compound label; wip

cyteonto/describe.py

@@ -131,59 +131,66 @@ def _build_prompt(label: str, use_pubmed: bool) -> str:
         if use_pubmed
         else "Do not call any tools."
     )
-    return dedent(
-        f"""
-        Task: produce a structured description of one cell type.
-
-        Input label:
-        {label}
-
-        The input label may be a single name, an abbreviation, or multiple
-        synonyms separated by semicolons. Treat them as referring to the same
-        cell type. If the label is ambiguous, choose the most likely
-        interpretation in a standard mammalian reference context and commit
-        to it throughout the response.
-
-        {tool_block}
-
-        Fill every field below. Write each field in third person, factual,
-        present tense. Do not add text outside the structured object. Do not
-        repeat the label text verbatim inside the other fields unless it is
-        the only correct wording.
-
-        initialLabel
-            Copy the input label verbatim, including punctuation and
-            semicolons. Do not normalise case, expand abbreviations, or
-            strip synonyms.
-
-        descriptiveName
-            One noun phrase, at most 12 words. A precise name that identifies
-            the cell type without relying on the initial label. Prefer lineage
-            and location when available (for example "tissue-resident alveolar
-            macrophage").
-
-        function
-            One to three sentences, at most 60 words total. State the primary
-            biological role of the cell. Mention key effector mechanisms only
-            when they are defining for this cell type.
-
-        diseaseRelevance
-            One to three sentences, at most 60 words total. State conditions
-            in which this cell type is pathologically involved, depleted,
-            expanded, or is a therapeutic target. If the cell type has no
-            well-established disease relevance, write "Not established".
-
-        developmentalStage
-            One sentence, at most 30 words. State the lineage, progenitor, or
-            developmental window the cell belongs to. Use "Terminally
-            differentiated" when applicable. Use "Not established" when the
-            stage is unknown.
-
-        Do not include marker genes, tissue lists, species ranges, or
-        citations inside any field. Do not hedge with phrases such as
-        "it is believed that" or "some studies suggest". State facts directly.
-        """
-    ).strip()
+    return (
+        dedent(
+            """
+            Task: produce a structured description of one cell type.
+
+            Input label:
+            {label}
+
+            The input label may be a single name, an abbreviation, or multiple
+            synonyms separated by semicolons. Treat them as referring to the same
+            cell type. If the label is ambiguous, choose the most likely
+            interpretation in a standard mammalian reference context and commit
+            to it throughout the response.
+
+            {tool_block}
+
+            Fill every field below. Write each field in third person, factual,
+            present tense. Do not add text outside the structured object. Do not
+            repeat the label text verbatim inside the other fields unless it is
+            the only correct wording.
+
+            initialLabel
+                Copy the input label verbatim, including punctuation and
+                semicolons. Do not normalise case, expand abbreviations, or
+                strip synonyms.
+
+            descriptiveName
+                One noun phrase, at most 12 words. A precise name that identifies
+                the cell type without relying on the initial label. Prefer lineage
+                and location when available (for example "tissue-resident alveolar
+                macrophage").
+
+            function
+                One to three sentences, at most 60 words total. State the primary
+                biological role of the cell. Mention key effector mechanisms only
+                when they are defining for this cell type.
+
+            diseaseRelevance
+                One to three sentences, at most 60 words total. State conditions
+                in which this cell type is pathologically involved, depleted,
+                expanded, or is a therapeutic target. If the cell type has no
+                well-established disease relevance, write "Not established".
+
+            developmentalStage
+                One sentence, at most 30 words. State the lineage, progenitor, or
+                developmental window the cell belongs to. Use "Terminally
+                differentiated" when applicable. Use "Not established" when the
+                stage is unknown.
+
+            Do not include marker genes, tissue lists, species ranges, or
+            citations inside any field. Do not hedge with phrases such as
+            "it is believed that" or "some studies suggest". State facts directly.
+            """
+        )
+        .format(
+            label=label,
+            tool_block=tool_block,
+        )
+        .strip()
+    )
 
