risk-network 0.0.6b10__py3-none-any.whl → 0.0.7__py3-none-any.whl

risk/neighborhoods/domains.py

@@ -4,6 +4,7 @@ risk/neighborhoods/domains
 """
 
 from contextlib import suppress
+from itertools import product
 from tqdm import tqdm
 from typing import Tuple
 
@@ -14,6 +15,7 @@ from sklearn.metrics import silhouette_score
 
 from risk.annotations import get_description
 from risk.constants import GROUP_LINKAGE_METHODS, GROUP_DISTANCE_METRICS
+from risk.log import logger
 
 
 def define_domains(
@@ -23,7 +25,8 @@ def define_domains(
     linkage_method: str,
     linkage_metric: str,
 ) -> pd.DataFrame:
-    """Define domains and assign nodes to these domains based on their enrichment scores and clustering.
+    """Define domains and assign nodes to these domains based on their enrichment scores and clustering,
+    handling errors by assigning unique domains when clustering fails.
 
     Args:
         top_annotations (pd.DataFrame): DataFrame of top annotations data for the network nodes.
@@ -35,31 +38,31 @@ def define_domains(
     Returns:
         pd.DataFrame: DataFrame with the primary domain for each node.
     """
-    # Check if there's more than one column in significant_neighborhoods_enrichment
-    if significant_neighborhoods_enrichment.shape[1] == 1:
-        print("Single annotation detected. Skipping clustering.")
-        top_annotations["domain"] = 1  # Assign a default domain or handle appropriately
-    else:
-        # Perform hierarchical clustering on the binary enrichment matrix
+    try:
+        # Transpose the matrix to cluster annotations
         m = significant_neighborhoods_enrichment[:, top_annotations["top attributes"]].T
         best_linkage, best_metric, best_threshold = _optimize_silhouette_across_linkage_and_metrics(
             m, linkage_criterion, linkage_method, linkage_metric
         )
-        try:
-            Z = linkage(m, method=best_linkage, metric=best_metric)
-        except ValueError as e:
-            raise ValueError("No significant annotations found.") from e
-
-        print(
+        # Perform hierarchical clustering
+        Z = linkage(m, method=best_linkage, metric=best_metric)
+        logger.warning(
             f"Linkage criterion: '{linkage_criterion}'\nLinkage method: '{best_linkage}'\nLinkage metric: '{best_metric}'"
         )
-        print(f"Optimal linkage threshold: {round(best_threshold, 3)}")
-
+        logger.debug(f"Optimal linkage threshold: {round(best_threshold, 3)}")
+        # Calculate the optimal threshold for clustering
         max_d_optimal = np.max(Z[:, 2]) * best_threshold
-        domains = fcluster(Z, max_d_optimal, criterion=linkage_criterion)
         # Assign domains to the annotations matrix
+        domains = fcluster(Z, max_d_optimal, criterion=linkage_criterion)
         top_annotations["domain"] = 0
         top_annotations.loc[top_annotations["top attributes"], "domain"] = domains
+    except ValueError:
+        # If a ValueError is encountered, handle it by assigning unique domains
+        n_rows = len(top_annotations)
+        logger.error(
+            f"Error encountered. Skipping clustering and assigning {n_rows} unique domains."
+        )
+        top_annotations["domain"] = range(1, n_rows + 1)  # Assign unique domains
 
     # Create DataFrames to store domain information
     node_to_enrichment = pd.DataFrame(
@@ -166,21 +169,20 @@ def _optimize_silhouette_across_linkage_and_metrics(
     total_combinations = len(linkage_methods) * len(linkage_metrics)
 
     # Evaluating optimal linkage method and metric
-    for method in tqdm(
-        linkage_methods,
+    for method, metric in tqdm(
+        product(linkage_methods, linkage_metrics),
         desc="Evaluating optimal linkage method and metric",
         total=total_combinations,
         bar_format="{l_bar}{bar}| {n_fmt}/{total_fmt} [{elapsed}<{remaining}]",
     ):
-        for metric in linkage_metrics:
-            with suppress(Exception):
-                Z = linkage(m, method=method, metric=metric)
-                threshold, score = _find_best_silhouette_score(Z, m, metric, linkage_criterion)
-                if score > best_overall_score:
-                    best_overall_score = score
-                    best_overall_threshold = threshold
-                    best_overall_method = method
-                    best_overall_metric = metric
+        with suppress(Exception):
+            Z = linkage(m, method=method, metric=metric)
+            threshold, score = _find_best_silhouette_score(Z, m, metric, linkage_criterion)
+            if score > best_overall_score:
+                best_overall_score = score
+                best_overall_threshold = threshold
+                best_overall_method = method
+                best_overall_metric = metric
 
     return best_overall_method, best_overall_metric, best_overall_threshold
 

risk/neighborhoods/neighborhoods.py

@@ -3,21 +3,24 @@ risk/neighborhoods/neighborhoods
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 """
 
