M3Drop 0.4.36__tar.gz → 0.4.37__tar.gz

{m3drop-0.4.36 → m3drop-0.4.37/M3Drop.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: M3Drop
-Version: 0.4.36
+Version: 0.4.37
 Summary: A Python implementation of the M3Drop single-cell RNA-seq analysis tool.
 Home-page: https://github.com/PragalvhaSharma/m3DropNew
 Author: Tallulah Andrews

{m3drop-0.4.36/M3Drop.egg-info → m3drop-0.4.37}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: M3Drop
-Version: 0.4.36
+Version: 0.4.37
 Summary: A Python implementation of the M3Drop single-cell RNA-seq analysis tool.
 Home-page: https://github.com/PragalvhaSharma/m3DropNew
 Author: Tallulah Andrews

m3drop-0.4.37/m3Drop/diagnosticsGPU.py

@@ -0,0 +1,466 @@
+import numpy as np
+import pandas as pd
+import cupy as cp
+import cupyx.scipy.sparse as csp
+import matplotlib.pyplot as plt
+import h5py
+import os
+import time
+import psutil
+import gc
+from scipy import sparse
+from scipy import stats
+import anndata
+
+# [GOVERNOR INTEGRATION]
+from .coreGPU import hidden_calc_valsGPU, NBumiFitModelGPU, NBumiFitDispVsMeanGPU, get_optimal_chunk_size
+from cupy.sparse import csr_matrix as cp_csr_matrix
+import scipy.sparse as sp
+from scipy.sparse import csr_matrix as sp_csr_matrix
+import statsmodels.api as sm
+
+def NBumiFitBasicModelGPU(
+    cleaned_filename: str,
+    stats: dict,
+    is_logged=False,
+    chunk_size: int = None 
+) -> dict:
+    """
+    Fits a simpler, unadjusted NB model out-of-core using a GPU-accelerated
+    algorithm. Designed to work with a standard (cell, gene) sparse matrix.
+    """
+    start_time = time.perf_counter()
+    print(f"FUNCTION: NBumiFitBasicModel() | FILE: {cleaned_filename}")
+
+    # [GOVERNOR INTEGRATION] Calculate optimal chunk size if not provided
+    if chunk_size is None:
+        chunk_size = get_optimal_chunk_size(cleaned_filename, multiplier=3.0, is_dense=True)
+
+    # --- Phase 1: Initialization ---
+    print("Phase [1/2]: Initializing parameters and arrays on GPU...")
+    tjs = stats['tjs'].values
+    nc, ng = stats['nc'], stats['ng']
+
+    tjs_gpu = cp.asarray(tjs, dtype=cp.float64)
+    sum_x_sq_gpu = cp.zeros(ng, dtype=cp.float64)
+    print("Phase [1/2]: COMPLETE")
+
+    # --- Phase 2: Calculate Variance from Data Chunks ---
+    print("Phase [2/2]: Calculating variance from data chunks...")
