pythonflex 0.3.4__py3-none-any.whl → 0.4__py3-none-any.whl

pythonflex/analysis.py

@@ -11,7 +11,6 @@ from pathlib import Path
 from art import tprint
 from bitarray import bitarray
 from joblib import Parallel, delayed, dump, load
-import matplotlib.pyplot as plt
 from numba import njit, prange
 import numpy as np
 import pandas as pd
@@ -20,8 +19,8 @@ from tqdm import tqdm
 
 # Local/application-specific imports
 from .logging_config import log
-from .preprocessing import filter_matrix_by_genes
-from .utils import dsave, dload, _sanitize
+from .preprocessing import filter_matrix_by_genes, filter_duplicate_terms
+from .utils import dsave, dload, _sanitize, normalize_analysis_genes
 
 import matplotlib as mpl
 
@@ -36,6 +35,8 @@ def deep_update(source, overrides):
 
 def initialize(config={}):
 
+    user_overrides = config if isinstance(config, dict) else {}
+
     default_config = {
         "min_genes_in_complex": 3,
         "min_genes_per_complex_analysis": 2,
@@ -43,8 +44,10 @@ def initialize(config={}):
         "gold_standard": "CORUM",
         "color_map": "RdYlBu",
         "jaccard": True,
-        "jaccard_threshold": 1.0,
-        "use_common_genes": True,
+        # Which genes are used for analysis (drives used_genes intersection)
+        # - 'shared'           : use genes common to all datasets (common_genes)
+        # - 'dataset_specific' : use genes present in each dataset individually
+        "analysis_genes": "shared",
         "plotting": {
             "save_plot": True,
             "show_plot": True,
@@ -56,6 +59,11 @@ def initialize(config={}):
             "drop_na": False,
         },
         "corr_function": "numpy",
+        "per_complex": {
+            "n_jobs": 4,
+            "chunk_size": 200,
+            "max_nbytes": "100M",
+        },
         "logging": {  # Added: Default logging config
             "visible_levels": ["DONE"]  # if needed #, "PROGRESS", "STARTED", "INFO"
         }
@@ -66,6 +74,26 @@ def initialize(config={}):
         config = deep_update(default_config, config)
     else:
         config = default_config
+
+    # Backward compatibility: if user provided legacy key but not the new one,
+    # map it to analysis_genes. (We must look at the original overrides, because
+    # defaults always include analysis_genes.)
+    analysis_genes_provided = (
+        isinstance(user_overrides, dict)
+        and "analysis_genes" in user_overrides
+        and user_overrides.get("analysis_genes") is not None
+        and str(user_overrides.get("analysis_genes")).strip() != ""
+    )
+    if (
+        isinstance(user_overrides, dict)
+        and "use_common_genes" in user_overrides
+        and not analysis_genes_provided
+    ):
+        config["analysis_genes"] = (
+            "shared" if bool(user_overrides.get("use_common_genes")) else "dataset_specific"
+        )
+
+    config["analysis_genes"] = normalize_analysis_genes(config.get("analysis_genes"))
     
     # Extract visible_levels from the merged config and set logging visibility immediately (before any logs)
     visible_levels = config.get("logging", {}).get("visible_levels", ["DONE"])
@@ -111,12 +139,15 @@ def update_matploblib_config(config=None, font_family="Arial", layout="single"):
     if config is None:
         config = {}
     # Fallback if chosen font missing
+    requested_font_family = font_family
     try:
         from matplotlib.font_manager import findfont, FontProperties
         findfont(FontProperties(family=font_family))
     except Exception:
         font_family = "Helvetica"  # Nature prefers Helvetica if Arial unavailable
-        print(f"Warning: '{font_family}' not found, falling back to 'Helvetica'.")
+        log.warning(
+            f"Font '{requested_font_family}' not found; falling back to '{font_family}'."
+        )
     
     # Figure size presets (Nature: single ≈ 89 mm, double ≈ 183 mm at 25.4 mm/inch)
     if isinstance(layout, tuple):
@@ -191,50 +222,114 @@ def update_matploblib_config(config=None, font_family="Arial", layout="single"):
         "svg.fonttype": "none",
     })
 
-def pra(dataset_name, matrix, is_corr=False):
-    log.info(f"******************** {dataset_name} ********************")
-    log.started(f"** Global Precision-Recall Analysis - {dataset_name} **")
+
+def _sort_ascending_for_dataset(dataset_name):
+    sorting = dload("input", "sorting")
+    if not isinstance(sorting, dict):
+        return False
+    sort_order = str(sorting.get(dataset_name, "high")).strip().lower()
+    return sort_order == "low"
+
+
+def prepare_terms_for_dataset(dataset_name, matrix):
+    """Prepare dataset-specific gold-standard terms and filtered matrix.
+
+        This computes:
+            - terms['used_genes'] as the intersection of terms['all_genes'] with either
+                shared genes (config['analysis_genes']=='shared') or the dataset genes
+                (config['analysis_genes']=='dataset_specific').
+      - genes_present_in_terms_<dataset_name>
+
+    Side effects:
+      - stores dataset-specific terms and genes list under:
+        dsave(..., 'common', f'terms_{dataset_name}')
+        dsave(..., 'common', f'genes_present_in_terms_{dataset_name}')
+
+    Returns:
+      (terms, genes_present, matrix_filtered)
+    """
     config = dload("config")
-    use_common_genes = config.get("use_common_genes", True)
+    if config is None:
+        raise RuntimeError(
+            "prepare_terms_for_dataset(): config not found. Run initialize() first."
+        )
 
     terms_data = dload("common", "terms")
     if terms_data is None or not isinstance(terms_data, pd.DataFrame):
-        raise ValueError("Expected 'terms' to be a DataFrame, but got None or invalid type.")
+        raise ValueError(
+            "prepare_terms_for_dataset(): expected 'terms' to be a DataFrame, but got None or invalid type. "
+            "Make sure to run load_gold_standard() first."
+        )
     terms = terms_data.copy()
-    sorting = dload("input", "sorting")
-    sort_order = sorting.get(dataset_name, "high")
-
-    if not is_corr:
-        matrix = perform_corr(matrix, config.get("corr_function"))
 
