masster 0.4.13__py3-none-any.whl → 0.4.16__py3-none-any.whl

masster/study/merge.py

@@ -0,0 +1,1607 @@
+"""
+Unified merge module for the Study class.
+Supports multiple merge methods: 'kd', 'qt', 'kd-nowarp', 'chunked'
+"""
+
+import time
+import numpy as np
+from collections import defaultdict
+from datetime import datetime
+from tqdm import tqdm
+import pyopenms as oms
+import polars as pl
+from masster.study.defaults import merge_defaults
+
+
+def merge(self, **kwargs) -> None:
+    """
+    Group features across samples into consensus features using various algorithms.
+
+    This function provides a unified interface to multiple feature grouping algorithms,
+    each optimized for different dataset sizes and analysis requirements.
+
+    Parameters
+    ----------
+    **kwargs : dict
+        Parameters from merge_defaults class:
+        - method : str, default 'kd'
+          Merge algorithm: 'kd', 'qt', 'kd-nowarp', 'chunked'
+        - min_samples : int, default 10  
+          Minimum number of samples for consensus feature
+        - rt_tol : float, default 2.0
+          RT tolerance in seconds
+        - mz_tol : float, default 0.01
+          m/z tolerance in Da (Daltons) for all methods
+        - chunk_size : int, default 500
+          Chunk size for 'chunked' method
+        - nr_partitions : int, default 500
+          Number of partitions in m/z dimension for KD algorithms
+        - min_rel_cc_size : float, default 0.3
+          Minimum relative connected component size for conflict resolution
+        - max_pairwise_log_fc : float, default 0.5
+          Maximum pairwise log fold change for conflict resolution
+        - max_nr_conflicts : int, default 0
+          Maximum number of conflicts allowed in consensus feature
+        - link_ms2 : bool, default True
+          Whether to link MS2 spectra to consensus features
+
+    Algorithm Guidelines
+    -------------------
+    - KD: Best general purpose, O(n log n), recommended default
+    - QT: Thorough but slow O(n²), good for <1000 samples  
+    - KD-NoWarp: Memory efficient KD without RT warping for large datasets
+    - Chunked: Memory-optimized KD algorithm for very large datasets (>5000 samples)
+      Uses optimized partitioning for better memory management while maintaining
+      full cross-sample consensus feature detection.
+    """
+    start_time = time.time()
+    
+    # Initialize with defaults and override with kwargs
+    params = merge_defaults()
+    
+    # Filter and apply only valid parameters
+    valid_params = set(params.list_parameters())
+    for key, value in kwargs.items():
+        if key in valid_params:
+            setattr(params, key, value)
+        else:
+            self.logger.warning(f"Unknown parameter '{key}' ignored")
+    
+    # Validate method
+    if params.method not in ['kd', 'qt', 'kd-nowarp', 'chunked']:
+        raise ValueError(f"Invalid method '{params.method}'. Must be one of: ['kd', 'qt', 'kd-nowarp', 'chunked']")
+    
+    # Persist last used params for diagnostics
+    try:
+        self._merge_params_last = params.to_dict()
+    except Exception:
+        self._merge_params_last = {}
+    
+    # Ensure feature maps are available for merging (regenerate if needed)
+    if len(self.features_maps) < len(self.samples_df):
+        self.features_maps = []
+        self.load_features()
+    
+    self.logger.info(
+        f"Merge: {params.method}, samples={params.min_samples}, rt_tol={params.rt_tol}s, mz_tol={params.mz_tol}Da, min_rel_cc_size={params.min_rel_cc_size}, max_pairwise_log_fc={params.max_pairwise_log_fc}, max_nr_conflicts={params.max_nr_conflicts}"
+    )
+    
+    # Initialize
+    self._reset_consensus_data()
+
+    # Cache adducts for performance (avoid repeated _get_adducts() calls)
+    cached_adducts_df = None
+    cached_valid_adducts = None
+    try:
+        cached_adducts_df = self._get_adducts()
+        if not cached_adducts_df.is_empty():
+            cached_valid_adducts = set(cached_adducts_df["name"].to_list())
+        else:
+            cached_valid_adducts = set()
+    except Exception as e:
+        self.logger.warning(f"Could not retrieve study adducts: {e}")
+        cached_valid_adducts = set()
+    
+    # Always allow '?' adducts
+    cached_valid_adducts.add("?")
+    
+    # Route to algorithm implementation  
+    if params.method == 'kd':
+        consensus_map = _merge_kd(self, params)
+        # Extract consensus features
+        self._extract_consensus_features(consensus_map, params.min_samples, cached_adducts_df, cached_valid_adducts)
+    elif params.method == 'qt':
+        consensus_map = _merge_qt(self, params)
+        # Extract consensus features
+        self._extract_consensus_features(consensus_map, params.min_samples, cached_adducts_df, cached_valid_adducts)
+    elif params.method == 'kd-nowarp':
+        consensus_map = _merge_kd_nowarp(self, params)
+        # Extract consensus features
+        self._extract_consensus_features(consensus_map, params.min_samples, cached_adducts_df, cached_valid_adducts)
+    elif params.method == 'chunked':
+        consensus_map = _merge_chunked(self, params, cached_adducts_df, cached_valid_adducts)
+        # Note: _merge_chunked populates consensus_df directly, no need to extract
+    
+    # Perform adduct grouping
+    self._perform_adduct_grouping(params.rt_tol, params.mz_tol)
+    
+    # Link MS2 if requested
+    if params.link_ms2:
+        self._finalize_merge(params.link_ms2, params.min_samples)
+    
+    # Log completion without the misleading feature count
+    elapsed = time.time() - start_time
+    self.logger.debug(f"Merge process completed in {elapsed:.1f}s")
+
+
+def _merge_kd(self, params: merge_defaults) -> oms.ConsensusMap:
+    """KD-tree based merge (fast, recommended)"""
+    
+    consensus_map = oms.ConsensusMap()
+    file_descriptions = consensus_map.getColumnHeaders()
+    
+    for i, feature_map in enumerate(self.features_maps):
+        file_description = file_descriptions.get(i, oms.ColumnHeader())
+        file_description.filename = self.samples_df.row(i, named=True)["sample_name"]
+        file_description.size = feature_map.size()
+        file_description.unique_id = feature_map.getUniqueId()
+        file_descriptions[i] = file_description
+    
+    consensus_map.setColumnHeaders(file_descriptions)
+    
+    # Configure KD algorithm
+    grouper = oms.FeatureGroupingAlgorithmKD()
+    params_oms = grouper.getParameters()
+    
+    params_oms.setValue("mz_unit", "Da")
+    params_oms.setValue("nr_partitions", params.nr_partitions)
+    params_oms.setValue("warp:enabled", "true")
+    params_oms.setValue("warp:rt_tol", params.rt_tol)
+    params_oms.setValue("warp:mz_tol", params.mz_tol)
+    params_oms.setValue("link:rt_tol", params.rt_tol)
+    params_oms.setValue("link:mz_tol", params.mz_tol)
+    params_oms.setValue("link:min_rel_cc_size", params.min_rel_cc_size)
+    params_oms.setValue("link:max_pairwise_log_fc", params.max_pairwise_log_fc)
+    params_oms.setValue("link:max_nr_conflicts", params.max_nr_conflicts)
+    #params_oms.setValue("link:charge_merging", "With_charge_zero") THIS LEADS TO A CRASH
+    
+    grouper.setParameters(params_oms)
+    grouper.group(self.features_maps, consensus_map)
+    
+    return consensus_map
+
+
+def _merge_qt(self, params: merge_defaults) -> oms.ConsensusMap:
+    """QT (Quality Threshold) based merge"""
+    
+    n_samples = len(self.features_maps)
+    if n_samples > 1000:
+        self.logger.warning(f"QT with {n_samples} samples may be slow [O(n²)]. Consider KD [O(n log n)]")
+    
+    consensus_map = oms.ConsensusMap()
+    file_descriptions = consensus_map.getColumnHeaders()
+    
+    for i, feature_map in enumerate(self.features_maps):
+        file_description = file_descriptions.get(i, oms.ColumnHeader())