+    with h5py.File(cleaned_filename, 'r') as f_in:
+        x_group = f_in['X']
+        h5_indptr = x_group['indptr']
+        h5_data = x_group['data']
+        h5_indices = x_group['indices']
+
+        for i in range(0, nc, chunk_size):
+            end_row = min(i + chunk_size, nc)
+            print(f"Phase [2/2]: Processing: {end_row} of {nc} cells.", end='\r')
+
+            start_idx, end_idx = h5_indptr[i], h5_indptr[end_row]
+            if start_idx == end_idx:
+                continue
+            
+            # Process in smaller sub-chunks if needed
+            max_elements = 5_000_000  # Process max 5M elements at a time
+            
+            if end_idx - start_idx > max_elements:
+                # Process in sub-chunks
+                for sub_start in range(start_idx, end_idx, max_elements):
+                    sub_end = min(sub_start + max_elements, end_idx)
+                    
+                    data_slice = h5_data[sub_start:sub_end]
+                    indices_slice = h5_indices[sub_start:sub_end]
+                    
+                    data_gpu = cp.asarray(data_slice, dtype=cp.float64)
+                    indices_gpu = cp.asarray(indices_slice)
+                    
+                    # Accumulate the sum of squares for each gene
+                    cp.add.at(sum_x_sq_gpu, indices_gpu, data_gpu**2)
+                    
+                    # Free GPU memory
+                    del data_gpu, indices_gpu
+                    cp.get_default_memory_pool().free_all_blocks()
+            else:
+                # Original processing for smaller chunks
+                data_slice = h5_data[start_idx:end_idx]
+                indices_slice = h5_indices[start_idx:end_idx]
+
+                data_gpu = cp.asarray(data_slice, dtype=cp.float64)
+                indices_gpu = cp.asarray(indices_slice)
+
+                # Accumulate the sum of squares for each gene
+                cp.add.at(sum_x_sq_gpu, indices_gpu, data_gpu**2)
+                
+                # Clean up
+                del data_gpu, indices_gpu
+                cp.get_default_memory_pool().free_all_blocks()
+    
+    print(f"Phase [2/2]: COMPLETE                                       ")
+
+    # --- Final calculations on GPU ---
+    if is_logged:
+        raise NotImplementedError("Logged data variance calculation is not implemented for out-of-core.")
+    else:
+        # Variance of raw data: Var(X) = E[X^2] - E[X]^2
+        mean_x_sq_gpu = sum_x_sq_gpu / nc
+        mean_mu_gpu = tjs_gpu / nc
+        my_rowvar_gpu = mean_x_sq_gpu - mean_mu_gpu**2
+        
+        # Calculate dispersion ('size')
+        size_gpu = mean_mu_gpu**2 / (my_rowvar_gpu - mean_mu_gpu)
+    
+    max_size_val = cp.nanmax(size_gpu) * 10
+    if cp.isnan(max_size_val): 
+        max_size_val = 1000
+    size_gpu[cp.isnan(size_gpu) | (size_gpu <= 0)] = max_size_val
+    size_gpu[size_gpu < 1e-10] = 1e-10
+    
+    # Move results to CPU
+    my_rowvar_cpu = my_rowvar_gpu.get()
+    sizes_cpu = size_gpu.get()
+
+    end_time = time.perf_counter()
+    print(f"Total time: {end_time - start_time:.2f} seconds.\n")
+
+    return {
+        'var_obs': pd.Series(my_rowvar_cpu, index=stats['tjs'].index),
+        'sizes': pd.Series(sizes_cpu, index=stats['tjs'].index),
+        'vals': stats
+    }
+
+def NBumiCheckFitFSGPU(
+    cleaned_filename: str,
+    fit: dict,
+    chunk_size: int = None,
+    suppress_plot=False,
+    plot_filename=None
+) -> dict:
+    """
+    Calculates the fit errors (gene_error, cell_error) for a given model.
+    """
+    start_time = time.perf_counter()
+    print(f"FUNCTION: NBumiCheckFitFS() | FILE: {cleaned_filename}")
+
+    # [GOVERNOR INTEGRATION] Adaptive chunk sizing
+    if chunk_size is None:
+        chunk_size = get_optimal_chunk_size(cleaned_filename, multiplier=5.0, is_dense=True)
+
+    # --- Phase 1: Initialization ---
+    print("Phase [1/2]: Initializing parameters and arrays on GPU...")
+    vals = fit['vals']
+    size_coeffs = NBumiFitDispVsMeanGPU(fit, suppress_plot=True)
+
+    # Must use float64 for precision
+    tjs_gpu = cp.asarray(vals['tjs'].values, dtype=cp.float64)
+    tis_gpu = cp.asarray(vals['tis'].values, dtype=cp.float64)
+    total = vals['total']
+    nc, ng = vals['nc'], vals['ng']
+
+    # Calculate smoothed size
+    mean_expression_gpu = tjs_gpu / nc
+    log_mean_expression_gpu = cp.log(mean_expression_gpu)
+    smoothed_size_gpu = cp.exp(size_coeffs[0] + size_coeffs[1] * log_mean_expression_gpu)
+
+    # Initialize result arrays
+    row_ps_gpu = cp.zeros(ng, dtype=cp.float64)
+    col_ps_gpu = cp.zeros(nc, dtype=cp.float64)
+    print("Phase [1/2]: COMPLETE")
+
+    # --- Phase 2: Calculate Expected Dropouts ---
+    print(f"Phase [2/2]: Calculating expected dropouts (Chunk: {chunk_size})...")