-    # Apply per-dataset gene filtering based on use_common_genes setting
-    if use_common_genes:
-        # Use common genes approach (current behavior)
-        common_genes = dload("common", "common_genes")
-        if not common_genes:
-            raise ValueError("Common genes not found.")
-        common_genes_set = set(common_genes)
-        terms["used_genes"] = terms["all_genes"].apply(lambda x: list(set(x) & common_genes_set))
-        log.info(f"Using common genes approach: {len(common_genes)} genes")
-    else:
-        # Use per-dataset approach (new behavior)
-        dataset_genes_set = set(matrix.index)
-        terms["used_genes"] = terms["all_genes"].apply(lambda x: list(set(x) & dataset_genes_set))
-        log.info(f"Using per-dataset approach for {dataset_name}: {len(dataset_genes_set)} genes in dataset")
+    analysis_genes = normalize_analysis_genes(config.get("analysis_genes"))
+
+    if analysis_genes == "shared":
+        common_genes = dload("common", "common_genes")
+        if common_genes is None:
+            raise ValueError(
+                "prepare_terms_for_dataset(): common genes not found. "
+                "Run get_common_genes() or set analysis_genes='dataset_specific'."
+            )
+
+        common_genes_list = list(common_genes)
+        if len(common_genes_list) == 0:
+            raise ValueError(
+                "prepare_terms_for_dataset(): common genes is empty. "
+                "Run get_common_genes() or set analysis_genes='dataset_specific'."
+            )
+
+        gene_universe = set(common_genes_list)
+        log.info(f"Using shared genes approach: {len(gene_universe)} genes")
+    else:
+        gene_universe = set(matrix.index)
+        log.info(
+            f"Using dataset-specific approach for {dataset_name}: {len(gene_universe)} genes in dataset"
+        )
+
+    terms["used_genes"] = terms["all_genes"].apply(
+        lambda genes: list(set(genes) & gene_universe)
+    )
 
-    # Filter terms by minimum genes after dataset-specific filtering
+    min_genes_raw = config.get("min_genes_in_complex", 3)
+    min_genes = int(min_genes_raw) if min_genes_raw is not None else 3
     terms["n_used_genes"] = terms["used_genes"].apply(len)
-    terms = terms[terms["n_used_genes"] >= config['min_genes_in_complex']]
-    
-    # Get genes present in terms for this specific dataset
-    genes_present = list(set([gene for genes_list in terms["used_genes"] for gene in genes_list]))
+    terms = terms[terms["n_used_genes"] >= min_genes]
+
+    if bool(config.get("jaccard", False)):
+        before = len(terms)
+        terms = filter_duplicate_terms(terms)
+        log.done(
+            f"After Jaccard duplicate used_genes filtering for {dataset_name}: "
+            f"{len(terms)} terms ({before - len(terms)} removed)"
+        )
+
+    genes_present = list(
+        set(gene for genes_list in terms["used_genes"] for gene in genes_list)
+    )
     log.info(f"Genes present in terms for {dataset_name}: {len(genes_present)}")
-    
-    matrix = filter_matrix_by_genes(matrix, genes_present)
+
+    matrix_filtered = filter_matrix_by_genes(matrix, genes_present)
+
+    dsave(terms, "common", f"terms_{dataset_name}")
+    dsave(genes_present, "common", f"genes_present_in_terms_{dataset_name}")
+
+    return terms, genes_present, matrix_filtered
+
+def pra(dataset_name, matrix, is_corr=False):
+    log.info(f"******************** {dataset_name} ********************")
+    log.started(f"** Global Precision-Recall Analysis - {dataset_name} **")
+    config = dload("config")
+    ascending = _sort_ascending_for_dataset(dataset_name)
+
+    if not is_corr:
+        matrix = perform_corr(matrix, config.get("corr_function"))
+
+    terms, _genes_present, matrix = prepare_terms_for_dataset(dataset_name, matrix)
     log.info(f"Matrix shape: {matrix.shape}")
     df = binary(matrix)
     log.info(f"Pair-wise shape: {df.shape}")
-    df = quick_sort(df, ascending=(sort_order == "low"))
+    df = quick_sort(df, ascending=ascending)
 
     log.started("Building gene-to-pair indices")
     gold_pair_to_complex = _build_gold_pair_to_complex(terms)  
@@ -252,23 +347,22 @@ def pra(dataset_name, matrix, is_corr=False):
     if df["prediction"].sum() == 0:
         log.info("No true positives found in dataset.")
         pr_auc = np.nan
+        df["tp"] = 0
+        df["precision"] = np.nan
+        df["recall"] = np.nan
     else:
         tp = df["prediction"].cumsum()
         df["tp"] = tp
         precision = tp / (np.arange(len(df)) + 1)
         recall = tp / tp.iloc[-1]
-        pr_auc = metrics.auc(recall, precision)
         df["precision"] = precision
         df["recall"] = recall
+        pr_auc = metrics.auc(recall, precision) if len(recall) >= 2 else np.nan
     
     log.info(f"PR-AUC: {pr_auc:.4f}, Number of true positives: {df['prediction'].sum()}")
     dsave(df, "pra", dataset_name)
     dsave(pr_auc, "pr_auc", dataset_name)
-    dsave( _corrected_auc(df) , "corrected_pr_auc", dataset_name)
-
-    # Save dataset-specific terms for per-complex analysis
-    dsave(terms, "common", f"terms_{dataset_name}")
-    dsave(genes_present, "common", f"genes_present_in_terms_{dataset_name}")
+    dsave(_corrected_auc(df), "corrected_pr_auc", dataset_name)
 
     log.done(f"Global PRA completed for {dataset_name}")
     return df
@@ -278,7 +372,12 @@ def pra(dataset_name, matrix, is_corr=False):
 # --------------------------------------------------------------------------
 
 def _corrected_auc(df: pd.DataFrame) -> float:
-    return np.trapz(df["precision"], df["recall"]) - df["precision"].iloc[-1]
+    if df.empty or "precision" not in df.columns or "recall" not in df.columns:
+        return np.nan
+    valid = df[["precision", "recall"]].replace([np.inf, -np.inf], np.nan).dropna()
+    if len(valid) < 2:
+        return np.nan
+    return np.trapz(valid["precision"], valid["recall"]) - valid["precision"].iloc[-1]
 
 def _build_gene_to_pair_indices(pairwise_df):
     indices = pairwise_df.index.values
@@ -395,28 +494,52 @@ def _process_chunk(chunk_terms, min_genes, memmap_path, gene_to_pair_indices):
         # Return error info for debugging
         return {'error': str(e), 'chunk_size': len(chunk_terms)}
 
