diff --git a/src/rapids_singlecell/pertpy_gpu/__init__.py b/src/rapids_singlecell/pertpy_gpu/__init__.py
new file mode 100644
index 00000000..90b92ca7
--- /dev/null
+++ b/src/rapids_singlecell/pertpy_gpu/__init__.py
@@ -0,0 +1,3 @@
+from __future__ import annotations
+
+from ._edistance import pertpy_edistance
diff --git a/src/rapids_singlecell/pertpy_gpu/_edistance.py b/src/rapids_singlecell/pertpy_gpu/_edistance.py
new file mode 100644
index 00000000..613c2384
--- /dev/null
+++ b/src/rapids_singlecell/pertpy_gpu/_edistance.py
@@ -0,0 +1,445 @@
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, NamedTuple
+
+import cupy as cp
+import numpy as np
+import pandas as pd
+
+from rapids_singlecell.pertpy_gpu._kernels._edistance import (
+ get_compute_group_distances_kernel,
+)
+from rapids_singlecell.preprocessing._harmony._helper import (
+ _create_category_index_mapping,
+)
+from rapids_singlecell.squidpy_gpu._utils import _assert_categorical_obs
+
+if TYPE_CHECKING:
+ from anndata import AnnData
+
+
+class EDistanceResult(NamedTuple):
+ distances: pd.DataFrame
+ distances_var: pd.DataFrame | None
+
+
+compute_group_distances_kernel = get_compute_group_distances_kernel()
+
+
+def pertpy_edistance(
+ adata: AnnData,
+ groupby: str,
+ *,
+ obsm_key: str = "X_pca",
+ groups: list[str] | None = None,
+ inplace: bool = False,
+ bootstrap: bool = False,
+ n_bootstrap: int = 100,
+ random_state: int = 0,
+) -> pd.DataFrame:
+ """
+ GPU-accelerated pairwise edistance computation with decomposed components.
+
+ Returns d_itself, d_other arrays and final edistance DataFrame where:
+ df[a,b] = 2*d_other[a,b] - d_itself[a] - d_itself[b]
+
+ Parameters
+ ----------
+ adata : AnnData
+ Annotated data matrix
+ groupby : str
+ Key in adata.obs for grouping
+ obsm_key : str
+ Key in adata.obsm for embeddings
+ groups : list[str] | None
+ Specific groups to compute (if None, use all)
+ copy : bool
+ Whether to return a copy
+
+ Returns
+ -------
+ df : pd.DataFrame
+ Final edistance matrix
+ """
+ _assert_categorical_obs(adata, key=groupby)
+
+ embedding = cp.array(adata.obsm[obsm_key]).astype(np.float32)
+ original_groups = adata.obs[groupby]
+ group_map = {v: i for i, v in enumerate(original_groups.cat.categories.values)}
+ group_labels = cp.array([group_map[c] for c in original_groups], dtype=cp.int32)
+
+ # Use harmony's category mapping
+ k = len(group_map)
+ cat_offsets, cell_indices = _create_category_index_mapping(group_labels, k)
+
+ groups_list = (
+ list(original_groups.cat.categories.values) if groups is None else groups
+ )
+ result = None
+ if not bootstrap:
+ df = _prepare_edistance_df(
+ embedding=embedding,
+ cat_offsets=cat_offsets,
+ cell_indices=cell_indices,
+ k=k,
+ groups_list=groups_list,
+ groupby=groupby,
+ )
+ result = EDistanceResult(distances=df, distances_var=None)
+
+ else:
+ df, df_var = _prepare_edistance_df_bootstrap(
+ embedding=embedding,
+ cat_offsets=cat_offsets,
+ cell_indices=cell_indices,
+ k=k,
+ groups_list=groups_list,
+ groupby=groupby,
+ n_bootstrap=n_bootstrap,
+ random_state=random_state,
+ )
+ result = EDistanceResult(distances=df, distances_var=df_var)
+
+ if inplace:
+ adata.uns[f"{groupby}_pairwise_edistance"] = dict(result)
+
+ return result
+
+
+def _pairwise_means(
+ embedding: cp.ndarray, cat_offsets: cp.ndarray, cell_indices: cp.ndarray, k: int
+) -> cp.ndarray:
+ """
+ Compute between-group mean distances for all group pairs.
+
+ Parameters
+ ----------
+ embedding : cp.ndarray
+ Cell embeddings [n_cells, n_features]
+ cat_offsets : cp.ndarray
+ Group start/end indices
+ cell_indices : cp.ndarray
+ Sorted cell indices by group
+ k : int
+ Number of groups
+
+ Returns
+ -------
+ d_other : cp.ndarray
+ Between-group mean distances [k, k]
+ """
+ _, n_features = embedding.shape
+
+ pair_left = []
+ pair_right = []
+ pair_indices = []
+ # only upper triangle
+ for a in range(k):
+ for b in range(a, k):
+ pair_left.append(a)
+ pair_right.append(b)
+ pair_indices.append(a * k + b) # Flatten matrix index
+
+ pair_left = cp.asarray(pair_left, dtype=cp.int32)
+ pair_right = cp.asarray(pair_right, dtype=cp.int32)
+ pair_indices = cp.asarray(pair_indices, dtype=cp.int32)
+
+ num_pairs = len(pair_left) # k * (k-1) pairs instead of k²
+
+ # Allocate output for off-diagonal distances only
+ d_other_offdiag = cp.zeros(num_pairs, dtype=embedding.dtype)
+
+ # Choose optimal block size
+ props = cp.cuda.runtime.getDeviceProperties(0)
+ max_smem = int(props.get("sharedMemPerBlock", 48 * 1024))
+
+ chosen_threads = None
+ shared_mem_size = 0 # TODO: think of a better way to do this
+ for tpb in (1024, 512, 256, 128, 64, 32):
+ required = tpb * cp.dtype(cp.float32).itemsize
+ if required <= max_smem:
+ chosen_threads = tpb
+ shared_mem_size = required
+ break
+
+ # Launch kernel - one block per OFF-DIAGONAL group pair only
+ grid = (num_pairs,)
+ block = (chosen_threads,)
+ compute_group_distances_kernel(
+ grid,
+ block,
+ (
+ embedding,
+ cat_offsets,
+ cell_indices,
+ pair_left,
+ pair_right,
+ d_other_offdiag,
+ k,
+ n_features,
+ ),
+ shared_mem=shared_mem_size,
+ )
+
+ # Build full k x k matrix
+ pairwise_means = cp.zeros((k, k), dtype=np.float32)
+
+ # Fill the full matrix
+ for i, idx in enumerate(pair_indices.get()):
+ a, b = divmod(idx, k)
+ pairwise_means[a, b] = d_other_offdiag[i]
+ pairwise_means[b, a] = d_other_offdiag[i]
+
+ return pairwise_means
+
+
+def _generate_bootstrap_indices(
+ cat_offsets: cp.ndarray,
+ k: int,
+ n_bootstrap: int = 100,
+ random_state: int = 0,
+) -> list[list[cp.ndarray]]:
+ """
+ Generate bootstrap indices for all groups and all bootstrap iterations.