+    
+    for i in range(0, nc, chunk_size):
+        end_col = min(i + chunk_size, nc)
+        print(f"Phase [2/2]: Processing: {end_col} of {nc} cells.", end='\r')
+
+        tis_chunk_gpu = tis_gpu[i:end_col]
+
+        # Standard calculation without errstate
+        mu_chunk_gpu = tjs_gpu[:, cp.newaxis] * tis_chunk_gpu[cp.newaxis, :] / total
+        
+        # Calculate p_is directly - CuPy handles overflow internally
+        base = 1 + mu_chunk_gpu / smoothed_size_gpu[:, cp.newaxis]
+        p_is_chunk_gpu = cp.power(base, -smoothed_size_gpu[:, cp.newaxis])
+        
+        # Handle any inf/nan values that might have occurred
+        p_is_chunk_gpu = cp.nan_to_num(p_is_chunk_gpu, nan=0.0, posinf=1.0, neginf=0.0)
+        
+        # Sum results
+        row_ps_gpu += p_is_chunk_gpu.sum(axis=1)
+        col_ps_gpu[i:end_col] = p_is_chunk_gpu.sum(axis=0)
+        
+        # Clean up
+        del mu_chunk_gpu, p_is_chunk_gpu, base, tis_chunk_gpu
+        cp.get_default_memory_pool().free_all_blocks()
+
+    print(f"Phase [2/2]: COMPLETE{' ' * 50}")
+
+    # Move results to CPU
+    row_ps_cpu = row_ps_gpu.get()
+    col_ps_cpu = col_ps_gpu.get()
+    djs_cpu = vals['djs'].values
+    dis_cpu = vals['dis'].values
+
+    # Plotting
+    if not suppress_plot:
+        plt.figure(figsize=(12, 5))
+        plt.subplot(1, 2, 1)
+        plt.scatter(djs_cpu, row_ps_cpu, alpha=0.5, s=10)
+        plt.title("Gene-specific Dropouts (Smoothed)")
+        plt.xlabel("Observed")
+        plt.ylabel("Fit")
+        lims = [min(plt.xlim()[0], plt.ylim()[0]), max(plt.xlim()[1], plt.ylim()[1])]
+        plt.plot(lims, lims, 'r-', alpha=0.75, zorder=0, label="y=x line")
+        plt.grid(True); plt.legend()
+
+        plt.subplot(1, 2, 2)
+        plt.scatter(dis_cpu, col_ps_cpu, alpha=0.5, s=10)
+        plt.title("Cell-specific Dropouts (Smoothed)")
+        plt.xlabel("Observed")
+        plt.ylabel("Expected")
+        lims = [min(plt.xlim()[0], plt.ylim()[0]), max(plt.xlim()[1], plt.ylim()[1])]
+        plt.plot(lims, lims, 'r-', alpha=0.75, zorder=0, label="y=x line")
+        plt.grid(True); plt.legend()
+        
+        plt.tight_layout()
+        if plot_filename:
+            plt.savefig(plot_filename, dpi=300, bbox_inches='tight')
+            print(f"STATUS: Diagnostic plot saved to '{plot_filename}'")
+        plt.show()
+        plt.close()
+
+    # Calculate errors
+    gene_error = np.sum((djs_cpu - row_ps_cpu)**2)
+    cell_error = np.sum((dis_cpu - col_ps_cpu)**2)
+    
+    end_time = time.perf_counter()
+    print(f"Total time: {end_time - start_time:.2f} seconds.\n")
+
+    return {
+        'gene_error': gene_error,
+        'cell_error': cell_error,
+        'rowPs': pd.Series(row_ps_cpu, index=fit['vals']['tjs'].index),
+        'colPs': pd.Series(col_ps_cpu, index=fit['vals']['tis'].index)
+    }
+
+def NBumiCompareModelsGPU(
+    raw_filename: str, # Kept for API compatibility, but functionally we use cleaned_filename for indices
+    cleaned_filename: str,
+    stats: dict,
+    fit_adjust: dict,
+    chunk_size: int = None,
+    suppress_plot=False,
+    plot_filename=None
+) -> dict:
+    """
+    OPTIMIZED VERSION (IN-MEMORY):
+    - Eliminates the 46GB '_basic_norm.h5ad' temporary file.