-def pra_percomplex(dataset_name, matrix, is_corr=False, chunk_size=200):
+def pra_percomplex(dataset_name, matrix, is_corr=False, chunk_size=None, n_jobs=None):
     log.started(f"*** Per-complex PRA started - {dataset_name} ***")
     config = dload("config")
-    
-    # Use dataset-specific terms and genes from pra function
-    terms = dload("common", f"terms_{dataset_name}")
-    genes_present = dload("common", f"genes_present_in_terms_{dataset_name}")
-    
-    if terms is None:
-        log.warning(f"No dataset-specific terms found for {dataset_name}, using global terms")
-        terms = dload("common", "terms")
-        genes_present = dload("common", "genes_present_in_terms")
-    
-    sorting = dload("input", "sorting")
-    sort_order = sorting.get(dataset_name, "high")
+    ascending = _sort_ascending_for_dataset(dataset_name)
+    per_complex_config = config.get("per_complex", {})
+    if not isinstance(per_complex_config, dict):
+        per_complex_config = {}
+    chunk_size_value = (
+        chunk_size if chunk_size is not None else per_complex_config.get("chunk_size", 200)
+    )
+    n_jobs_value = n_jobs if n_jobs is not None else per_complex_config.get("n_jobs", 4)
+    max_nbytes = per_complex_config.get("max_nbytes", "100M")
+
+    try:
+        effective_chunk_size = int(chunk_size_value)
+        effective_n_jobs = int(n_jobs_value)
+    except (TypeError, ValueError) as exc:
+        raise ValueError(
+            "per-complex chunk_size and n_jobs must be integer-compatible values."
+        ) from exc
+
+    if effective_chunk_size <= 0:
+        raise ValueError("per-complex chunk_size must be greater than 0.")
+    if effective_n_jobs <= 0:
+        raise ValueError("per-complex n_jobs must be greater than 0.")
+
     if not is_corr:
         matrix = perform_corr(matrix, config.get("corr_function"))
-    matrix = filter_matrix_by_genes(matrix, genes_present)
+
+    # Prefer terms prepared by pra(); if absent, prepare them here so direct
+    # pra_percomplex() calls use the same dataset-specific gene universe.
+    terms = dload("common", f"terms_{dataset_name}")
+    genes_present = dload("common", f"genes_present_in_terms_{dataset_name}")
+
+    if not isinstance(terms, pd.DataFrame) or genes_present is None:
+        log.warning(
+            f"No dataset-specific terms found for {dataset_name}; preparing them now."
+        )
+        terms, genes_present, matrix = prepare_terms_for_dataset(dataset_name, matrix)
+    else:
+        matrix = filter_matrix_by_genes(matrix, genes_present)
+
     log.info(f"Matrix shape: {matrix.shape}")
     df = binary(matrix)
     log.info(f"Pair-wise shape: {df.shape}")
-    df = quick_sort(df, ascending=(sort_order == "low"))
+    df = quick_sort(df, ascending=ascending)
     pairwise_df = df.copy()
     pairwise_df['gene1'] = pairwise_df['gene1'].astype("category")
     pairwise_df['gene2'] = pairwise_df['gene2'].astype("category")
@@ -434,26 +557,42 @@ def pra_percomplex(dataset_name, matrix, is_corr=False, chunk_size=200):
     pairwise_df = _precompute_complex_ids(pairwise_df, gold_pair_to_complex)
     log.done("Precomputing complex IDs")  # 
 
+    chunks = [
+        terms.iloc[i:i + effective_chunk_size]
+        for i in range(0, len(terms), effective_chunk_size)
+    ]
+    min_genes = config["min_genes_per_complex_analysis"]
+
+    if not chunks:
+        terms["auc_score"] = pd.Series(dtype=float)
+        terms["corrected_auc_score"] = pd.Series(dtype=float)
+        dsave(terms, "pra_percomplex", dataset_name)
+        log.done("Per-complex PRA completed with no eligible terms.")
+        return terms
+
     log.info('Dumping pairwise_df to memmap')
     memmap_path = _dump_pairwise_memmap(pairwise_df, dataset_name)
     log.done('Dumping pairwise_df to memmap')
 
-    # choose smaller chunks now that pickling cost is gone
-    chunks = [terms.iloc[i:i+chunk_size] for i in range(0, len(terms), chunk_size)]
-    min_genes = config["min_genes_per_complex_analysis"]
-
     # Initialize results variable
     results = None
     
     try:
         # Compatible parallel execution for older joblib versions
         log.started("Processing chunks in parallel")
-        
+        actual_n_jobs = min(effective_n_jobs, len(chunks))
+        log.info(
+            "Per-complex parallel config: "
+            f"n_jobs={actual_n_jobs}, requested_n_jobs={effective_n_jobs}, "
+            f"chunk_size={effective_chunk_size}, chunks={len(chunks)}, "
+            f"max_nbytes={max_nbytes}"
+        )
+
         # Use a more conservative approach with older joblib
         results = Parallel(
-            n_jobs=min(4, len(chunks)),  # Limit to 4 workers or number of chunks
+            n_jobs=actual_n_jobs,
             temp_folder=os.path.dirname(memmap_path),     
-            max_nbytes='100M',  # Set memory limit
+            max_nbytes=max_nbytes,
             verbose=1  # Show progress
         )(delayed(_process_chunk)(chunk, min_genes, memmap_path, gene_to_pair_indices) 
           for chunk in tqdm(chunks, desc="Per-complex PRA"))
@@ -484,11 +623,13 @@ def pra_percomplex(dataset_name, matrix, is_corr=False, chunk_size=200):
     # Merge results with enhanced error handling
     auc_scores = {}
     corrected_auc_scores = {}
+    errors = []
     if results:
         for i, res in enumerate(results):
             if isinstance(res, dict):
                 if 'error' in res:
                     log.error(f"Error in chunk {i}: {res['error']}")
+                    errors.append(f"chunk {i}: {res['error']}")
                 elif 'auc' in res and 'corrected_auc' in res:
                     # New format with both AUC types
                     auc_scores.update(res['auc'])
@@ -498,8 +639,15 @@ def pra_percomplex(dataset_name, matrix, is_corr=False, chunk_size=200):
                     auc_scores.update(res)
             elif isinstance(res, tuple) and len(res) >= 2 and res[0] is None:
                 log.error(f"Chunk {i} error: {res[1]}")
+                errors.append(f"chunk {i}: {res[1]}")
             else:
                 log.warning(f"Unexpected result type from chunk {i}: {type(res)} - {res}")
+                errors.append(f"chunk {i}: unexpected result type {type(res)}")
+
+    if errors:
+        preview = "; ".join(errors[:3])
+        extra = f" ({len(errors) - 3} more)" if len(errors) > 3 else ""
+        raise RuntimeError(f"Per-complex PRA failed in worker chunks: {preview}{extra}")
     
     # Add the computed AUC scores to the terms DataFrame.
     terms["auc_score"] = pd.Series(auc_scores)
@@ -511,10 +659,20 @@ def pra_percomplex(dataset_name, matrix, is_corr=False, chunk_size=200):
 def complex_contributions(name):
     log.info(f"Computing complex contributions (Greedy) for dataset: {name}")
     pra = dload("pra", name)
-    terms = dload("common", "terms")
+    terms = dload("common", f"terms_{name}")
+    if not isinstance(terms, pd.DataFrame):
+        # Fallback for backward compatibility
+        terms = dload("common", "terms")
+    if not isinstance(pra, pd.DataFrame) or pra.empty:
+        raise RuntimeError(f"complex_contributions(): PRA data for dataset '{name}' not found.")
+    if not isinstance(terms, pd.DataFrame) or terms.empty:
+        raise RuntimeError(f"complex_contributions(): terms for dataset '{name}' not found.")
     