+ This matches the CPU implementation's random sampling logic for reproducibility.
+
+ Parameters
+ ----------
+ cat_offsets : cp.ndarray
+ Group start/end indices
+ k : int
+ Number of groups
+ n_bootstrap : int
+ Number of bootstrap samples
+ random_state : int
+ Random seed for reproducibility
+
+ Returns
+ -------
+ bootstrap_indices : list[list[cp.ndarray]]
+ For each bootstrap iteration, list of indices arrays for each group
+ Shape: [n_bootstrap][k] where each element is cp.ndarray of group_size
+ """
+ import numpy as np
+
+ # Use same RNG logic as CPU code
+ rng = np.random.default_rng(random_state)
+
+ # Convert to numpy for CPU-based random generation
+ cat_offsets_np = cat_offsets.get()
+
+ bootstrap_indices = []
+
+ for bootstrap_iter in range(n_bootstrap):
+ group_indices = []
+
+ for group_idx in range(k):
+ start_idx = cat_offsets_np[group_idx]
+ end_idx = cat_offsets_np[group_idx + 1]
+ group_size = end_idx - start_idx
+
+ if group_size > 0:
+ # Generate bootstrap indices using same logic as CPU code
+ # rng.choice(a=X.shape[0], size=X.shape[0], replace=True)
+ bootstrap_group_indices = rng.choice(
+ group_size, size=group_size, replace=True
+ )
+ # Convert to CuPy array
+ group_indices.append(cp.array(bootstrap_group_indices, dtype=cp.int32))
+ else:
+ # Empty group
+ group_indices.append(cp.array([], dtype=cp.int32))
+
+ bootstrap_indices.append(group_indices)
+
+ return bootstrap_indices
+
+
+def _bootstrap_sample_cells_from_indices(
+ *,
+ cat_offsets: cp.ndarray,
+ cell_indices: cp.ndarray,
+ k: int,
+ bootstrap_group_indices: list[cp.ndarray],
+) -> tuple[cp.ndarray, cp.ndarray]:
+ """
+ Bootstrap sample cells using pre-generated indices.
+
+ Parameters
+ ----------
+ cat_offsets : cp.ndarray
+ Group start/end indices
+ cell_indices : cp.ndarray
+ Sorted cell indices by group
+ k : int
+ Number of groups
+ bootstrap_group_indices : list[cp.ndarray]
+ Pre-generated bootstrap indices for each group
+
+ Returns
+ -------
+ new_cat_offsets, new_cell_indices : tuple[cp.ndarray, cp.ndarray]
+ New category structure with bootstrapped cells
+ """
+ new_cell_indices = []
+ new_cat_offsets = cp.zeros(k + 1, dtype=cp.int32)
+
+ for group_idx in range(k):
+ start_idx = cat_offsets[group_idx]
+ end_idx = cat_offsets[group_idx + 1]
+ group_size = end_idx - start_idx
+
+ if group_size > 0:
+ # Get original cell indices for this group
+ group_cells = cell_indices[start_idx:end_idx]
+
+ # Use pre-generated bootstrap indices
+ bootstrap_indices = bootstrap_group_indices[group_idx]
+ bootstrap_cells = group_cells[bootstrap_indices]
+
+ new_cell_indices.extend(bootstrap_cells.get().tolist())
+
+ new_cat_offsets[group_idx + 1] = len(new_cell_indices)
+
+ return new_cat_offsets, cp.array(new_cell_indices, dtype=cp.int32)
+
+
+def _pairwise_means_bootstrap(
+ embedding: cp.ndarray,
+ *,
+ cat_offsets: cp.ndarray,
+ cell_indices: cp.ndarray,
+ k: int,
+ n_bootstrap: int = 100,
+ random_state: int = 0,
+) -> tuple[cp.ndarray, cp.ndarray]:
+ """
+ Compute bootstrap statistics for between-group distances.
+ Uses CPU-compatible random generation for reproducibility.
