Why is the Welch t-test used for ranking velocity genes when data is not normally distributed

[paula-tataru]

Hi,

According to the API, you use the Welch t-test to rank the velocity genes.
https://scvelo.readthedocs.io/scvelo.tl.rank_velocity_genes/#scvelo.tl.rank_velocity_genes

This test assumes that data is normally distributed. I got curious to see if the data meets this requirement, so I looked at the pancreas data set and tested for normality the velocity estimates for the first cluster for the top velocity gene according to the ranking.

import scvelo as scv
import numpy as np
from scipy.stats import shapiro
from statsmodels.graphics.gofplots import qqplot
from matplotlib import pyplot

adata = scv.datasets.pancreas()
scv.pp.filter_and_normalize(adata, min_shared_counts=20, n_top_genes=2000)
scv.pp.moments(adata, n_neighbors=30)
scv.tl.velocity(adata, mode='deterministic')
scv.tl.rank_velocity_genes(adata, groupby='clusters')

df = scv.DataFrame(adata.uns['rank_velocity_genes']['names'])
cluster = df.columns[0]
g = np.where(adata.var_names == df[cluster][0])
p = np.where(adata.obs['clusters'] == cluster)
# velocity values for the cells in the first cluster and its top gene
v = adata.layers["velocity"][p, g][0]

pyplot.hist(v)
qqplot(v, line='s')

stat, p = shapiro(v)
print("P-value:", p)

Here are the two plots:

And the p-value is about 1.75e-9.

The data looks to be far from being normally distributed. Why was the Welch t-test used?

/Paula

[WeilerP]

@VolkerBergen, any comments/thoughts on this?