HiCImpute

HiCImpute distinguishes structural zeros from sampling zeros and imputes single cell Hi-C (scHi-C) contact matrices to improve analysis of chromatin interactions, clustering, and cellular subtype discovery.

Key Features:

• Bayesian hierarchical model: Implements a Bayesian hierarchical framework for modeling scHi-C data.
• Structural vs sampling zero classification: Differentiates structural zeros (true absence of interaction) from sampling zeros (dropouts) in sparse scHi-C matrices.
• Spatial dependency modeling: Leverages spatial dependencies inherent in 2D scHi-C contact data.
• Information borrowing across cells: Borrows information from similar single cells to improve estimates and imputation.
• Bulk Hi-C integration: Integrates bulk Hi-C datasets when available to inform imputation and zero classification.
• Dropout imputation: Imputes dropout values to recover missing or under-sampled contacts.
• Improved sensitivity for structural zeros: Achieves high sensitivity in identifying structural zeros.
• Validation on synthetic and real data: Demonstrated performance on both simulated and empirical scHi-C datasets.

Scientific Applications:

• Accurate downstream analysis: Enables more reliable clustering and downstream analyses by correcting dropout-induced sparsity.
• Cell type and subtype identification: Facilitates discovery and refinement of cell types and subtypes from scHi-C data.
• Neuronal subtype discovery: Applied to identify distinct excitatory neuronal subtypes in layers L4 and L5 of the prefrontal cortex.
• Benchmarking and method evaluation: Used for performance assessment on synthetic and real scHi-C datasets.

Methodology:

Uses a Bayesian hierarchical model that leverages spatial dependencies in 2D scHi-C data, borrows information across similar single cells, and integrates bulk Hi-C when available to distinguish structural versus sampling zeros and impute dropout values.