+        file_description.filename = self.samples_df.row(i, named=True)["sample_name"]
+        file_description.size = feature_map.size()
+        file_description.unique_id = feature_map.getUniqueId()
+        file_descriptions[i] = file_description
+    
+    consensus_map.setColumnHeaders(file_descriptions)
+    
+    # Configure QT algorithm
+    grouper = oms.FeatureGroupingAlgorithmQT()
+    params_oms = grouper.getParameters()
+    
+    params_oms.setValue("distance_RT:max_difference", params.rt_tol)
+    params_oms.setValue("distance_MZ:max_difference", params.mz_tol)
+    params_oms.setValue("distance_MZ:unit", "Da")  # QT now uses Da like all other methods
+    params_oms.setValue("ignore_charge", "true")
+    params_oms.setValue("min_rel_cc_size", params.min_rel_cc_size)
+    params_oms.setValue("max_pairwise_log_fc", params.max_pairwise_log_fc)
+    params_oms.setValue("max_nr_conflicts", params.max_nr_conflicts)
+    params_oms.setValue("nr_partitions", params.nr_partitions)
+
+    grouper.setParameters(params_oms)
+    grouper.group(self.features_maps, consensus_map)
+    
+    return consensus_map
+
+
+def _merge_kd_nowarp(self, params: merge_defaults) -> oms.ConsensusMap:
+    """KD-tree based merge without RT warping"""
+    
+    consensus_map = oms.ConsensusMap()
+    file_descriptions = consensus_map.getColumnHeaders()
+    
+    for i, feature_map in enumerate(self.features_maps):
+        file_description = file_descriptions.get(i, oms.ColumnHeader())
+        file_description.filename = self.samples_df.row(i, named=True)["sample_name"]
+        file_description.size = feature_map.size()
+        file_description.unique_id = feature_map.getUniqueId()
+        file_descriptions[i] = file_description
+    
+    consensus_map.setColumnHeaders(file_descriptions)
+    
+    # Configure KD algorithm with warping disabled for memory efficiency
+    grouper = oms.FeatureGroupingAlgorithmKD()
+    params_oms = grouper.getParameters()
+    
+    params_oms.setValue("mz_unit", "Da")
+    params_oms.setValue("nr_partitions", params.nr_partitions)
+    params_oms.setValue("warp:enabled", "false")  # Disabled for memory efficiency
+    params_oms.setValue("link:rt_tol", params.rt_tol)
+    params_oms.setValue("link:mz_tol", params.mz_tol)
+    params_oms.setValue("link:min_rel_cc_size", params.min_rel_cc_size)
+    params_oms.setValue("link:max_pairwise_log_fc", params.max_pairwise_log_fc)
+    params_oms.setValue("link:max_nr_conflicts", params.max_nr_conflicts)
+    #params_oms.setValue("link:charge_merging", "Any")
+    
+    grouper.setParameters(params_oms)
+    grouper.group(self.features_maps, consensus_map)
+    
+    return consensus_map
+
+
+def _merge_chunked(self, params: merge_defaults, cached_adducts_df=None, cached_valid_adducts=None) -> oms.ConsensusMap:
+    """Chunked merge with proper cross-chunk consensus building"""
+    
+    n_samples = len(self.features_maps)
+    if n_samples <= params.chunk_size:
+        self.logger.info(f"Dataset size ({n_samples}) ≤ chunk_size, using KD merge")
+        consensus_map = _merge_kd(self, params)
+        # Extract consensus features to populate consensus_df for chunked method consistency
+        self._extract_consensus_features(consensus_map, params.min_samples, cached_adducts_df, cached_valid_adducts)
+        return consensus_map
+    
+    # Process in chunks
+    chunks = []
+    for i in range(0, n_samples, params.chunk_size):
+        chunk_end = min(i + params.chunk_size, n_samples)
+        chunks.append((i, self.features_maps[i:chunk_end]))
+    
+    self.logger.debug(f"Processing {len(chunks)} chunks of max {params.chunk_size} samples")
+    
+    # Process each chunk to create chunk consensus maps
+    chunk_consensus_maps = []
+    
+    for chunk_idx, (chunk_start_idx, chunk_maps) in enumerate(tqdm(chunks, desc="Chunk", disable=self.log_level not in ["TRACE", "DEBUG", "INFO"])):
+        chunk_consensus_map = oms.ConsensusMap()
+        
+        # Set up file descriptions for chunk
+        file_descriptions = chunk_consensus_map.getColumnHeaders()
+        for j, feature_map in enumerate(chunk_maps):
+            file_description = file_descriptions.get(j, oms.ColumnHeader())
+            file_description.filename = self.samples_df.row(chunk_start_idx + j, named=True)["sample_name"]
+            file_description.size = feature_map.size()
+            file_description.unique_id = feature_map.getUniqueId()
+            file_descriptions[j] = file_description
+        
+        chunk_consensus_map.setColumnHeaders(file_descriptions)
+        
+        # Use KD algorithm for chunk
+        grouper = oms.FeatureGroupingAlgorithmKD()
+        chunk_params = grouper.getParameters()
+        chunk_params.setValue("mz_unit", "Da")
+        chunk_params.setValue("nr_partitions", params.nr_partitions)
+        chunk_params.setValue("warp:enabled", "true")
+        chunk_params.setValue("warp:rt_tol", params.rt_tol)
+        chunk_params.setValue("warp:mz_tol", params.mz_tol)
+        chunk_params.setValue("link:rt_tol", params.rt_tol)
+        chunk_params.setValue("link:mz_tol", params.mz_tol)
+        chunk_params.setValue("link:min_rel_cc_size", params.min_rel_cc_size)
+        chunk_params.setValue("link:max_pairwise_log_fc", params.max_pairwise_log_fc)
+        chunk_params.setValue("link:max_nr_conflicts", params.max_nr_conflicts)
+        
+        grouper.setParameters(chunk_params)
+        grouper.group(chunk_maps, chunk_consensus_map)
+        
+        chunk_consensus_maps.append((chunk_start_idx, chunk_consensus_map))
+    
+    # Merge chunk results with proper cross-chunk consensus building
+    _merge_chunk_results(self, chunk_consensus_maps, params, cached_adducts_df, cached_valid_adducts)
+    
+    # Create a dummy consensus map for compatibility (since other functions expect it)
+    consensus_map = oms.ConsensusMap()
+    return consensus_map
+
+
+def _merge_chunk_results(self, chunk_consensus_maps: list, params: merge_defaults, cached_adducts_df=None, cached_valid_adducts=None) -> None:
+    """
+    Scalable aggregation of chunk consensus maps into final consensus_df.
+    
+    This function implements cross-chunk consensus building by:
+    1. Extracting feature_uids from each chunk consensus map
+    2. Aggregating features close in RT/m/z across chunks
+    3. Building consensus_df and consensus_mapping_df directly
+    """
+    
+    if len(chunk_consensus_maps) == 1:
+        # Single chunk case - just extract using the true global min_samples.
+        # No need for permissive threshold because we are not discarding singletons pre-aggregation.
+        self._extract_consensus_features(
+            chunk_consensus_maps[0][1],
+            params.min_samples,
+            cached_adducts_df,
+            cached_valid_adducts,
+        )
+        return
+    
+    # Build feature_uid to feature_data lookup for fast access
+    feature_uid_map = {
+        row["feature_id"]: row["feature_uid"]
+        for row in self.features_df.iter_rows(named=True)
+    }
+    
+    features_lookup = _optimized_feature_lookup(self, self.features_df)
+    
+    # Extract all consensus features from chunks with their feature_uids
+    all_chunk_consensus = []
+    consensus_id_counter = 0
+    
+    for chunk_idx, (chunk_start_idx, chunk_consensus_map) in enumerate(chunk_consensus_maps):
+        for consensus_feature in chunk_consensus_map:
+            # ACCEPT ALL consensus features (size >=1) here.
+            # Reason: A feature that is globally present in many samples can still
+            # appear only once inside a given sample chunk. Early filtering at
+            # size>=2 causes irreversible loss and underestimates the final
+            # consensus count (observed ~296 vs 950 for KD). We defer filtering
+            # strictly to the final global min_samples.
