TOAST

TOAST identifies topologically associated domains (TADs) from Hi-C contact matrices by applying graph auto-encoders and HDBSCAN to generate embeddings, cluster genomic bins, delineate TAD boundaries, and assess their association with genomic markers.

Key Features:

• Graph representation: Conceptualizes each genomic bin as a node in a graph with an adjacency matrix derived from Hi-C contact matrices.
• Graph auto-encoders: Learns informative embeddings that capture complex relationships between genomic regions using graph auto-encoders.
• Unsupervised clustering (HDBSCAN): Processes auto-encoder-derived embeddings with HDBSCAN for unsupervised clustering of genomic bins.
• Label-based TAD calling: Assigns labels to each genomic bin and defines contiguous regions with identical labels as TADs.
• Evaluation on simulated data: Benchmarks TAD boundary detection on diverse simulated Hi-C contact matrices, reporting improvements in speed and accuracy compared to existing algorithms.
• Anchoring ratio analysis: Computes anchoring ratios of TAD boundaries for genomic markers including CTCF, SMC3, RAD21, POLR2A, H3K36me3, H3K9me3, H3K4me3, H3K4me1, Enhancer, and Promoter.

Scientific Applications:

• TAD identification: Identification and delineation of TADs from Hi-C contact matrices to study three-dimensional genome organization.
• Boundary-marker association: Assessing associations between TAD boundaries and genomic markers such as CTCF, SMC3, RAD21, POLR2A, histone marks (H3K36me3, H3K9me3, H3K4me3, H3K4me1), enhancers, and promoters.
• Algorithm benchmarking: Comparative benchmarking of TAD-calling algorithms on simulated Hi-C datasets for speed and accuracy.

Methodology:

Derives a graph adjacency matrix from Hi-C contact matrices, trains graph auto-encoders to produce embeddings, clusters embeddings with HDBSCAN to assign labels to genomic bins, defines contiguous identical labels as TADs, and evaluates performance on simulated Hi-C contact matrices while computing anchoring ratios for specified genomic markers.