-    # Ensure pra is sorted by score descending (matches R's order by predicted descending)
-    pra = pra.sort_values(by='score', ascending=False).reset_index(drop=True)
+    # Respect the dataset's score direction: high scores by default, low scores if configured.
+    pra = pra.sort_values(
+        by='score',
+        ascending=_sort_ascending_for_dataset(name),
+    ).reset_index(drop=True)
     
     # Compute cumulative TP and precision (matches R's TP.count = cumsum(true), Precision = TP / (1:n))
     pra['cumTP'] = pra['prediction'].cumsum()
@@ -809,6 +967,9 @@ def binary(corr):
     
     stack = corr.stack().rename_axis(index=['gene1', 'gene2']).\
             reset_index().rename(columns={0: 'score'})
+    if stack.empty:
+        log.done("Pair-wise conversion.")
+        return stack
     if has_mirror_of_first_pair(stack):
         log.info("Mirror pairs detected. Dropping them to ensure unique gene pairs.")
         stack = drop_mirror_pairs(stack)
@@ -859,7 +1020,7 @@ def save_results_to_csv(categories = ["complex_contributions", "pr_auc", "pra_pe
             continue
 
         if category == "mpr_complexes_auc" and isinstance(data, dict):
-            # Dict[dataset_name -> Dict[filter_key -> auc]]
+            # Dict[dataset_name -> Dict[variant_key -> auc]]
             try:
                 df = pd.DataFrame.from_dict(data, orient="index")
                 df.index.name = "Dataset"
@@ -902,21 +1063,10 @@ def save_results_to_csv(categories = ["complex_contributions", "pr_auc", "pra_pe
 
     log.done("Results saved to CSV files in the output folder.")
 
-################### mPR
-################### mPR ###################
-
-
-
-################### mPR ###################
-################### mPR ###################
-
 # -----------------------------------------------------------------------------
 # mPR preparation (module-level precision–recall, Fig. 1E / 1F)
 # -----------------------------------------------------------------------------
 
-import numpy as np
-import pandas as pd
-
 
 def _mpr_get_mtRibo_ETCI_ids(terms_like):
     """
@@ -971,34 +1121,16 @@ def _mpr_get_small_high_auprc_ids(
 # Helpers implementing the FLEX stepwise module-level PR logic
 # -------------------------------------------------------------------------
 
-"""
-CORRECT FIX for _mpr_build_pairs in analysis.py
-
-The issue: The current code marks filtered TPs as true=0, which makes them
-count as False Positives and dramatically lowers precision.
-
-The R code (getSubsetOfCoAnnRemoveIDs with replace=FALSE) REMOVES the 
-filtered positive pairs entirely from the dataset.
-
-This is the key difference:
-- Current Python: Keeps all rows, filtered TPs become FPs → precision tanks
-- R code: Removes filtered TP rows → they don't affect precision at all
-"""
-
-import numpy as np
-import pandas as pd
-
-
-def _mpr_build_pairs(pra, removed_ids=None):
+def _mpr_build_pairs(pra, removed_ids=None, ascending=False):
     """
     Build a Pairs.in.data-like table for mPR / stepwise contributions.
-    
-    FIXED: Removes rows that contain filtered complex IDs (matching R behavior)
-    instead of marking them as true=0.
+
+    Rows containing filtered positive complex IDs are removed from the ranking,
+    matching the FLEX stepwise module-level precision-recall behavior.
 
     Input:
       pra : DataFrame with at least columns
-            - 'score'       : ranking score (higher = better)
+            - 'score'       : ranking score
             - 'complex_id'  : complex annotations
       removed_ids : set[int] of complexes to remove
 
@@ -1057,11 +1189,9 @@ def _mpr_build_pairs(pra, removed_ids=None):
     out["complex_ids"] = df[cid_col].apply(normalize_ids)
     out["true"] = out["complex_ids"].apply(lambda ids: 1 if len(ids) > 0 else 0)
     
-    # KEY FIX: Remove rows that are TPs AND contain a removed complex ID
-    # This matches the R behavior of getSubsetOfCoAnnRemoveIDs with replace=FALSE
+    # Remove rows that are TPs and contain a removed complex ID.
     if removed_ids:
         should_remove_mask = df[cid_col].apply(should_remove)
-        # Only remove if it's a TP (true=1)
         remove_mask = should_remove_mask & (out["true"] == 1)
         out = out[~remove_mask].copy()
     
@@ -1071,41 +1201,11 @@ def _mpr_build_pairs(pra, removed_ids=None):
             lambda ids: [cid for cid in ids if cid not in removed_ids]
         )
 
-    # Sort by predicted descending
-    out = out.sort_values("predicted", ascending=False).reset_index(drop=True)
+    # Sort by the dataset's configured score direction.
+    out = out.sort_values("predicted", ascending=ascending).reset_index(drop=True)
     return out
 
 
-# ============================================================================
-# HOW TO APPLY THIS FIX
-# ============================================================================
-# 
-# In analysis.py, replace the _mpr_build_pairs function (around line 962-1025)
-# with the _mpr_build_pairs_fixed function above.
-#
-# The key changes are:
-#
-# 1. REMOVE rows instead of marking true=0:
-#    
-#    OLD CODE:
-#        if removed_ids:
-#            ids = [cid for cid in ids if cid not in removed_ids]
-#        return ids
-#        ...
-#        out["true"] = out["complex_ids"].apply(lambda ids: 1 if len(ids) > 0 else 0)
-#    
-#    NEW CODE:
-#        # First compute true normally
-#        out["true"] = out["complex_ids"].apply(lambda ids: 1 if len(ids) > 0 else 0)
-#        
-#        # Then REMOVE rows that are TPs and contain removed IDs
-#        if removed_ids:
-#            should_remove_mask = df[cid_col].apply(should_remove)
-#            remove_mask = should_remove_mask & (out["true"] == 1)
-#            out = out[~remove_mask].copy()
-#
-# ============================================================================
-
 def _mpr_precision_cutoffs_from_pairs(pairs, step=0.025):
     """
     Choose precision cutoffs similar to FLEX:
@@ -1137,7 +1237,7 @@ def _mpr_precision_cutoffs_from_pairs(pairs, step=0.025):
     return np.array(cuts, dtype=float)
 
 
-def _mpr_stepwise_contributions(pairs, precision_cutoffs):
+def _mpr_stepwise_contributions(pairs, precision_cutoffs, ascending=False):
     """
     Greedy, stepwise TP allocation per complex at each precision cutoff.
 
