Interpretation of allele-specific chromatin accessibility using cell state-aware deep learning.

Area of Science:

• Genomics
• Cancer Biology
• Bioinformatics

Background:

• Genomic variations in regulatory regions like enhancers and promoters significantly affect cellular function and phenotype.
• Identifying specific variants impacting cis-regulatory function within large personal or cancer genomes is a complex challenge.
• Explainable artificial intelligence (AI) offers a promising approach to predict and interpret the effects of noncoding genome variations on gene regulation.

Purpose of the Study:

• To develop and validate a specialized deep learning model for predicting the impact of genomic variants on gene regulation in melanoma.
• To identify and interpret allele-specific chromatin accessibility variants (ASCAVs) in melanoma genomes.
• To integrate multi-omics data for a comprehensive understanding of regulatory element function in cancer.

Main Methods:

• Generation of phased whole genomes with matched chromatin accessibility, histone modifications, and gene expression data for 10 melanoma cell lines.
• Training a deep learning model (DeepMEL2) on melanoma chromatin accessibility data to capture cell-specific regulatory programs.
• Comparison of DeepMEL2 performance against motif-based scoring and generic deep learning models.
• Utilizing ChIP-seq data to validate transcription factor binding site alterations and allele-specific binding.

Main Results:

• DeepMEL2 accurately captures regulatory programs in melanocytic and mesenchymal-like melanoma cell states, outperforming existing methods.
• Hundreds to thousands of ASCAVs were detected per genome, with 15-20% attributable to transcription factor binding site changes.
• A significant portion of ASCAVs resulted from altered AP-1 binding, confirmed by allele-specific binding of JUN and FOSL1.
• Augmenting DeepMEL2 with ChIP-seq data enabled high-confidence identification of TERT promoter mutations and ETS motif gains.

Conclusions:

• An integrative genomics approach combined with a novel deep learning model (DeepMEL2) can effectively identify and interpret functional enhancer mutations.
• The study demonstrates the utility of deep learning in analyzing allelic imbalance in chromatin accessibility and gene expression for variant interpretation.
• This methodology provides a powerful tool for understanding the functional consequences of genomic variation in cancer.