Why is only the upper percentile used when fitting an offest?

[paula-tataru]

Hi,

I have been running scVelo on the pancreas example data set, trying to understand the model details.

I have noticed that it seems that the lower percentile is not used when calculating the steady-state velocity with the default parameters.

Running the code below

import scvelo as scv
import numpy as np

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)

p = [5, 10, 20]
result_all = []
type_all = []
for l in p:
  result_p = []
  type_p = []
  for u in p:
    r = scv.tl.velocity(adata, mode='deterministic', copy=True, perc = [l, 100 - u])
    result_p.append(r.var["velocity_gamma"])
    type_p.append([l, 100 - u])
  result_all.append(result_p)
  type_all.append(type_p)
  
for k in range(len(p)):
  for i in range(len(p) - 1):
    for j in range(i + 1, len(p)):
      print(type_all[i][k], "vs", type_all[j][k], np.array_equal(result_all[i][k], result_all[j][k]))
      print(type_all[k][i], "vs", type_all[k][j], np.array_equal(result_all[k][i], result_all[k][j]))

results in the following

[5, 95] vs [10, 95] True
[5, 95] vs [5, 90] False
[5, 95] vs [20, 95] True
[5, 95] vs [5, 80] False
[10, 95] vs [20, 95] True
[5, 90] vs [5, 80] False
[5, 90] vs [10, 90] True
[10, 95] vs [10, 90] False
[5, 90] vs [20, 90] True
[10, 95] vs [10, 80] False
[10, 90] vs [20, 90] True
[10, 90] vs [10, 80] False
[5, 80] vs [10, 80] True
[20, 95] vs [20, 90] False
[5, 80] vs [20, 80] True
[20, 95] vs [20, 80] False
[10, 80] vs [20, 80] True
[20, 90] vs [20, 80] False

and it illustrates that gamma is the same regardless of the lower percentile..

Looking into the scVelo code, I saw that when fit_offset is set to False, which is the default value, only the upper percentile is used.

Why is that?

Thanks,
Paula

[WeilerP]

@paula-tataru, the splicing process begins/ends in the origin. As such, we do not need to infer the lower percentile for the fit since it is know a priori. Only the upper quantile needs to be inferred.

[paula-tataru]

Thanks for the prompt answer

Your article states "Steady states are expected at the lower and upper quantiles in phase space [...] hence, the ratio can be approximated by a linear regression on these extreme quantiles" under the "Steady-state model" methods section.

Maybe clarifying this in the API might help other users :)