+                
+            # Extract feature_uids from this consensus feature
+            feature_uids = []
+            feature_data_list = []
+            sample_uids = []
+            
+            for feature_handle in consensus_feature.getFeatureList():
+                fuid = str(feature_handle.getUniqueId())
+                if fuid not in feature_uid_map:
+                    continue
+                    
+                feature_uid = feature_uid_map[fuid]
+                feature_data = features_lookup.get(feature_uid)
+                if feature_data:
+                    feature_uids.append(feature_uid)
+                    feature_data_list.append(feature_data)
+                    sample_uids.append(chunk_start_idx + feature_handle.getMapIndex() + 1)
+
+            if not feature_data_list:
+                # No retrievable feature metadata (possible stale map reference) -> skip
+                continue            # Derive RT / m/z ranges from underlying features (used for robust cross-chunk stitching)
+            rt_vals_local = [fd.get("rt") for fd in feature_data_list if fd.get("rt") is not None]
+            mz_vals_local = [fd.get("mz") for fd in feature_data_list if fd.get("mz") is not None]
+            if rt_vals_local:
+                rt_min_local = min(rt_vals_local)
+                rt_max_local = max(rt_vals_local)
+            else:
+                rt_min_local = rt_max_local = consensus_feature.getRT()
+            if mz_vals_local:
+                mz_min_local = min(mz_vals_local)
+                mz_max_local = max(mz_vals_local)
+            else:
+                mz_min_local = mz_max_local = consensus_feature.getMZ()
+                
+            # Store chunk consensus with feature tracking
+            chunk_consensus_data = {
+                'consensus_id': consensus_id_counter,
+                'chunk_idx': chunk_idx,
+                'chunk_start_idx': chunk_start_idx,
+                'mz': consensus_feature.getMZ(),
+                'rt': consensus_feature.getRT(),
+                'mz_min': mz_min_local,
+                'mz_max': mz_max_local,
+                'rt_min': rt_min_local,
+                'rt_max': rt_max_local,
+                'intensity': consensus_feature.getIntensity(),
+                'quality': consensus_feature.getQuality(),
+                'feature_uids': feature_uids,
+                'feature_data_list': feature_data_list,
+                'sample_uids': sample_uids,
+                'sample_count': len(feature_data_list)
+            }
+            
+            all_chunk_consensus.append(chunk_consensus_data)
+            consensus_id_counter += 1
+
+    if not all_chunk_consensus:
+        # No valid consensus features found
+        self.consensus_df = pl.DataFrame()
+        self.consensus_mapping_df = pl.DataFrame()
+        return
+    
+    # Perform cross-chunk clustering using optimized spatial indexing
+    def _cluster_chunk_consensus(chunk_consensus_list: list, rt_tol: float, mz_tol: float) -> list:
+        """Cluster chunk consensus features using interval overlap (no over-relaxation).
+
+        A union is formed if either centroids are within tolerance OR their RT / m/z
+        intervals (expanded by tolerance) overlap, and they originate from different chunks.
+        """
+        if not chunk_consensus_list:
+            return []
+
+        n_features = len(chunk_consensus_list)
+
+        # Spatial bins using strict tolerances (improves candidate reduction without recall loss)
+        rt_bin_size = rt_tol if rt_tol > 0 else 1.0
+        mz_bin_size = mz_tol if mz_tol > 0 else 0.01
+        features_by_bin = defaultdict(list)
+
+        for i, cf in enumerate(chunk_consensus_list):
+            rt_bin = int(cf['rt'] / rt_bin_size)
+            mz_bin = int(cf['mz'] / mz_bin_size)
+            features_by_bin[(rt_bin, mz_bin)].append(i)
+
+        class UF:
+            def __init__(self, n):
+                self.p = list(range(n))
+                self.r = [0]*n
+            def find(self, x):
+                if self.p[x] != x:
+                    self.p[x] = self.find(self.p[x])
+                return self.p[x]
+            def union(self, a,b):
+                pa, pb = self.find(a), self.find(b)
+                if pa == pb:
+                    return
+                if self.r[pa] < self.r[pb]:
+                    pa, pb = pb, pa
+                self.p[pb] = pa
+                if self.r[pa] == self.r[pb]:
+                    self.r[pa] += 1
+
+        uf = UF(n_features)
+        checked = set()
+        for (rtb, mzb), idxs in features_by_bin.items():
+            for dr in (-1,0,1):
+                for dm in (-1,0,1):
+                    neigh = (rtb+dr, mzb+dm)
+                    if neigh not in features_by_bin:
+                        continue
+                    for i in idxs:
+                        for j in features_by_bin[neigh]:
+                            if i >= j:
+                                continue
+                            pair = (i,j)
+                            if pair in checked:
+                                continue
+                            checked.add(pair)
+                            a = chunk_consensus_list[i]
+                            b = chunk_consensus_list[j]
+                            if a['chunk_idx'] == b['chunk_idx']:
+                                continue
+                            # Centroid checks
+                            centroid_close = (abs(a['rt']-b['rt']) <= rt_tol and abs(a['mz']-b['mz']) <= mz_tol)
+                            # Interval overlap checks (expanded by tolerance)
+                            rt_overlap = (a['rt_min'] - rt_tol) <= (b['rt_max'] + rt_tol) and (b['rt_min'] - rt_tol) <= (a['rt_max'] + rt_tol)
+                            mz_overlap = (a['mz_min'] - mz_tol) <= (b['mz_max'] + mz_tol) and (b['mz_min'] - mz_tol) <= (a['mz_max'] + mz_tol)
+                            if centroid_close or (rt_overlap and mz_overlap):
+                                uf.union(i,j)
+
+        groups_by_root = defaultdict(list)
+        for i in range(n_features):
+            groups_by_root[uf.find(i)].append(chunk_consensus_list[i])
+        return list(groups_by_root.values())
+    # (Obsolete relaxed + centroid stitching code removed.)
+
+    # --- Stage 1: initial cross-chunk clustering of chunk consensus features ---
+    initial_groups = _cluster_chunk_consensus(all_chunk_consensus, params.rt_tol, params.mz_tol)
+
+    # --- Stage 2: centroid refinement (lightweight second pass) ---
+    def _refine_groups(groups: list, rt_tol: float, mz_tol: float) -> list:
+        """Refine groups by clustering group centroids (single-link) under same tolerances.
+
+        This reconciles borderline splits left after interval-overlap clustering without
+        re-introducing broad over-merging. Works on group centroids only (low cost).
+        """
+        if len(groups) <= 1:
+            return groups
+        # Build centroid list
+        centroids = []  # (idx, rt, mz)
+        for gi, g in enumerate(groups):
+            if not g:
+                continue
+            rt_vals = [cf['rt'] for cf in g]
+            mz_vals = [cf['mz'] for cf in g]
+            if not rt_vals or not mz_vals:
+                continue
+            centroids.append((gi, float(np.mean(rt_vals)), float(np.mean(mz_vals))))
+        if len(centroids) <= 1:
+            return groups
+
+        # Spatial binning for centroid clustering
+        rt_bin = rt_tol if rt_tol > 0 else 1.0
+        mz_bin = mz_tol if mz_tol > 0 else 0.01
+        bins = defaultdict(list)
+        for idx, rt_c, mz_c in centroids:
+            bins[(int(rt_c/rt_bin), int(mz_c/mz_bin))].append((idx, rt_c, mz_c))
+
+        # Union-Find over group indices
+        parent = list(range(len(groups)))
+        rank = [0]*len(groups)
+        def find(x):
+            if parent[x] != x:
+                parent[x] = find(parent[x])
+            return parent[x]
+        def union(a,b):
+            pa, pb = find(a), find(b)
+            if pa == pb:
+                return
+            if rank[pa] < rank[pb]:
+                pa, pb = pb, pa
+            parent[pb] = pa
+            if rank[pa] == rank[pb]:
+                rank[pa] += 1
+
+        checked = set()
+        for (rb, mb), items in bins.items():
+            for dr in (-1,0,1):
+                for dm in (-1,0,1):
+                    neigh_key = (rb+dr, mb+dm)
+                    if neigh_key not in bins:
+                        continue
+                    for (gi, rt_i, mz_i) in items:
+                        for (gj, rt_j, mz_j) in bins[neigh_key]:
+                            if gi >= gj:
+                                continue
+                            pair = (gi, gj)
+                            if pair in checked:
+                                continue
+                            checked.add(pair)
+                            if abs(rt_i-rt_j) <= rt_tol and abs(mz_i-mz_j) <= mz_tol:
+                                union(gi, gj)
+
+        merged = defaultdict(list)
+        for gi, g in enumerate(groups):
+            merged[find(gi)].extend(g)
+        return list(merged.values())
+
+    refined_groups = _refine_groups(initial_groups, params.rt_tol, params.mz_tol)
+
+    # --- Stage 3: build final consensus feature metadata and mapping ---
+    consensus_metadata = []
+    consensus_mapping_list = []
+    consensus_uid_counter = 0
+
+    for group in refined_groups:
+        if not group:
+            continue
+        
+        # Aggregate underlying feature data (deduplicated by feature_uid)
+        feature_data_acc = {}
+        sample_uids_acc = set()
+        rt_values_chunk = []  # use chunk-level centroids for statistic helper
+        mz_values_chunk = []
+        intensity_values_chunk = []
+        quality_values_chunk = []
+
+        for cf in group:
+            rt_values_chunk.append(cf['rt'])
+            mz_values_chunk.append(cf['mz'])
+            intensity_values_chunk.append(cf.get('intensity', 0.0) or 0.0)
+            quality_values_chunk.append(cf.get('quality', 1.0) or 1.0)
+            
+            for fd, samp_uid in zip(cf['feature_data_list'], cf['sample_uids']):
+                fid = fd.get('feature_uid') or fd.get('uid') or fd.get('feature_id')
+                # feature_uid expected in fd under 'feature_uid'; fallback attempts just in case
+                if fid is None:
+                    continue
+                if fid not in feature_data_acc:
+                    feature_data_acc[fid] = fd
+                sample_uids_acc.add(samp_uid)
+                
+        if not feature_data_acc:
+            continue
+
+        number_samples = len(sample_uids_acc)
+        
+        # NOTE: Don't filter by min_samples here - let _finalize_merge handle it
+        # This allows proper cross-chunk consensus building before final filtering
+
+        metadata = _calculate_consensus_statistics(
+            self,
+            consensus_uid_counter,
+            list(feature_data_acc.values()),
+            rt_values_chunk,
+            mz_values_chunk,
+            intensity_values_chunk,
+            quality_values_chunk,
+            number_features=len(feature_data_acc),
+            number_samples=number_samples,
+            cached_adducts_df=cached_adducts_df,
+            cached_valid_adducts=cached_valid_adducts,
+        )
+        consensus_metadata.append(metadata)
+
+        # Build mapping rows (deduplicated)
+        for fid, fd in feature_data_acc.items():
+            samp_uid = fd.get('sample_uid') or fd.get('sample_id') or fd.get('sample')
+            # If absent we attempt to derive from original group sample_uids pairing
+            # but most feature_data rows should include sample_uid already.
