gengeneeval 0.2.1__tar.gz → 0.4.0__tar.gz

{gengeneeval-0.2.1 → gengeneeval-0.4.0}/PKG-INFO

@@ -1,10 +1,10 @@
 Metadata-Version: 2.4
 Name: gengeneeval
-Version: 0.2.1
-Summary: Comprehensive evaluation of generated gene expression data. Computes metrics between real and generated datasets with support for condition matching, train/test splits, memory-efficient lazy loading, and publication-quality visualizations.
+Version: 0.4.0
+Summary: Comprehensive evaluation of generated gene expression data. Computes metrics between real and generated datasets with support for condition matching, DEG-focused evaluation, per-context analysis, train/test splits, memory-efficient lazy loading, CPU parallelization, GPU acceleration, and publication-quality visualizations.
 License: MIT
 License-File: LICENSE
-Keywords: gene expression,evaluation,metrics,single-cell,generative models,benchmarking,memory-efficient
+Keywords: gene expression,evaluation,metrics,single-cell,generative models,benchmarking,memory-efficient,DEG,perturbation
 Author: GenEval Team
 Author-email: geneval@example.com
 Requires-Python: >=3.8,<4.0
@@ -24,6 +24,7 @@ Provides-Extra: full
 Provides-Extra: gpu
 Requires-Dist: anndata (>=0.8.0)
 Requires-Dist: geomloss (>=0.2.1) ; extra == "full" or extra == "gpu"
+Requires-Dist: joblib (>=1.0.0)
 Requires-Dist: matplotlib (>=3.5.0)
 Requires-Dist: numpy (>=1.21.0)
 Requires-Dist: pandas (>=1.3.0)
@@ -77,8 +78,12 @@ All metrics are computed **per-gene** (returning a vector) and **aggregated**:
 - ✅ Condition-based matching (perturbation, cell type, etc.)
 - ✅ Train/test split support
 - ✅ Per-gene and aggregate metrics
+- ✅ **DEG-focused evaluation** with per-context (covariate × perturbation) support
+- ✅ **Fast DEG detection** via vectorized Welch's t-test, Student's t-test, Wilcoxon
 - ✅ **Memory-efficient lazy loading** for large datasets
 - ✅ **Batched evaluation** to avoid OOM errors
+- ✅ **CPU parallelization** via joblib (multi-core speedup)
+- ✅ **GPU acceleration** via PyTorch (10-100x speedup)
 - ✅ Modular, extensible architecture
 - ✅ Command-line interface
 - ✅ Publication-quality visualizations
@@ -173,6 +178,162 @@ with load_data_lazy("real.h5ad", "gen.h5ad", ["perturbation"]) as loader:
         pass
 ```
 
+### Accelerated Evaluation (CPU Parallelization & GPU)
+
+GenEval supports CPU parallelization and GPU acceleration for significant speedups:
+
+```python
+from geneval import evaluate, get_available_backends
+
+# Check available backends
+print(get_available_backends())
+# {'joblib': True, 'torch': True, 'geomloss': True, 'cuda': True, 'mps': False}
+
+# Parallel CPU evaluation (use all cores)
+results = evaluate(
+    real_path="real.h5ad",
+    generated_path="generated.h5ad",
+    condition_columns=["perturbation"],
+    n_jobs=-1,  # Use all available CPU cores
+)
+
+# GPU-accelerated evaluation
+results = evaluate(
+    real_path="real.h5ad",
+    generated_path="generated.h5ad",
+    condition_columns=["perturbation"],
+    device="cuda",  # Use NVIDIA GPU
+)
+
+# Combined: parallel CPU + auto device selection
+results = evaluate(..., n_jobs=8, device="auto")
+```
+
+#### Low-level Accelerated API
+
+For custom workflows, use the accelerated metrics directly:
+
+```python
+from geneval.metrics.accelerated import (
+    compute_metrics_accelerated,
+    GPUWasserstein1,
+    GPUMMD,
+    vectorized_wasserstein1,
+)
+import numpy as np
+
+# Load your data
+real = np.random.randn(1000, 5000)      # 1000 cells, 5000 genes
+generated = np.random.randn(1000, 5000)
+
+# Compute multiple metrics with acceleration
+results = compute_metrics_accelerated(
+    real, generated,
+    metrics=["wasserstein_1", "wasserstein_2", "mmd", "energy"],
+    n_jobs=8,          # CPU parallelization
+    device="cuda",     # GPU acceleration
+    verbose=True,
+)
+
+# Access results
+print(f"W1: {results['wasserstein_1'].aggregate_value:.4f}")
+print(f"MMD: {results['mmd'].aggregate_value:.4f}")
+```
+
+#### Performance Tips
+
+| Optimization | Speedup | When to Use |
+|--------------|---------|-------------|
+| `n_jobs=-1` (all cores) | 4-16x | Always (if joblib available) |
