predict_phenotype¶
Out-of-fold prediction of a sample phenotype from the first n_pcs PCs of the sample embedding X_DR_sample (0.5.0+). Standardized PCs feed an L2 logistic regression (binary or multiclass) or, for a continuous phenotype, a ridge regression. The penalty is chosen by an inner 3-fold grouped cross-validation inside each training split; the outer split is stratified, grouped and seeded. With group_col, every sample of a group (for example a donor) stays in one fold, so no donor is in training and test at once.
Source: sampledisco/sample_prediction/prediction.py:31
Signature¶
def predict_phenotype(
pseudo_adata: AnnData,
label_col: str,
*,
group_col: Optional[str] = None,
n_pcs: Optional[int] = None,
task: str = "classification",
score_by: str = "sample",
n_splits: int = 5,
random_state: int = 0,
) -> Tuple[pd.DataFrame, Dict]
Parameters¶
| Name | Type | Default | Description |
|---|---|---|---|
pseudo_adata |
AnnData | — | Sample-level AnnData (build_sample_adata) with X_DR_sample in .uns (DataFrame) or .obsm. |
label_col |
str | — | Phenotype column in .obs. Samples with a missing value are skipped. |
group_col |
str, optional | None |
Column whose groups must never be split across folds (donor, or sample for multi-omics units). None = every sample is its own group. |
n_pcs |
int, optional | None |
Use the first n_pcs PCs; None = all. |
task |
str | "classification" |
"classification" (logistic regression, C in {10, 1, 0.1, 0.01}) or "regression" (ridge, alpha in {0.1, 1, 10, 100}). Ties go to the weakest penalty. |
score_by |
str | "sample" |
"group" averages the out-of-fold probabilities (or predictions) and labels per group_col group before scoring (majority label for classification). |
n_splits |
int | 5 |
Number of outer folds. |
random_state |
int | 0 |
Seed of the outer split. |
Returns¶
Tuple[DataFrame, dict] — (pred, metrics).
pred: indexed by sample;label,predicted,fold, andprob_<class>per class for classification (out-of-fold).metrics:task,score_by,n_samples,n_groups,n_pcs, the penalty chosen per fold (Coralpha), andauc(binary; second class in sorted order is positive),accuracy,balanced_accuracy(classification) orr2(regression).
Usage¶
from sampledisco import predict_phenotype
pred, metrics = predict_phenotype(pseudo_adata, "cmv", group_col="donor", score_by="group")
See the tutorial.