This package implements maximum likelihood estimation of the zero-inflated negative binomial model described in:
Abhishek K Sarkar, Po-Yuan Tung, John D. Blischak, Jonathan E. Burnett, Yang I. Li, Matthew Stephens, Yoav Gilad. “Discovery and characterization of variance QTLs in human induced pluripotent stem cells”. PLoS Genetics (2019). https://doi.org/10.1371/journal.pgen.1008045
The key idea is that we learn latent point-Gamma distributions for expression per individual/condition per gene, and then perform all downstream analysis on (the parameters of) those distributions.
pip install git+https://www.github.com/aksarkar/scqtl.git#egg=scqtl
The package has been tested with:
numpy==1.14.3 scipy==1.0.0 tensorflow==1.3.0
In the analysis pipeline for the paper, we used a common Conda environment for all R and python code which includes these exact versions.
import numpy as np
import scqtl
# Generate some ZINB-distributed counts
num_samples = 1000
umi = np.concatenate([scqtl.simulation.simulate(
num_samples=num_samples,
size=1e5,
seed=trial)[0][:,:1] for trial in range(10)], axis=1)
size_factor = 1e5 * np.ones((num_samples, 1))
# Generate a null design matrix
design = np.zeros((num_samples, 1))
# Map all samples to one individual/condition, i.e. one set of ZINB parameters
onehot = np.ones((num_samples, 1))
# Find the NB MLE
# Important: casting to float32 is required
init = scqtl.tf.fit(
umi=umi.astype(np.float32),
onehot=onehot.astype(np.float32),
design=design.astype(np.float32),
size_factor=size_factor.astype(np.float32),
learning_rate=1e-3,
max_epochs=20000,
verbose=True,
)
# Find the ZINB MLE, starting from the NB MLE
log_mu, log_phi, logodds, nb_llik, zinb_llik = scqtl.tf.fit(
umi=umi.astype(np.float32),
onehot=onehot.astype(np.float32),
design=design.astype(np.float32),
size_factor=size_factor.astype(np.float32),
learning_rate=1e-3,
max_epochs=20000,
warm_start=init[:3],
verbose=True)