+
+ Returns:
+ means: [k, k] matrix of bootstrap means
+ variances: [k, k] matrix of bootstrap variances
+ """
+ # Generate all bootstrap indices upfront using CPU-compatible logic
+ bootstrap_indices = _generate_bootstrap_indices(
+ cat_offsets, k, n_bootstrap, random_state
+ )
+
+ bootstrap_results = []
+
+ for bootstrap_iter in range(n_bootstrap):
+ # Use pre-generated indices for this bootstrap iteration
+ boot_cat_offsets, boot_cell_indices = _bootstrap_sample_cells_from_indices(
+ cat_offsets=cat_offsets,
+ cell_indices=cell_indices,
+ k=k,
+ bootstrap_group_indices=bootstrap_indices[bootstrap_iter],
+ )
+
+ # Compute distances with bootstrapped samples
+ pairwise_means = _pairwise_means(
+ embedding=embedding,
+ cat_offsets=boot_cat_offsets,
+ cell_indices=boot_cell_indices,
+ k=k,
+ )
+ bootstrap_results.append(pairwise_means.get())
+
+ # Compute statistics across bootstrap samples
+ bootstrap_stack = cp.array(bootstrap_results) # [n_bootstrap, k, k]
+ means = cp.mean(bootstrap_stack, axis=0)
+ variances = cp.var(bootstrap_stack, axis=0)
+
+ return means, variances
+
+
+def _prepare_edistance_df_bootstrap(
+ embedding: cp.ndarray,
+ *,
+ cat_offsets: cp.ndarray,
+ cell_indices: cp.ndarray,
+ k: int,
+ groups_list: list[str],
+ groupby: str,
+ n_bootstrap: int = 100,
+ random_state: int = 0,
+) -> tuple[pd.DataFrame, pd.DataFrame]:
+ # Bootstrap computation
+ pairwise_means_boot, pairwise_vars_boot = _pairwise_means_bootstrap(
+ embedding=embedding,
+ cat_offsets=cat_offsets,
+ cell_indices=cell_indices,
+ k=k,
+ n_bootstrap=n_bootstrap,
+ random_state=random_state,
+ )
+
+ # Compute final edistance for means and variances
+ edistance_means = cp.zeros((k, k), dtype=np.float32)
+ edistance_vars = cp.zeros((k, k), dtype=np.float32)
+
+ for a in range(k):
+ for b in range(a + 1, k):
+ # Bootstrap mean edistance
+ edistance_means[a, b] = (
+ 2 * pairwise_means_boot[a, b]
+ - pairwise_means_boot[a, a]
+ - pairwise_means_boot[b, b]
+ )
+ edistance_means[b, a] = edistance_means[a, b]
+
+ # Bootstrap variance edistance (using delta method approximation)
+ # Var(2*X - Y - Z) = 4*Var(X) + Var(Y) + Var(Z) (assuming independence)
+ edistance_vars[a, b] = (
+ 4 * pairwise_vars_boot[a, b]
+ + pairwise_vars_boot[a, a]
+ + pairwise_vars_boot[b, b]
+ )
+ edistance_vars[b, a] = edistance_vars[a, b]
+
+ # 5. Create output DataFrames
+
+ df_mean = pd.DataFrame(
+ edistance_means.get(), index=groups_list, columns=groups_list
+ )
+ df_mean.index.name = groupby
+ df_mean.columns.name = groupby
+ df_mean.name = "pairwise edistance"
+
+ df_var = pd.DataFrame(edistance_vars.get(), index=groups_list, columns=groups_list)
+ df_var.index.name = groupby
+ df_var.columns.name = groupby
+ df_var.name = "pairwise edistance variance"
+
+ return df_mean, df_var
+
+
+def _prepare_edistance_df(
+ embedding: cp.ndarray,
+ *,
+ cat_offsets: cp.ndarray,
+ cell_indices: cp.ndarray,
+ k: int,
+ groups_list: list[str],
+ groupby: str,
+) -> pd.DataFrame:
+ # Compute means
+ pairwise_means = _pairwise_means(embedding, cat_offsets, cell_indices, k)
+
+ # Compute final edistance: df[a,b] = 2*d_other[a,b] - d_itself[a] - d_itself[b]
+ edistance_matrix = cp.zeros((k, k), dtype=np.float32)
+ for a in range(k):
+ for b in range(a + 1, k):
+ edistance_matrix[a, b] = (
+ 2 * pairwise_means[a, b] - pairwise_means[a, a] - pairwise_means[b, b]
+ )
+ edistance_matrix[b, a] = edistance_matrix[a, b]
+
+ # Create output DataFrame
+ df = pd.DataFrame(edistance_matrix.get(), index=groups_list, columns=groups_list)
+ df.index.name = groupby
+ df.columns.name = groupby
+ df.name = "pairwise edistance"
+
+ return df
diff --git a/src/rapids_singlecell/pertpy_gpu/_kernels/_edistance.py b/src/rapids_singlecell/pertpy_gpu/_kernels/_edistance.py
new file mode 100644
index 00000000..d669c2da
--- /dev/null
+++ b/src/rapids_singlecell/pertpy_gpu/_kernels/_edistance.py
@@ -0,0 +1,72 @@
+from __future__ import annotations
+
+from cuml.common.kernel_utils import cuda_kernel_factory
+
+_compute_group_distances_kernel = r"""
+(const float* __restrict__ embedding,
+ const int* __restrict__ cat_offsets,
+ const int* __restrict__ cell_indices,
+ const int* __restrict__ pair_left,
+ const int* __restrict__ pair_right,
+ float* __restrict__ d_other,
+ int k,
+ int n_features) {
+ extern __shared__ float shared_sums[];
+
+ const int thread_id = threadIdx.x;
+ const int block_id = blockIdx.x;
+ const int block_size = blockDim.x;
+
+ float local_sum = 0.0f;
+
+ const int a = pair_left[block_id];
+ const int b = pair_right[block_id];
+
+ const int start_a = cat_offsets[a];
+ const int end_a = cat_offsets[a + 1];
+ const int start_b = cat_offsets[b];
+ const int end_b = cat_offsets[b + 1];
+
