grassp.tl.svm_tune_hyperparameters

svm_tune_hyperparameters(data, gt_col, C_range=array([0.0625, 0.125, 0.25, 0.5, 1., 2., 4., 8., 16.]), gamma_range=array([1.e-03, 1.e-02, 1.e-01, 1.e+00, 1.e+01, 1.e+02]), class_weight='balanced', cv_splits=5, cv_repeats=20, n_jobs=-1, random_state=None, key_added='svm', inplace=True)

Grid-search the SVM’s C and gamma over marker proteins.

Despite the name it replaces, this does not train a reusable model: it scores a grid of (C, gamma) by repeated stratified cross-validation and stores the winning pair in .uns for svm_annotation(), which fits its own estimator from scratch. Nothing about the fitted classifier survives the call, so calling it “training” promised a model you could apply elsewhere – unlike tagm_map_train(), which genuinely persists one.

Parameters:

data AnnData

anndata.AnnData object with proteins as observations.

gt_col str

Observation column containing marker annotations. Proteins with NaN values are considered unknown and excluded from training.

C_range ndarray (default: array([ 0.0625,  0.125 ,  0.25  ,  0.5   ,  1.    ,  2.    ,  4.    ,         8.    , 16.    ]))

Array of C (regularization) values to search. Default: 2^-4 to 2^4.

gamma_range ndarray (default: array([1.e-03, 1.e-02, 1.e-01, 1.e+00, 1.e+01, 1.e+02]))

Array of gamma (kernel coefficient) values. Default: 10^-3 to 10^2.

cv_splits int (default: 5)

Number of cross-validation folds (default 5).

cv_repeats int (default: 20)

Number of CV repetitions (default 20). Total fits per parameter combination: cv_splits × cv_repeats.

n_jobs int (default: -1)

Number of parallel jobs. -1 uses all available cores.

random_state int | None (default: None)

Random seed for reproducibility.

key_added str (default: 'svm')

Key prefix for storing results in .uns (default "svm"), so the winning parameters land in uns[f"{key_added}_params"].

inplace bool (default: True)

If True store results in .uns; if False return the grid-search object and the parameter dictionary, which is useful for inspecting the full CV grid.

class_weight None | dict | Literal['balanced']

Return type:

tuple[GridSearchCV, dict] | None

Returns:

None or tuple If inplace=False, returns (grid_search, params). Otherwise writes data.uns[f"{key_added}_params"] and returns None.

Examples

>>> import grassp as gr
>>> adata = gr.ds.hein_2024(enrichment="enriched")

# When actually training, increase cv_repeats and cv_splits # We recommend >20 repeats with 5 splits >>> gr.tl.svm_tune_hyperparameters(adata, gt_col=”hein2024_gt_component”, cv_repeats=2, cv_splits=2, random_state=42) Fitting 4 folds for each of 54 candidates, totalling 216 fits >>> adata.uns[“svm_params”][“best_params”] {‘C’: 2.0, ‘gamma’: 0.01}