+    - Performs depth normalization and variance calculation on-the-fly in GPU VRAM.
+    - PRESERVED SCIENTIFIC LOGIC: Var(X) = E[X^2] - (E[X])^2 on normalized data.
+    """
+    pipeline_start_time = time.time()
+    print(f"FUNCTION: NBumiCompareModels() | Comparing models for {cleaned_filename}")
+
+    # [GOVERNOR] High multiplier (12.0) because we hold Raw + Norm + Square in VRAM
+    if chunk_size is None:
+        chunk_size = get_optimal_chunk_size(cleaned_filename, multiplier=12.0, is_dense=False)
+
+    # --- Phase 1: In-Memory "Basic Fit" (Normalization + Variance) ---
+    print("Phase [1/3]: Calculating Basic Model (Depth-Normalized) variance on-the-fly...")
+    
+    # 1. Prepare Size Factors (CPU)
+    tjs = stats['tjs'].values # Gene sums
+    tis = stats['tis'].values # Cell sums
+    nc, ng = stats['nc'], stats['ng']
+    
+    median_sum = np.median(tis[tis > 0])
+    size_factors = np.ones_like(tis, dtype=np.float32)
+    non_zero_mask = tis > 0
+    size_factors[non_zero_mask] = tis[non_zero_mask] / median_sum
+    
+    # 2. Prepare GPU Arrays
+    sum_x_sq_gpu = cp.zeros(ng, dtype=cp.float64)
+    sum_x_gpu = cp.zeros(ng, dtype=cp.float64) # Need sum(x) to calc mean(x) for variance
+    
+    # 3. GPU Loop (Raw Data -> Normalize -> Accumulate)
+    # CRITICAL: We read CLEANED_FILENAME to ensure indices match 'stats'
+    with h5py.File(cleaned_filename, 'r') as f_in:
+        h5_indptr = f_in['X']['indptr']
+        h5_data = f_in['X']['data']
+        h5_indices = f_in['X']['indices']
+
+        for i in range(0, nc, chunk_size):
+            end_row = min(i + chunk_size, nc)
+            print(f"Phase [1/3]: Processing: {end_row} of {nc} cells.", end='\r')
+            
+            start_idx, end_idx = h5_indptr[i], h5_indptr[end_row]
+            if start_idx == end_idx: continue
+
+            # Load Raw Chunk
+            data_gpu = cp.asarray(h5_data[start_idx:end_idx], dtype=cp.float32)
+            indices_gpu = cp.asarray(h5_indices[start_idx:end_idx])
+            indptr_gpu = cp.asarray(h5_indptr[i:end_row + 1] - start_idx)
+
+            # Expand Size Factors to match Data Structure
+            nnz_in_chunk = indptr_gpu[-1].item()
+            cell_boundary_markers = cp.zeros(nnz_in_chunk, dtype=cp.int32)
+            if len(indptr_gpu) > 1:
+                cell_boundary_markers[indptr_gpu[:-1]] = 1
+            # row_indices maps every data point to its cell index (0 to chunk_size)
+            row_indices = cp.cumsum(cell_boundary_markers, axis=0) - 1
+            
+            # Get size factors for this chunk
+            sf_chunk = cp.asarray(size_factors[i:end_row])
+            
+            # --- THE MAGIC: On-the-Fly Normalization ---
+            # data_norm = data_raw / size_factor
+            data_gpu /= sf_chunk[row_indices]
+            
+            # Accumulate for Variance: E[X^2] and E[X]
+            cp.add.at(sum_x_sq_gpu, indices_gpu, data_gpu**2)
+            cp.add.at(sum_x_gpu, indices_gpu, data_gpu)
+            
+            # Clean up VRAM
+            del data_gpu, indices_gpu, indptr_gpu, row_indices, sf_chunk, cell_boundary_markers
+            cp.get_default_memory_pool().free_all_blocks()
+
+    print(f"Phase [1/3]: COMPLETE{' '*50}")
+
+    # 4. Finalize Basic Statistics
+    # Var(X) = E[X^2] - (E[X])^2