+            if samp_uid is None:
+                # fallback: search for cf containing this fid
+                for cf in group:
+                    for fd2, samp2 in zip(cf['feature_data_list'], cf['sample_uids']):
+                        f2id = fd2.get('feature_uid') or fd2.get('uid') or fd2.get('feature_id')
+                        if f2id == fid:
+                            samp_uid = samp2
+                            break
+                    if samp_uid is not None:
+                        break
+            if samp_uid is None:
+                continue
+            consensus_mapping_list.append({
+                'consensus_uid': consensus_uid_counter,
+                'sample_uid': samp_uid,
+                'feature_uid': fid,
+            })
+
+        consensus_uid_counter += 1
+
+    # Assign DataFrames
+    self.consensus_df = pl.DataFrame(consensus_metadata, strict=False)
+    self.consensus_mapping_df = pl.DataFrame(consensus_mapping_list, strict=False)
+
+    # Ensure mapping only contains features from retained consensus_df
+    if len(self.consensus_df) > 0:
+        valid_consensus_ids = set(self.consensus_df['consensus_uid'].to_list())
+        self.consensus_mapping_df = self.consensus_mapping_df.filter(
+            pl.col('consensus_uid').is_in(list(valid_consensus_ids))
+        )
+    else:
+        self.consensus_mapping_df = pl.DataFrame()
+
+    # Attach empty consensus_map placeholder for downstream compatibility
+    self.consensus_map = oms.ConsensusMap()
+    return
+
+
+def _calculate_consensus_statistics(study_obj, consensus_uid: int, feature_data_list: list, 
+                                  rt_values: list, mz_values: list, 
+                                  intensity_values: list, quality_values: list,
+                                  number_features: int = None, number_samples: int = None,
+                                  cached_adducts_df=None, cached_valid_adducts=None) -> dict:
+    """
+    Calculate comprehensive statistics for a consensus feature from aggregated feature data.
+    
+    Args:
+        consensus_uid: Unique ID for this consensus feature
+        feature_data_list: List of individual feature dictionaries
+        rt_values: RT values from chunk consensus features
+        mz_values: m/z values from chunk consensus features  
+        intensity_values: Intensity values from chunk consensus features
+        quality_values: Quality values from chunk consensus features
+        
+    Returns:
+        Dictionary with consensus feature metadata
+    """
+    if not feature_data_list:
+        return {}
+    
+    # Convert feature data to numpy arrays for vectorized computation
+    rt_feat_values = np.array([fd.get("rt", 0) for fd in feature_data_list if fd.get("rt") is not None])
+    mz_feat_values = np.array([fd.get("mz", 0) for fd in feature_data_list if fd.get("mz") is not None])
+    rt_start_values = np.array([fd.get("rt_start", 0) for fd in feature_data_list if fd.get("rt_start") is not None])
+    rt_end_values = np.array([fd.get("rt_end", 0) for fd in feature_data_list if fd.get("rt_end") is not None])
+    rt_delta_values = np.array([fd.get("rt_delta", 0) for fd in feature_data_list if fd.get("rt_delta") is not None])
+    mz_start_values = np.array([fd.get("mz_start", 0) for fd in feature_data_list if fd.get("mz_start") is not None])
+    mz_end_values = np.array([fd.get("mz_end", 0) for fd in feature_data_list if fd.get("mz_end") is not None])
+    inty_values = np.array([fd.get("inty", 0) for fd in feature_data_list if fd.get("inty") is not None])
+    coherence_values = np.array([fd.get("chrom_coherence", 0) for fd in feature_data_list if fd.get("chrom_coherence") is not None])
+    prominence_values = np.array([fd.get("chrom_prominence", 0) for fd in feature_data_list if fd.get("chrom_prominence") is not None])
+    prominence_scaled_values = np.array([fd.get("chrom_prominence_scaled", 0) for fd in feature_data_list if fd.get("chrom_prominence_scaled") is not None])
+    height_scaled_values = np.array([fd.get("chrom_height_scaled", 0) for fd in feature_data_list if fd.get("chrom_height_scaled") is not None])
+    iso_values = np.array([fd.get("iso", 0) for fd in feature_data_list if fd.get("iso") is not None])
+    charge_values = np.array([fd.get("charge", 0) for fd in feature_data_list if fd.get("charge") is not None])
+    
+    # Process adducts with cached validation
+    all_adducts = []
+    valid_adducts = cached_valid_adducts if cached_valid_adducts is not None else set()
+    valid_adducts.add("?")  # Always allow '?' adducts
+    
+    for fd in feature_data_list:
+        adduct = fd.get("adduct")
+        if adduct is not None:
+            # Only include adducts that are valid (from cached study adducts or contain '?')