+| `device="cuda"` | 10-100x | Large datasets, NVIDIA GPU available |
+| `device="mps"` | 5-20x | Apple Silicon Macs |
+| Vectorized NumPy | 2-5x | Automatic fallback |
+
+### DEG-Focused Evaluation
+
+GenEval supports **Differentially Expressed Genes (DEG)-focused evaluation**, computing metrics only on biologically relevant DEGs rather than all genes. This provides more meaningful evaluation for perturbation prediction tasks.
+
+#### Key Features
+
+- **Fast DEG detection**: Vectorized Welch's t-test, Student's t-test, or Wilcoxon rank-sum
+- **Per-context evaluation**: Automatically evaluates each (covariate × perturbation) combination
+- **GPU acceleration**: DEG detection and metrics on GPU for large datasets
+- **Comprehensive reporting**: Aggregated and expanded results with visualizations
+
+#### Quick Start
+
+```python
+from geneval import evaluate_degs
+import pandas as pd
+
+# Evaluate with DEG-focused metrics
+results = evaluate_degs(
+    real_data=real_adata.X,           # (n_samples, n_genes)
+    generated_data=gen_adata.X,
+    real_obs=real_adata.obs,
+    generated_obs=gen_adata.obs,
+    condition_columns=["cell_type", "perturbation"],  # Context columns
+    control_key="control",            # Value indicating control samples
+    perturbation_column="perturbation",
+    deg_method="welch",               # or "student", "wilcoxon", "logfc"
+    pval_threshold=0.05,
+    lfc_threshold=0.5,                # log2 fold change threshold
+    device="cuda",                    # GPU acceleration
+)
+
+# Access results
+print(results.aggregated_metrics)     # Summary across all contexts
+print(results.expanded_metrics)       # Per-context metrics
+print(results.deg_summary)            # DEG counts per context
+
+# Save results with plots
+results.save("deg_evaluation/")
+```
+
+#### Per-Context Evaluation
+
+When multiple condition columns are provided (e.g., `["cell_type", "perturbation"]`), GenEval evaluates **every combination** separately:
+
+| Context | n_DEGs | W1 (DEGs only) | MMD (DEGs only) |
+|---------|--------|----------------|-----------------|
+| TypeA_drug1 | 234 | 0.42 | 0.031 |
+| TypeA_drug2 | 189 | 0.38 | 0.027 |
+| TypeB_drug1 | 312 | 0.51 | 0.045 |
+
+If only `perturbation` column is provided, evaluation is done per-perturbation.
+
+#### Available DEG Methods
+
+| Method | Description | Speed |
+|--------|-------------|-------|
+| `welch` | Welch's t-test (unequal variance) | ⚡ Fast |
+| `student` | Student's t-test (equal variance) | ⚡ Fast |
+| `wilcoxon` | Wilcoxon rank-sum (non-parametric) | 🐢 Slower |
+| `logfc` | Log fold change only (no p-value) | ⚡⚡ Fastest |
+
+#### Visualization
+
+```python
+from geneval.deg import (
+    plot_deg_distributions,
+    plot_context_heatmap,
+    plot_deg_counts,
+    create_deg_report,
+)
+
+# Distribution of metrics across contexts
+plot_deg_distributions(results, save_path="dist.png")
+
+# Heatmap: context × metric
+plot_context_heatmap(results, save_path="heatmap.png")
+
+# DEG counts per context (up/down regulated)
+plot_deg_counts(results, save_path="deg_counts.png")
+
+# Generate comprehensive report
+create_deg_report(results, "report/", include_plots=True)
+```
+
 ## Expected Data Format
 
 GenEval expects AnnData (h5ad) files with:

{gengeneeval-0.2.1 → gengeneeval-0.4.0}/README.md

@@ -38,8 +38,12 @@ All metrics are computed **per-gene** (returning a vector) and **aggregated**:
 - ✅ Condition-based matching (perturbation, cell type, etc.)
 - ✅ Train/test split support
 - ✅ Per-gene and aggregate metrics
+- ✅ **DEG-focused evaluation** with per-context (covariate × perturbation) support
+- ✅ **Fast DEG detection** via vectorized Welch's t-test, Student's t-test, Wilcoxon
 - ✅ **Memory-efficient lazy loading** for large datasets
 - ✅ **Batched evaluation** to avoid OOM errors
+- ✅ **CPU parallelization** via joblib (multi-core speedup)
+- ✅ **GPU acceleration** via PyTorch (10-100x speedup)
 - ✅ Modular, extensible architecture
 - ✅ Command-line interface
 - ✅ Publication-quality visualizations
@@ -134,6 +138,162 @@ with load_data_lazy("real.h5ad", "gen.h5ad", ["perturbation"]) as loader:
         pass
 ```
 