+import random
 import warnings
 from typing import Any, Dict, List, Tuple
 
 import networkx as nx
 import numpy as np
 from sklearn.exceptions import DataConversionWarning
+from sklearn.metrics.pairwise import cosine_similarity
 
 from risk.neighborhoods.community import (
-    calculate_dijkstra_neighborhoods,
+    calculate_greedy_modularity_neighborhoods,
     calculate_label_propagation_neighborhoods,
     calculate_louvain_neighborhoods,
     calculate_markov_clustering_neighborhoods,
     calculate_spinglass_neighborhoods,
     calculate_walktrap_neighborhoods,
 )
+from risk.log import logger
 
 # Suppress DataConversionWarning
 warnings.filterwarnings(action="ignore", category=DataConversionWarning)
@@ -25,7 +28,7 @@ warnings.filterwarnings(action="ignore", category=DataConversionWarning)
 
 def get_network_neighborhoods(
     network: nx.Graph,
-    distance_metric: str = "dijkstra",
+    distance_metric: str = "louvain",
     edge_length_threshold: float = 1.0,
     louvain_resolution: float = 1.0,
     random_seed: int = 888,
@@ -34,8 +37,8 @@ def get_network_neighborhoods(
 
     Args:
         network (nx.Graph): The network graph.
-        distance_metric (str): The distance metric to use ('euclidean', 'dijkstra', 'louvain', 'affinity_propagation',
-            'label_propagation', 'markov_clustering', 'walktrap', 'spinglass').
+        distance_metric (str): The distance metric to use ('greedy_modularity', 'louvain', 'label_propagation',
+            'markov_clustering', 'walktrap', 'spinglass').
         edge_length_threshold (float): The edge length threshold for the neighborhoods.
         louvain_resolution (float, optional): Resolution parameter for the Louvain method. Defaults to 1.0.
         random_seed (int, optional): Random seed for methods requiring random initialization. Defaults to 888.
@@ -43,12 +46,19 @@ def get_network_neighborhoods(
     Returns:
         np.ndarray: Neighborhood matrix calculated based on the selected distance metric.
     """
-    network = _create_percentile_limited_subgraph(network, edge_length_threshold)
+    # Set random seed for reproducibility in all methods besides Louvain, which requires a separate seed
+    random.seed(random_seed)
+    np.random.seed(random_seed)
+
+    # Create a subgraph based on the edge length percentile threshold
+    network = _create_percentile_limited_subgraph(
+        network, edge_length_percentile=edge_length_threshold
+    )
 
-    if distance_metric == "dijkstra":
-        return calculate_dijkstra_neighborhoods(network)
     if distance_metric == "louvain":
         return calculate_louvain_neighborhoods(network, louvain_resolution, random_seed=random_seed)
+    if distance_metric == "greedy_modularity":
+        return calculate_greedy_modularity_neighborhoods(network)
     if distance_metric == "label_propagation":
         return calculate_label_propagation_neighborhoods(network)
     if distance_metric == "markov_clustering":
@@ -59,41 +69,51 @@ def get_network_neighborhoods(
         return calculate_spinglass_neighborhoods(network)
 
     raise ValueError(
-        "Incorrect distance metric specified. Please choose from 'dijkstra', 'louvain',"
+        "Incorrect distance metric specified. Please choose from 'greedy_modularity', 'louvain',"
         "'label_propagation', 'markov_clustering', 'walktrap', 'spinglass'."
     )
 
 
 def _create_percentile_limited_subgraph(G: nx.Graph, edge_length_percentile: float) -> nx.Graph:
-    """Calculate the edge length corresponding to the given percentile of edge lengths in the graph
-    and create a subgraph with all nodes and edges below this length.
+    """Create a subgraph containing all nodes and edges where the edge length is below the
+    specified percentile of all edge lengths in the input graph.
 
     Args:
-        G (nx.Graph): The input graph.
-        edge_length_percentile (float): The percentile to calculate (between 0 and 1).
+        G (nx.Graph): The input graph with 'length' attributes on edges.
+        edge_length_percentile (float): The percentile (between 0 and 1) to filter edges by length.
 