@@ -1152,7 +1252,7 @@ def _mpr_stepwise_contributions(pairs, precision_cutoffs):
       contrib_df : DataFrame [complex_id x cutoff] with TP counts
     """
     pairs = pairs.copy()
-    pairs = pairs.sort_values("predicted", ascending=False).reset_index(drop=True)
+    pairs = pairs.sort_values("predicted", ascending=ascending).reset_index(drop=True)
 
     true = pairs["true"].to_numpy(dtype=int)
     n = len(true)
@@ -1272,16 +1372,29 @@ def _mpr_module_coverage(contrib_df, terms, tp_th=1, percent_th=0.1):
         row = terms.loc[cid_int]
 
         n_genes = None
-        
-        # FIXED: Handle all_genes as list (how it's stored in preprocessing)
-        if "all_genes" in row.index:
+
+        # Prefer used_genes (genes actually in the dataset) for a fair coverage
+        # fraction. This matters for GOBP/PATHWAY where all_genes >> used_genes.
+        if "used_genes" in row.index:
+            genes = row["used_genes"]
+            if isinstance(genes, (list, np.ndarray)) and len(genes) > 0:
+                n_genes = len(genes)
+        if n_genes is None and "n_used_genes" in row.index:
+            try:
+                v = int(row["n_used_genes"])
+                if v > 0:
+                    n_genes = v
+            except (ValueError, TypeError):
+                pass
+
+        # Fallback: all_genes (how it's stored in preprocessing)
+        if n_genes is None and "all_genes" in row.index:
             genes = row["all_genes"]
-            if isinstance(genes, list):
+            if isinstance(genes, (list, np.ndarray)):
                 n_genes = len(genes)
             elif isinstance(genes, str):
-                # Fallback if stored as string
                 n_genes = len([g for g in genes.split(";") if g])
-        
+
         # Fallback to Genes column (original string from CORUM)
         if n_genes is None and "Genes" in row.index:
             genes_str = row["Genes"]
@@ -1338,7 +1451,7 @@ def _mpr_complexes_auc(
 
     We compute a normalized AUC by integrating precision over the *normalized*
     coverage axis:
-        AUC = \int y \, d(x/max_complexes)
+        AUC = integral y d(x/max_complexes)
 
     This yields a score in [0, 1] (or NaN if insufficient data).
     """
@@ -1348,7 +1461,7 @@ def _mpr_complexes_auc(
     if cov.size == 0 or prec.size == 0:
         return 0.0
 
-    # Match plot_mpr_complexes_multi(): only count cov>0 (log-x cannot show 0)
+    # Match plot_mpr_complex_coverage_curve(): only count cov>0 (log-x cannot show 0)
     mask = (
         np.isfinite(cov)
         & np.isfinite(prec)
@@ -1410,26 +1523,31 @@ def mpr_prepare(
     """
     pra = dload("pra", name)
     pra_percomplex = dload("pra_percomplex", name)
-    terms = dload("common", "terms")
+    terms = dload("common", f"terms_{name}")
+    if not isinstance(terms, pd.DataFrame):
+        # Fallback for backward compatibility
+        terms = dload("common", "terms")
 
-    if pra is None:
+    if pra is None or not isinstance(pra, pd.DataFrame) or pra.empty:
         raise RuntimeError(
             f"mpr_prepare(): PRA data for dataset '{name}' not found "
             "(dload('pra', name))."
         )
-    if pra_percomplex is None:
+    if pra_percomplex is None or not isinstance(pra_percomplex, pd.DataFrame) or pra_percomplex.empty:
         raise RuntimeError(
             f"mpr_prepare(): per-complex PRA data for dataset '{name}' not found "
             "(dload('pra_percomplex', name))."
         )
-    if terms is None:
+    if terms is None or not isinstance(terms, pd.DataFrame) or terms.empty:
         raise RuntimeError(
             "mpr_prepare(): CORUM 'terms' table not found (dload('common', 'terms'))."
         )
 
-    # sort by score descending (ranking)
+    ascending = _sort_ascending_for_dataset(name)
+
+    # Sort by the dataset's configured score direction.
     if "score" in pra.columns:
-        pra = pra.sort_values("score", ascending=False).reset_index(drop=True)
+        pra = pra.sort_values("score", ascending=ascending).reset_index(drop=True)
     else:
         pra = pra.reset_index(drop=True)
 
@@ -1455,7 +1573,7 @@ def mpr_prepare(
 
     for label, removed in filter_sets.items():
         # 1) Build pairs table after removing complexes in `removed`
-        pairs = _mpr_build_pairs(pra, removed_ids=removed)
+        pairs = _mpr_build_pairs(pra, removed_ids=removed, ascending=ascending)
 
         true = pairs["true"].to_numpy(dtype=int)
         n = len(true)
@@ -1482,13 +1600,24 @@ def mpr_prepare(
         if precision_cutoffs is None:
             precision_cutoffs = _mpr_precision_cutoffs_from_pairs(pairs)
 
-        contrib_df = _mpr_stepwise_contributions(pairs, precision_cutoffs)
+        contrib_df = _mpr_stepwise_contributions(
+            pairs,
+            precision_cutoffs,
+            ascending=ascending,
+        )
         cov = _mpr_module_coverage(
             contrib_df,
             terms,
             tp_th=tp_th,
             percent_th=percent_th,
         )
+        # precision_cutoffs are sorted ascending (low → high).
+        # Coverage must be non-increasing in that direction: a more permissive
+        # threshold (lower precision) should never yield fewer covered terms.
+        # The independent greedy allocation per cutoff can violate this, so
+        # enforce monotonicity by propagating the max from right to left.
+        if cov.size > 0:
+            cov = np.maximum.accumulate(cov[::-1])[::-1]
         coverage_curves[label] = cov
         complexes_auc[label] = _mpr_complexes_auc(
             cov,
@@ -1516,424 +1645,3 @@ def mpr_prepare(
 
