holmrenser / GPcounts

GPcounts

GPcounts is Gaussian process regression package for counts data with negative binomial and zero-inflated negative binomial likelihoods described in the paper "Non-parametric modelling of temporal and spatial counts data from RNA-seq experiments". It is implemented in python, using the tensorflow and GPflow.

This is now published in Bioinformatics.

Installation:

1. Clone GPcounts repository:

git clone https://github.com/ManchesterBioinference/GPcounts.git

2. Install:

• Install requirements and package

cd GPcounts
pip install -r requirements.txt
cd 
git clone https://github.com/markvdw/RobustGP
cd RobustGP
python setup.py install
cd 
cd GPcounts
python setup.py install
cd 

In order to reproduce the paper results we have recorded the original packages used in a different requirements file

cd GPcounts
pip install -r paper_requirements.txt
cd
git clone https://github.com/markvdw/RobustGP
cd RobustGP
python setup.py install
cd 
cd GPcounts
python setup.py install
cd  

Notebooks to demonstrate GPcounts features:

Run the GPcounts/demo-notebooks

cd GPcounts/demo-notebooks
jupyter notebook

File name	Description
bulk_time_series	Applying GPcounts with negative binomial likelihood on bulk RNA-Seq time course data. We compare with Gaussian likelihoood results and show how to infer trajectories and carry out one-sample and two-samples tests
scRNA-Seq_time_series	Applying GPcounts with negative binomial likelihood on scRNA-seq gene expression data to find DE genes. We also demonstrate the use of sparse inference to improve computational efficiency.
GPcounts_spatial	Applying GPcounts with negative binomial likelihood to identify spatially expressed genes on spatial data from Mouse Olfactory Bulb. We demonstrate how to use the 'scaled' version which is based on data normalisation via multiplication of the NB mean by a location specific scale factor.
GPcounts_spatial_smf_scales	Applying GPcounts with negative binomial likelihood to identify spatially expressed genes on spatial data from Mouse Olfactory Bulb. We show how to calculate the scales factor using python's 'statsmodels' module instead of R code used in the above notebook. This way is easier and faster.
Branching_GPcounts	Applying GPcounts on the single-cell data to estimate the most probable branching locations for individual genes. This notebook demonstrates how to build a GPcounts model and plot the posterior model fit and posterior branching times. The application of this approach can be extended to the bulk time series data to identify the differentiation or the perturbation points,at which the two time-courses start to diverge for the first time.

Notebooks to reproduce additional paper results:

Run the GPcounts/paper-notebooks

cd GPcounts/paper-notebooks
jupyter notebook

File name	Description
Simulate_synthetic_counts	Simulate synthetic bulk RNA-seq timeseries
Anscombe_transformation	A modified version of Anscombe transformation notebook from SpatialDE packages
One_sample_test	One sample test on simulated bulk RNA-seq datasets and ROC curves to compare different likelihood functions
DESeq2_scRNA-seq	Run DESeq2 R package to normalize scRNA-seq Islet 𝛼 cells gene expression data
Precision_recall_spearman_correlation	Correlate DESeq2 and GPcounts with NB and Gaussian likelihood results on scRNA-seq Islet 𝛼 cells gene expression data
scRNA-Seq_time_series	Applying GPcounts with zero-inflated negative binomial, negative binomial and Gaussian likelihoods on scRNA-seq gene expression data to find DE genes. We also demonstrate the use of sparse inference to improve computational efficiency.
Sparse_Precision_recall_spearman_correlation	Correlate GPcounts using NB and Gaussian likelihood results on scRNA-seq Islet 𝛼 cells gene expression data with sparse GPcounts using same likelihoods comapring different methods to select the the location of inducing points.
20190403_evaluateCyclicAll	Modified R script from tradeSeq to run tradeSeq and read GPcounts with NB and Gaussian likelihoods results.
performanceTIPlotCyclic_acrossSimulations	Modified R script from tradeSeq to show the results of tradeSeq and GPcounts with NB and Gaussian likelihoods.
time_evaluation	GPcounts with one sample test on the tenth simulated cyclic dataset from tradeSeq package to compare full GP versus sparse GP using different number of inducing points.
evaluate_genes	Example of genes from the first simulated cyclic dataset from tradeSeq package fitted using GPcounts one-sample test with pseudotime estimated using slingshot package and slingshot and using the true simulated time
spatial_data_pvalue_hist	p-value histrogram plots for p-values as calculated via a permutation test and p-values as calculated assuming that the null follows a chi-squared with one degree distribution
MOUSE_OB_plots	Results from null which follows a chi-squared with one degree of freedom distribution and results from permuted null distribution.