     Returns:
-        nx.Graph: A subgraph with all nodes and edges below the edge length corresponding to the given percentile.
+        nx.Graph: A subgraph with all nodes and edges where the edge length is below the
+        calculated threshold length.
     """
-    # Extract edge lengths from the graph
+    # Extract edge lengths and handle missing lengths
     edge_lengths = [d["length"] for _, _, d in G.edges(data=True) if "length" in d]
+    if not edge_lengths:
+        raise ValueError(
+            "No edge lengths found in the graph. Ensure edges have 'length' attributes."
+        )
+
     # Calculate the specific edge length for the given percentile
     percentile_length = np.percentile(edge_lengths, edge_length_percentile * 100)
-    # Create a new graph with all nodes from the original graph
+    # Create the subgraph by directly filtering edges during iteration
     subgraph = nx.Graph()
-    subgraph.add_nodes_from(G.nodes(data=True))
-    # Add edges to the subgraph if they are below the specified percentile length
+    subgraph.add_nodes_from(G.nodes(data=True))  # Retain all nodes from the original graph
+    # Add edges below the specified percentile length in a single pass
     for u, v, d in G.edges(data=True):
         if d.get("length", 1) <= percentile_length:
             subgraph.add_edge(u, v, **d)
 
+    # Return the subgraph; optionally check if it's too sparse
+    if subgraph.number_of_edges() == 0:
+        raise Warning("The resulting subgraph has no edges. Consider adjusting the percentile.")
+
     return subgraph
 
 
 def process_neighborhoods(
     network: nx.Graph,
     neighborhoods: Dict[str, Any],
-    impute_depth: int = 1,
+    impute_depth: int = 0,
     prune_threshold: float = 0.0,
 ) -> Dict[str, Any]:
     """Process neighborhoods based on the imputation and pruning settings.
@@ -101,7 +121,7 @@ def process_neighborhoods(
     Args:
         network (nx.Graph): The network data structure used for imputing and pruning neighbors.
         neighborhoods (dict): Dictionary containing 'enrichment_matrix', 'binary_enrichment_matrix', and 'significant_enrichment_matrix'.
-        impute_depth (int, optional): Depth for imputing neighbors. Defaults to 1.
+        impute_depth (int, optional): Depth for imputing neighbors. Defaults to 0.
         prune_threshold (float, optional): Distance threshold for pruning neighbors. Defaults to 0.0.
 
     Returns:
@@ -110,7 +130,7 @@ def process_neighborhoods(
     enrichment_matrix = neighborhoods["enrichment_matrix"]
     binary_enrichment_matrix = neighborhoods["binary_enrichment_matrix"]
     significant_enrichment_matrix = neighborhoods["significant_enrichment_matrix"]
-    print(f"Imputation depth: {impute_depth}")
+    logger.debug(f"Imputation depth: {impute_depth}")
     if impute_depth:
         (
             enrichment_matrix,
@@ -123,7 +143,7 @@ def process_neighborhoods(
             max_depth=impute_depth,
         )
 
-    print(f"Pruning threshold: {prune_threshold}")
+    logger.debug(f"Pruning threshold: {prune_threshold}")
     if prune_threshold:
         (
             enrichment_matrix,
@@ -167,55 +187,134 @@ def _impute_neighbors(
             - np.ndarray: The imputed alpha threshold matrix.
             - np.ndarray: The significant enrichment matrix with non-significant entries set to zero.
     """
-    # Calculate shortest distances for each node to determine the distance threshold
-    shortest_distances = []
-    for node in network.nodes():
-        try:
-            neighbors = [
-                n for n in network.neighbors(node) if binary_enrichment_matrix[n].sum() != 0
-            ]
-        except IndexError as e:
-            raise IndexError(
-                f"Failed to find neighbors for node '{node}': Ensure that the node exists in the network and that the binary enrichment matrix is correctly indexed."
-            ) from e
-
-        # Calculate the shortest distance to a neighbor
-        if neighbors:
-            shortest_distance = min([_get_euclidean_distance(node, n, network) for n in neighbors])
-            shortest_distances.append(shortest_distance)
+    # Calculate the distance threshold value based on the shortest distances
+    enrichment_matrix, binary_enrichment_matrix = _impute_neighbors_with_similarity(
+        network, enrichment_matrix, binary_enrichment_matrix, max_depth=max_depth
+    )
+    # Create a matrix where non-significant entries are set to zero
+    significant_enrichment_matrix = np.where(binary_enrichment_matrix == 1, enrichment_matrix, 0)
+
+    return enrichment_matrix, binary_enrichment_matrix, significant_enrichment_matrix
+
 