 
 def _build_descriptor_agent(
@@ -381,51 +388,56 @@ def _normalize_decomposition(
         return LabelDecomposition.single(label)
     if not output.isCompound:
         return LabelDecomposition.single(label)
-    cleaned = [p.strip() for p in output.parts if p and p.strip()]
+    cleaned = [p.strip().lower() for p in output.parts if p and p.strip()]
     if len(cleaned) < 2:
         return LabelDecomposition.single(label)
     return LabelDecomposition(initialLabel=label, isCompound=True, parts=cleaned)
 
 
 def _build_decompose_prompt(label: str) -> str:
-    return dedent(
-        f"""
-        Task: decide whether a cell-type annotation label refers to one cell
-        type or multiple distinct cell types.
-
-        Input label:
-        {label}
-
-        Context: labels come from single-cell RNA-seq cluster annotations.
-        Some labels name a single cell type; others name a mixture such as a
-        doublet or mixed population.
-
-        Rules:
-        - Semicolon-separated text lists synonyms for one cell type. Do not
-          split those.
-        - Commas and hyphens inside a single cell-type name do not indicate
-          multiple types.
-        - When the label names multiple distinct cell types, set isCompound to
-          true and list each type in parts as a short sanitized name suitable
-          for ontology lookup. Remove mixture qualifiers such as "doublet",
-          "mixed population", or "contamination" from part names.
-        - When the label names one cell type, set isCompound to false and set
-          parts to a list containing only the original label verbatim.
-
-        initialLabel
-            Copy the input label exactly as given.
-
-        isCompound
-            true when the label names multiple distinct cell types; false
-            otherwise.
-
-        parts
-            When isCompound is false, must contain exactly one element equal
-            to initialLabel.
-            When isCompound is true, must contain two or more sanitized
-            cell-type names.
+    return (
+        dedent(
+            """
+            Task: decide whether a cell-type annotation label refers to one cell
+            type or multiple distinct cell types.
+
+            Input label:
+            {label}
+
+            Context: labels come from single-cell RNA-seq cluster annotations.
+            Some labels name a single cell type; others name a mixture such as a
+            doublet or mixed population.
+
+            Rules:
+            - Semicolon-separated text lists synonyms for one cell type. Do not
+            split those.
+            - Commas and hyphens inside a single cell-type name do not indicate
+            multiple types.
+            - When the label names multiple distinct cell types, set isCompound to
+            true and list each type in parts as a short sanitized name suitable
+            for ontology lookup. Remove mixture qualifiers such as "doublet",
+            "mixed population", or "contamination" from part names.
+            - When isCompound is true, return every part name in lowercase.
+            - When the label names one cell type, set isCompound to false and set
+            parts to a list containing only the original label verbatim.
+
+            initialLabel
+                Copy the input label exactly as given.
+
+            isCompound
+                true when the label names multiple distinct cell types; false
+                otherwise.
+
+            parts
+                When isCompound is false, must contain exactly one element equal
+                to initialLabel.
+                When isCompound is true, must contain two or more sanitized
+                cell-type names.
         """
-    ).strip()
+        )
+        .format(label=label)
+        .strip()
+    )
 
 
 def _build_decompose_agent(base_agent: Agent, reasoning: bool = False) -> Agent:

