ENCODE

ENCODE systematically maps functional elements across the human genome using high-throughput assays. The consortium generates genome-wide data for:

• Chromatin accessibility via DNase-seq and ATAC-seq
• Histone modifications such as H3K4me3 (active promoters) and H3K27ac (active enhancers)
• Transcription factor binding sites through ChIP-seq for hundreds of regulatory proteins
• DNA methylation patterns via whole-genome bisulfite sequencing
• RNA expression across diverse cell lines and tissues

These multi-modal epigenomic tracks provide the labeled training data required for supervised deep learning models like Enformer and Basenji to learn the regulatory grammar of the genome.

ENCODE's power lies in its tissue-specific resolution. Rather than providing a single static map, the consortium profiles functional elements across hundreds of biologically distinct contexts. Key cell lines include:

• GM12878 (lymphoblastoid) — the most deeply characterized, used as a benchmark for genomic models
• K562 (myelogenous leukemia) — widely used for studying enhancer-promoter dynamics
• HepG2 (liver carcinoma) — critical for metabolic gene regulation studies
• H1-hESC (embryonic stem cells) — essential for understanding pluripotency networks

This diversity enables models to learn context-dependent regulatory logic, predicting how the same DNA sequence can drive different expression patterns depending on cellular state.

ENCODE introduced a standardized registry of candidate cis-regulatory elements — discrete genomic regions with biochemical signatures of regulatory activity. The classification system includes:

• Promoter-like signatures (PLS): regions near transcription start sites with high H3K4me3 signal
• Enhancer-like signatures (ELS): distal elements marked by H3K27ac and H3K4me1, subdivided into proximal and distal enhancers
• CTCF-only elements: insulator regions bound by the architectural protein CTCF
• DNase-H3K4me3 elements: accessible regions at promoters with lower activity signals

This structured ontology transforms raw epigenomic data into a machine-readable regulatory map, providing discrete prediction targets for sequence-to-function models.

ENCODE data serves as the ground truth labels for state-of-the-art genomic deep learning models. The typical training workflow includes:

• Input: One-hot encoded DNA sequences spanning 100-200 kilobase windows
• Targets: ENCODE's BigWig tracks of chromatin accessibility, histone marks, and CAGE expression data
• Architecture: Dilated convolutional networks or transformers that map sequence to multi-track epigenomic predictions
• Evaluation: Pearson correlation and Spearman rank correlation between predicted and ENCODE-observed signal tracks

Models like Enformer are trained on 5,313 ENCODE tracks simultaneously using multi-task learning, learning shared representations that generalize across assays and cell types.

ENCODE maintains rigorous data quality standards and provides unified access through multiple portals:

• ENCODE Portal: Primary access point with metadata-rich search and visual browsing at encodeproject.org
• UCSC Genome Browser: Integrated ENCODE tracks alongside reference annotations for visual exploration
• BigWig format: Standardized binary format for dense continuous signal tracks, enabling efficient random access during model training
• BED format: Tab-delimited files defining discrete element coordinates (peaks, cCREs) for classification tasks

All data undergoes a uniform processing pipeline to correct for batch effects and technical artifacts, ensuring that models trained on ENCODE learn biological signal rather than experimental noise.

ENCODE has progressed through distinct phases, each expanding its scope and resolution:

• Pilot Phase (2003-2007): Focused on 1% of the genome (44 regions), establishing proof-of-concept that biochemical signatures could identify functional elements
• Production Phase (2007-2012): Expanded to genome-wide mapping across 147 cell types, discovering that 80% of the genome exhibits biochemical activity
• Phase 3 (2012-2017): Added RNA-binding protein maps, long-range chromatin interaction data (Hi-C), and expanded cell-type diversity
• Phase 4 (2017-present): Deep characterization of additional biosamples, integration with GTEx eQTL data, and development of the cCRE registry

Each phase has increased the depth and breadth of training data available for predictive genomic models.