+    mean_x_sq_gpu = sum_x_sq_gpu / nc
+    mean_mu_gpu = sum_x_gpu / nc
+    my_rowvar_gpu = mean_x_sq_gpu - mean_mu_gpu**2
+    
+    # Dispersion = Mean^2 / (Var - Mean)
+    size_gpu = mean_mu_gpu**2 / (my_rowvar_gpu - mean_mu_gpu)
+    
+    # Safety Clamping
+    max_size_val = cp.nanmax(size_gpu) * 10
+    if cp.isnan(max_size_val): max_size_val = 1000
+    size_gpu[cp.isnan(size_gpu) | (size_gpu <= 0)] = max_size_val
+    size_gpu[size_gpu < 1e-10] = 1e-10
+    
+    # Construct "Basic Fit" Object
+    fit_basic = {
+        'sizes': pd.Series(size_gpu.get(), index=stats['tjs'].index),
+        'vals': stats,
+        'var_obs': pd.Series(my_rowvar_gpu.get(), index=stats['tjs'].index)
+    }
+    
+    # --- Phase 2: Check Fit (Calculate Errors) ---
+    print("Phase [2/3]: Evaluating fit errors on ORIGINAL data...")
+    
+    # Check Adjust (M3Drop) - uses its own governor
+    check_adjust = NBumiCheckFitFSGPU(
+        cleaned_filename, fit_adjust, suppress_plot=True
+    )
+    
+    # Check Basic (Depth-Norm) - uses its own governor
+    check_basic = NBumiCheckFitFSGPU(
+        cleaned_filename, fit_basic, suppress_plot=True
+    )
+    print("Phase [2/3]: COMPLETE")
+
+    # --- Phase 3: Plotting & Comparison ---
+    print("Phase [3/3]: Generating comparison...")
+    
+    nc_data = stats['nc']
+    mean_expr = stats['tjs'] / nc_data
+    observed_dropout = stats['djs'] / nc_data
+    
+    adj_dropout_fit = check_adjust['rowPs'] / nc_data
+    bas_dropout_fit = check_basic['rowPs'] / nc_data
+    
+    err_adj = np.sum(np.abs(adj_dropout_fit - observed_dropout))
+    err_bas = np.sum(np.abs(bas_dropout_fit - observed_dropout))
+    
+    comparison_df = pd.DataFrame({
+        'mean_expr': mean_expr,
+        'observed': observed_dropout,
+        'adj_fit': adj_dropout_fit,
+        'bas_fit': bas_dropout_fit
+    })
+    
+    plt.figure(figsize=(10, 6))
+    sorted_idx = np.argsort(mean_expr.values)
+    
+    plt.scatter(mean_expr.iloc[sorted_idx], observed_dropout.iloc[sorted_idx], 
+                c='black', s=3, alpha=0.5, label='Observed')
+    plt.scatter(mean_expr.iloc[sorted_idx], bas_dropout_fit.iloc[sorted_idx], 
+                c='purple', s=3, alpha=0.6, label=f'Basic Fit (Error: {err_bas:.2f})')
+    plt.scatter(mean_expr.iloc[sorted_idx], adj_dropout_fit.iloc[sorted_idx], 
+                c='goldenrod', s=3, alpha=0.7, label=f'Depth-Adjusted Fit (Error: {err_adj:.2f})')
+    
+    plt.xscale('log')
+    plt.xlabel("Mean Expression")
+    plt.ylabel("Dropout Rate")
+    plt.title("M3Drop Model Comparison")
+    plt.legend()
+    plt.grid(True, linestyle='--', alpha=0.3)
+
+    if plot_filename:
+        plt.savefig(plot_filename, dpi=300, bbox_inches='tight')
+        print(f"STATUS: Model comparison plot saved to '{plot_filename}'")
+
+    if not suppress_plot:
+        plt.show()
+    
+    plt.close()
+    
+    pipeline_end_time = time.time()
+    print(f"Total time: {pipeline_end_time - pipeline_start_time:.2f} seconds.\n")
+    
+    return {
+        "errors": {"Depth-Adjusted": err_adj, "Basic": err_bas},
+        "comparison_df": comparison_df
+    }
+
+def NBumiPlotDispVsMeanGPU(
+    fit: dict,
+    suppress_plot: bool = False,
+    plot_filename: str = None
+):
+    """
+    Generates a diagnostic plot of the dispersion vs. mean expression.
