gengeneeval 0.1.1__tar.gz → 0.2.0__tar.gz

{gengeneeval-0.1.1 → gengeneeval-0.2.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: gengeneeval
-Version: 0.1.1
+Version: 0.2.0
 Summary: Comprehensive evaluation of generated gene expression data. Computes metrics between real and generated datasets with support for condition matching, train/test splits, and publication-quality visualizations.
 License: MIT
 License-File: LICENSE
@@ -42,7 +42,7 @@ Description-Content-Type: text/markdown
 [![PyPI version](https://badge.fury.io/py/gengeneeval.svg)](https://badge.fury.io/py/gengeneeval)
 [![Python 3.8+](https://img.shields.io/badge/python-3.8+-blue.svg)](https://www.python.org/downloads/)
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
-[![Tests](https://github.com/AndreaRubbi/GenGeneEval/actions/workflows/tests.yml/badge.svg)](https://github.com/AndreaRubbi/GenGeneEval/actions)
+[![Tests](https://github.com/AndreaRubbi/GenGeneEval/actions/workflows/test.yml/badge.svg)](https://github.com/AndreaRubbi/GenGeneEval/actions)
 
 **Comprehensive evaluation of generated gene expression data against real datasets.**
 
@@ -55,6 +55,10 @@ All metrics are computed **per-gene** (returning a vector) and **aggregated**:
 
 | Metric | Description | Direction |
 |--------|-------------|-----------|
+| **MSE** | Mean Squared Error | Lower is better |
+| **RMSE** | Root Mean Squared Error | Lower is better |
+| **MAE** | Mean Absolute Error | Lower is better |
+| **R²** | Coefficient of Determination | Higher is better |
 | **Pearson Correlation** | Linear correlation between expression profiles | Higher is better |
 | **Spearman Correlation** | Rank correlation (robust to outliers) | Higher is better |
 | **Wasserstein-1** | Earth Mover's Distance (L1) | Lower is better |

{gengeneeval-0.1.1 → gengeneeval-0.2.0}/README.md

@@ -3,7 +3,7 @@
 [![PyPI version](https://badge.fury.io/py/gengeneeval.svg)](https://badge.fury.io/py/gengeneeval)
 [![Python 3.8+](https://img.shields.io/badge/python-3.8+-blue.svg)](https://www.python.org/downloads/)
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
-[![Tests](https://github.com/AndreaRubbi/GenGeneEval/actions/workflows/tests.yml/badge.svg)](https://github.com/AndreaRubbi/GenGeneEval/actions)
+[![Tests](https://github.com/AndreaRubbi/GenGeneEval/actions/workflows/test.yml/badge.svg)](https://github.com/AndreaRubbi/GenGeneEval/actions)
 
 **Comprehensive evaluation of generated gene expression data against real datasets.**
 
@@ -16,6 +16,10 @@ All metrics are computed **per-gene** (returning a vector) and **aggregated**:
 
 | Metric | Description | Direction |
 |--------|-------------|-----------|
+| **MSE** | Mean Squared Error | Lower is better |
+| **RMSE** | Root Mean Squared Error | Lower is better |
+| **MAE** | Mean Absolute Error | Lower is better |
+| **R²** | Coefficient of Determination | Higher is better |
 | **Pearson Correlation** | Linear correlation between expression profiles | Higher is better |
 | **Spearman Correlation** | Rank correlation (robust to outliers) | Higher is better |
 | **Wasserstein-1** | Earth Mover's Distance (L1) | Lower is better |

{gengeneeval-0.1.1 → gengeneeval-0.2.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "gengeneeval"
-version = "0.1.1"
+version = "0.2.0"
 description = "Comprehensive evaluation of generated gene expression data. Computes metrics between real and generated datasets with support for condition matching, train/test splits, and publication-quality visualizations."
 authors = ["GenEval Team <geneval@example.com>"]
 license = "MIT"

{gengeneeval-0.1.1 → gengeneeval-0.2.0}/src/geneval/__init__.py

@@ -25,7 +25,7 @@ CLI Usage:
               --conditions perturbation cell_type --output results/
 """
 
-__version__ = "0.1.1"
+__version__ = "0.2.0"
 __author__ = "GenEval Team"
 
 # Main evaluation interface
@@ -69,6 +69,12 @@ from .metrics.distances import (
     MultivariateWasserstein,
     MultivariateMMD,
 )
+from .metrics.reconstruction import (
+    MSEDistance,
+    RMSEDistance,
+    MAEDistance,
+    R2Score,
+)
 
 # Visualization
 from .visualization.visualizer import (

{gengeneeval-0.1.1 → gengeneeval-0.2.0}/src/geneval/evaluator.py

@@ -27,11 +27,15 @@ from .metrics.distances import (
     MultivariateWasserstein,
     MultivariateMMD,
 )
+from .metrics.reconstruction import (
+    MSEDistance,
+)
 from .results import EvaluationResult, SplitResult, ConditionResult
 
 
 # Default metrics to compute
 DEFAULT_METRICS = [
+    MSEDistance,
     PearsonCorrelation,
     SpearmanCorrelation,
     MeanPearsonCorrelation,

{gengeneeval-0.1.1 → gengeneeval-0.2.0}/src/geneval/metrics/__init__.py

@@ -2,6 +2,7 @@
 Metrics module for gene expression evaluation.
 
 Provides per-gene and aggregate metrics for comparing distributions:
+- Reconstruction metrics (MSE, RMSE, MAE, R²)
 - Correlation metrics (Pearson, Spearman)
 - Distribution distances (Wasserstein, MMD, Energy)
 - Multivariate distances
@@ -27,13 +28,26 @@ from .distances import (
     MultivariateWasserstein,
     MultivariateMMD,
 )
+from .reconstruction import (
+    MSEDistance,
+    RMSEDistance,
+    MAEDistance,
+    R2Score,
+)
 
 # All available metrics
 ALL_METRICS = [
+    # Reconstruction
+    MSEDistance,
+    RMSEDistance,
+    MAEDistance,
+    R2Score,
+    # Correlation
     PearsonCorrelation,
     SpearmanCorrelation,
     MeanPearsonCorrelation,
     MeanSpearmanCorrelation,
+    # Distribution
     Wasserstein1Distance,
     Wasserstein2Distance,
     MMDDistance,
@@ -48,6 +62,11 @@ __all__ = [
     "MetricResult",
     "DistributionMetric",
     "CorrelationMetric",
+    # Reconstruction metrics
+    "MSEDistance",
+    "RMSEDistance",
+    "MAEDistance",
+    "R2Score",
     # Correlation metrics
     "PearsonCorrelation",
     "SpearmanCorrelation",

gengeneeval-0.2.0/src/geneval/metrics/reconstruction.py

@@ -0,0 +1,243 @@
+"""
+Reconstruction metrics for gene expression evaluation.
+
+Provides MSE (Mean Squared Error) and related reconstruction quality metrics.
+"""
+from __future__ import annotations
+
+import numpy as np
+from typing import Optional
+
+from .base_metric import BaseMetric
+
+
+def _ensure_2d(arr: np.ndarray) -> np.ndarray:
+    """Ensure array is 2D (samples x genes)."""
+    arr = np.asarray(arr, dtype=np.float64)
+    if arr.ndim == 1:
+        arr = arr.reshape(-1, 1)
+    return arr
+
+
+class MSEDistance(BaseMetric):
+    """
+    Mean Squared Error (MSE) between real and generated distributions.
+    
+    Computes the average squared difference between samples. When sample
+    sizes differ, compares mean expression profiles.
+    
+    Lower values indicate better reconstruction.
+    
+    Parameters
+    ----------
+    compare_means : bool
+        If True, always compare mean profiles regardless of sample sizes.
+        If False, compute sample-wise MSE when sizes match.
+        
+    Examples
+    --------
+    >>> mse = MSEDistance()
+    >>> result = mse.compute(real_data, generated_data, gene_names)
+    >>> print(f"MSE: {result.aggregate_value:.4f}")
+    """
+    
+    def __init__(self, compare_means: bool = False):
+        super().__init__(
+            name="mse",
+            description="Mean Squared Error per gene",
+            higher_is_better=False,
+            requires_distribution=True,
+        )
+        self.compare_means = compare_means
+    
+    def compute_per_gene(
+        self,
+        real: np.ndarray,
+        generated: np.ndarray,
+    ) -> np.ndarray:
+        """
+        Compute MSE for each gene.
+        
+        Parameters
+        ----------
+        real : np.ndarray
+            Real data, shape (n_samples_real, n_genes)
+        generated : np.ndarray
+            Generated data, shape (n_samples_gen, n_genes)
+            
+        Returns
+        -------
+        np.ndarray
+            MSE per gene, shape (n_genes,)
+        """
+        real = _ensure_2d(real)
+        generated = _ensure_2d(generated)
+        n_genes = real.shape[1]
+        
+        mse = np.zeros(n_genes)
+        
+        # Compare mean profiles when sample sizes differ or compare_means is True
+        if self.compare_means or real.shape[0] != generated.shape[0]:
+            real_mean = np.mean(real, axis=0)
+            gen_mean = np.mean(generated, axis=0)
+            mse = (real_mean - gen_mean) ** 2
+        else:
+            # Sample-wise MSE when sizes match
+            for i in range(n_genes):
+                r_vals = real[:, i]
+                g_vals = generated[:, i]
+                
+                # Filter NaN values
+                valid = ~(np.isnan(r_vals) | np.isnan(g_vals))
+                if not valid.any():
+                    mse[i] = np.nan
+                    continue
+                
+                mse[i] = np.mean((r_vals[valid] - g_vals[valid]) ** 2)
+        
+        return mse
+
+
+class RMSEDistance(BaseMetric):
+    """
+    Root Mean Squared Error (RMSE) between real and generated distributions.
+    
+    Square root of MSE, in the same units as the original data.
+    Lower values indicate better reconstruction.
+    """
+    
+    def __init__(self, compare_means: bool = False):
+        super().__init__(
+            name="rmse",
+            description="Root Mean Squared Error per gene",
+            higher_is_better=False,
+            requires_distribution=True,
+        )
+        self.compare_means = compare_means
+        self._mse = MSEDistance(compare_means=compare_means)
+    
+    def compute_per_gene(
+        self,
+        real: np.ndarray,
+        generated: np.ndarray,
+    ) -> np.ndarray:
+        """
+        Compute RMSE for each gene.
+        """
+        mse = self._mse.compute_per_gene(real, generated)
+        return np.sqrt(mse)
+
+
+class MAEDistance(BaseMetric):
+    """
+    Mean Absolute Error (MAE) between real and generated distributions.
+    
+    More robust to outliers than MSE.
+    Lower values indicate better reconstruction.
+    """
+    
+    def __init__(self, compare_means: bool = False):
+        super().__init__(
+            name="mae",
+            description="Mean Absolute Error per gene",
+            higher_is_better=False,
+            requires_distribution=True,
+        )
+        self.compare_means = compare_means
+    
+    def compute_per_gene(
+        self,
+        real: np.ndarray,
+        generated: np.ndarray,
+    ) -> np.ndarray:
+        """
+        Compute MAE for each gene.
+        """
+        real = _ensure_2d(real)
+        generated = _ensure_2d(generated)
+        n_genes = real.shape[1]
+        
+        mae = np.zeros(n_genes)
+        
+        if self.compare_means or real.shape[0] != generated.shape[0]:
+            real_mean = np.mean(real, axis=0)
+            gen_mean = np.mean(generated, axis=0)
+            mae = np.abs(real_mean - gen_mean)
+        else:
+            for i in range(n_genes):
+                r_vals = real[:, i]
+                g_vals = generated[:, i]
+                
+                valid = ~(np.isnan(r_vals) | np.isnan(g_vals))
+                if not valid.any():
+                    mae[i] = np.nan
+                    continue
+                
+                mae[i] = np.mean(np.abs(r_vals[valid] - g_vals[valid]))
+        
+        return mae
+
+
+class R2Score(BaseMetric):
+    """
+    Coefficient of Determination (R²) between real and generated data.
+    
+    Measures the proportion of variance explained. Values close to 1
+    indicate good fit, 0 means no better than mean prediction.
+    
+    Higher values indicate better reconstruction.
+    """
+    
+    def __init__(self):
+        super().__init__(
+            name="r2",
+            description="R² (coefficient of determination) per gene",
+            higher_is_better=True,
+            requires_distribution=True,
+        )
+    
+    def compute_per_gene(
+        self,
+        real: np.ndarray,
+        generated: np.ndarray,
+    ) -> np.ndarray:
+        """
+        Compute R² for each gene.
+        """
+        real = _ensure_2d(real)
+        generated = _ensure_2d(generated)
+        n_genes = real.shape[1]
+        
+        r2 = np.zeros(n_genes)
+        
+        # R² only makes sense when sample sizes match
+        if real.shape[0] != generated.shape[0]:
+            # Fall back to using mean comparison
+            real_mean = np.mean(real, axis=0)
+            gen_mean = np.mean(generated, axis=0)
+            
+            ss_tot = np.var(real, axis=0) * real.shape[0]
+            ss_res = (real_mean - gen_mean) ** 2
+            
+            with np.errstate(invalid='ignore', divide='ignore'):
+                r2 = 1 - ss_res / (ss_tot / real.shape[0] + 1e-10)
+            r2 = np.nan_to_num(r2, nan=0.0)
+        else:
+            for i in range(n_genes):
+                r_vals = real[:, i]
+                g_vals = generated[:, i]
+                
+                valid = ~(np.isnan(r_vals) | np.isnan(g_vals))
+                if not valid.any():
+                    r2[i] = np.nan
+                    continue
+                
+                ss_tot = np.sum((r_vals[valid] - np.mean(r_vals[valid])) ** 2)
+                ss_res = np.sum((r_vals[valid] - g_vals[valid]) ** 2)
+                
+                if ss_tot < 1e-10:
+                    r2[i] = 1.0 if ss_res < 1e-10 else 0.0
+                else:
+                    r2[i] = 1 - ss_res / ss_tot
+        
+        return r2