of a minimal promoter
• Readout: RNA barcode counts quantify enhancer activity per sequence
• Scale: 10,000–100,000+ sequences tested in a single experiment
• ML integration: MPRA data provides high-throughput training labels for fine-tuning computational models
• Limitations: Tests sequences in an episomal context; may miss chromatin-dependent regulation

A strategy where a model pre-trained on epigenomic track prediction is fine-tuned for enhancer classification. The model's internal representations capture cis-regulatory syntax that transfers across tasks.

• Pre-training: Models like Basenji or Enformer learn to predict thousands of epigenomic tracks from sequence
• Feature extraction: Bottleneck layers encode a rich representation of regulatory potential
• Fine-tuning: A classifier head is added and trained on binary enhancer/non-enhancer labels
• Benefit: Requires far fewer labeled enhancers than training from scratch
• Cross-species transfer: Models trained on human data often generalize to mouse regulatory elements