+    """
+    print("FUNCTION: NBumiPlotDispVsMean()")
+
+    # --- 1. Extract data and regression coefficients ---
+    mean_expression = fit['vals']['tjs'].values / fit['vals']['nc']
+    sizes = fit['sizes'].values
+    coeffs = NBumiFitDispVsMeanGPU(fit, suppress_plot=True)
+    intercept, slope = coeffs[0], coeffs[1]
+
+    # --- 2. Calculate the fitted line for plotting ---
+    log_mean_expr_range = np.linspace(
+        np.log(mean_expression[mean_expression > 0].min()),
+        np.log(mean_expression.max()),
+        100
+    )
+    log_fitted_sizes = intercept + slope * log_mean_expr_range
+    fitted_sizes = np.exp(log_fitted_sizes)
+
+    # --- 3. Create the plot ---
+    plt.figure(figsize=(8, 6))
+    plt.scatter(mean_expression, sizes, label='Observed Dispersion', alpha=0.5, s=8)
+    plt.plot(np.exp(log_mean_expr_range), fitted_sizes, color='red', label='Regression Fit', linewidth=2)
+
+    plt.xscale('log')
+    plt.yscale('log')
+    plt.xlabel('Mean Expression')
+    plt.ylabel('Dispersion Parameter (Sizes)')
+    plt.title('Dispersion vs. Mean Expression')
+    plt.legend()
+    plt.grid(True, which="both", linestyle='--', alpha=0.6)
+
+    if plot_filename:
+        plt.savefig(plot_filename, dpi=300, bbox_inches='tight')
+        print(f"STATUS: Diagnostic plot saved to '{plot_filename}'")
+
+    if not suppress_plot:
+        plt.show()
+
+    plt.close()
+    print("FUNCTION: NBumiPlotDispVsMean() COMPLETE\n")

{m3drop-0.4.36 → m3drop-0.4.37}/setup.py

@@ -5,7 +5,7 @@ with open("README.md", "r", encoding="utf-8") as fh:
 
 setuptools.setup(
     name="M3Drop",
-    version="0.4.36",  # Version bump
+    version="0.4.37",  # Version bump
     author="Tallulah Andrews",
     author_email="tandrew6@uwo.ca",
     description="A Python implementation of the M3Drop single-cell RNA-seq analysis tool.",

m3drop-0.4.36/m3Drop/diagnosticsGPU.py

@@ -1,171 +0,0 @@
-def NBumiCompareModelsGPU(
-    raw_filename: str,
-    cleaned_filename: str,
-    stats: dict,
-    fit_adjust: dict,
-    chunk_size: int = None,
-    suppress_plot=False,
-    plot_filename=None
-) -> dict:
-    """
-    OPTIMIZED VERSION (IN-MEMORY):
-    - Eliminates the 46GB '_basic_norm.h5ad' temporary file.
-    - Performs depth normalization and variance calculation on-the-fly in GPU VRAM.
