mantispy.pp.feature_select_chatterjee

mantispy.pp.feature_select_chatterjee#

mantispy.pp.feature_select_chatterjee(adata, groupby='Metadata_Perturbation', threshold=0.1, m=1, seed=0, min_finite=40, key_added='selected_chatterjee', copy=False)#

Keep features whose values depend on the group, monotonically or otherwise.

Parameters:

  • adata (AnnData) – Object to select features on.

  • groupby (str (default: 'Metadata_Perturbation')) – obs column the features are tested against.

  • threshold (float (default: 0.1)) – Keep features scoring above this. xi is near zero under independence and approaches one when the feature is a deterministic function of the group, so the threshold is comparable across datasets in a way a correlation cutoff is not.

  • m (int (default: 1)) – Right nearest neighbors [Lin and Han, 2022]. m=1 is the coefficient of Chatterjee [2020] up to a term of order 1/n; larger values lower the noise floor without changing what the statistic converges to.

  • seed (int (default: 0)) – Seed for breaking ties between rows of the same group.

  • min_finite (int (default: 40)) – Fewest finite values a feature may be scored on. A feature measured fewer times than this scores NaN, which is above no threshold and so is never selected.

  • key_added (str (default: 'selected_chatterjee')) – Name of the boolean var column written.

  • copy (bool (default: False)) – Return a modified copy instead of mutating in place.

Return type:

AnnData | None

Returns:

None, or the modified copy. Writes var[key_added] and the statistic itself to var["chatterjee_xi"], which is NaN for a feature that is constant or too sparsely measured to score.

Raises:

ValueError – If groupby has a single group, so that no feature can depend on it.

Notes

Run it after feature_select(), which drops redundant or unmeasurable features; this keeps the features that carry information about the perturbation. Subset with mt.pp.subset_features(adata, key="selected_chatterjee").

A feature is scored on the rows where it was measured, so one that is missing in a whole group is scored against the groups that do have it rather than against the pattern of what is missing.

xi reaches one only for a noiseless function of the group, so real values are much lower. Over pki’s 899 selected features and 38 treatments the largest was 0.32, and the default threshold of 0.1 kept about half of them. Check the distribution in var["chatterjee_xi"] before relying on a fixed cutoff.

With shuffled group labels on the same data, the largest xi is 0.039 at m=1 and 0.027 at m=5, while the largest real value barely changes (0.322 and 0.321). m=1 is the default because the threshold was calibrated there. scmorph uses m=5, and on input free of ties the two implementations agree to 1e-9. scmorph breaks ties in y at random, so on tied input the two differ.