DECODE

DECODE refines enhancer predictions and boundary localization using a deep neural network trained on direct enhancer activity assays (e.g., STARR-seq) and a weakly-supervised object detection framework to improve mapping of distal regulatory elements.

Key Features:

• Enhancer Prediction Accuracy: DECODE employs a deep neural network classifier trained on direct enhancer activity readouts from functional assays like STARR-seq, achieving cell-type-specific predictions and a reported 24% improvement in transgenic mouse validation compared with previous methods.
• Boundary Localization Precision: The framework implements a weakly-supervised object detection approach to refine enhancer boundaries to approximately 10 bp resolution versus roughly 500 bp for traditional methods.
• Condensation of Annotations: DECODE reduces the size of enhancer annotations to 12.6% of their original extent while maintaining or improving conservation scores and enrichments for genome-wide association study (GWAS) variants, supporting downstream analyses such as causal variant mapping and functional validation.
• Efficiency in Computational Processing: The framework leverages graphic processing units (GPUs) to optimize deep-learning computations for regulatory element mapping.

Scientific Applications:

• Genome Evolution and Functional Genomics: More accurate enhancer annotations enable studies of genome evolution and the role of distal regulatory elements in biological functions.
• Causal Variant Mapping and GWAS Interpretation: Condensed, high-confidence enhancer calls improve power for mapping causal noncoding variants and interpreting GWAS signals.
• Disease Variant Investigation: Precise, cell-type-specific enhancer delineation facilitates investigation of how genetic variation influences disease mechanisms.
• Experimental Functional Validation and Fine-Mapping: High-resolution boundary localization supports targeted functional validations and fine-mapping of regulatory elements.

Methodology:

DECODE trains a deep neural network classifier using direct enhancer activity data from assays such as STARR-seq and applies a weakly-supervised object detection method to achieve detailed boundary localization and annotation condensation.