+ const int n_a = end_a - start_a;
+ const int n_b = end_b - start_b;
+
+ for (int ia = start_a + thread_id; ia < end_a; ia += block_size) {
+ const int idx_i = cell_indices[ia];
+
+ for (int jb = start_b; jb < end_b; ++jb) {
+ const int idx_j = cell_indices[jb];
+
+ double dist_sq = 0.0;
+ for (int feat = 0; feat < n_features; ++feat) {
+ double diff = (double)embedding[idx_i * n_features + feat] -
+ (double)embedding[idx_j * n_features + feat];
+ dist_sq += diff * diff;
+ }
+
+ local_sum += (float)sqrt(dist_sq);
+ }
+ }
+
+ shared_sums[thread_id] = local_sum;
+ __syncthreads();
+
+ for (int stride = block_size / 2; stride > 0; stride >>= 1) {
+ if (thread_id < stride) {
+ shared_sums[thread_id] += shared_sums[thread_id + stride];
+ }
+ __syncthreads();
+ }
+
+ if (thread_id == 0) {
+ d_other[block_id] = shared_sums[0] / (float)(n_a * n_b);
+ }
+}
+"""
+
+
+def get_compute_group_distances_kernel():
+ return cuda_kernel_factory(
+ _compute_group_distances_kernel,
+ (),
+ "compute_group_distances",
+ )
diff --git a/src/rapids_singlecell/ptg.py b/src/rapids_singlecell/ptg.py
new file mode 100644
index 00000000..2358f441
--- /dev/null
+++ b/src/rapids_singlecell/ptg.py
@@ -0,0 +1,3 @@
+from __future__ import annotations
+
+from .pertpy_gpu import *
diff --git a/tests/pertpy/test_distances.py b/tests/pertpy/test_distances.py
new file mode 100644
index 00000000..23a7a557
--- /dev/null
+++ b/tests/pertpy/test_distances.py
@@ -0,0 +1,126 @@
+from __future__ import annotations
+
+import cupy as cp
+import numpy as np
+import pandas as pd
+import pytest
+from anndata import AnnData
+
+from rapids_singlecell.pertpy_gpu._edistance import EDistanceResult, pertpy_edistance
+
+
+@pytest.fixture
+def small_adata() -> AnnData:
+ rng = np.random.default_rng(0)
+ n_groups = 3
+ cells_per_group = 4
+ n_features = 5
+ total_cells = n_groups * cells_per_group
+
+ cpu_embedding = rng.normal(size=(total_cells, n_features)).astype(np.float32)
+ groups = [f"g{idx}" for idx in range(n_groups) for _ in range(cells_per_group)]
+ obs = pd.DataFrame(
+ {"group": pd.Categorical(groups, categories=[f"g{i}" for i in range(n_groups)])}
+ )
+
+ adata = AnnData(cpu_embedding.copy(), obs=obs)
+ adata.obsm["X_pca"] = cp.asarray(cpu_embedding, dtype=cp.float32)
+ return adata
+
+
+def _compute_cpu_reference(
+ adata: AnnData, obsm_key: str, group_key: str
+) -> tuple[np.ndarray, np.ndarray]:
+ embedding = adata.obsm[obsm_key].get()
+ group_series = adata.obs[group_key]
+ categories = list(group_series.cat.categories)
+ k = len(categories)
+
+ pair_means = np.zeros((k, k), dtype=np.float32)
+
+ for i, gi in enumerate(categories):
+ idx_i = np.where(group_series == gi)[0]
+ for j, gj in enumerate(categories[i:], start=i):
+ idx_j = np.where(group_series == gj)[0]
+ if len(idx_i) == 0 or len(idx_j) == 0:
+ mean_distance = 0.0
+ else:
+ distances = []
+ for idx in idx_i:
+ diffs = embedding[idx] - embedding[idx_j]
+ distances.append(np.sqrt(np.sum(diffs**2, axis=1)))
+ stacked = np.concatenate(distances)
+ mean_distance = stacked.mean(dtype=np.float64)
+ pair_means[i, j] = pair_means[j, i] = np.float32(mean_distance)
+
+ edistance = np.zeros((k, k), dtype=np.float32)
+ for i in range(k):
+ for j in range(i + 1, k):
+ value = 2 * pair_means[i, j] - pair_means[i, i] - pair_means[j, j]
+ edistance[i, j] = edistance[j, i] = np.float32(value)
+
+ return pair_means, edistance
+
+
+def test_pertpy_edistance_matches_cpu_reference(small_adata: AnnData) -> None:
+ result = pertpy_edistance(small_adata, groupby="group", obsm_key="X_pca")
+
+ assert isinstance(result, EDistanceResult)
+ assert result.distances_var is None
+
+ _, cpu_edistance = _compute_cpu_reference(small_adata, "X_pca", "group")
+
+ assert result.distances.shape == cpu_edistance.shape
+ np.testing.assert_allclose(result.distances.values, cpu_edistance, atol=1e-5)
+ assert np.allclose(result.distances.values, result.distances.values.T)
+
+
+def test_pertpy_edistance_inplace_populates_uns(small_adata: AnnData) -> None:
+ key = "group_pairwise_edistance"
+ result = pertpy_edistance(
+ small_adata,
+ groupby="group",
+ obsm_key="X_pca",
+ inplace=True,
+ )
+
+ assert isinstance(result, EDistanceResult)
+ assert key in small_adata.uns
+ stored = small_adata.uns[key]
+ assert set(stored.keys()) == {"distances", "distances_var"}
+ np.testing.assert_allclose(stored["distances"].values, result.distances.values)
+ assert stored["distances_var"] is None
+
+
+def test_pertpy_edistance_bootstrap_returns_variance(small_adata: AnnData) -> None:
+ result = pertpy_edistance(
+ small_adata,
+ groupby="group",
+ obsm_key="X_pca",
+ bootstrap=True,
+ n_bootstrap=8,
+ random_state=11,
+ )
+
+ assert isinstance(result, EDistanceResult)
