pylimma.decide_tests

pylimma.decide_tests(data, method='separate', adjust_method='BH', p_value=0.05, lfc=0.0, coefficients=None, cor_matrix=None, tstat=None, df=inf, genewise_p_value=None, *, key='pylimma')[source]

Classify genes as differentially expressed.

Applies multiple testing correction and classifies each gene-coefficient combination as up-regulated (1), down-regulated (-1), or not significant (0).

Parameters:

  • data (AnnData, dict, or ndarray) – Fit object from e_bayes(), or a matrix of p-values. If AnnData, reads from adata.uns[key].

  • method (str, default "separate") –

    Method for multiple testing correction:

    • ”separate”: adjust p-values for each coefficient separately

    • ”global”: adjust all p-values together

    • ”hierarchical”: first test overall significance (F-test), then adjust within significant genes

    • ”nestedF”: use nested F-tests for multiple contrasts

  • adjust_method (str, default "BH") – P-value adjustment method: “BH”, “bonferroni”, “holm”, “BY”, “none”.

  • p_value (float, default 0.05) – Significance threshold for adjusted p-values.

  • lfc (float, default 0.0) – Log fold-change threshold. Genes with |logFC| < lfc are set to 0.

  • key (str, default "pylimma") – Key for fit results in adata.uns (AnnData input only).

  • coefficients (ndarray | None)

  • cor_matrix (ndarray | None)

  • tstat (ndarray | None)

  • df (float | ndarray)

  • genewise_p_value (ndarray | None)

Returns:

Matrix of test results with values -1 (down), 0 (not significant), 1 (up). Shape is (n_genes, n_coefficients).

Return type:

ndarray

Notes

The “separate” method is the default and most commonly used. It adjusts p-values within each coefficient independently.

The “hierarchical” and “nestedF” methods first perform an overall F-test for each gene, then test individual contrasts only for genes that pass the F-test. This can increase power when there are many contrasts.

Examples

>>> fit = e_bayes(lm_fit(expr, design))
>>> results = decide_tests(fit, p_value=0.05, lfc=1)
>>> np.sum(results == 1, axis=0)  # count up-regulated per coefficient