+def _impute_neighbors_with_similarity(
+    network: nx.Graph,
+    enrichment_matrix: np.ndarray,
+    binary_enrichment_matrix: np.ndarray,
+    max_depth: int = 3,
+) -> Tuple[np.ndarray, np.ndarray]:
+    """Impute non-enriched nodes based on the closest enriched neighbors' profiles and their similarity.
+
+    Args:
+        network (nx.Graph): The network graph with nodes having IDs matching the matrix indices.
+        enrichment_matrix (np.ndarray): The enrichment matrix with rows to be imputed.
+        binary_enrichment_matrix (np.ndarray): The alpha threshold matrix to be imputed similarly.
+        max_depth (int): Maximum depth of nodes to traverse for imputing values.
+
+    Returns:
+        Tuple[np.ndarray, np.ndarray]: A tuple containing:
+            - The imputed enrichment matrix.
+            - The imputed alpha threshold matrix.
+    """
     depth = 1
     rows_to_impute = np.where(binary_enrichment_matrix.sum(axis=1) == 0)[0]
     while len(rows_to_impute) and depth <= max_depth:
-        next_rows_to_impute = []
-        for row_index in rows_to_impute:
-            neighbors = nx.single_source_shortest_path_length(network, row_index, cutoff=depth)
-            valid_neighbors = [
-                n
-                for n in neighbors
-                if n != row_index
-                and binary_enrichment_matrix[n].sum() != 0
-                and enrichment_matrix[n].sum() != 0
-            ]
-            if valid_neighbors:
-                closest_neighbor = min(
-                    valid_neighbors, key=lambda n: _get_euclidean_distance(row_index, n, network)
+        # Iterate over all enriched nodes
+        for row_index in range(binary_enrichment_matrix.shape[0]):
+            if binary_enrichment_matrix[row_index].sum() != 0:
+                enrichment_matrix, binary_enrichment_matrix = _process_node_imputation(
+                    row_index, network, enrichment_matrix, binary_enrichment_matrix, depth
                 )
-                # Impute the row with the closest valid neighbor's data
-                enrichment_matrix[row_index] = enrichment_matrix[closest_neighbor]
-                binary_enrichment_matrix[row_index] = binary_enrichment_matrix[
-                    closest_neighbor
-                ] / np.sqrt(depth + 1)
-            else:
-                next_rows_to_impute.append(row_index)
 
-        rows_to_impute = next_rows_to_impute
+        # Update rows to impute for the next iteration
+        rows_to_impute = np.where(binary_enrichment_matrix.sum(axis=1) == 0)[0]
         depth += 1
 
-    # Create a matrix where non-significant entries are set to zero
-    significant_enrichment_matrix = np.where(binary_enrichment_matrix == 1, enrichment_matrix, 0)
+    return enrichment_matrix, binary_enrichment_matrix
 