+ assert result.distances_var is not None
+ assert result.distances.shape == result.distances_var.shape
+ assert np.allclose(result.distances.values, result.distances.values.T)
+ assert np.allclose(result.distances_var.values, result.distances_var.values.T)
+ assert np.all(result.distances_var.values >= 0)
+
+
+def test_pertpy_edistance_requires_categorical_obs(small_adata: AnnData) -> None:
+ bad = small_adata.copy()
+ bad.obs["group"] = bad.obs["group"].astype(str)
+
+ with pytest.raises(TypeError):
+ pertpy_edistance(bad, groupby="group", obsm_key="X_pca")
+
+
+@pytest.mark.parametrize("missing_key", ["missing", "other"])
+def test_pertpy_edistance_missing_group_raises(
+ small_adata: AnnData, missing_key: str
+) -> None:
+ with pytest.raises(KeyError):
+ pertpy_edistance(small_adata, groupby=missing_key, obsm_key="X_pca")
diff --git a/tmp_scripts/a.ipynb b/tmp_scripts/a.ipynb
new file mode 100644
index 00000000..a1cb90b9
--- /dev/null
+++ b/tmp_scripts/a.ipynb
@@ -0,0 +1,1259 @@
+{
+ "cells": [
+ {
+ "cell_type": "code",
+ "execution_count": 15,
+ "id": "e1d4de19",
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "from __future__ import annotations\n",
+ "\n",
+ "import os\n",
+ "\n",
+ "import anndata as ad\n",
+ "import cupy as cp\n",
+ "import pertpy as pt\n",
+ "import rmm\n",
+ "from rmm.allocators.cupy import rmm_cupy_allocator\n",
+ "\n",
+ "import rapids_singlecell as rsc\n",
+ "from rapids_singlecell.ptg import Distance\n",
+ "\n",
+ "rmm.reinitialize(\n",
+ " managed_memory=False, # Allows oversubscription\n",
+ " pool_allocator=True, # default is False\n",
+ " devices=0, # GPU device IDs to register. By default registers only GPU 0.\n",
+ ")\n",
+ "cp.cuda.set_allocator(rmm_cupy_allocator)"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 16,
+ "id": "e9c83bff",
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "save_dir = os.path.join(\n",
+ " os.path.expanduser(\"~\"), \"data\", \"adamson_2016_upr_epistasis_pca.h5ad\"\n",
+ ")"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 17,
+ "id": "ea369599",
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "adata = ad.read_h5ad(save_dir)"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "df4921a8",
+ "metadata": {},
+ "outputs": [
+ {
+ "data": {
+ "text/plain": [
+ "['3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " '*',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'PERK_only_pMJ146',\n",
+ " nan,\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " nan,\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " nan,\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " nan,\n",
+ " 'ATF6_only_pMJ145',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " nan,\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " nan,\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " nan,\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " nan,\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " nan,\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '*',\n",
+ " 'XBP1_pBA578',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " nan,\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " nan,\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_only_pMJ146',\n",
+ " nan,\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " nan,\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_only_pMJ145',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_PERK_pMJ150',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'PERK_only_pMJ146',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_IRE1_pMJ154',\n",
+ " 'IRE1_only_pMJ148',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " '3x_neg_ctrl_pMJ144-2',\n",
+ " '3x_neg_ctrl_pMJ144-1',\n",
+ " 'PERK_only_pMJ146',\n",
+ " 'ATF6_PERK_IRE1_pMJ158',\n",
+ " 'ATF6_IRE1_pMJ152',\n",
+ " ...]"
+ ]
+ },
+ "execution_count": 23,
+ "metadata": {},
+ "output_type": "execute_result"
+ }
+ ],
+ "source": []
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 4,
+ "id": "48beeb95",
+ "metadata": {},
+ "outputs": [],
+ "source": [
+ "rsc.get.anndata_to_GPU(adata, convert_all=True)"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 9,
+ "id": "553326ea",
+ "metadata": {},
+ "outputs": [
+ {
+ "name": "stderr",
+ "output_type": "stream",
+ "text": [
+ "/home/shadeform/rapids_singlecell/src/rapids_singlecell/pertpy_gpu/_distances.py:730: RuntimeWarning: Mean of empty slice.\n",
+ " sigma_Y = P[~idx, :][:, ~idx].mean()\n",
+ "/home/shadeform/miniforge3/envs/pertpy/lib/python3.13/site-packages/numpy/_core/_methods.py:145: RuntimeWarning: invalid value encountered in divide\n",
+ " ret = ret.dtype.type(ret / rcount)\n",
+ "/home/shadeform/rapids_singlecell/src/rapids_singlecell/pertpy_gpu/_distances.py:731: RuntimeWarning: Mean of empty slice.\n",