+            if adduct in valid_adducts or "?" in adduct:
+                all_adducts.append(adduct)
+    
+    # Calculate adduct consensus
+    adduct_values = []
+    adduct_top = None
+    adduct_charge_top = None
+    adduct_mass_neutral_top = None
+    adduct_mass_shift_top = None
+    
+    if all_adducts:
+        adduct_counts = {adduct: all_adducts.count(adduct) for adduct in set(all_adducts)}
+        total_count = sum(adduct_counts.values())
+        for adduct, count in adduct_counts.items():
+            percentage = (count / total_count) * 100 if total_count > 0 else 0
+            adduct_values.append([str(adduct), int(count), float(round(percentage, 2))])
+        
+        adduct_values.sort(key=lambda x: x[1], reverse=True)
+        
+        if adduct_values:
+            adduct_top = adduct_values[0][0]
+            # Try to get charge and mass shift from cached study adducts
+            adduct_found = False
+            if cached_adducts_df is not None and not cached_adducts_df.is_empty():
+                matching_adduct = cached_adducts_df.filter(
+                    pl.col("name") == adduct_top,
+                )
+                if not matching_adduct.is_empty():
+                    adduct_row = matching_adduct.row(0, named=True)
+                    adduct_charge_top = adduct_row["charge"]
+                    adduct_mass_shift_top = adduct_row["mass_shift"]
+                    adduct_found = True
+            
+            if not adduct_found:
+                # Set default charge and mass shift for top adduct
+                adduct_charge_top = 1
+                adduct_mass_shift_top = 1.007825
+    else:
+        # Default adduct based on study polarity
+        study_polarity = getattr(study_obj, "polarity", "positive")
+        if study_polarity in ["negative", "neg"]:
+            adduct_top = "[M-?]1-"
+            adduct_charge_top = -1
+            adduct_mass_shift_top = -1.007825
+        else:
+            adduct_top = "[M+?]1+"
+            adduct_charge_top = 1
+            adduct_mass_shift_top = 1.007825
+        
+        adduct_values = [[adduct_top, 1, 100.0]]
+    
+    # Calculate neutral mass
+    consensus_mz = round(float(np.mean(mz_values)), 4) if len(mz_values) > 0 else 0.0
+    if adduct_charge_top and adduct_mass_shift_top is not None:
+        adduct_mass_neutral_top = consensus_mz * abs(adduct_charge_top) - adduct_mass_shift_top
+    
+    # Calculate MS2 count
+    ms2_count = 0
+    for fd in feature_data_list:
+        ms2_scans = fd.get("ms2_scans")
+        if ms2_scans is not None:
+            ms2_count += len(ms2_scans)
+    
+    # Build consensus metadata
+    return {
+        "consensus_uid": int(consensus_uid),
+        "consensus_id": str(consensus_uid),  # Use simple string ID
+        "quality": round(float(np.mean(quality_values)), 3) if len(quality_values) > 0 else 1.0,
+        "number_samples": number_samples if number_samples is not None else len(feature_data_list),
+        "rt": round(float(np.mean(rt_values)), 4) if len(rt_values) > 0 else 0.0,
+        "mz": consensus_mz,
+        "rt_min": round(float(np.min(rt_feat_values)), 3) if len(rt_feat_values) > 0 else 0.0,
+        "rt_max": round(float(np.max(rt_feat_values)), 3) if len(rt_feat_values) > 0 else 0.0,
+        "rt_mean": round(float(np.mean(rt_feat_values)), 3) if len(rt_feat_values) > 0 else 0.0,
+        "rt_start_mean": round(float(np.mean(rt_start_values)), 3) if len(rt_start_values) > 0 else 0.0,
+        "rt_end_mean": round(float(np.mean(rt_end_values)), 3) if len(rt_end_values) > 0 else 0.0,
+        "rt_delta_mean": round(float(np.mean(rt_delta_values)), 3) if len(rt_delta_values) > 0 else 0.0,
+        "mz_min": round(float(np.min(mz_feat_values)), 4) if len(mz_feat_values) > 0 else 0.0,
+        "mz_max": round(float(np.max(mz_feat_values)), 4) if len(mz_feat_values) > 0 else 0.0,
+        "mz_mean": round(float(np.mean(mz_feat_values)), 4) if len(mz_feat_values) > 0 else 0.0,
+        "mz_start_mean": round(float(np.mean(mz_start_values)), 4) if len(mz_start_values) > 0 else 0.0,
+        "mz_end_mean": round(float(np.mean(mz_end_values)), 4) if len(mz_end_values) > 0 else 0.0,
+        "inty_mean": round(float(np.mean(inty_values)), 0) if len(inty_values) > 0 else 0.0,
+        "bl": -1.0,
+        "chrom_coherence_mean": round(float(np.mean(coherence_values)), 3) if len(coherence_values) > 0 else 0.0,
+        "chrom_prominence_mean": round(float(np.mean(prominence_values)), 0) if len(prominence_values) > 0 else 0.0,
+        "chrom_prominence_scaled_mean": round(float(np.mean(prominence_scaled_values)), 3) if len(prominence_scaled_values) > 0 else 0.0,
+        "chrom_height_scaled_mean": round(float(np.mean(height_scaled_values)), 3) if len(height_scaled_values) > 0 else 0.0,
+        "iso_mean": round(float(np.mean(iso_values)), 2) if len(iso_values) > 0 else 0.0,
+        "charge_mean": round(float(np.mean(charge_values)), 2) if len(charge_values) > 0 else 0.0,
+        "number_ms2": int(ms2_count),
+        "adducts": adduct_values,
+        "adduct_top": adduct_top,
+        "adduct_charge_top": adduct_charge_top,
+        "adduct_mass_neutral_top": round(adduct_mass_neutral_top, 6) if adduct_mass_neutral_top is not None else None,
+        "adduct_mass_shift_top": round(adduct_mass_shift_top, 6) if adduct_mass_shift_top is not None else None,
+        "id_top_name": None,
+        "id_top_class": None,
+        "id_top_adduct": None,
+        "id_top_score": None,
+    }
+
+
+def _cluster_consensus_features(features: list, rt_tol: float, mz_tol: float) -> list:
+    """
+    Cluster consensus features from different chunks based on RT and m/z similarity.
+    
+    Args:
+        features: List of feature dictionaries with 'mz', 'rt', 'id' keys
+        rt_tol: RT tolerance in seconds
+        mz_tol: m/z tolerance in Da
+        
+    Returns:
+        List of groups, where each group is a list of feature dictionaries
+    """
+    if not features:
+        return []
+    
+    # Use Union-Find for efficient clustering
+    class UnionFind:
+        def __init__(self, n):
+            self.parent = list(range(n))
+            self.rank = [0] * n
+        
+        def find(self, x):
+            if self.parent[x] != x:
+                self.parent[x] = self.find(self.parent[x])
+            return self.parent[x]
+        
+        def union(self, x, y):
+            px, py = self.find(x), self.find(y)
+            if px == py:
+                return
+            if self.rank[px] < self.rank[py]:
+                px, py = py, px
+            self.parent[py] = px
+            if self.rank[px] == self.rank[py]:
+                self.rank[px] += 1
+    
+    n_features = len(features)
+    uf = UnionFind(n_features)
+    
+    # Build distance matrix and cluster features within tolerance
+    for i in range(n_features):
+        for j in range(i + 1, n_features):
+            feat_i = features[i]
+            feat_j = features[j]
+            
+            # Skip if features are from the same chunk (they're already processed)
+            if feat_i['chunk_idx'] == feat_j['chunk_idx']:
+                continue
+            
+            mz_diff = abs(feat_i['mz'] - feat_j['mz'])
+            rt_diff = abs(feat_i['rt'] - feat_j['rt'])
+            
+            # Cluster if within tolerance
+            if mz_diff <= mz_tol and rt_diff <= rt_tol:
+                uf.union(i, j)
+    
+    # Extract groups
+    groups_by_root = {}
+    for i in range(n_features):
+        root = uf.find(i)
+        if root not in groups_by_root:
+            groups_by_root[root] = []
+        groups_by_root[root].append(features[i])
+    
+    return list(groups_by_root.values())
+
+
+# Note: Restored proper chunked implementation with cross-chunk consensus clustering
+
+
+def _reset_consensus_data(self):
+    """Reset consensus-related DataFrames at the start of merge."""
+    self.consensus_df = pl.DataFrame()
+    self.consensus_ms2 = pl.DataFrame()
+    self.consensus_mapping_df = pl.DataFrame()
+
+
+def _extract_consensus_features(self, consensus_map, min_samples, cached_adducts_df=None, cached_valid_adducts=None):
+    """Extract consensus features and build metadata."""
+    # create a dict to map uid to feature_uid using self.features_df
+    feature_uid_map = {
+        row["feature_id"]: row["feature_uid"]
+        for row in self.features_df.iter_rows(named=True)
+    }
+    imax = consensus_map.size()
+
+    self.logger.debug(f"Found {imax} feature groups by clustering.")