     # Convenience: store AUCs as their own category for easy export / plotting.
     dsave(complexes_auc, "mpr_complexes_auc", name)
-
-
-
-### OLD FUNCTIONS
-
-# new but withoutparallel
-
-# def pra_percomplex(dataset_name, matrix, is_corr=False):
-#     log.started(f"*** Per-complex PRA started - {dataset_name} ***")
-#     config = dload("config")
-#     terms = dload("tmp", "terms")
-#     genes_present = dload("tmp", "genes_present_in_terms")
-#     sorting = dload("input", "sorting")
-#     sort_order = sorting.get(dataset_name, "high")
-#     if not is_corr:
-#         matrix = perform_corr(matrix, config.get("corr_function"))
-#     matrix = filter_matrix_by_genes(matrix, genes_present)
-#     log.info(f"Matrix shape: {matrix.shape}")
-#     df = binary(matrix)
-#     log.info(f"Pair-wise shape: {df.shape}")
-#     df = quick_sort(df, ascending=(sort_order == "low"))
-#     pairwise_df = df.copy()
-#     pairwise_df['gene1'] = pairwise_df['gene1'].astype("category")
-#     pairwise_df['gene2'] = pairwise_df['gene2'].astype("category")
-    
-#     # Precompute a mapping from each gene to the row indices in the pairwise DataFrame where it appears.
-#     gene_to_pair_indices = {}
-#     for i, (gene_a, gene_b) in enumerate(zip(pairwise_df["gene1"], pairwise_df["gene2"])):
-#         gene_to_pair_indices.setdefault(gene_a, []).append(i)
-#         gene_to_pair_indices.setdefault(gene_b, []).append(i)
-#     log.done
-    
-#     # Build gold_pair_to_complex using sets for efficiency
-#     gold_pair_to_complex = defaultdict(set)
-#     for idx, row in terms.iterrows():
-#         genes = row.used_genes
-#         if len(genes) < 2:
-#             continue
-#         for i, g1 in enumerate(genes):
-#             for g2 in genes[i + 1:]:
-#                 pair = tuple(sorted((g1, g2)))
-#                 gold_pair_to_complex[pair].add(idx)
-    
-#     # Precompute complex_ids as semicolon-separated strings in pairwise_df
-#     pairs = [tuple(sorted((g1, g2))) for g1, g2 in zip(pairwise_df["gene1"], pairwise_df["gene2"])]
-#     pairwise_df['complex_ids'] = [';'.join(map(str, sorted(gold_pair_to_complex.get(pair, set())))) for pair in pairs]
-    
-#     # Initialize AUC scores
-#     auc_scores = {}
-#     # Loop over each gene complex
-#     for idx, row in tqdm(terms.iterrows()):
-#         gene_set = set(row.used_genes)
-#         if config["min_genes_per_complex_analysis"] > len(gene_set):  
-#             continue
-#         # Collect all row indices in the pairwise data where either gene belongs to the complex.
-#         candidate_indices = bitarray(len(pairwise_df))
-#         for gene in gene_set:
-#             if gene in gene_to_pair_indices:
-#                 candidate_indices[gene_to_pair_indices[gene]] = True
-        
-#         if not candidate_indices.any():
-#             continue
-        
-#         # Select only the relevant pairwise comparisons.
-#         selected_rows = np.unpackbits(candidate_indices).view(bool)[:len(pairwise_df)]
-#         sub_df = pairwise_df.iloc[selected_rows]
-        
-#         # Get current complex ID (assuming idx is the ID; adjust if row['ID'] is different)
-#         complex_id = str(idx)  # Or str(row['ID']) if available
-        
-#         # Create true_label: 1 if complex_id in complex_ids (vectorized with str.contains)
-#         #true_label = sub_df['complex_ids'].str.contains(complex_id, regex=False).astype(int)
-
-#         # Inside the loop, for each complex:
-#         # Inside the loop:
-#         complex_id = str(idx)
-#         # Use (?:^|;) and (?:;|$) to avoid capturing groups
-#         pattern = r'(?:^|;)' + re.escape(complex_id) + r'(?:;|$)'
-#         true_label = sub_df['complex_ids'].str.contains(pattern, regex=True).astype(int)
-#         # Filter to keep verified negatives (complex_ids == "") or positives for this complex (true_label == 1)
-#         complex_mask = (sub_df['complex_ids'] == "") | (true_label == 1)
-        
-#         # Use the masked true labels for AUPRC (avoids SettingWithCopyWarning)
-#         predictions = true_label[complex_mask]
-        
-#         if predictions.sum() == 0:
-#             continue
-#         # Compute cumulative true positives and derive precision and recall.
-#         true_positive_cumsum = predictions.cumsum()
-#         precision = true_positive_cumsum / (np.arange(len(predictions)) + 1)
-#         recall = true_positive_cumsum / true_positive_cumsum.iloc[-1]
-        
-#         if len(recall) < 2 or recall.iloc[-1] == 0:
-#             continue
-#         auc_scores[idx] = metrics.auc(recall, precision)
-    
-#     # Add the computed AUC scores to the terms DataFrame.
-#     terms["auc_score"] = pd.Series(auc_scores)
-#     terms.drop(columns=["hash"], inplace=True)
-#     dsave(terms, "pra_percomplex", dataset_name)
-#     log.done(f"Per-complex PRA completed.")
-#     return terms
-
-# it works quick but only maps 1 complex to each pair
-
-# def pra_percomplex_old_type_filtering(dataset_name, matrix, is_corr=False):
-#     log.started(f"*** Per-complex PRA started - {dataset_name} ***")
-#     config = dload("config")
-#     terms = dload("tmp", "terms")
-#     genes_present = dload("tmp", "genes_present_in_terms")
-#     sorting = dload("input", "sorting")
-#     sort_order = sorting.get(dataset_name, "high")
-#     if not is_corr:
-#         matrix = perform_corr(matrix, config.get("corr_function"))
-#     matrix = filter_matrix_by_genes(matrix, genes_present)
-#     log.info(f"Matrix shape: {matrix.shape}")
-#     df = binary(matrix)
-#     log.info(f"Pair-wise shape: {df.shape}")
-#     df = quick_sort(df, ascending=(sort_order == "low"))
-#     pairwise_df = df.copy()
-#     pairwise_df['gene1'] = pairwise_df['gene1'].astype("category")
-#     pairwise_df['gene2'] = pairwise_df['gene2'].astype("category")  
-#     # Precompute a mapping from each gene to the row indices in the pairwise DataFrame where it appears.
-#     gene_to_pair_indices = {}
-#     for i, (gene_a, gene_b) in enumerate(zip(pairwise_df["gene1"], pairwise_df["gene2"])):
-#         gene_to_pair_indices.setdefault(gene_a, []).append(i)
-#         gene_to_pair_indices.setdefault(gene_b, []).append(i)  
-#     # Initialize AUC scores (one for each complex) with NaNs.
-#     #auc_scores = np.full(len(terms), np.nan)