+ " delta = P[idx, :][:, ~idx].mean()\n"
+ ]
+ },
+ {
+ "name": "stdout",
+ "output_type": "stream",
+ "text": [
+ "CPU times: user 6.32 s, sys: 1.88 s, total: 8.2 s\n",
+ "Wall time: 8.19 s\n"
+ ]
+ },
+ {
+ "name": "stderr",
+ "output_type": "stream",
+ "text": [
+ "/home/shadeform/rapids_singlecell/src/rapids_singlecell/pertpy_gpu/_distances.py:729: RuntimeWarning: Mean of empty slice.\n",
+ " sigma_X = P[idx, :][:, idx].mean()\n"
+ ]
+ }
+ ],
+ "source": [
+ "%%time\n",
+ "\n",
+ "obs_key = \"perturbation\"\n",
+ "dist = Distance(obsm_key=\"X_pca\", metric=\"edistance\")\n",
+ "df = dist.pairwise(adata, groupby=obs_key)"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": 14,
+ "id": "bd845994",
+ "metadata": {},
+ "outputs": [
+ {
+ "data": {
+ "application/vnd.jupyter.widget-view+json": {
+ "model_id": "9423ba54408542468fc5e8eec76894de",
+ "version_major": 2,
+ "version_minor": 0
+ },
+ "text/plain": [
+ "Output()"
+ ]
+ },
+ "metadata": {},
+ "output_type": "display_data"
+ },
+ {
+ "data": {
+ "text/html": [
+ "/home/shadeform/miniforge3/envs/pertpy/lib/python3.13/site-packages/pertpy/tools/_distances/_distances.py:675: \n",
+ "RuntimeWarning: Mean of empty slice.\n",
+ " sigma_Y = P[~idx, :][:, ~idx].mean()\n",
+ "
\n"
+ ],
+ "text/plain": [
+ "/home/shadeform/miniforge3/envs/pertpy/lib/python3.13/site-packages/pertpy/tools/_distances/_distances.py:675: \n",
+ "RuntimeWarning: Mean of empty slice.\n",
+ " sigma_Y = P[~idx, :][:, ~idx].mean()\n"
+ ]
+ },
+ "metadata": {},
+ "output_type": "display_data"
+ },
+ {
+ "data": {
+ "text/html": [
+ "/home/shadeform/miniforge3/envs/pertpy/lib/python3.13/site-packages/numpy/_core/_methods.py:145: RuntimeWarning: \n",
+ "invalid value encountered in divide\n",
+ " ret = ret.dtype.type(ret / rcount)\n",
+ "
\n"
+ ],
+ "text/plain": [
+ "/home/shadeform/miniforge3/envs/pertpy/lib/python3.13/site-packages/numpy/_core/_methods.py:145: RuntimeWarning: \n",
+ "invalid value encountered in divide\n",
+ " ret = ret.dtype.type(ret / rcount)\n"
+ ]
+ },
+ "metadata": {},
+ "output_type": "display_data"
+ },
+ {
+ "data": {
+ "text/html": [
+ "/home/shadeform/miniforge3/envs/pertpy/lib/python3.13/site-packages/pertpy/tools/_distances/_distances.py:676: \n",
+ "RuntimeWarning: Mean of empty slice.\n",
+ " delta = P[idx, :][:, ~idx].mean()\n",
+ "
\n"
+ ],
+ "text/plain": [
+ "/home/shadeform/miniforge3/envs/pertpy/lib/python3.13/site-packages/pertpy/tools/_distances/_distances.py:676: \n",
+ "RuntimeWarning: Mean of empty slice.\n",
+ " delta = P[idx, :][:, ~idx].mean()\n"
+ ]
+ },
+ "metadata": {},
+ "output_type": "display_data"
+ },
+ {
+ "data": {
+ "text/html": [
+ "/home/shadeform/miniforge3/envs/pertpy/lib/python3.13/site-packages/pertpy/tools/_distances/_distances.py:674: \n",
+ "RuntimeWarning: Mean of empty slice.\n",
+ " sigma_X = P[idx, :][:, idx].mean()\n",
+ "
\n"
+ ],
+ "text/plain": [
+ "/home/shadeform/miniforge3/envs/pertpy/lib/python3.13/site-packages/pertpy/tools/_distances/_distances.py:674: \n",
+ "RuntimeWarning: Mean of empty slice.\n",
+ " sigma_X = P[idx, :][:, idx].mean()\n"
+ ]
+ },
+ "metadata": {},
+ "output_type": "display_data"
+ },
+ {
+ "data": {
+ "text/html": [
+ "\n"
+ ],
+ "text/plain": []
+ },
+ "metadata": {},
+ "output_type": "display_data"
+ }
+ ],
+ "source": [
+ "dist = pt.tl.Distance(obsm_key=\"X_pca\", metric=\"edistance\")\n",
+ "df = dist.pairwise(adata, groupby=obs_key)"
+ ]
+ },
+ {
+ "cell_type": "code",
+ "execution_count": null,
+ "id": "e92bee29",
+ "metadata": {},
+ "outputs": [],
+ "source": []
+ }
+ ],
+ "metadata": {
+ "kernelspec": {
+ "display_name": "pertpy",
+ "language": "python",
+ "name": "python3"
+ },
+ "language_info": {
+ "codemirror_mode": {
+ "name": "ipython",
+ "version": 3
+ },
+ "file_extension": ".py",
+ "mimetype": "text/x-python",
+ "name": "python",
+ "nbconvert_exporter": "python",
+ "pygments_lexer": "ipython3",
+ "version": "3.13.7"
+ }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}
diff --git a/tmp_scripts/compute_edistance.py b/tmp_scripts/compute_edistance.py
new file mode 100644
index 00000000..e325e7a8
--- /dev/null
+++ b/tmp_scripts/compute_edistance.py
@@ -0,0 +1,37 @@
+from __future__ import annotations
+
+import os
+import time
+from pathlib import Path
+
+import anndata as ad
+import cupy as cp
+import rmm
+from rmm.allocators.cupy import rmm_cupy_allocator
+
+import rapids_singlecell as rsc
+from rapids_singlecell.ptg import Distance
+
+rmm.reinitialize(
+ managed_memory=False, # Allows oversubscription
+ pool_allocator=True, # default is False
+ devices=0, # GPU device IDs to register. By default registers only GPU 0.