+
+    # Pre-build fast lookup tables for features_df data using optimized approach
+    features_lookup = _optimized_feature_lookup(self, self.features_df)
+
+    # create a list to store the consensus mapping
+    consensus_mapping = []
+    metadata_list = []
+
+    tqdm_disable = self.log_level not in ["TRACE", "DEBUG"]
+
+    for i, feature in enumerate(
+        tqdm(
+            consensus_map,
+            total=imax,
+            disable=tqdm_disable,
+            desc=f"{datetime.now().strftime('%Y-%m-%d %H:%M:%S.%f')[:-3]} | INFO     | {self.log_label}Extract metadata",
+        ),
+    ):
+        # get all features in the feature map with the same unique id as the consensus feature
+        features_list = feature.getFeatureList()
+        uids = []
+        feature_data_list = []
+
+        for _j, f in enumerate(features_list):
+            fuid = str(f.getUniqueId())
+            if fuid not in feature_uid_map:
+                # this is a feature that was removed but is still in the feature maps
+                continue
+            fuid = feature_uid_map[fuid]
+            consensus_mapping.append(
+                {
+                    "consensus_uid": i,
+                    "sample_uid": f.getMapIndex() + 1,
+                    "feature_uid": fuid,
+                },
+            )
+            uids.append(fuid)
+
+            # Get feature data from lookup instead of DataFrame filtering
+            feature_data = features_lookup.get(fuid)
+            if feature_data:
+                feature_data_list.append(feature_data)
+
+        if not feature_data_list:
+            # Skip this consensus feature if no valid features found
+            continue
+
+        # Compute statistics using vectorized operations on collected data
+        # Convert to numpy arrays for faster computation
+        rt_values = np.array(
+            [fd.get("rt", 0) for fd in feature_data_list if fd.get("rt") is not None],
+        )
+        mz_values = np.array(
+            [fd.get("mz", 0) for fd in feature_data_list if fd.get("mz") is not None],
+        )
+        rt_start_values = np.array(
+            [
+                fd.get("rt_start", 0)
+                for fd in feature_data_list
+                if fd.get("rt_start") is not None
+            ],
+        )
+        rt_end_values = np.array(
+            [
+                fd.get("rt_end", 0)
+                for fd in feature_data_list
+                if fd.get("rt_end") is not None
+            ],
+        )
+        rt_delta_values = np.array(
+            [
+                fd.get("rt_delta", 0)
+                for fd in feature_data_list
+                if fd.get("rt_delta") is not None
+            ],
+        )
+        mz_start_values = np.array(
+            [
+                fd.get("mz_start", 0)
+                for fd in feature_data_list
+                if fd.get("mz_start") is not None
+            ],
+        )
+        mz_end_values = np.array(
+            [
+                fd.get("mz_end", 0)
+                for fd in feature_data_list
+                if fd.get("mz_end") is not None
+            ],
+        )
+        inty_values = np.array(
+            [
+                fd.get("inty", 0)
+                for fd in feature_data_list
+                if fd.get("inty") is not None
+            ],
+        )
+        coherence_values = np.array(
+            [
+                fd.get("chrom_coherence", 0)
+                for fd in feature_data_list
+                if fd.get("chrom_coherence") is not None
+            ],
+        )
+        prominence_values = np.array(
+            [
+                fd.get("chrom_prominence", 0)
+                for fd in feature_data_list
+                if fd.get("chrom_prominence") is not None
+            ],
+        )
+        prominence_scaled_values = np.array(
+            [
+                fd.get("chrom_prominence_scaled", 0)
+                for fd in feature_data_list
+                if fd.get("chrom_prominence_scaled") is not None
+            ],
+        )
+        height_scaled_values = np.array(
+            [
+                fd.get("chrom_height_scaled", 0)
+                for fd in feature_data_list
+                if fd.get("chrom_height_scaled") is not None
+            ],
+        )
+        iso_values = np.array(
+            [fd.get("iso", 0) for fd in feature_data_list if fd.get("iso") is not None],
+        )
+        charge_values = np.array(
+            [
+                fd.get("charge", 0)
+                for fd in feature_data_list
+                if fd.get("charge") is not None
+            ],
+        )
+
+        # adduct_values
+        # Collect all adducts from feature_data_list to create consensus adduct information
+        # Only consider adducts that are in study._get_adducts() plus items with '?'
+        all_adducts = []
+        adduct_masses = {}
+
+        # Get valid adducts from cached result (avoid repeated _get_adducts() calls)
+        valid_adducts = cached_valid_adducts if cached_valid_adducts is not None else set()
+        valid_adducts.add("?")  # Always allow '?' adducts
+
+        for fd in feature_data_list:
+            # Get individual adduct and mass from each feature data (fd)
+            adduct = fd.get("adduct")
+            adduct_mass = fd.get("adduct_mass")
+
+            if adduct is not None:
+                # Only include adducts that are valid (from study._get_adducts() or contain '?')
+                if adduct in valid_adducts or "?" in adduct:
+                    all_adducts.append(adduct)
+                    if adduct_mass is not None:
+                        adduct_masses[adduct] = adduct_mass
+
+        # Calculate adduct_values for the consensus feature
+        adduct_values = []
+        if all_adducts:
+            adduct_counts = {
+                adduct: all_adducts.count(adduct) for adduct in set(all_adducts)
+            }
+            total_count = sum(adduct_counts.values())
+            for adduct, count in adduct_counts.items():
+                percentage = (count / total_count) * 100 if total_count > 0 else 0
+                # Store as list with [name, num, %] format for the adducts column
+                adduct_values.append(
+                    [
+                        str(adduct),
+                        int(count),
+                        float(round(percentage, 2)),
+                    ],
+                )
+
+        # Sort adduct_values by count in descending order
+        adduct_values.sort(key=lambda x: x[1], reverse=True)  # Sort by count (index 1)
+        # Store adduct_values for use in metadata
+        consensus_adduct_values = adduct_values
+
+        # Extract top adduct information for new columns
+        adduct_top = None
+        adduct_charge_top = None
+        adduct_mass_neutral_top = None
+        adduct_mass_shift_top = None
+
+        if consensus_adduct_values:
+            top_adduct_name = consensus_adduct_values[0][0]  # Get top adduct name
+            adduct_top = top_adduct_name
+
+            # Parse adduct information to extract charge and mass shift
+            # Handle "?" as "H" and parse common adduct formats
+            if top_adduct_name == "?" or top_adduct_name == "[M+?]+":
+                adduct_charge_top = 1
+                adduct_mass_shift_top = 1.007825  # H mass
+            elif top_adduct_name == "[M+?]-":
+                adduct_charge_top = -1
+                adduct_mass_shift_top = -1.007825  # -H mass
+            else:
+                # Try to get charge and mass shift from cached study adducts
+                adduct_found = False
+                if cached_adducts_df is not None and not cached_adducts_df.is_empty():
+                    # Look for exact match in study adducts
+                    matching_adduct = cached_adducts_df.filter(
+                        pl.col("name") == top_adduct_name,
+                    )
+                    if not matching_adduct.is_empty():
+                        adduct_row = matching_adduct.row(0, named=True)
+                        adduct_charge_top = adduct_row["charge"]
+                        adduct_mass_shift_top = adduct_row["mass_shift"]
+                        adduct_found = True
+
+                if not adduct_found:
+                    # Fallback to regex parsing
+                    import re
+
+                    # Pattern for adducts like [M+H]+, [M-H]-, [M+Na]+, etc.