+### Accelerated Evaluation (CPU Parallelization & GPU)
+
+GenEval supports CPU parallelization and GPU acceleration for significant speedups:
+
+```python
+from geneval import evaluate, get_available_backends
+
+# Check available backends
+print(get_available_backends())
+# {'joblib': True, 'torch': True, 'geomloss': True, 'cuda': True, 'mps': False}
+
+# Parallel CPU evaluation (use all cores)
+results = evaluate(
+    real_path="real.h5ad",
+    generated_path="generated.h5ad",
+    condition_columns=["perturbation"],
+    n_jobs=-1,  # Use all available CPU cores
+)
+
+# GPU-accelerated evaluation
+results = evaluate(
+    real_path="real.h5ad",
+    generated_path="generated.h5ad",
+    condition_columns=["perturbation"],
+    device="cuda",  # Use NVIDIA GPU
+)
+
+# Combined: parallel CPU + auto device selection
+results = evaluate(..., n_jobs=8, device="auto")
+```
+
+#### Low-level Accelerated API
+
+For custom workflows, use the accelerated metrics directly:
+
+```python
+from geneval.metrics.accelerated import (
+    compute_metrics_accelerated,
+    GPUWasserstein1,
+    GPUMMD,
+    vectorized_wasserstein1,
+)
+import numpy as np
+
+# Load your data
+real = np.random.randn(1000, 5000)      # 1000 cells, 5000 genes
+generated = np.random.randn(1000, 5000)
+
+# Compute multiple metrics with acceleration
+results = compute_metrics_accelerated(
+    real, generated,
+    metrics=["wasserstein_1", "wasserstein_2", "mmd", "energy"],
+    n_jobs=8,          # CPU parallelization
+    device="cuda",     # GPU acceleration
+    verbose=True,
+)
+
+# Access results
+print(f"W1: {results['wasserstein_1'].aggregate_value:.4f}")
+print(f"MMD: {results['mmd'].aggregate_value:.4f}")
+```
+
+#### Performance Tips
+
+| Optimization | Speedup | When to Use |
+|--------------|---------|-------------|
+| `n_jobs=-1` (all cores) | 4-16x | Always (if joblib available) |
+| `device="cuda"` | 10-100x | Large datasets, NVIDIA GPU available |
+| `device="mps"` | 5-20x | Apple Silicon Macs |
+| Vectorized NumPy | 2-5x | Automatic fallback |
+
+### DEG-Focused Evaluation
+
+GenEval supports **Differentially Expressed Genes (DEG)-focused evaluation**, computing metrics only on biologically relevant DEGs rather than all genes. This provides more meaningful evaluation for perturbation prediction tasks.
+
+#### Key Features
+
+- **Fast DEG detection**: Vectorized Welch's t-test, Student's t-test, or Wilcoxon rank-sum
+- **Per-context evaluation**: Automatically evaluates each (covariate × perturbation) combination
+- **GPU acceleration**: DEG detection and metrics on GPU for large datasets
+- **Comprehensive reporting**: Aggregated and expanded results with visualizations
+
+#### Quick Start
+
+```python
+from geneval import evaluate_degs
+import pandas as pd
+
+# Evaluate with DEG-focused metrics
+results = evaluate_degs(
+    real_data=real_adata.X,           # (n_samples, n_genes)
+    generated_data=gen_adata.X,
+    real_obs=real_adata.obs,
+    generated_obs=gen_adata.obs,
+    condition_columns=["cell_type", "perturbation"],  # Context columns
+    control_key="control",            # Value indicating control samples
+    perturbation_column="perturbation",
+    deg_method="welch",               # or "student", "wilcoxon", "logfc"
+    pval_threshold=0.05,
+    lfc_threshold=0.5,                # log2 fold change threshold
+    device="cuda",                    # GPU acceleration
+)
+
+# Access results
+print(results.aggregated_metrics)     # Summary across all contexts
+print(results.expanded_metrics)       # Per-context metrics
+print(results.deg_summary)            # DEG counts per context
+
+# Save results with plots
+results.save("deg_evaluation/")
+```
+
+#### Per-Context Evaluation
+
+When multiple condition columns are provided (e.g., `["cell_type", "perturbation"]`), GenEval evaluates **every combination** separately:
+
+| Context | n_DEGs | W1 (DEGs only) | MMD (DEGs only) |
+|---------|--------|----------------|-----------------|
+| TypeA_drug1 | 234 | 0.42 | 0.031 |
+| TypeA_drug2 | 189 | 0.38 | 0.027 |
+| TypeB_drug1 | 312 | 0.51 | 0.045 |
+
+If only `perturbation` column is provided, evaluation is done per-perturbation.
+
+#### Available DEG Methods
+
+| Method | Description | Speed |
+|--------|-------------|-------|
+| `welch` | Welch's t-test (unequal variance) | ⚡ Fast |
+| `student` | Student's t-test (equal variance) | ⚡ Fast |
+| `wilcoxon` | Wilcoxon rank-sum (non-parametric) | 🐢 Slower |
+| `logfc` | Log fold change only (no p-value) | ⚡⚡ Fastest |
+
+#### Visualization
+
+```python
+from geneval.deg import (
+    plot_deg_distributions,
+    plot_context_heatmap,
+    plot_deg_counts,
+    create_deg_report,
+)
+
+# Distribution of metrics across contexts
+plot_deg_distributions(results, save_path="dist.png")
+
+# Heatmap: context × metric
+plot_context_heatmap(results, save_path="heatmap.png")
+
+# DEG counts per context (up/down regulated)
+plot_deg_counts(results, save_path="deg_counts.png")
+
+# Generate comprehensive report
+create_deg_report(results, "report/", include_plots=True)
+```
+
 ## Expected Data Format
 