-    """
-    pipeline_start_time = time.time()
-    print(f"FUNCTION: NBumiCompareModels() | Comparing models for {cleaned_filename}")
-
-    # [GOVERNOR] High multiplier (12.0) because we hold Raw + Norm + Square in VRAM
-    if chunk_size is None:
-        chunk_size = get_optimal_chunk_size(raw_filename, multiplier=12.0, is_dense=False)
-
-    # --- Phase 1: In-Memory "Basic Fit" (Normalization + Variance) ---
-    print("Phase [1/3]: Calculating Basic Model (Depth-Normalized) variance on-the-fly...")
-    
-    # 1. Prepare Size Factors (CPU)
-    tjs = stats['tjs'].values # Gene sums (needed for final dataframe)
-    tis = stats['tis'].values # Cell sums (needed for size factors)
-    nc, ng = stats['nc'], stats['ng']
-    
-    median_sum = np.median(tis[tis > 0])
-    size_factors = np.ones_like(tis, dtype=np.float32)
-    non_zero_mask = tis > 0
-    size_factors[non_zero_mask] = tis[non_zero_mask] / median_sum
-    
-    # 2. Prepare GPU Arrays
-    sum_x_sq_gpu = cp.zeros(ng, dtype=cp.float64)
-    sum_x_gpu = cp.zeros(ng, dtype=cp.float64) # Need sum(x) to calc mean(x) for variance
-    
-    # 3. GPU Loop (Raw Data -> Normalize -> Accumulate)
-    with h5py.File(raw_filename, 'r') as f_in:
-        h5_indptr = f_in['X']['indptr']
-        h5_data = f_in['X']['data']
-        h5_indices = f_in['X']['indices']
-
-        for i in range(0, nc, chunk_size):
-            end_row = min(i + chunk_size, nc)
-            print(f"Phase [1/3]: Processing: {end_row} of {nc} cells.", end='\r')
-            
-            start_idx, end_idx = h5_indptr[i], h5_indptr[end_row]
-            if start_idx == end_idx: continue
-
-            # Load Raw Chunk
-            data_gpu = cp.asarray(h5_data[start_idx:end_idx], dtype=cp.float32)
-            indices_gpu = cp.asarray(h5_indices[start_idx:end_idx])
-            indptr_gpu = cp.asarray(h5_indptr[i:end_row + 1] - start_idx)
-
-            # Expand Size Factors to match Data Structure
-            # (Map cell's size factor to every non-zero gene in that cell)
-            nnz_in_chunk = indptr_gpu[-1].item()
-            cell_boundary_markers = cp.zeros(nnz_in_chunk, dtype=cp.int32)
-            if len(indptr_gpu) > 1:
-                cell_boundary_markers[indptr_gpu[:-1]] = 1
-            # row_indices maps every data point to its cell index (0 to chunk_size)
-            row_indices = cp.cumsum(cell_boundary_markers, axis=0) - 1
-            
-            # Get size factors for this chunk
-            sf_chunk = cp.asarray(size_factors[i:end_row])
-            
-            # --- THE MAGIC: On-the-Fly Normalization ---
-            # data_norm = data_raw / size_factor
-            data_gpu /= sf_chunk[row_indices]
-            
-            # Accumulate for Variance: E[X^2] and E[X]
-            cp.add.at(sum_x_sq_gpu, indices_gpu, data_gpu**2)
-            cp.add.at(sum_x_gpu, indices_gpu, data_gpu)
-            
-            # Clean up VRAM
-            del data_gpu, indices_gpu, indptr_gpu, row_indices, sf_chunk, cell_boundary_markers
-            cp.get_default_memory_pool().free_all_blocks()
-
-    print(f"Phase [1/3]: COMPLETE{' '*50}")
-
-    # 4. Finalize Basic Statistics
-    # Var(X) = E[X^2] - (E[X])^2
-    mean_x_sq_gpu = sum_x_sq_gpu / nc
-    mean_mu_gpu = sum_x_gpu / nc
-    my_rowvar_gpu = mean_x_sq_gpu - mean_mu_gpu**2
-    
-    # Dispersion = Mean^2 / (Var - Mean)
-    size_gpu = mean_mu_gpu**2 / (my_rowvar_gpu - mean_mu_gpu)
-    
-    # Safety Clamping (Same as original)
-    max_size_val = cp.nanmax(size_gpu) * 10
-    if cp.isnan(max_size_val): max_size_val = 1000
-    size_gpu[cp.isnan(size_gpu) | (size_gpu <= 0)] = max_size_val
-    size_gpu[size_gpu < 1e-10] = 1e-10
-    
-    # Construct "Basic Fit" Object
-    fit_basic = {
-        'sizes': pd.Series(size_gpu.get(), index=stats['tjs'].index),
-        'vals': stats,
-        'var_obs': pd.Series(my_rowvar_gpu.get(), index=stats['tjs'].index)
-    }
-    
-    # --- Phase 2: Check Fit (Calculate Errors) ---
-    print("Phase [2/3]: Evaluating fit errors on ORIGINAL data...")