+                    pattern = r"\[M([+\-])([A-Za-z0-9]+)\]([0-9]*)([+\-])"
+                    match = re.match(pattern, top_adduct_name)
+
+                    if match:
+                        sign = match.group(1)
+                        element = match.group(2)
+                        multiplier_str = match.group(3)
+                        charge_sign = match.group(4)
+
+                        multiplier = int(multiplier_str) if multiplier_str else 1
+                        charge = multiplier if charge_sign == "+" else -multiplier
+                        adduct_charge_top = charge
+
+                        # Calculate mass shift based on element
+                        element_masses = {
+                            "H": 1.007825,
+                            "Na": 22.989769,
+                            "K": 38.963708,
+                            "NH4": 18.033823,
+                            "Li": 7.016930,
+                            "Cl": 34.969401,
+                            "Br": 78.918885,
+                            "HCOO": 44.998201,
+                            "CH3COO": 59.013851,
+                            "H2O": 18.010565,
+                        }
+
+                        base_mass = element_masses.get(
+                            element,
+                            1.007825,
+                        )  # Default to H if unknown
+                        mass_shift = (
+                            base_mass * multiplier
+                            if sign == "+"
+                            else -base_mass * multiplier
+                        )
+                        adduct_mass_shift_top = mass_shift
+                    else:
+                        # Default fallback
+                        adduct_charge_top = 1
+                        adduct_mass_shift_top = 1.007825
+        else:
+            # No valid adducts found - assign default based on study polarity
+            study_polarity = getattr(self, "polarity", "positive")
+            if study_polarity in ["negative", "neg"]:
+                # Negative mode default
+                adduct_top = "[M-?]1-"
+                adduct_charge_top = -1
+                adduct_mass_shift_top = -1.007825  # -H mass (loss of proton)
+            else:
+                # Positive mode default (includes 'positive', 'pos', or any other value)
+                adduct_top = "[M+?]1+"
+                adduct_charge_top = 1
+                adduct_mass_shift_top = 1.007825  # H mass (gain of proton)
+
+            # Create a single default adduct entry in the adducts list for consistency
+            consensus_adduct_values = [[adduct_top, 1, 100.0]]
+
+        # Calculate neutral mass from consensus mz (for both cases)
+        consensus_mz = (
+            round(float(np.mean(mz_values)), 4) if len(mz_values) > 0 else 0.0
+        )
+        if adduct_charge_top and adduct_mass_shift_top is not None:
+            adduct_mass_neutral_top = (
+                consensus_mz * abs(adduct_charge_top) - adduct_mass_shift_top
+            )
+
+        # Calculate number of MS2 spectra
+        ms2_count = 0
+        for fd in feature_data_list:
+            ms2_scans = fd.get("ms2_scans")
+            if ms2_scans is not None:
+                ms2_count += len(ms2_scans)
+
+        metadata_list.append(
+            {
+                "consensus_uid": int(i),  # "consensus_id": i,
+                "consensus_id": str(feature.getUniqueId()),
+                "quality": round(float(feature.getQuality()), 3),
+                "number_samples": len(feature_data_list),
+                # "number_ext": int(len(features_list)),
+                "rt": round(float(np.mean(rt_values)), 4)
+                if len(rt_values) > 0
+                else 0.0,
+                "mz": round(float(np.mean(mz_values)), 4)
+                if len(mz_values) > 0
+                else 0.0,
+                "rt_min": round(float(np.min(rt_values)), 3)
+                if len(rt_values) > 0
+                else 0.0,
+                "rt_max": round(float(np.max(rt_values)), 3)
+                if len(rt_values) > 0
+                else 0.0,
+                "rt_mean": round(float(np.mean(rt_values)), 3)
+                if len(rt_values) > 0
+                else 0.0,
+                "rt_start_mean": round(float(np.mean(rt_start_values)), 3)
+                if len(rt_start_values) > 0
+                else 0.0,
+                "rt_end_mean": round(float(np.mean(rt_end_values)), 3)
+                if len(rt_end_values) > 0
+                else 0.0,
+                "rt_delta_mean": round(float(np.ptp(rt_delta_values)), 3)
+                if len(rt_delta_values) > 0
+                else 0.0,
+                "mz_min": round(float(np.min(mz_values)), 4)
+                if len(mz_values) > 0
+                else 0.0,
+                "mz_max": round(float(np.max(mz_values)), 4)
+                if len(mz_values) > 0
+                else 0.0,
+                "mz_mean": round(float(np.mean(mz_values)), 4)
+                if len(mz_values) > 0
+                else 0.0,
+                "mz_start_mean": round(float(np.mean(mz_start_values)), 4)
+                if len(mz_start_values) > 0
+                else 0.0,
+                "mz_end_mean": round(float(np.mean(mz_end_values)), 4)
+                if len(mz_end_values) > 0
+                else 0.0,
+                "inty_mean": round(float(np.mean(inty_values)), 0)
+                if len(inty_values) > 0
+                else 0.0,
+                "bl": -1.0,
+                "chrom_coherence_mean": round(float(np.mean(coherence_values)), 3)
+                if len(coherence_values) > 0
+                else 0.0,
+                "chrom_prominence_mean": round(float(np.mean(prominence_values)), 0)
+                if len(prominence_values) > 0
+                else 0.0,
+                "chrom_prominence_scaled_mean": round(
+                    float(np.mean(prominence_scaled_values)),
+                    3,
+                )
+                if len(prominence_scaled_values) > 0
+                else 0.0,
+                "chrom_height_scaled_mean": round(
+                    float(np.mean(height_scaled_values)),
+                    3,
+                )
+                if len(height_scaled_values) > 0
+                else 0.0,
+                "iso_mean": round(float(np.mean(iso_values)), 2)
+                if len(iso_values) > 0
+                else 0.0,
+                "charge_mean": round(float(np.mean(charge_values)), 2)
+                if len(charge_values) > 0
+                else 0.0,
+                "number_ms2": int(ms2_count),
+                "adducts": consensus_adduct_values
+                if consensus_adduct_values
+                else [],  # Ensure it's always a list
+                # New columns for top-ranked adduct information
+                "adduct_top": adduct_top,
+                "adduct_charge_top": adduct_charge_top,
+                "adduct_mass_neutral_top": round(adduct_mass_neutral_top, 6)
+                if adduct_mass_neutral_top is not None
+                else None,
+                "adduct_mass_shift_top": round(adduct_mass_shift_top, 6)
+                if adduct_mass_shift_top is not None
+                else None,
+                # New columns for top-scoring identification results
+                "id_top_name": None,
+                "id_top_class": None,
+                "id_top_adduct": None,
+                "id_top_score": None,
+            },
+        )
+
+    consensus_mapping_df = pl.DataFrame(consensus_mapping)
+    # remove all rows in consensus_mapping_df where consensus_id is not in self.featured_df['uid']
+    l1 = len(consensus_mapping_df)
+    consensus_mapping_df = consensus_mapping_df.filter(
+        pl.col("feature_uid").is_in(self.features_df["feature_uid"].to_list()),
+    )
+    self.logger.debug(
+        f"Filtered {l1 - len(consensus_mapping_df)} orphan features from maps.",
+    )
+    self.consensus_mapping_df = consensus_mapping_df
+    self.consensus_df = pl.DataFrame(metadata_list, strict=False)
+
+    if min_samples is None:
+        min_samples = 1
+    if min_samples < 1:
+        min_samples = int(min_samples * len(self.samples_df))
+
+    # Validate that min_samples doesn't exceed the number of samples
+    if min_samples > len(self.samples_df):
+        self.logger.warning(
+            f"min_samples ({min_samples}) exceeds the number of samples ({len(self.samples_df)}). "
+            f"Setting min_samples to {len(self.samples_df)}.",
+        )
+        min_samples = len(self.samples_df)
+
+    # filter out consensus features with less than min_samples features
+    l1 = len(self.consensus_df)
+    self.consensus_df = self.consensus_df.filter(
+        pl.col("number_samples") >= min_samples,
+    )
+    self.logger.debug(
+        f"Filtered {l1 - len(self.consensus_df)} consensus features with less than {min_samples} samples.",
+    )
+    # filter out consensus mapping with less than min_samples features
+    self.consensus_mapping_df = self.consensus_mapping_df.filter(
+        pl.col("consensus_uid").is_in(self.consensus_df["consensus_uid"].to_list()),
+    )
+
+    self.consensus_map = consensus_map
+
+
+def _perform_adduct_grouping(self, rt_tol, mz_tol):
+    """Perform adduct grouping on consensus features."""
+    import polars as pl
+    
+    # Add adduct grouping and adduct_of assignment
+    if len(self.consensus_df) > 0:
+        # Get relevant columns for grouping
+        consensus_data = []
+        for row in self.consensus_df.iter_rows(named=True):
+            consensus_data.append(
+                {
+                    "consensus_uid": row["consensus_uid"],
+                    "rt": row["rt"],
+                    "adduct_mass_neutral_top": row.get("adduct_mass_neutral_top"),
+                    "adduct_top": row.get("adduct_top"),
+                    "inty_mean": row.get("inty_mean", 0),
+                },
+            )
+
+        # Use optimized adduct grouping
+        adduct_group_list, adduct_of_list = _optimized_adduct_grouping(
+            self, consensus_data, rt_tol, mz_tol
+        )
+
+        # Add the new columns to consensus_df
+        self.consensus_df = self.consensus_df.with_columns(
+            [
+                pl.Series("adduct_group", adduct_group_list, dtype=pl.Int64),
+                pl.Series("adduct_of", adduct_of_list, dtype=pl.Int64),
+            ],
+        )
+
+
+def _finalize_merge(self, link_ms2, min_samples):
+    """Complete the merge process with final calculations and cleanup."""
+    import polars as pl
+    
+    # Check if consensus_df is empty or missing required columns
+    if len(self.consensus_df) == 0 or "number_samples" not in self.consensus_df.columns:
+        self.logger.debug("No consensus features found or consensus_df is empty. Skipping finalize merge.")
