Adversarial attack of sequence-free enhancer prediction identifies chromatin architecture

Area of Science:

• Genomics
• Computational Biology
• Epigenetics

Background:

• Multicellular complexity arises from intricate gene regulation.
• Enhancers are key regulatory elements controlling gene programs.
• Previous enhancer prediction relied on DNA sequence and epigenomic data.

Purpose of the Study:

• To assess the feasibility of enhancer prediction using only epigenomic data.
• To develop a cell-type invariant enhancer prediction platform.
• To apply novel explainable AI methods for genomic network interpretation.

Main Methods:

• Utilized minimal epigenomic datasets for enhancer prediction.
• Trained neural networks on multi-cell type chromatin data.
• Employed adversarial particle swarm optimization (APSO) for network deconvolution.

Main Results:

• Demonstrated high-accuracy, genome-wide enhancer identification using chromatin data.
• Developed a cell-type invariant prediction platform based on protein binding patterns.
• Showcased APSO as a network-architecture independent method for feature attribution.

Conclusions:

• Epigenomic data is sufficient for accurate enhancer prediction.
• The developed platform offers a robust, cell-type invariant approach.
• APSO provides a versatile tool for understanding genomic model predictions.