cyteonto/ontology.py

@@ -26,12 +26,39 @@ def load(self) -> bool:
             return True
         try:
             df = pd.read_csv(self.csv_path)
+            if "label_normalized" not in df.columns:
+                df["label_normalized"] = df["label"].astype(str).str.lower()
+            else:
+                df["label_normalized"] = df["label_normalized"].astype(str).str.lower()
+            dup_mask = df.duplicated(
+                subset=["ontology_id", "label_normalized"], keep=False
+            )
+            if dup_mask.any():
+                for (oid, norm), grp in df[dup_mask].groupby(
+                    ["ontology_id", "label_normalized"]
+                ):
+                    originals = grp["label"].astype(str).tolist()
+                    logger.warning(
+                        f"Normalized label collision for {oid} {norm!r}: "
+                        f"original labels {originals}; keeping {originals[0]!r}"
+                    )
+            before = len(df)
+            df = df.drop_duplicates(
+                subset=["ontology_id", "label_normalized"], keep="first"
+            )
+            dropped = before - len(df)
+            if dropped:
+                logger.info(
+                    f"Dropped {dropped} duplicate ontology rows "
+                    "(same ontology_id and label_normalized)"
+                )
             self._df = df
             for _, row in df.iterrows():
-                label = str(row["label"])
+                norm = str(row["label_normalized"])
+                original = str(row["label"])
                 oid = str(row["ontology_id"])
-                self._id_to_labels.setdefault(oid, []).append(label)
-                self._label_to_id.setdefault(label, oid)
+                self._id_to_labels.setdefault(oid, []).append(original)
+                self._label_to_id.setdefault(norm, oid)
             self._loaded = True
             logger.info(f"Loaded {len(df)} ontology mappings from {self.csv_path}")
             return True
@@ -49,7 +76,7 @@ def df(self) -> pd.DataFrame:
     def label_to_id(self, label: str) -> str | None:
         if not self._loaded:
             self.load()
-        return self._label_to_id.get(label)
+        return self._label_to_id.get(label.lower())
 
     def labels_for_id(self, ontology_id: str) -> list[str]:
         if not self._loaded:
@@ -61,8 +88,12 @@ def ids_and_joined_labels(self) -> tuple[list[str], list[str]]:
         if not self._loaded:
             self.load()
         assert self._df is not None
-        grouped = self._df.groupby("ontology_id")["label"].apply(";".join).reset_index()
-        return grouped["ontology_id"].tolist(), grouped["label"].tolist()
+        grouped = (
+            self._df.groupby("ontology_id")["label_normalized"]
+            .apply(lambda s: ";".join(dict.fromkeys(s.astype(str))))
+            .reset_index()
+        )
+        return grouped["ontology_id"].tolist(), grouped["label_normalized"].tolist()
 
 
 class OntologySimilarity:

cyteonto/paths.py

@@ -19,6 +19,7 @@ class PathConfig:
     Layout produced under ``data_dir``::
 
         cell_ontology/cell_to_cell_ontology.csv
+        cell_ontology/cell_to_cell_ontology_enriched.csv
         cell_ontology/cl.owl
         embedding/cell_ontology/embeddings_<llmKey>_<embdKey>.npz
         embedding/descriptions/descriptions_<llmKey>.json
@@ -43,6 +44,18 @@ def __init__(
     def ontology_csv(self) -> Path:
         return self.data_dir / "cell_ontology" / "cell_to_cell_ontology.csv"
 
+    @property
+    def ontology_enriched_csv(self) -> Path:
+        return self.data_dir / "cell_ontology" / "cell_to_cell_ontology_enriched.csv"
+
+    @property
+    def ontology_mapping_csv(self) -> Path:
+        """CSV used for lookups: enriched file when present, else shipped original."""
+        enriched = self.ontology_enriched_csv
+        if enriched.exists():
+            return enriched
+        return self.ontology_csv
+
     @property
     def ontology_owl(self) -> Path:
         return self.data_dir / "cell_ontology" / "cl.owl"
@@ -113,5 +126,6 @@ def run_kind_dirs(self, run_id: str, kind: str) -> list[Path]:
     def core_files_present(self) -> dict[str, bool]:
         return {
             "ontology_csv": self.ontology_csv.exists(),
+            "ontology_enriched_csv": self.ontology_enriched_csv.exists(),
             "ontology_owl": self.ontology_owl.exists(),
         }