-    return enrichment_matrix, binary_enrichment_matrix, significant_enrichment_matrix
+
+def _process_node_imputation(
+    row_index: int,
+    network: nx.Graph,
+    enrichment_matrix: np.ndarray,
+    binary_enrichment_matrix: np.ndarray,
+    depth: int,
+) -> Tuple[np.ndarray, np.ndarray]:
+    """Process the imputation for a single node based on its enriched neighbors.
+
+    Args:
+        row_index (int): The index of the enriched node being processed.
+        network (nx.Graph): The network graph with nodes having IDs matching the matrix indices.
+        enrichment_matrix (np.ndarray): The enrichment matrix with rows to be imputed.
+        binary_enrichment_matrix (np.ndarray): The alpha threshold matrix to be imputed similarly.
+        depth (int): Current depth for traversal.
+
+    Returns:
+        Tuple[np.ndarray, np.ndarray]: The modified enrichment matrix and binary threshold matrix.
+    """
+    # Check neighbors at the current depth
+    neighbors = nx.single_source_shortest_path_length(network, row_index, cutoff=depth)
+    # Filter annotated neighbors (already enriched)
+    annotated_neighbors = [
+        n
+        for n in neighbors
+        if n != row_index
+        and binary_enrichment_matrix[n].sum() != 0
+        and enrichment_matrix[n].sum() != 0
+    ]
+    # Filter non-enriched neighbors
+    valid_neighbors = [
+        n
+        for n in neighbors
+        if n != row_index
+        and binary_enrichment_matrix[n].sum() == 0
+        and enrichment_matrix[n].sum() == 0
+    ]
+    # If there are valid non-enriched neighbors
+    if valid_neighbors and annotated_neighbors:
+        # Calculate distances to annotated neighbors
+        distances_to_annotated = [
+            _get_euclidean_distance(row_index, n, network) for n in annotated_neighbors
+        ]
+        # Calculate the IQR to identify outliers
+        q1, q3 = np.percentile(distances_to_annotated, [25, 75])
+        iqr = q3 - q1
+        lower_bound = q1 - 1.5 * iqr
+        upper_bound = q3 + 1.5 * iqr
+        # Filter valid non-enriched neighbors that fall within the IQR bounds
+        valid_neighbors_within_iqr = [
+            n
+            for n in valid_neighbors
+            if lower_bound <= _get_euclidean_distance(row_index, n, network) <= upper_bound
+        ]
+        # If there are any valid neighbors within the IQR
+        if valid_neighbors_within_iqr:
+            # If more than one valid neighbor is within the IQR, compute pairwise cosine similarities
+            if len(valid_neighbors_within_iqr) > 1:
+                # Find the most similar neighbor based on pairwise cosine similarities
+                def sum_pairwise_cosine_similarities(neighbor):
+                    return sum(
+                        cosine_similarity(
+                            enrichment_matrix[neighbor].reshape(1, -1),
+                            enrichment_matrix[other_neighbor].reshape(1, -1),
+                        )[0][0]
+                        for other_neighbor in valid_neighbors_within_iqr
+                        if other_neighbor != neighbor
+                    )
+
+                most_similar_neighbor = max(
+                    valid_neighbors_within_iqr, key=sum_pairwise_cosine_similarities
+                )
+            else:
+                most_similar_neighbor = valid_neighbors_within_iqr[0]
+
+            # Impute the most similar non-enriched neighbor with the enriched node's data, scaled by depth
+            enrichment_matrix[most_similar_neighbor] = enrichment_matrix[row_index] / np.sqrt(
+                depth + 1
+            )
+            binary_enrichment_matrix[most_similar_neighbor] = binary_enrichment_matrix[row_index]
+
+    return enrichment_matrix, binary_enrichment_matrix
 
 
 def _prune_neighbors(
@@ -240,27 +339,27 @@ def _prune_neighbors(
     """
     # Identify indices with non-zero rows in the binary enrichment matrix
     non_zero_indices = np.where(binary_enrichment_matrix.sum(axis=1) != 0)[0]
-    average_distances = []
+    median_distances = []
     for node in non_zero_indices:
         neighbors = [n for n in network.neighbors(node) if binary_enrichment_matrix[n].sum() != 0]
         if neighbors:
-            average_distance = np.mean(
+            median_distance = np.median(
                 [_get_euclidean_distance(node, n, network) for n in neighbors]
             )
-            average_distances.append(average_distance)
+            median_distances.append(median_distance)
 
     # Calculate the distance threshold value based on rank
-    distance_threshold_value = _calculate_threshold(average_distances, 1 - distance_threshold)
+    distance_threshold_value = _calculate_threshold(median_distances, 1 - distance_threshold)
     # Prune nodes that are outliers based on the distance threshold
     for row_index in non_zero_indices:
         neighbors = [
             n for n in network.neighbors(row_index) if binary_enrichment_matrix[n].sum() != 0
         ]
         if neighbors:
-            average_distance = np.mean(
+            median_distance = np.median(
                 [_get_euclidean_distance(row_index, n, network) for n in neighbors]
             )
-            if average_distance >= distance_threshold_value:
+            if median_distance >= distance_threshold_value:
                 enrichment_matrix[row_index] = 0
                 binary_enrichment_matrix[row_index] = 0
 
@@ -305,18 +404,18 @@ def _get_node_position(network: nx.Graph, node: Any) -> np.ndarray:
     )
 
 
-def _calculate_threshold(average_distances: List, distance_threshold: float) -> float:
-    """Calculate the distance threshold based on the given average distances and a percentile threshold.
+def _calculate_threshold(median_distances: List, distance_threshold: float) -> float:
+    """Calculate the distance threshold based on the given median distances and a percentile threshold.
 
     Args:
-        average_distances (list): An array of average distances.
+        median_distances (list): An array of median distances.
         distance_threshold (float): A percentile threshold (0 to 1) used to determine the distance cutoff.
 
     Returns:
         float: The calculated distance threshold value.
     """
-    # Sort the average distances
-    sorted_distances = np.sort(average_distances)
+    # Sort the median distances
+    sorted_distances = np.sort(median_distances)
     # Compute the rank percentiles for the sorted distances
     rank_percentiles = np.linspace(0, 1, len(sorted_distances))
     # Interpolating the ranks to 1000 evenly spaced percentiles