HiCPotts

HiCPotts provides a comprehensive Bayesian framework for analyzing Hi-C interaction data, integrating both spatial and genomic biases within a probabilistic modeling framework to segment the genomic landscape into meaningful components and quantify the impact of genomic biases on observed interaction frequencies.

Key Features:

• Spatial dependency modeling: Leverages the Potts model to capture and quantify spatial dependencies across interaction loci arranged on a genomic lattice by treating each interaction as a spatially correlated random variable.
• Genomic bias correction: Employs a regression-based approach incorporating multiple covariates including genomic distance (D), GC-content (GC), transposable elements (TEs), and accessibility score (Acc).
• Hierarchical mixture modeling: Embeds covariates into a hierarchical mixture model to characterize each interaction's probability of belonging to one of several latent components.
• Probabilistic classification: Computes posterior probabilities that assign each interaction to a specific component, distinguishing genuine high-confidence interactions (signal) from background noise and potential false signals.
• Posterior inference: Computes posterior means of regression parameters and other quantities of interest from MCMC samples for flexible posterior inference on model parameters and hidden states.

Scientific Applications:

• Hi-C data analysis: Enables robust segmentation of the genomic landscape into noise, true signals, and false signals in Hi-C interaction data.
• Bias quantification: Quantifies the impact of genomic biases including genomic distance, GC-content, transposable elements, and chromatin accessibility on observed Hi-C interaction frequencies.
• Interaction state assignment: Provides interpretable, model-based assignments of interaction states to better understand the interplay between genomic organization, biases, and spatial correlation in Hi-C data.

Methodology:

The Potts model captures spatial dependencies across interaction loci on a genomic lattice. A regression-based approach incorporates covariates (genomic distance, GC-content, transposable elements, accessibility score) into a hierarchical mixture model. Model parameters including regression coefficients, zero-inflation parameters (for ZIP/ZINB distributions), and dispersion terms (for NB/ZINB distributions) are inferred via an MCMC sampler that draws samples from the joint posterior distribution.