 GenEval expects AnnData (h5ad) files with:

{gengeneeval-0.2.1 → gengeneeval-0.4.0}/pyproject.toml

@@ -1,13 +1,13 @@
 [tool.poetry]
 name = "gengeneeval"
-version = "0.2.1"
-description = "Comprehensive evaluation of generated gene expression data. Computes metrics between real and generated datasets with support for condition matching, train/test splits, memory-efficient lazy loading, and publication-quality visualizations."
+version = "0.4.0"
+description = "Comprehensive evaluation of generated gene expression data. Computes metrics between real and generated datasets with support for condition matching, DEG-focused evaluation, per-context analysis, train/test splits, memory-efficient lazy loading, CPU parallelization, GPU acceleration, and publication-quality visualizations."
 authors = ["GenEval Team <geneval@example.com>"]
 license = "MIT"
 readme = "README.md"
 homepage = "https://github.com/AndreaRubbi/GenGeneEval"
 repository = "https://github.com/AndreaRubbi/GenGeneEval"
-keywords = ["gene expression", "evaluation", "metrics", "single-cell", "generative models", "benchmarking", "memory-efficient"]
+keywords = ["gene expression", "evaluation", "metrics", "single-cell", "generative models", "benchmarking", "memory-efficient", "DEG", "perturbation"]
 classifiers = [
     "Development Status :: 4 - Beta",
     "Intended Audience :: Science/Research",
@@ -29,6 +29,7 @@ scipy = ">=1.7.0"
 torch = ">=1.9.0"
 matplotlib = ">=3.5.0"
 seaborn = ">=0.11.0"
+joblib = ">=1.0.0"
 geomloss = {version = ">=0.2.1", optional = true}
 pykeops = {version = ">=1.4.0", optional = true}
 umap-learn = {version = ">=0.5.0", optional = true}

{gengeneeval-0.2.1 → gengeneeval-0.4.0}/src/geneval/__init__.py

@@ -7,8 +7,10 @@ and generated gene expression datasets stored in AnnData (h5ad) format.
 Features:
 - Multiple distance and correlation metrics (per-gene and aggregate)
 - Condition-based matching (perturbation, cell type, etc.)
+- DEG-focused evaluation with per-context (covariate × perturbation) support
 - Train/test split support
 - Memory-efficient lazy loading for large datasets
+- CPU parallelization and GPU acceleration
 - Publication-quality visualizations
 - Command-line interface
 