+        return
+    
+    # Validate min_samples parameter
+    if min_samples is None:
+        min_samples = 1
+    if min_samples < 1:
+        min_samples = int(min_samples * len(self.samples_df))
+
+    # Validate that min_samples doesn't exceed the number of samples
+    if min_samples > len(self.samples_df):
+        self.logger.warning(
+            f"min_samples ({min_samples}) exceeds the number of samples ({len(self.samples_df)}). "
+            f"Setting min_samples to {len(self.samples_df)}.",
+        )
+        min_samples = len(self.samples_df)
+
+    # Filter out consensus features with less than min_samples features
+    l1 = len(self.consensus_df)
+    self.consensus_df = self.consensus_df.filter(
+        pl.col("number_samples") >= min_samples,
+    )
+    self.logger.debug(
+        f"Filtered {l1 - len(self.consensus_df)} consensus features with less than {min_samples} samples.",
+    )
+    
+    # Filter out consensus mapping with less than min_samples features
+    self.consensus_mapping_df = self.consensus_mapping_df.filter(
+        pl.col("consensus_uid").is_in(self.consensus_df["consensus_uid"].to_list()),
+    )
+
+    # Calculate the completeness of the consensus map
+    if len(self.consensus_df) > 0 and len(self.samples_df) > 0:
+        c = (
+            len(self.consensus_mapping_df)
+            / len(self.consensus_df)
+            / len(self.samples_df)
+        )
+        self.logger.info(
+            f"Merging completed. Consensus features: {len(self.consensus_df)}. Completeness: {c:.2f}.",
+        )
+    else:
+        self.logger.warning(
+            f"Merging completed with empty result. Consensus features: {len(self.consensus_df)}. "
+            f"This may be due to min_samples ({min_samples}) being too high for the available data.",
+        )
+    
+    if link_ms2:
+        self.find_ms2()
+
+
+def _optimized_feature_lookup(study_obj, features_df):
+    """
+    Optimized feature lookup creation using Polars operations.
+    """
+    study_obj.logger.debug("Creating optimized feature lookup...")
+    start_time = time.time()
+    
+    # Use Polars select for faster conversion
+    feature_columns = [
+        "feature_uid", "rt", "mz", "rt_start", "rt_end", "rt_delta", 
+        "mz_start", "mz_end", "inty", "chrom_coherence", "chrom_prominence", 
+        "chrom_prominence_scaled", "chrom_height_scaled", "iso", "charge", 
+        "ms2_scans", "adduct", "adduct_mass"
+    ]
+    
+    # Filter to only existing columns
+    existing_columns = [col for col in feature_columns if col in features_df.columns]
+    
+    # Convert to dictionary more efficiently
+    selected_df = features_df.select(existing_columns)
+    
+    features_lookup = {}
+    for row in selected_df.iter_rows(named=True):
+        feature_uid = row["feature_uid"]
+        # Keep feature_uid in the dictionary for chunked merge compatibility
+        features_lookup[feature_uid] = {k: v for k, v in row.items()}
+    
+    lookup_time = time.time() - start_time
+    if len(features_lookup) > 50000:
+        study_obj.logger.debug(f"Feature lookup created in {lookup_time:.2f}s for {len(features_lookup)} features")
+    return features_lookup
+
+
+def _optimized_adduct_grouping(study_obj, consensus_data, rt_tol, mz_tol):
+    """
+    Optimized O(n log n) adduct grouping using spatial indexing.
+    
+    Args:
+        study_obj: Study object with logger
+        consensus_data: List of consensus feature dictionaries
+        rt_tol: RT tolerance in minutes
+        mz_tol: m/z tolerance in Da
+        
+    Returns:
+        Tuple of (adduct_group_list, adduct_of_list)
+    """
+    if not consensus_data:
+        return [], []
+    
+    n_features = len(consensus_data)
+    if n_features > 10000:
+        study_obj.logger.info(f"Adduct grouping for {n_features} consensus features...")
+    else:
+        study_obj.logger.debug(f"Adduct grouping for {n_features} consensus features...")
+    
+    # Build spatial index using RT and neutral mass as coordinates
+    features_by_mass = defaultdict(list)
+    mass_bin_size = mz_tol * 2  # 2x tolerance for conservative binning
+    
+    valid_features = []
+    for feature in consensus_data:
+        consensus_uid = feature["consensus_uid"]
+        rt = feature["rt"]
+        neutral_mass = feature.get("adduct_mass_neutral_top")
+        intensity = feature.get("inty_mean", 0)
+        adduct = feature.get("adduct_top", "")
+        
+        if neutral_mass is not None:
+            mass_bin = int(neutral_mass / mass_bin_size)
+            features_by_mass[mass_bin].append((consensus_uid, rt, neutral_mass, intensity, adduct))
+            valid_features.append((consensus_uid, rt, neutral_mass, intensity, adduct, mass_bin))
+    
+    # Union-Find for efficient grouping
+    class UnionFind:
+        def __init__(self, n):
+            self.parent = list(range(n))
+            self.rank = [0] * n
+        
+        def find(self, x):
+            if self.parent[x] != x:
+                self.parent[x] = self.find(self.parent[x])
+            return self.parent[x]
+        
+        def union(self, x, y):
+            px, py = self.find(x), self.find(y)
+            if px == py:
+                return
+            if self.rank[px] < self.rank[py]:
+                px, py = py, px
+            self.parent[py] = px
+            if self.rank[px] == self.rank[py]:
+                self.rank[px] += 1
+    
+    uid_to_idx = {feature[0]: i for i, feature in enumerate(valid_features)}
+    uf = UnionFind(len(valid_features))
+    
+    # Find groups using spatial index
+    checked_pairs = set()
+    for i, (uid1, rt1, mass1, inty1, adduct1, bin1) in enumerate(valid_features):
+        for bin_offset in [-1, 0, 1]:
+            check_bin = bin1 + bin_offset
+            if check_bin not in features_by_mass:
+                continue
+                
+            for uid2, rt2, mass2, inty2, adduct2 in features_by_mass[check_bin]:
+                if uid1 >= uid2:
+                    continue
+                
+                pair = (min(uid1, uid2), max(uid1, uid2))
+                if pair in checked_pairs:
+                    continue
+                checked_pairs.add(pair)
+                
+                mass_diff = abs(mass1 - mass2)
+                rt_diff = abs(rt1 - rt2) / 60.0  # Convert to minutes
+                
+                if mass_diff <= mz_tol and rt_diff <= rt_tol:
+                    j = uid_to_idx[uid2]
+                    uf.union(i, j)
+    
+    # Extract groups
+    groups_by_root = defaultdict(list)
+    for i, (uid, rt, mass, inty, adduct, _) in enumerate(valid_features):
+        root = uf.find(i)
+        groups_by_root[root].append((uid, rt, mass, inty, adduct))
+    
+    groups = {}
+    group_id = 1
+    assigned_groups = {}
+    
+    for group_members in groups_by_root.values():
+        member_uids = [uid for uid, _, _, _, _ in group_members]
+        
+        for uid in member_uids:
+            assigned_groups[uid] = group_id
+        groups[group_id] = member_uids
+        group_id += 1
+    
+    # Handle features without neutral mass
+    for feature in consensus_data:
+        uid = feature["consensus_uid"]
+        if uid not in assigned_groups:
+            assigned_groups[uid] = group_id
+            groups[group_id] = [uid]
+            group_id += 1
+    
+    # Determine adduct_of for each group
+    group_adduct_of = {}
+    for grp_id, member_uids in groups.items():
+        best_uid = None
+        best_priority = -1
+        best_intensity = 0
+        
+        for uid in member_uids:
+            feature_data = next((f for f in consensus_data if f["consensus_uid"] == uid), None)
+            if not feature_data:
+                continue
+                
+            adduct = feature_data.get("adduct_top", "")
+            intensity = feature_data.get("inty_mean", 0)
+            
+            priority = 0
+            if adduct and ("[M+H]" in adduct or adduct == "H" or adduct == "?"):
+                priority = 3
+            elif adduct and "[M-H]" in adduct:
+                priority = 2
+            elif adduct and "M" in adduct:
+                priority = 1
+            
+            if priority > best_priority or (priority == best_priority and intensity > best_intensity):
+                best_uid = uid
+                best_priority = priority
+                best_intensity = intensity
+        
+        group_adduct_of[grp_id] = best_uid if best_uid else member_uids[0]
+    
+    # Build final lists in same order as consensus_data
+    adduct_group_list = []
+    adduct_of_list = []
+    
+    for feature in consensus_data:
+        uid = feature["consensus_uid"]
+        group = assigned_groups.get(uid, 0)
+        adduct_of = group_adduct_of.get(group, uid)
+        
+        adduct_group_list.append(group)
+        adduct_of_list.append(adduct_of)
+    
+    if n_features > 10000:
+        study_obj.logger.info("Adduct grouping completed.")
+    else:
+        study_obj.logger.debug("Adduct grouping completed.")
+
+    return adduct_group_list, adduct_of_list