-#     auc_scores = {}
-#     # Loop over each gene complex
-#     for idx, row in tqdm(terms.iterrows()):
-#         gene_set = set(row.used_genes)
-
-#         if config["min_genes_per_complex_analysis"] > len(gene_set):  
-#             continue
-#         # Collect all row indices in the pairwise data where either gene belongs to the complex.
-#         candidate_indices = bitarray(len(pairwise_df))
-#         for gene in gene_set:
-#             if gene in gene_to_pair_indices:
-#                 candidate_indices[gene_to_pair_indices[gene]] = True      
-#         if not candidate_indices.any():
-#             continue     
-#         # Select only the relevant pairwise comparisons.
-#         selected_rows = np.unpackbits(candidate_indices).view(bool)[:len(pairwise_df)]
-#         sub_df = pairwise_df.iloc[selected_rows]
-#         # A prediction is 1 if both genes in the pair are in the complex; otherwise 0.
-#         predictions = (sub_df["gene1"].isin(gene_set) & sub_df["gene2"].isin(gene_set)).astype(int)
-#         if predictions.sum() == 0:
-#             continue
-#         # Compute cumulative true positives and derive precision and recall.
-#         true_positive_cumsum = predictions.cumsum()
-#         precision = true_positive_cumsum / (np.arange(len(predictions)) + 1)
-#         recall = true_positive_cumsum / true_positive_cumsum.iloc[-1]
-        
-#         if len(recall) < 2 or recall.iloc[-1] == 0:
-#             continue
-#         auc_scores[idx] = metrics.auc(recall, precision)   
-#     # Add the computed AUC scores to the terms DataFrame.
-#     terms["auc_score"] = pd.Series(auc_scores)
-#     terms.drop(columns=["hash"], inplace=True)
-#     dsave(terms, "pra_percomplex", dataset_name)
-#     log.done(f"Per-complex PRA completed.")
-#     return terms
-
-# OLD
-# def pra_percomplex(dataset_name, matrix, is_corr=False):
-#     log.started(f"*** Per-complex PRA started for {dataset_name} ***")
-#     config = dload("config")
-#     terms = dload("tmp", "terms")
-#     genes_present = dload("tmp", "genes_present_in_terms")
-#     sorting = dload("input", "sorting")
-#     sort_order = sorting.get(dataset_name, "high")
-
-#     if not is_corr:
-#         matrix = perform_corr(matrix, "numpy")
-#     matrix = filter_matrix_by_genes(matrix, genes_present)
-#     log.info(f"Matrix shape: {matrix.shape}")
-#     df = binary(matrix)
-#     log.info(f"Pair-wise shape: {df.shape}")
-#     df = quick_sort(df, ascending=(sort_order == "low"))
-#     # Precompute gene → row indices
-#     gene_to_rows = {}
-#     for i, (g1, g2) in enumerate(zip(df["gene1"], df["gene2"])):
-#         gene_to_rows.setdefault(g1, []).append(i)
-#         gene_to_rows.setdefault(g2, []).append(i)
-#     aucs = np.full(len(terms), np.nan)
-#     N = len(df)
-#     for idx, row in tqdm(terms.iterrows()):
-#         genes = set(row.used_genes)
-#         if len(genes) < config["min_complex_size_for_percomplex"]:  # Skip small complexes
-#             continue
-#         # Get all row indices where either gene is in the complex
-#         candidate_idxs = set()
-#         for g in genes:
-#             candidate_idxs.update(gene_to_rows.get(g, []))
-#         candidate_idxs = sorted(candidate_idxs)
-#         if not candidate_idxs:
-#             continue
-#         # Use only relevant rows for prediction
-#         sub = df.loc[candidate_idxs]
-#         preds = (sub["gene1"].isin(genes) & sub["gene2"].isin(genes)).astype(int)
-#         if preds.sum() == 0:
-#             continue
-#         tp = preds.cumsum()
-#         prec = tp / (np.arange(len(preds)) + 1)
-#         recall = tp / tp.iloc[-1]
-#         if len(recall) < 2 or recall.iloc[-1] == 0:
-#             continue
-#         aucs[idx] = metrics.auc(recall, prec)
-#     terms["auc_score"] = aucs
-#     terms.drop(columns=["list", "set", "hash"], inplace=True)
-#     dsave(terms, "pra_percomplex", dataset_name)
-#     log.done(f"Per-complex PRA completed.")
-#     return terms
-
-# without greedy
-# def complex_contributions(name):
-#     log.info(f"Computing complex contributions for dataset: {name}")
-
-#     pra = dload("pra", name)
-#     terms = dload("tmp", "terms")
-#     d = pra.query('prediction == 1').drop(columns=['gene1', 'gene2'])
-#     results = {}
-#     thresholds = [round(i, 2) for i in np.arange(1, 0.0001, -0.025)]
-#     for cid in terms.ID.to_list():
-#         arr = []
-#         for threshold in thresholds:
-#             r = d[d.complex_id == cid].query('precision >= @threshold')
-#             arr.append(r.shape[0])
-#         results[cid] = arr
-
-#     r = pd.DataFrame(results, index=thresholds).T
-#     t = terms[['ID', 'Name']].set_index('ID')
-#     r['Name'] = r.index.map(t.Name)
-#     r = r[list(reversed(list(r.columns)))]
-#     r = r.reset_index(drop=True)
-#     dsave(r, "complex_contributions", name)
-#     log.info(f"Complex contributions computation completed for dataset: {name}")
-#     return r
-
-# # new
-# def complex_contributions(name):
-#     log.info(f"Computing complex contributions using R-style greedy logic for dataset: {name}")
-#     pra = dload("pra", name)
-#     terms = dload("common", "terms")
-    
-#     # Ensure pra is sorted by score descending
-#     pra = pra.sort_values(by='score', ascending=False).reset_index(drop=True)
-    
-#     # Compute cumulative TP and precision if not present
-#     pra['cumTP'] = pra['prediction'].cumsum()
-#     pra['rank'] = pra.index + 1
-#     pra['precision'] = pra['cumTP'] / pra['rank']
-    
-#     # R-style precision thresholds
-#     prec_min = pra['precision'].min()
-#     prec_max = pra['precision'].max()
-#     precision_cutoffs = [round(prec_min, 3)]
-#     cutoffs_range = np.arange(0.1, prec_max + 0.001, 0.025)
-#     precision_cutoffs += [round(t, 3) for t in cutoffs_range if t > prec_min]
-#     thresholds = sorted(set(precision_cutoffs))  # Ensure unique and sorted
-    
-#     results = {}
-#     for t in thresholds:
-#         if pra['precision'].max() < t:
-#             continue
-#         cand = pra[pra['precision'] >= t]
-#         if cand.empty:
-#             continue
-#         k = cand.index.max()  # rightmost index where precision >= t
-#         tp_target = pra.loc[k, 'cumTP']
-#         # Find the smallest m where cumTP[m] >= tp_target