-    
-    # Check Adjust (M3Drop)
-    check_adjust = NBumiCheckFitFSGPU(
-        cleaned_filename, fit_adjust, suppress_plot=True, chunk_size=chunk_size
-    )
-    
-    # Check Basic (Depth-Norm)
-    check_basic = NBumiCheckFitFSGPU(
-        cleaned_filename, fit_basic, suppress_plot=True, chunk_size=chunk_size
-    )
-    print("Phase [2/3]: COMPLETE")
-
-    # --- Phase 3: Plotting & Comparison ---
-    print("Phase [3/3]: Generating comparison...")
-    
-    nc_data = stats['nc']
-    mean_expr = stats['tjs'] / nc_data
-    observed_dropout = stats['djs'] / nc_data
-    
-    adj_dropout_fit = check_adjust['rowPs'] / nc_data
-    bas_dropout_fit = check_basic['rowPs'] / nc_data
-    
-    err_adj = np.sum(np.abs(adj_dropout_fit - observed_dropout))
-    err_bas = np.sum(np.abs(bas_dropout_fit - observed_dropout))
-    
-    comparison_df = pd.DataFrame({
-        'mean_expr': mean_expr,
-        'observed': observed_dropout,
-        'adj_fit': adj_dropout_fit,
-        'bas_fit': bas_dropout_fit
-    })
-    
-    plt.figure(figsize=(10, 6))
-    sorted_idx = np.argsort(mean_expr.values)
-    
-    plt.scatter(mean_expr.iloc[sorted_idx], observed_dropout.iloc[sorted_idx], 
-                c='black', s=3, alpha=0.5, label='Observed')
-    plt.scatter(mean_expr.iloc[sorted_idx], bas_dropout_fit.iloc[sorted_idx], 
-                c='purple', s=3, alpha=0.6, label=f'Basic Fit (Error: {err_bas:.2f})')
-    plt.scatter(mean_expr.iloc[sorted_idx], adj_dropout_fit.iloc[sorted_idx], 
-                c='goldenrod', s=3, alpha=0.7, label=f'Depth-Adjusted Fit (Error: {err_adj:.2f})')
-    
-    plt.xscale('log')
-    plt.xlabel("Mean Expression")
-    plt.ylabel("Dropout Rate")
-    plt.title("M3Drop Model Comparison")
-    plt.legend()
-    plt.grid(True, linestyle='--', alpha=0.3)
-
-    if plot_filename:
-        plt.savefig(plot_filename, dpi=300, bbox_inches='tight')
-        print(f"STATUS: Model comparison plot saved to '{plot_filename}'")
-
-    if not suppress_plot:
-        plt.show()
-    
-    plt.close()
-    
-    pipeline_end_time = time.time()
-    print(f"Total time: {pipeline_end_time - pipeline_start_time:.2f} seconds.\n")
-    
-    return {
-        "errors": {"Depth-Adjusted": err_adj, "Basic": err_bas},
-        "comparison_df": comparison_df
-    }