PCA feature compression for organelle attribution

.gitignore

@@ -105,4 +105,6 @@ ENV/
 .vscode/
 
 
-configs/
+configs/
+scratch/
+scripts/compare_dino_cp_map.py

pyproject.toml

@@ -27,6 +27,7 @@ classifiers = [
 dynamic = ["version"]
 dependencies = [
     "ops_utils",
+    "scikit-learn",
     "torch==2.7.1",
     "lightning",
     "byol-pytorch",
@@ -35,7 +36,7 @@ dependencies = [
     "cytoself @ git+https://github.com/royerlab/cytoself.git@version_1.5",
     "pytorch-metric-learning",
     "cp-measure",
-    "iohub==0.3.0a6",
+    "iohub>=0.3.0a7",
     "viscy",
     "monai",
     "diffusers[torch]",
@@ -64,9 +65,8 @@ dev = [
 homepage = "https://github.com/ahillsley/ops_model"
 repository = "https://github.com/ahillsley/ops_model"
 
-# same as console_scripts entry point
-# [project.scripts]
-# spam-cli = "spam:main_cli"
+[project.scripts]
+run_eval = "ops_model.eval.run_eval:main"
 
 # Entry points
 # https://peps.python.org/pep-0621/#entry-points

src/ops_model/data/data_loader.py

@@ -37,6 +37,19 @@
 
 warnings.filterwarnings("ignore", category=zarr.errors.ZarrUserWarning)
 
+# Maps experiment-specific column names → canonical internal names.
+# Add new entries here to support additional experiment types.
+COLUMN_ALIASES: dict[str, str] = {
+    "minibinder_perturbation": "gene_name",  # minibinder experiments
+    "AA_sequence": "sgRNA",                  # minibinder experiments
+}
+
+
+def normalize_link_csv(df: pd.DataFrame) -> pd.DataFrame:
+    """Rename experiment-specific column variants to canonical names."""
+    rename = {src: dst for src, dst in COLUMN_ALIASES.items() if src in df.columns}
+    return df.rename(columns=rename)
+
 
 class BaseDataset(Dataset):
 
@@ -509,6 +522,7 @@ def get_labels(self):
                 if self.verbose:
                     print("reading labels for", exp_name, f"links_{w[0]}{w[2]}")
                 labels_tmp = pd.read_csv(OpsPaths(exp_name, well=w).links["training"])
+                labels_tmp = normalize_link_csv(labels_tmp)
 
                 # remove rows with NaN segmentation_id
                 labels_tmp = labels_tmp.dropna(subset=["segmentation_id"])

src/ops_model/data/paths.py

@@ -14,7 +14,7 @@ def __init__(self, experiment: str, well: str = None):
         else:
             self.well_prefix = None
 
-        self.base = Path("/hpc/projects/intracellular_dashboard/ops")
+        self.base = Path("/hpc/projects/icd.fast.ops")
 
         self.stores = {
             "phenotyping": self.base / self.experiment / "3-assembly/phenotyping.zarr",

src/ops_model/eval/__init__.py

@@ -0,0 +1 @@
+"""Evaluation suite for OPS embedding models."""

src/ops_model/eval/eval_io.py

@@ -0,0 +1,91 @@
+"""Save evaluation outputs (CSVs and plots) to disk."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+import matplotlib
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+import pandas as pd
+
+from ops_utils.analysis.map_scores import plot_map_scatter
+
+
+def save_guide_eval(
+    activity_map: pd.DataFrame,
+    distinctiveness_map: pd.DataFrame,
+    metrics: dict,
+    output_dir: str | Path,
+) -> None:
+    """Save guide-level evaluation outputs to output_dir.
+
+    Writes:
+    - ``phenotypic_activity.csv``
+    - ``phenotypic_distinctiveness.csv``
+    - ``map_activity_distinctiveness.png``
+
+    Parameters
+    ----------
+    activity_map : DataFrame
+        Per-perturbation mAP results from ``phenotypic_activity_assesment``.
+    distinctiveness_map : DataFrame
+        Per-perturbation mAP results from ``phenotypic_distinctivness``.
+    metrics : dict
+        Scalar metrics returned by ``evaluate_guide_level`` (used for plot titles).
+    output_dir : str or Path
+        Directory to write outputs into (created if it does not exist).
+    """
+    output_dir = Path(output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    activity_map.to_csv(output_dir / "phenotypic_activity.csv", index=False)
+    distinctiveness_map.to_csv(output_dir / "phenotypic_distinctiveness.csv", index=False)
+
+    fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(16, 7))
+    plot_map_scatter(ax1, activity_map, "Activity", metrics["pct_perturbations_active"])
+    plot_map_scatter(ax2, distinctiveness_map, "Distinctiveness", metrics["pct_perturbations_distinct"])
+    fig.tight_layout()
+    fig.savefig(output_dir / "map_activity_distinctiveness.png", dpi=150, bbox_inches="tight")
+    plt.close(fig)
+
+
+def save_gene_eval(
+    consistency_corum_map: pd.DataFrame,
+    consistency_manual_map: pd.DataFrame,
+    metrics: dict,
+    output_dir: str | Path,
+) -> None:
+    """Save gene-level evaluation outputs to output_dir.
+
+    Writes:
+    - ``phenotypic_consistency_corum.csv``
+    - ``phenotypic_consistency_manual.csv``
+    - ``map_consistency.png``
+
+    Parameters
+    ----------
+    consistency_corum_map : DataFrame
+        Per-complex mAP results from ``phenotypic_consistency_corum``.
+    consistency_manual_map : DataFrame
+        Per-complex mAP results from ``phenotypic_consistency_manual_annotation``.
+    metrics : dict
+        Scalar metrics returned by ``evaluate_gene_level`` (used for plot titles).
+    output_dir : str or Path
+        Directory to write outputs into (created if it does not exist).
+    """
+    output_dir = Path(output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    consistency_corum_map.to_csv(output_dir / "phenotypic_consistency_corum.csv", index=False)
+    consistency_manual_map.to_csv(output_dir / "phenotypic_consistency_manual.csv", index=False)
+
+    corum_entity_col = "complex_id" if "complex_id" in consistency_corum_map.columns else consistency_corum_map.columns[0]
+    chad_entity_col = "complex_num" if "complex_num" in consistency_manual_map.columns else consistency_manual_map.columns[0]
+
+    fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(16, 7))
+    plot_map_scatter(ax1, consistency_corum_map, "Consistency (CORUM)", metrics["pct_complexes_significant_corum"])
+    plot_map_scatter(ax2, consistency_manual_map, "Consistency (CHAD)", metrics["pct_complexes_significant_manual"])
+    fig.tight_layout()
+    fig.savefig(output_dir / "map_consistency.png", dpi=150, bbox_inches="tight")
+    plt.close(fig)