-#         ind = pra[pra['cumTP'] >= tp_target].index.min()
-#         if pd.isna(ind):
-#             continue
-#         # Select top (ind+1) rows
-#         tmp = pra.iloc[0:ind + 1].copy()
-#         # Filter for predicted positives (true == 1)
-#         tmp = tmp[tmp['prediction'] == 1]
-#         tmp = tmp[tmp["complex_id"].notnull()]
-#         tmp["ID"] = tmp["complex_id"].apply(lambda ids: ";".join(str(int(i)) for i in ids if pd.notnull(i)))
-#         # Now greedy logic
-#         final_contrib = {}
-#         while not tmp.empty:
-#             all_ids = tmp["ID"].str.split(";").explode()
-#             contrib = all_ids.value_counts()
-#             if contrib.empty:
-#                 break
-#             top_id = contrib.idxmax()
-#             final_contrib[top_id] = contrib[top_id]
-#             tmp = tmp[~tmp["ID"].str.contains(rf"\b{top_id}\b", regex=True)]
-#         for cid, count in final_contrib.items():
-#             if cid not in results:
-#                 results[cid] = [0] * len(thresholds)
-#             results[cid][thresholds.index(t)] = count
-    
-#     # Add back gold standard complexes with 0 contribution
-#     gold_ids = set(terms.index.astype(str))
-#     all_ids = set(results.keys())
-#     missing_ids = gold_ids - all_ids
-#     for cid in missing_ids:
-#         results[cid] = [0] * len(thresholds)
-    
-#     # Build result DataFrame
-#     r = pd.DataFrame(results, index=thresholds).T
-#     r['Name'] = r.index.astype(int).map(terms['Name'])
-#     r = r[['Name'] + [c for c in r.columns if c != 'Name']]  # Name as first col
-#     r = r[(r.drop(columns="Name").sum(axis=1) > 0)]
-#     # Move ID to first column, keep Name second, then precision columns in order
-#     dsave(r, "complex_contributions", name)
-#     log.info(f"Greedy R-style complex contribution completed for dataset: {name}")
-#     return r
-
-# def pra(dataset_name, matrix, is_corr=False):
-#     log.info(f"******************** {dataset_name} ********************")
-#     log.started(f"** Global Precision-Recall Analysis - {dataset_name} **")
-#     config = dload("config")
-
-#     terms_data = dload("tmp", "terms")
-#     if terms_data is None or not isinstance(terms_data, pd.DataFrame):
-#         raise ValueError("Expected 'terms' to be a DataFrame, but got None or invalid type.")
-#     terms = terms_data
-#     genes_present = dload("tmp", "genes_present_in_terms")
-#     sorting = dload("input", "sorting")
-#     sort_order = sorting.get(dataset_name, "high")
-
-#     if not is_corr:
-#         matrix = perform_corr(matrix, config.get("corr_function"))
-        
-#     matrix = filter_matrix_by_genes(matrix, genes_present)
-
-#     log.info(f"Matrix shape: {matrix.shape}")
-#     df = binary(matrix)
-#     log.info(f"Pair-wise shape: {df.shape}")
-#     df = quick_sort(df, ascending=(sort_order == "low"))
-
-#     gold_pair_to_complex = defaultdict(list)
-#     for idx, row in terms.iterrows():
-#         genes = row.used_genes
-#         if len(genes) < 2:
-#             continue
-#         for i, g1 in enumerate(genes):
-#             for g2 in genes[i + 1:]:
-#                 pair = tuple(sorted((g1, g2)))
-#                 gold_pair_to_complex[pair].append(idx)
-
-#     # Label predictions and complex IDs
-#     complex_ids = []
-#     predictions = []
-#     for g1, g2 in zip(df["gene1"], df["gene2"]):
-#         pair = tuple(sorted((g1, g2)))
-#         ids = gold_pair_to_complex.get(pair, [])
-#         if ids:
-#             predictions.append(1)
-#             complex_ids.append(ids)
-#         else:
-#             predictions.append(0)
-#             complex_ids.append([])
-
-#     df["prediction"] = predictions
-#     df["complex_id"] = complex_ids
-
-#     if df["prediction"].sum() == 0:
-#         log.info("No true positives found in dataset.")
-#         pr_auc = np.nan
-#     else:
-#         tp = df["prediction"].cumsum()
-#         df["tp"] = tp
-#         precision = tp / (np.arange(len(df)) + 1)
-#         recall = tp / tp.iloc[-1]
-#         pr_auc = metrics.auc(recall, precision)
-#         df["precision"] = precision
-#         df["recall"] = recall
-
-#     log.info(f"PR-AUC: {pr_auc:.4f}, Number of true positives: {df['prediction'].sum()}")
-#     dsave(df, "pra", dataset_name)
-#     dsave(pr_auc, "pr_auc", dataset_name)
-#     log.done(f"Global PRA completed for {dataset_name}")
-#     return df, pr_auc
-
-# def compute_pra(df):
-#     log.info("Calculating precision-recall and AUC score.")
-#     if df.empty:
-#         log.warning("Empty DataFrame encountered in compute_pra. Returning empty DataFrame.")
-#         return df  
-#     df["tp"] = df["prediction"].cumsum()
-#     df.reset_index(drop=True, inplace=True)
-#     df["precision"] = df["tp"] / (df.index + 1)
-#     df["recall"] = df["tp"] / df["tp"].iloc[-1]
-#     log.info("DONE: Calculating precision-recall AUC score.")
-#     return df
-
-# def pra(dataset_name, matrix, is_corr=False):
-#     log.info(f"PRA computation started for {dataset_name}.")
-#     genes_present_in_terms = dload("tmp", "genes_present_in_terms")
-#     #terms_hash_table = dload("tmp", "terms_hash_table")
-#     sorting_prefs = dload("input", "sorting")
-#     sort_order = sorting_prefs.get(dataset_name, "high") 
-#     if not is_corr: matrix = perform_corr(matrix, "numpy")
-#     matrix = filter_matrix_by_genes(matrix, genes_present_in_terms)
-#     stack = binary(matrix)
-
-#     log.info("Checking gene pairs against the gold standard.")
-#     gene_pairs = list(zip(stack["gene1"], stack["gene2"]))
-#     hashed_pairs = [hash(pair) for pair in gene_pairs]
-#     stack["complex_id"] = [terms_hash_table.get(h, 0) for h in hashed_pairs]
-#     stack["prediction"] = [1 if h in terms_hash_table else 0 for h in hashed_pairs]
-
-#     annotated = stack.copy()
-#     if sort_order == "low":
-#         ann_sorted = quick_sort(annotated, ascending=True) 
-#     else:
-#         ann_sorted = quick_sort(annotated) 
-
-#     pra = compute_pra(ann_sorted)
-#     pr_auc = metrics.auc(pra.recall, pra.precision)
-#     dsave(pra, "pra", dataset_name)
-#     dsave(pr_auc, "pr_auc", dataset_name)
-#     log.info(f"PRA computation completed for {dataset_name} (Sorting: {sort_order}).")
-#     return pra, pr_auc