@@ -21,6 +23,17 @@ Quick Start:
     ...     output_dir="output/"
     ... )
 
+DEG-Focused Evaluation:
+    >>> from geneval import evaluate_degs
+    >>> results = evaluate_degs(
+    ...     real_data, generated_data,
+    ...     real_obs, generated_obs,
+    ...     condition_columns=["cell_type", "perturbation"],
+    ...     control_key="control",
+    ...     deg_method="welch",
+    ...     device="cuda",  # GPU acceleration
+    ... )
+
 Memory-Efficient Mode (for large datasets):
     >>> from geneval import evaluate_lazy
     >>> results = evaluate_lazy(
@@ -36,7 +49,7 @@ CLI Usage:
               --conditions perturbation cell_type --output results/
 """
 
-__version__ = "0.2.1"
+__version__ = "0.4.0"
 __author__ = "GenEval Team"
 
 # Main evaluation interface
@@ -101,6 +114,30 @@ from .metrics.reconstruction import (
     R2Score,
 )
 
+# Accelerated computation
+from .metrics.accelerated import (
+    AccelerationConfig,
+    ParallelMetricComputer,
+    get_available_backends,
+    compute_metrics_accelerated,
+)
+
+# DEG-focused evaluation
+from .deg import (
+    DEGEvaluator,
+    DEGResult,
+    DEGEvaluationResult,
+    ContextEvaluator,
+    ContextResult,
+    compute_degs_fast,
+    compute_degs_gpu,
+    get_contexts,
+    plot_deg_distributions,
+    plot_context_heatmap,
+    create_deg_report,
+)
+from .deg.evaluator import evaluate_degs
+
 # Visualization
 from .visualization.visualizer import (
     EvaluationVisualizer,
@@ -161,6 +198,24 @@ __all__ = [
     "RMSEDistance",
     "MAEDistance",
     "R2Score",
+    # Acceleration
+    "AccelerationConfig",
+    "ParallelMetricComputer",
+    "get_available_backends",
+    "compute_metrics_accelerated",
+    # DEG evaluation
+    "DEGEvaluator",
+    "DEGResult",
+    "DEGEvaluationResult",
+    "ContextEvaluator",
+    "ContextResult",
+    "compute_degs_fast",
+    "compute_degs_gpu",
+    "evaluate_degs",
+    "get_contexts",
+    "plot_deg_distributions",
+    "plot_context_heatmap",
+    "create_deg_report",
     # Visualization
     "EvaluationVisualizer",
     "visualize",

gengeneeval-0.4.0/src/geneval/deg/__init__.py

@@ -0,0 +1,65 @@
+"""
+Differentially Expressed Genes (DEG) module for GenGeneEval.
+
+This module provides:
+- Fast DEG detection using vectorized statistical tests
+- Per-context evaluation (covariates × perturbations)
+- DEG-focused metrics computation
+- Integration with GPU acceleration
+
+Example usage:
+    >>> from geneval.deg import DEGEvaluator, compute_degs_fast
+    >>> degs = compute_degs_fast(control_data, perturbed_data, method="welch")
+    >>> evaluator = DEGEvaluator(loader, deg_method="welch", pval_threshold=0.05)
+    >>> results = evaluator.evaluate()
+"""
+
+from .detection import (
+    compute_degs_fast,
+    compute_degs_gpu,
+    compute_degs_auto,
+    DEGResult,
+    DEGMethod,
+)
+from .context import (
+    ContextEvaluator,
+    ContextResult,
+    get_contexts,
+    get_context_id,
+    filter_by_context,
+)
+from .evaluator import (
+    DEGEvaluator,
+    DEGEvaluationResult,
+    evaluate_degs,
+)
+from .visualization import (
+    plot_deg_distributions,
+    plot_context_heatmap,
+    plot_deg_counts,
+    create_deg_report,
+)
+
+__all__ = [
+    # Detection
+    "compute_degs_fast",
+    "compute_degs_gpu",
+    "compute_degs_auto",
+    "DEGResult",
+    "DEGMethod",
+    # Context
+    "ContextEvaluator",
+    "ContextResult",
+    "get_contexts",
+    "get_context_id",
+    "filter_by_context",
+    # Evaluator
+    "DEGEvaluator",
+    "DEGEvaluationResult",
+    "evaluate_degs",
+    # Visualization
+    "plot_deg_distributions",
+    "plot_context_heatmap",
+    "plot_deg_counts",
+    "create_deg_report",
+]