+)
+cp.cuda.set_allocator(rmm_cupy_allocator)
+
+
+if __name__ == "__main__":
+ obs_key = "perturbation"
+
+ # homedir/data/adamson_2016_upr_epistasis
+ save_dir = os.path.join(
+ os.path.expanduser("~"),
+ "data",
+ )
+ adata = ad.read_h5ad(Path(save_dir) / "adamson_2016_upr_epistasis_pca.h5ad")
+ rsc.get.anndata_to_GPU(adata, convert_all=True)
+ dist = Distance(obsm_key="X_pca", metric="edistance")
+ start_time = time.time()
+ df = dist.pairwise(adata, groupby=obs_key)
+ end_time = time.time()
+ print(f"Time taken: {end_time - start_time} seconds")
diff --git a/tmp_scripts/compute_edistance_cpu.py b/tmp_scripts/compute_edistance_cpu.py
new file mode 100644
index 00000000..694224d1
--- /dev/null
+++ b/tmp_scripts/compute_edistance_cpu.py
@@ -0,0 +1,38 @@
+from __future__ import annotations
+
+import os
+import time
+from argparse import ArgumentParser
+from pathlib import Path
+
+import anndata as ad
+from pertpy.tools import Distance
+
+if __name__ == "__main__":
+ parser = ArgumentParser()
+ parser.add_argument("--bootstrap", action="store_true")
+ args = parser.parse_args()
+ obs_key = "perturbation"
+ bootstrap = args.bootstrap
+
+ # homedir/data/adamson_2016_upr_epistasis
+ save_dir = os.path.join(
+ os.path.expanduser("~"),
+ "data",
+ )
+ adata = ad.read_h5ad(Path(save_dir) / "adamson_2016_upr_epistasis_pca.h5ad")
+ dist = Distance(obsm_key="X_pca", metric="edistance")
+ start_time = time.time()
+ if bootstrap:
+ df, df_var = dist.pairwise(
+ adata, groupby=obs_key, bootstrap=True, n_bootstrap=100
+ )
+ else:
+ df = dist.pairwise(adata, groupby=obs_key)
+ end_time = time.time()
+ print(f"Time taken: {end_time - start_time} seconds")
+ if bootstrap:
+ df_var.to_csv(Path(save_dir) / "df_cpu_bootstrap_var.csv")
+ df.to_csv(Path(save_dir) / "df_cpu_bootstrap.csv")
+ else:
+ df.to_csv(Path(save_dir) / "df_cpu.csv")
diff --git a/tmp_scripts/compute_edistance_standalone.py b/tmp_scripts/compute_edistance_standalone.py
new file mode 100644
index 00000000..ef1d12dd
--- /dev/null
+++ b/tmp_scripts/compute_edistance_standalone.py
@@ -0,0 +1,109 @@
+from __future__ import annotations
+
+import os
+import time
+from argparse import ArgumentParser
+from pathlib import Path
+
+import anndata as ad
+import cupy as cp
+import numpy as np
+import pandas as pd
+import rmm
+from rmm.allocators.cupy import rmm_cupy_allocator
+
+import rapids_singlecell as rsc
+from rapids_singlecell.pertpy_gpu._edistance import pertpy_edistance
+
+rmm.reinitialize(
+ managed_memory=False, # Allows oversubscription
+ pool_allocator=True, # default is False
+ devices=0, # GPU device IDs to register. By default registers only GPU 0.
+)
+cp.cuda.set_allocator(rmm_cupy_allocator)
+
+if __name__ == "__main__":
+ parser = ArgumentParser()
+ parser.add_argument("--bootstrap", action="store_true")
+ args = parser.parse_args()
+ bootstrap = args.bootstrap
+ obs_key = "perturbation"
+ home_dir = Path(os.path.expanduser("~")) / "data"
+
+ # homedir/data/adamson_2016_upr_epistasis
+ save_dir = home_dir / "adamson_2016_upr_epistasis_pca.h5ad"
+ adata = ad.read_h5ad(save_dir)
+ rsc.get.anndata_to_GPU(adata, convert_all=True)
+
+ df_expected = None
+ if bootstrap:
+ df_cpu_bootstrap_var = pd.read_csv(
+ home_dir / "df_cpu_bootstrap_var.csv", index_col=0
+ )
+ df_cpu_bootstrap = pd.read_csv(home_dir / "df_cpu_bootstrap.csv", index_col=0)
+ df_expected = df_cpu_bootstrap
+ else:
+ df_cpu_float64 = pd.read_csv(home_dir / "df_cpu_float64.csv", index_col=0)
+ df_expected = df_cpu_float64
+
+ start_time = time.time()
+ if not bootstrap:
+ res = pertpy_edistance(
+ adata, groupby=obs_key, obsm_key="X_pca", bootstrap=bootstrap
+ )
+ df_gpu = res.distances
+ else:
+ res = pertpy_edistance(
+ adata,
+ groupby=obs_key,
+ obsm_key="X_pca",
+ bootstrap=bootstrap,
+ n_bootstrap=100,
+ )
+ df_gpu = res.distances
+ df_gpu_var = res.distances_var
+ end_time = time.time()
+ print(f"Time taken: {end_time - start_time} seconds")
+
+ is_not_close = []
+ groups = adata.obs[obs_key].unique()
+ k = len(groups)
+ atol = 1e-8 if not bootstrap else 1e-2
+ for idx1 in range(k):
+ for idx2 in range(idx1 + 1, k):
+ group_x = groups[idx1]
+ group_y = groups[idx2]
+ if group_x == group_y:
+ assert df_gpu.loc[group_x, group_y] == 0
+ else:
+ if not np.isclose(
+ df_gpu.loc[group_x, group_y],
+ df_expected.loc[group_x, group_y],
+ atol=atol,
+ ):
+ is_not_close.append(
+ (
+ (group_x, group_y),
+ df_expected.loc[group_x, group_y],
+ df_gpu.loc[group_x, group_y],
+ np.abs(
