GPcounts

GPcounts models temporal and spatial RNA-seq count data using Gaussian process regression with negative binomial and zero-inflated negative binomial likelihoods to account for over-dispersion and dropout in bulk and single-cell datasets.

Key Features:

• Non-parametric Gaussian process regression: Models changes in gene expression over time or space using Gaussian processes to provide flexibility beyond parametric models.
• Negative binomial likelihood: Employs the negative binomial distribution to capture over-dispersion typical of RNA-seq counts in bulk and single-cell data.
• Zero-inflated negative binomial option: Offers a zero-inflated NB likelihood to model dropout events in single-cell RNA-seq.
• Variational Bayesian inference: Uses variational Bayesian inference for scalable approximate posterior inference on large transcriptomics datasets.
• Logarithmic link and dispersion estimation: Models the NB mean via a logarithmic link function and estimates the dispersion parameter by maximum likelihood.

Scientific Applications:

• Temporal inference: Identifies changes in gene expression in time course and pseudotime analyses, including application to single-cell datasets after pseudotime and branching inference.
• Spatial inference: Detects genes with spatially variable expression in spatial transcriptomics data, exemplified by analyses of the mouse olfactory bulb, using NB/ZINB models for more realistic counts modeling compared to Gaussian or Poisson likelihoods.

Methodology:

Implements Gaussian process regression for counts within the GPflow library in Python, using negative binomial and zero-inflated negative binomial likelihoods, a logarithmic link for the mean, dispersion estimated by maximum likelihood, and variational Bayesian inference for scalable approximation.