recall method

scallops/features/map_eval.py

@@ -1,3 +1,4 @@
+from collections.abc import Callable, Sequence
 from typing import Literal
 
 import anndata
@@ -6,6 +7,71 @@
 from sklearn.metrics.pairwise import cosine_similarity
 
 
+def recall(
+    true_positives_df: pd.DataFrame,
+    similarity_df: pd.DataFrame,
+    similarity_column_true_positives: str = "value",
+    similarity_column: str = "value",
+    quantiles: Sequence[float] = (0.01, 0.05),
+    two_sided: bool = True,
+    n_true_positives: Callable[[pd.DataFrame], int] = len,
+) -> pd.DataFrame:
+    """Compute recall at the specified quantiles.
+
+    :param true_positives_df: Dataframe containing true positive pairwise similarities
+    (e.g. relationships from CORUM)
+    :param similarity_df: Dataframe containing all pairwise similarities from which
+    quantiles are computed.
+    :param similarity_column_true_positives: Column in `true_positives_df` containing
+    similarity scores.
+    :param similarity_column: Column in `similarity_df` containing similarity scores.
+    :param quantiles: List of quantiles to extract relevant gene pairs.
+    :param two_sided: If two-sided, recall is computed at specific quantiles
+    and 1-quantiles.
+    :param n_true_positives: Function that accepts a dataframe and returns the
+    number of true positives. Default is the length of the dataframe
+    :return: Dataframe containing recall results.
+    """
+    results = []
+
+    n_relevant = n_true_positives(true_positives_df)
+    quantiles_ = quantiles
+    if two_sided:
+        quantiles = np.array(quantiles)
+        one_minus_quantiles = 1 - np.array(quantiles)
+        quantiles_ = [
+            (quantiles[i], one_minus_quantiles[i]) for i in range(len(quantiles))
+        ]
+        quantiles = np.concatenate((quantiles, one_minus_quantiles))
+    recall_thresholds = similarity_df[similarity_column].quantile(quantiles)
+    for quantile in quantiles_:
+        if two_sided:
+            threshold_low, threshold_high = (
+                recall_thresholds[quantile[0]],
+                recall_thresholds[quantile[1]],
+            )
+
+            df_retrieved = true_positives_df[
+                (true_positives_df[similarity_column_true_positives] >= threshold_high)
+                | (true_positives_df[similarity_column_true_positives] <= threshold_low)
+            ]
+        elif quantile <= 0.5:
+            df_retrieved = true_positives_df[
+                true_positives_df[similarity_column_true_positives]
+                <= recall_thresholds[quantile]
+            ]
+        else:
+            df_retrieved = true_positives_df[
+                true_positives_df[similarity_column_true_positives]
+                >= recall_thresholds[quantile]
+            ]
+        n_relevant_retrieved = n_true_positives(df_retrieved)
+        results.append(
+            [quantile[0] if two_sided else quantile, n_relevant_retrieved / n_relevant]
+        )
+    return pd.DataFrame(results, columns=["quantile", "recall"])
+
+
 def pairwise_similarities(
     data: anndata.AnnData, metric: Literal["cosine", "pearson"] = "cosine"
 ) -> np.ndarray: