napistu 0.4.1__py3-none-any.whl → 0.4.2__py3-none-any.whl

tests/test_network_net_propagation.py

@@ -1,89 +1,380 @@
 import pytest
-import igraph as ig
 import numpy as np
-from napistu.network.net_propagation import personalized_pagerank_by_attribute
-
-
-def test_personalized_pagerank_by_attribute_basic():
-    g = ig.Graph.Full(3)
-    g.vs["name"] = ["A", "B", "C"]
-    g.vs["score"] = [1, 0, 2]
-    df = personalized_pagerank_by_attribute(g, "score")
-    assert set(df.columns) == {
-        "name",
-        "pagerank_by_attribute",
-        "pagerank_uniform",
-        "score",
-    }
-    assert np.isclose(df["score"].sum(), 3)
-    assert np.isclose(df["pagerank_by_attribute"].sum(), 1)
-    assert np.isclose(df["pagerank_uniform"].sum(), 1)
-    # Uniform should only include A and C
-    assert df.loc[df["name"] == "B", "pagerank_uniform"].values[0] > 0
-
-
-def test_personalized_pagerank_by_attribute_no_uniform():
-    g = ig.Graph.Full(3)
-    g.vs["score"] = [1, 0, 2]
-    df = personalized_pagerank_by_attribute(g, "score", calculate_uniform_dist=False)
-    assert "pagerank_uniform" not in df.columns
-    assert np.isclose(df["pagerank_by_attribute"].sum(), 1)
-
-
-def test_personalized_pagerank_by_attribute_missing_and_negative():
-    g = ig.Graph.Full(3)
-    g.vs["score"] = [1, None, 2]
-    # None should be treated as 0
-    df = personalized_pagerank_by_attribute(g, "score")
-    assert np.isclose(df["score"].sum(), 3)
-    # Negative values should raise
-    g.vs["score"] = [1, -1, 2]
-    with pytest.raises(ValueError):
-        personalized_pagerank_by_attribute(g, "score")
-
-
-def test_personalized_pagerank_by_attribute_additional_args_directed():
-    # create an asymmetric directed graph to test whether additional_propagation_args is respected
-    g = ig.Graph(directed=True)
-    g.add_vertices(3)
-    g.add_edges([(0, 1), (1, 2)])
-    g.vs["score"] = [1, 0, 2]
-    # Run with directed=False, which should treat the graph as undirected
-    df_directed = personalized_pagerank_by_attribute(
-        g, "score", additional_propagation_args={"directed": True}
+import pandas as pd
+import igraph as ig
+from napistu.network.net_propagation import (
+    net_propagate_attributes,
+    uniform_null,
+    parametric_null,
+    node_permutation_null,
+    edge_permutation_null,
+    NULL_GENERATORS,
+    network_propagation_with_null,
+)
+from napistu.network.constants import (
+    NAPISTU_GRAPH_VERTICES,
+    NULL_STRATEGIES,
+)
+
+
+def test_network_propagation_with_null():
+    """Test the main orchestrator function with different null strategies."""
+    # Create test graph
+    graph = ig.Graph(5)
+    graph.vs[NAPISTU_GRAPH_VERTICES.NAME] = ["A", "B", "C", "D", "E"]
+    graph.vs["attr1"] = [1.0, 0.0, 2.0, 0.0, 1.5]  # Non-negative, not all zero
+    graph.add_edges([(0, 1), (1, 2), (2, 3), (3, 4)])
+
+    attributes = ["attr1"]
+
+    # Test 1: Uniform null (should return ratios)
+    result_uniform = network_propagation_with_null(
+        graph, attributes, null_strategy=NULL_STRATEGIES.UNIFORM
     )
-    df_undirected = personalized_pagerank_by_attribute(
-        g, "score", additional_propagation_args={"directed": False}
+
+    # Check structure
+    assert isinstance(result_uniform, pd.DataFrame)
+    assert result_uniform.shape == (5, 1)
+    assert list(result_uniform.columns) == attributes
+    assert list(result_uniform.index) == ["A", "B", "C", "D", "E"]
+
+    # Should be ratios (can be > 1)
+    assert (result_uniform.values > 0).all(), "Ratios should be positive"
+    # Some ratios should be > 1 since observed scores concentrate on fewer nodes
+    assert (result_uniform.values > 1).any(), "Some ratios should be > 1"
+
+    # Test 2: Node permutation null (should return quantiles)
+    result_permutation = network_propagation_with_null(
+        graph,
+        attributes,
+        null_strategy=NULL_STRATEGIES.NODE_PERMUTATION,
+        n_samples=10,  # Small for testing
     )
-    # The results should differ for directed vs undirected
-    assert not np.allclose(
-        df_directed["pagerank_by_attribute"], df_undirected["pagerank_by_attribute"]
+
+    # Check structure
+    assert isinstance(result_permutation, pd.DataFrame)
+    assert result_permutation.shape == (5, 1)
+    assert list(result_permutation.columns) == attributes
+
+    # Should be quantiles (0 to 1)
+    assert (result_permutation.values >= 0).all(), "Quantiles should be >= 0"
+    assert (result_permutation.values <= 1).all(), "Quantiles should be <= 1"
+
+    # Test 3: Edge permutation null
+    result_edge = network_propagation_with_null(
+        graph,
+        attributes,
+        null_strategy=NULL_STRATEGIES.EDGE_PERMUTATION,
+        n_samples=5,
+        burn_in_ratio=2,  # Small for testing
+        sampling_ratio=0.2,
+    )
+
+    # Check structure
+    assert isinstance(result_edge, pd.DataFrame)
+    assert result_edge.shape == (5, 1)
+    assert (result_edge.values >= 0).all()
+    assert (result_edge.values <= 1).all()
+
+    # Test 4: Gaussian null
+    result_parametric = network_propagation_with_null(
+        graph, attributes, null_strategy=NULL_STRATEGIES.PARAMETRIC, n_samples=8
+    )
+
+    # Check structure
+    assert isinstance(result_parametric, pd.DataFrame)
+    assert result_parametric.shape == (5, 1)
+    assert (result_parametric.values >= 0).all()
+    assert (result_parametric.values <= 1).all()
+
+    # Test 5: Custom propagation parameters
+    result_custom = network_propagation_with_null(
+        graph,
+        attributes,
+        null_strategy=NULL_STRATEGIES.UNIFORM,
+        additional_propagation_args={"damping": 0.7},
     )
-    # Uniform should also be affected
+
+    # Should be different from default
     assert not np.allclose(
-        df_directed["pagerank_uniform"], df_undirected["pagerank_uniform"]
+        result_uniform.values, result_custom.values
+    ), "Different propagation parameters should give different results"
+
+    # Test 6: Custom null parameters (mask)
+    mask_array = np.array([True, False, True, False, True])
+    result_masked = network_propagation_with_null(
+        graph,
+        attributes,
+        null_strategy=NULL_STRATEGIES.NODE_PERMUTATION,
+        n_samples=5,
+        mask=mask_array,
     )
 
+    # Should work without error
+    assert isinstance(result_masked, pd.DataFrame)
+    assert result_masked.shape == (5, 1)
 
-def test_personalized_pagerank_by_attribute_additional_args_invalid():
-    g = ig.Graph.Full(3)
-    g.vs["score"] = [1, 0, 2]
-    # Passing an invalid argument should raise ValueError
-    with pytest.raises(ValueError):
-        personalized_pagerank_by_attribute(
-            g, "score", additional_propagation_args={"not_a_real_arg": 123}
+    # Test 7: Error handling - invalid null strategy
+    with pytest.raises(ValueError, match="Unknown null strategy"):
+        network_propagation_with_null(
+            graph, attributes, null_strategy="invalid_strategy"
         )
 
 
-def test_personalized_pagerank_by_attribute_all_missing():
-    g = ig.Graph.Full(3)
-    # No 'score' attribute at all
-    with pytest.raises(ValueError, match="missing for all vertices"):
-        personalized_pagerank_by_attribute(g, "score")
+def test_net_propagate_attributes():
+    """Test net_propagate_attributes with multiple attributes and various scenarios."""
+    # Create test graph with edges for realistic propagation
+    graph = ig.Graph(4)
+    graph.vs[NAPISTU_GRAPH_VERTICES.NAME] = ["node1", "node2", "node3", "node4"]
+    graph.vs["attr1"] = [1.0, 0.0, 2.0, 0.0]  # Non-negative, not all zero
+    graph.vs["attr2"] = [0.5, 1.5, 0.0, 1.0]  # Non-negative, not all zero
+    graph.add_edges([(0, 1), (1, 2), (2, 3), (0, 3)])  # Create connected graph
+
+    # Test 1: Basic functionality with two attributes
+    result = net_propagate_attributes(graph, ["attr1", "attr2"])
+
+    # Check structure
+    assert isinstance(result, pd.DataFrame)
+    assert result.shape == (4, 2)
+    assert list(result.index) == ["node1", "node2", "node3", "node4"]
+    assert list(result.columns) == ["attr1", "attr2"]
+
+    # Check that values are valid probabilities (PPR returns probabilities)
+    assert np.all(result.values >= 0)
+    assert np.all(result.values <= 1)
+    # Each column should sum to approximately 1 (PPR property)
+    assert np.allclose(result.sum(axis=0), [1.0, 1.0], atol=1e-10)
+
+    # Test 2: Single attribute
+    result_single = net_propagate_attributes(graph, ["attr1"])
+    assert result_single.shape == (4, 1)
+    assert list(result_single.columns) == ["attr1"]
+
+    # Test 3: Graph without names (should use indices)
+    graph_no_names = ig.Graph(3)
+    graph_no_names.vs["attr1"] = [1.0, 2.0, 1.0]
+    graph_no_names.add_edges([(0, 1), (1, 2)])
+
+    result_no_names = net_propagate_attributes(graph_no_names, ["attr1"])
+    assert list(result_no_names.index) == [0, 1, 2]  # Should use integer indices
 
+    # Test 4: Invalid propagation method
+    with pytest.raises(ValueError, match="Invalid propagation method"):
+        net_propagate_attributes(graph, ["attr1"], propagation_method="invalid_method")
 
-def test_personalized_pagerank_by_attribute_all_zero():
-    g = ig.Graph.Full(3)
-    g.vs["score"] = [0, 0, 0]
+    # Test 5: Additional arguments (test damping parameter)
+    result_default = net_propagate_attributes(graph, ["attr1"])
+    result_damped = net_propagate_attributes(
+        graph, ["attr1"], additional_propagation_args={"damping": 0.5}  # Lower damping
+    )
+
+    # Results should be different with different damping
+    assert not np.allclose(result_default.values, result_damped.values)
+
+    # Test 6: Invalid attribute (should be caught by internal validation)
+    graph.vs["bad_attr"] = [-1.0, 1.0, 2.0, 0.0]  # Has negative values
+    with pytest.raises(ValueError, match="contains negative values"):
+        net_propagate_attributes(graph, ["bad_attr"])
+
+    # Test 7: Zero attribute (should be caught by internal validation)
+    graph.vs["zero_attr"] = [0.0, 0.0, 0.0, 0.0]
     with pytest.raises(ValueError, match="zero for all vertices"):
-        personalized_pagerank_by_attribute(g, "score")
+        net_propagate_attributes(graph, ["zero_attr"])
+
+
+def test_all_null_generators_structure():
+    """Test all null generators with default options and validate output structure."""
+    # Create test graph with edges for realistic propagation
+    graph = ig.Graph(5)
+    graph.vs[NAPISTU_GRAPH_VERTICES.NAME] = ["A", "B", "C", "D", "E"]
+    graph.vs["attr1"] = [1.0, 0.0, 2.0, 0.0, 1.5]  # Non-negative, not all zero
+    graph.vs["attr2"] = [0.5, 1.0, 0.0, 2.0, 0.0]  # Non-negative, not all zero
+    graph.add_edges([(0, 1), (1, 2), (2, 3), (3, 4)])
+
+    attributes = ["attr1", "attr2"]
+    n_samples = 3  # Small for testing
+
+    for generator_name, generator_func in NULL_GENERATORS.items():
+        print(f"Testing {generator_name}")
+
+        if generator_name == NULL_STRATEGIES.UNIFORM:
+            # Uniform null doesn't take n_samples
+            result = generator_func(graph, attributes)
+            expected_rows = 5  # One row per node
+        elif generator_name == NULL_STRATEGIES.EDGE_PERMUTATION:
+            # Edge permutation has different parameters
+            result = generator_func(graph, attributes, n_samples=n_samples)
+            expected_rows = n_samples * 5  # n_samples rows per node
+        else:
+            # Gaussian and node_permutation
+            result = generator_func(graph, attributes, n_samples=n_samples)
+            expected_rows = n_samples * 5  # n_samples rows per node
+
+        # Validate structure
+        assert isinstance(
+            result, pd.DataFrame
+        ), f"{generator_name} should return DataFrame"
+        assert result.shape == (
+            expected_rows,
+            2,
+        ), f"{generator_name} wrong shape: {result.shape}"
+        assert list(result.columns) == attributes, f"{generator_name} wrong columns"
+
+        # Validate index structure
+        if generator_name == NULL_STRATEGIES.UNIFORM:
+            assert list(result.index) == [
+                "A",
+                "B",
+                "C",
+                "D",
+                "E",
+            ], f"{generator_name} wrong index"
+        else:
+            expected_index = ["A", "B", "C", "D", "E"] * n_samples
+            assert (
+                list(result.index) == expected_index
+            ), f"{generator_name} wrong repeated index"
+
+        # Validate values are numeric and finite (propagated outputs should be valid probabilities)
+        assert result.isna().sum().sum() == 0, f"{generator_name} contains NaN values"
+        assert np.isfinite(
+            result.values
+        ).all(), f"{generator_name} contains infinite values"
+        assert (result.values >= 0).all(), f"{generator_name} contains negative values"
+        assert (
+            result.values <= 1
+        ).all(), f"{generator_name} should contain probabilities <= 1"
+
+        # Each sample should sum to approximately 1 (PPR property)
+        if generator_name == NULL_STRATEGIES.UNIFORM:
+            assert np.allclose(
+                result.sum(axis=0), [1.0, 1.0], atol=1e-10
+            ), f"{generator_name} doesn't sum to 1"
+        else:
+            # For multiple samples, each individual sample should sum to 1
+            for i in range(n_samples):
+                start_idx = i * 5
+                end_idx = (i + 1) * 5
+                sample_data = result.iloc[start_idx:end_idx]
+                assert np.allclose(
+                    sample_data.sum(axis=0), [1.0, 1.0], atol=1e-10
+                ), f"{generator_name} sample {i} doesn't sum to 1"
+
+
+def test_mask_application():
+    """Test that masks are correctly applied across all null generators."""
+    # Create test graph
+    graph = ig.Graph(6)
+    graph.vs[NAPISTU_GRAPH_VERTICES.NAME] = ["A", "B", "C", "D", "E", "F"]
+    graph.vs["attr1"] = [1.0, 0.0, 2.0, 0.0, 1.5, 0.0]  # Nonzero at indices 0, 2, 4
+    graph.vs["attr2"] = [0.0, 1.0, 0.0, 2.0, 0.0, 1.0]  # Nonzero at indices 1, 3, 5
+    graph.add_edges([(0, 1), (1, 2), (2, 3), (3, 4), (4, 5)])
+
+    attributes = ["attr1", "attr2"]
+
+    # Test mask that includes nodes with nonzero values for both attributes
+    # Use nodes 0, 1, 2, 3 which covers nonzero values for both attributes
+    mask_array = np.array([True, True, True, True, False, False])  # Nodes 0, 1, 2, 3
+
+    for generator_name, generator_func in NULL_GENERATORS.items():
+        print(f"Testing mask application for {generator_name}")
+
+        if generator_name == NULL_STRATEGIES.UNIFORM:
+            result = generator_func(graph, attributes, mask=mask_array)
+
+            # For uniform null with mask, verify structure is correct
+            assert result.shape == (6, 2), f"{generator_name} wrong shape with mask"
+            # After propagation, all nodes will have some value due to network effect
+            assert (
+                result.values > 0
+            ).all(), "All nodes should have positive values after propagation"
+
+        elif generator_name == NULL_STRATEGIES.EDGE_PERMUTATION:
+            # Edge permutation ignores mask, just test it doesn't crash
+            result = generator_func(graph, attributes, n_samples=2)
+            assert result.shape[0] == 12  # 2 samples * 6 nodes
+
+        else:
+            # Gaussian and node_permutation with mask
+            result = generator_func(graph, attributes, mask=mask_array, n_samples=2)
+
+            # Check that structure is maintained
+            assert result.shape == (12, 2)  # 2 samples * 6 nodes
+
+
+def test_edge_cases_and_errors():
+    """Test edge cases and error conditions for null generators."""
+    # Create minimal test graph
+    graph = ig.Graph(3)
+    graph.vs["attr1"] = [1.0, 2.0, 0.0]
+    graph.vs["bad_attr"] = [0.0, 0.0, 0.0]  # All zeros
+    graph.add_edges([(0, 1), (1, 2)])
+
+    # Test 1: All zero attribute should raise error for all generators
+    with pytest.raises(ValueError):
+        uniform_null(graph, ["bad_attr"])
+
+    with pytest.raises(ValueError):
+        parametric_null(graph, ["bad_attr"])
+
+    with pytest.raises(ValueError):
+        node_permutation_null(graph, ["bad_attr"])
+
+    with pytest.raises(ValueError):
+        edge_permutation_null(graph, ["bad_attr"])
+
+    # Test 2: Empty mask should raise error
+    empty_mask = np.array([False, False, False])
+    with pytest.raises(ValueError, match="No nodes in mask"):
+        uniform_null(graph, ["attr1"], mask=empty_mask)
+
+    # Test 3: Single node mask (edge case)
+    single_mask = np.array([True, False, False])
+    result = uniform_null(graph, ["attr1"], mask=single_mask)
+    assert result.shape == (3, 1)  # Should work
+
+    # Test 4: Replace parameter in node permutation
+    result_no_replace = node_permutation_null(
+        graph, ["attr1"], replace=False, n_samples=2
+    )
+    result_replace = node_permutation_null(graph, ["attr1"], replace=True, n_samples=2)
+
+    # Both should have same structure
+    assert result_no_replace.shape == result_replace.shape
+
+
+def test_propagation_method_parameters():
+    """Test that propagation method and additional arguments are properly passed through."""
+    # Create test graph
+    graph = ig.Graph(4)
+    graph.vs["attr1"] = [1.0, 2.0, 0.0, 1.5]
+    graph.add_edges([(0, 1), (1, 2), (2, 3)])
+
+    # Test different damping parameters produce different results
+    result_default = uniform_null(graph, ["attr1"])
+    result_damped = uniform_null(
+        graph, ["attr1"], additional_propagation_args={"damping": 0.5}
+    )
+
+    # Results should be different with different damping
+    assert not np.allclose(
+        result_default.values, result_damped.values
+    ), "Different damping should produce different results"
+
+    # Test that all generators accept method parameters
+    for generator_name, generator_func in NULL_GENERATORS.items():
+        if generator_name == NULL_STRATEGIES.UNIFORM:
+            result = generator_func(
+                graph, ["attr1"], additional_propagation_args={"damping": 0.8}
+            )
+        else:
+            result = generator_func(
+                graph,
+                ["attr1"],
+                additional_propagation_args={"damping": 0.8},
+                n_samples=2,
+            )
+
+        # Should produce valid results
+        assert isinstance(result, pd.DataFrame)
+        assert not result.empty

tests/test_statistics_quantiles.py

@@ -0,0 +1,133 @@
+import pytest
+import numpy as np
+import pandas as pd
+from napistu.statistics import quantiles
+
+
+def test_calculate_quantiles_valid_inputs():
+    """Test calculate_quantiles with valid, well-formed inputs."""
+    # Create observed data: 4 features x 3 attributes
+    observed = pd.DataFrame(
+        [[0.8, 0.3, 0.9], [0.2, 0.7, 0.1], [0.5, 0.5, 0.5], [0.1, 0.9, 0.2]],
+        index=["gene1", "gene2", "gene3", "gene4"],
+        columns=["attr1", "attr2", "attr3"],
+    )
+
+    # Create null data: 2 samples per feature (8 rows total)
+    null_index = ["gene1", "gene2", "gene3", "gene4"] * 2
+    null_data = pd.DataFrame(
+        [
+            [0.1, 0.2, 0.3],  # gene1 sample 1
+            [0.4, 0.5, 0.6],  # gene2 sample 1
+            [0.7, 0.8, 0.9],  # gene3 sample 1
+            [0.0, 0.1, 0.2],  # gene4 sample 1
+            [0.2, 0.3, 0.4],  # gene1 sample 2
+            [0.5, 0.6, 0.7],  # gene2 sample 2
+            [0.8, 0.9, 1.0],  # gene3 sample 2
+            [0.1, 0.2, 0.3],  # gene4 sample 2
+        ],
+        index=null_index,
+        columns=["attr1", "attr2", "attr3"],
+    )
+
+    # Calculate quantiles
+    result = quantiles.calculate_quantiles(observed, null_data)
+
+    # Verify output structure
+    assert result.shape == observed.shape
+    assert list(result.index) == list(observed.index)
+    assert list(result.columns) == list(observed.columns)
+
+    # Check specific quantile calculations
+    # gene1, attr1: observed=0.8, nulls=[0.1, 0.2] -> quantile = 1.0 (100%)
+    assert result.loc["gene1", "attr1"] == 1.0
+
+    # gene2, attr2: observed=0.7, nulls=[0.5, 0.6] -> quantile = 1.0 (100%)
+    assert result.loc["gene2", "attr2"] == 1.0
+
+    # gene3, attr3: observed=0.5, nulls=[0.9, 1.0] -> quantile = 0.0 (0%)
+    assert result.loc["gene3", "attr3"] == 0.0
+
+    # gene4, attr1: observed=0.1, nulls=[0.0, 0.1]
+    # With ≤: 0.0 ≤ 0.1 (True), 0.1 ≤ 0.1 (True) → 2/2 = 1.0
+    assert result.loc["gene4", "attr1"] == 1.0
+
+
+def test_calculate_quantiles_error_cases():
+    """Test calculate_quantiles with invalid inputs that should raise errors or warnings."""
+    # Base observed data
+    observed = pd.DataFrame(
+        [[0.8, 0.3], [0.2, 0.7]], index=["gene1", "gene2"], columns=["attr1", "attr2"]
+    )
+
+    # Test 1: Mismatched columns
+    null_wrong_cols = pd.DataFrame(
+        [[0.1, 0.2, 0.3], [0.4, 0.5, 0.6]],
+        index=["gene1", "gene2"],
+        columns=["attr1", "attr2", "attr3"],  # Extra column
+    )
+
+    with pytest.raises((KeyError, ValueError)):
+        quantiles.calculate_quantiles(observed, null_wrong_cols)
+
+    # Test 2: Missing features in null data
+    null_missing_feature = pd.DataFrame(
+        [[0.1, 0.2]], index=["gene1"], columns=["attr1", "attr2"]  # Missing gene2
+    )
+
+    # Current implementation doesn't validate - it will likely fail in groupby or indexing
+    # This test verifies current behavior (may change if validation added)
+    try:
+        result = quantiles.calculate_quantiles(observed, null_missing_feature)
+        # If it succeeds, gene2 quantiles will be invalid/error
+        assert True  # Just check it doesn't crash for now
+    except (KeyError, ValueError, IndexError):
+        assert True  # Expected behavior
+
+    # Test 3: Unequal null samples per feature
+    null_unequal_samples = pd.DataFrame(
+        [
+            [0.1, 0.2],  # gene1 sample 1
+            [0.3, 0.4],  # gene1 sample 2
+            [0.5, 0.6],  # gene2 sample 1 (only 1 sample)
+        ],
+        index=["gene1", "gene1", "gene2"],
+        columns=["attr1", "attr2"],
+    )
+
+    # This should still work but may give different results
+    result = quantiles.calculate_quantiles(observed, null_unequal_samples)
+    assert result.shape == observed.shape
+
+    # Test 4: Empty null data
+    null_empty = pd.DataFrame(columns=["attr1", "attr2"])
+
+    with pytest.raises((ValueError, IndexError)):
+        quantiles.calculate_quantiles(observed, null_empty)
+
+    # Test 5: Single null sample (edge case)
+    null_single = pd.DataFrame(
+        [[0.1, 0.2], [0.5, 0.6]], index=["gene1", "gene2"], columns=["attr1", "attr2"]
+    )
+
+    result = quantiles.calculate_quantiles(observed, null_single)
+    assert result.shape == observed.shape
+    # With single sample, results should be binary (0 or 1)
+    assert all(val in [0.0, 1.0] for val in result.values.flatten())
+
+    # Test 6: NaN values in data
+    observed_with_nan = observed.copy()
+    observed_with_nan.loc["gene1", "attr1"] = np.nan
+
+    null_with_nan = pd.DataFrame(
+        [[np.nan, 0.2], [0.4, 0.5], [0.1, 0.3], [0.6, 0.7]],
+        index=["gene1", "gene2", "gene1", "gene2"],
+        columns=["attr1", "attr2"],
+    )
+
+    # Should raise ValueError for NaN values
+    with pytest.raises(ValueError, match="NaN values found in observed data"):
+        quantiles.calculate_quantiles(observed_with_nan, null_single)
+
+    with pytest.raises(ValueError, match="NaN values found in null data"):
+        quantiles.calculate_quantiles(observed, null_with_nan)