tGPLVM: A robust nonlinear manifold model for single cell RNA-seq data.

Intro

Dimension reduction is a common and critical first step in analysis of high throughput singe cell RNA sequencing. tGPLVM is a nonparametric, generative model for nonlinear manifold learning; that is a flexible, nearly assumption-free model that doesn't require setting parameters a priori (e.g. number of dimensions, perplexity, etc.) and provides uncertainty estimates for sample mappings. tGPLVM can be used for visualization of high-dimensional data or as part of a pipeline for cell type identification or pseudotime reconstruction.

We provide a script for fitting the model with Black Box Variational Inference for speed and scabality. A batch learning implementation is also provided for larger datasets that need to be fit under memory restriction.

Usage

Requirements

tGPLVM is implemented in python 2.7 with the following packages:

1. numpy 1.14.5
2. pandas 0.23.3
3. h5py 2.8.0
4. tensorflow 1.6.0
5. edwards 1.3.5
6. sklearn 0.19.2

Running

Input: A numpy array or sparse csr/csc matrix of scRNA counts (or other types data) with format N cells (samples) as rows by p genes (features) as columns (loaded to y_train). Input this directly into the code.

Options: The following parameters can be adjusted in the script to adjust inference:

1. Degrees of freedom (--df) - default: 4
2. Use t-Distribution error model (otherwise normal error) (--T) - default: True
3. Initial Number of Dimensions (--Q) - default: 3
4. Kernel Function
   • Matern 1/2, 3/2, 5/2 (--m12, --m32, --m52) - default: True
   • Periodic (--per_bool) - default: False
5. Number of Inducing Points (--m) - default: 30
6. Batch size (--M) - default: 250
7. Max iterations (--iterations) - default: 5000
8. Save frequency (--save_freq): - default: 250
9. Sparse data type (is CSC or CSR) (--sparse): - default: False
10. PCA Initialization (otherwise random initialization) (--pca_init): - default: True
11. Output directory (--out): - default: ./test

Output: hdf5 file with

1. Latent mapping posterior (mean and variance)
2. Gene-specific noise
3. Kernel hyperparameters (variance, lengthscale)
4. Inducing points in latent and high-dimensional space

Example:

When the input is Test_3_Pollen.h5, the following code runs 250 iterations with the full dataset

python tGPLVM-batch.py --Q 2 --M 249 --p 6982 --m12 True --m32 True --m52 True --iterations 250 --out ./test

We provide the input code for two other files:

1. tapio_tcell_tpm.txt - Data from Lonnberg gpfates. Data is available at https://github.com/Teichlab/GPfates
2. 1M_neurons_filtered_gene_bc_matrices_h5.h5 - 1 million 10x mice brains cell. Data is available at https://support.10xgenomics.com/single-cell-gene-expression/datasets/1.3.0/1M_neurons. Make sure to set --sparse True for this data.

The final data from the paper is available here: https://support.10xgenomics.com/single-cell-gene-expression/datasets/1.1.0/cd34