+ df_expected.loc[group_x, group_y]
+ - df_gpu.loc[group_x, group_y]
+ ),
+ )
+ )
+ print(
+ f"Group df_gpu: {df_gpu.loc[group_x, group_y]}, Group df: {df_expected.loc[group_x, group_y]}, idx: ({idx1}, {idx2})"
+ )
+
+ print(
+ "Out of",
+ int(k * (k - 1) / 2),
+ "pairs,",
+ len(is_not_close),
+ "pairs are not close with atol=",
+ atol,
+ )
+ # print(df.equals(df_gpu))
+ # print(df)
+ # print(df_gpu)
diff --git a/tmp_scripts/generate_float64_reference.py b/tmp_scripts/generate_float64_reference.py
new file mode 100644
index 00000000..e3751137
--- /dev/null
+++ b/tmp_scripts/generate_float64_reference.py
@@ -0,0 +1,93 @@
+from __future__ import annotations
+
+import os
+
+import anndata as ad
+import numpy as np
+import pandas as pd
+from sklearn.metrics import pairwise_distances as sklearn_pairwise_distances
+
+
+def compute_edistance_sklearn(X, Y, dtype=np.float64):
+ """Compute edistance using sklearn's pairwise_distances with specified precision"""
+ X = np.array(X, dtype=dtype)
+ Y = np.array(Y, dtype=dtype)
+
+ # Compute pairwise distances using sklearn
+ sigma_X = sklearn_pairwise_distances(X, X, metric="euclidean").mean()
+ sigma_Y = sklearn_pairwise_distances(Y, Y, metric="euclidean").mean()
+ delta = sklearn_pairwise_distances(X, Y, metric="euclidean").mean()
+
+ return 2 * delta - sigma_X - sigma_Y
+
+
+def compute_edistance_pairwise_sklearn(
+ adata, groupby, obsm_key="X_pca", dtype=np.float64
+):
+ """Compute pairwise edistance matrix using sklearn with specified precision"""
+ # Get data and convert to CPU numpy
+ embedding = np.array(adata.obsm[obsm_key], dtype=dtype)
+
+ groups = adata.obs[groupby].cat.categories
+ k = len(groups)
+
+ print(f"Computing edistance for {k} groups with dtype {dtype}...")
+
+ # Build edistance matrix
+ edistance_matrix = np.zeros((k, k), dtype=dtype)
+
+ for i, group_a in enumerate(groups):
+ mask_a = adata.obs[groupby] == group_a
+ X = embedding[mask_a]
+
+ for j, group_b in enumerate(groups):
+ if i == j:
+ edistance_matrix[i, j] = 0.0
+ elif i < j: # Only compute upper triangle
+ mask_b = adata.obs[groupby] == group_b
+ Y = embedding[mask_b]
+
+ edist = compute_edistance_sklearn(X, Y, dtype=dtype)
+ edistance_matrix[i, j] = edist
+ edistance_matrix[j, i] = edist # Symmetric
+
+ if (i + 1) % 5 == 0:
+ print(f" Processed {i + 1}/{k} groups")
+
+ return pd.DataFrame(edistance_matrix, index=groups, columns=groups)
+
+
+if __name__ == "__main__":
+ obs_key = "perturbation"
+
+ # Load the data
+ save_dir = os.path.join(os.path.expanduser("~"), "data")
+ adata_path = os.path.join(save_dir, "adamson_2016_upr_epistasis_pca.h5ad")
+
+ print(f"Loading data from {adata_path}...")
+ adata = ad.read_h5ad(adata_path)
+
+ print(f"Data shape: {adata.shape}")
+ print(f"Groups: {len(adata.obs[obs_key].cat.categories)}")
+
+ # Generate the float64 reference CSV using sklearn
+ print("\nGenerating float64 reference using sklearn...")
+ df_reference = compute_edistance_pairwise_sklearn(
+ adata, obs_key, obsm_key="X_pca", dtype=np.float64
+ )
+
+ # Save the float64 reference
+ output_path = os.path.join(save_dir, "df_cpu_float64.csv")
+ df_reference.to_csv(output_path)
+ print(f"Saved float64 reference to: {output_path}")
+
+ # Show a sample of values
+ print("\nSample values:")
+ groups = df_reference.index[:3]
+ for i, group_a in enumerate(groups):
+ for j, group_b in enumerate(groups):
+ if i < j:
+ val = df_reference.loc[group_a, group_b]
+ print(f" {group_a} vs {group_b}: {val:.10f}")
+
+ print("Float64 reference generation complete!")
diff --git a/tmp_scripts/prepare_data.py b/tmp_scripts/prepare_data.py
new file mode 100644
index 00000000..2557eeb8
--- /dev/null
+++ b/tmp_scripts/prepare_data.py
@@ -0,0 +1,26 @@
+from __future__ import annotations
+
+import os
+
+import pertpy as pt
+import scanpy as sc
+
+import rapids_singlecell as rsc
+
+if __name__ == "__main__":
+ adata = pt.data.adamson_2016_upr_epistasis()
+ obs_key = "perturbation"
+
+ # remove genes with 0 expression
+ sc.pp.filter_genes(adata, min_counts=1)
+ sc.pp.filter_cells(adata, min_counts=1)
+ # fill na obskeys
+ # set categories first
+ adata.obs[obs_key] = adata.obs[obs_key].cat.add_categories("control")
+ adata.obs[obs_key] = adata.obs[obs_key].fillna("control")
+ rsc.pp.pca(adata, n_comps=50)
+ # save dir as
+ # homedir/data/adamson_2016_upr_epistasis
+ save_dir = os.path.join(os.path.expanduser("~"), "data")
+ os.makedirs(save_dir, exist_ok=True)
+ adata.write(os.path.join(save_dir, "adamson_2016_upr_epistasis_pca.h5ad"))