feat(data): add clean PlantMolecularTasteDB loader; document FartDB exclusion

Diligence on richer taste data, with the licensing line held:

- EXCLUDE FartDB (HuggingFace, ~31k tastants, the largest public set): MIT badge
  over a composite of FlavorDB (CC-BY-NC-SA) + SciFinder (CAS, proprietary) +
  Tas2R-DB (ACS). The badge can't relicense data FartLabs doesn't own — the same
  trap as OpenPOM's GS-LF. Documented in SOURCES.md / DATA-SOURCES.md.
- ADD PlantMolecularTasteDB (Gradinaru 2022, Frontiers Pharmacol., CC-BY): taste-
  active phytochemicals — clean, and exactly the natural-product space a flavor
  house cares about. `load_pmtdb` wired into build_taste_dataset.py, reusing the
  tolerant `_find` + `parse_taste_column` logic (skips cleanly until the file is
  dropped in; columns marked [VERIFY] per issue #25).

No data committed (gitignored as always); this readies the next clean taste source
for a drop-in retrain.

Signed-off-by: Austin L. <86896075+rvnminers-A-and-N@users.noreply.github.com>

Author: rvnminers-A-and-N

docs/DATA-SOURCES.md

@@ -12,6 +12,8 @@ What feeds the models, what each source unlocks, its license, and how to get it.
 | Source | Unlocks | License | Status | How to get |
 |--------|---------|---------|--------|-----------|
 | **ChemTastesDB v2.0** | sweet/bitter/umami training + sour/salty rule data (primary) | CC-BY-4.0 | ✅ in use | Zenodo record `14963136`, direct download (see `training/SETUP.md`). Column mapping verified. |
+| **PlantMolecularTasteDB** | more sweet/bitter/umami/sour — **natural-product / phytochemical** tastants (complements ChemTastesDB; squarely a flavor house's world) | **CC-BY** (Frontiers 2022 article + supplementary) | 🟡 candidate (clean) | Gradinaru et al., *Front. Pharmacol.* `10.3389/fphar.2021.751712` — supplementary data (CC-BY) → `pmtdb_taste.csv` (SMILES + taste). `load_pmtdb` merges it; verify columns/provenance on obtain (#25). |
+| **FartDB** | largest public tastant set (~31k molecules) | **MIT badge — composite NOT clean** | ❌ **excluded** | HuggingFace `FartLabs/FartDB` is labeled MIT, but it composites **FlavorDB (CC-BY-NC-SA)** + **SciFinder (CAS, proprietary)** + Tas2R-DB (ACS). The MIT badge can't relicense data FartLabs doesn't own — the **OpenPOM GS-LF trap again**. NonCommercial/proprietary upstream → not usable commercially. |
 | **cosylab/bittersweet** | more sweet/bitter volume | AGPL-3.0 | ⬜ to get | `git clone github.com/cosylabiiit/bittersweet` → `bittersweet/data/*.tsv`. AGPL: keep a CC-BY-clean build with `INCLUDE_COSYLAB=False`, or opt in deliberately. |
 | **FlavorDB** | ~25k molecules taste+odor + natural-source mapping | **CC BY-NC-SA 3.0** | ❌ skip | NonCommercial — incompatible with a commercial product. |
 | **UMP442 / BIOPEP-UWM** | more umami examples | none / unclear | ❌ skip | BIOPEP-UWM is web-only; the `Shoombuatong/Dataset-Code` repost has **no LICENSE** (all-rights-reserved) and is umami *peptide* data (different class from our small-molecule head). |
@@ -29,8 +31,11 @@ What feeds the models, what each source unlocks, its license, and how to get it.
   loaders (the `[VERIFY]` markers) — see issue **#25**. ChemTastesDB's mapping is
   resolved (both the coarse `Class taste` and the granular `Taste` columns are parsed).
 - **Suggested priority.** (1) `Taste`-column mining — **done** ✅ → (2) SweetenersDB
-  intensity head — **done** ✅ → (3) aroma model (see the data caveat below) →
-  (4) cosylab — **skip** (AGPL). **FlavorDB is CC BY-NC-SA (NonCommercial) — skip.**
+  intensity head — **done** ✅ → (3) **PlantMolecularTasteDB** (CC-BY phytochemicals —
+  the next clean taste add; loader wired, just obtain the file) → (4) aroma model (see
+  the data caveat below) → (5) cosylab — **skip** (AGPL). **FlavorDB is CC BY-NC-SA
+  (NonCommercial) — skip. FartDB (~31k) is excluded — MIT badge over an NC/proprietary
+  composite (FlavorDB + SciFinder); the same "downloadable ≠ usable" trap as GS-LF.**
 - **Aroma data — COMMERCIAL-CLEAN ONLY (decision).** Every *rich* odor-descriptor set
   (Leffingwell, GoodScents, Arctander, Flavornet © Datu Inc., FlavorDB / FooDB NC) is
   restricted or NonCommercial — **we do NOT use any of them, not even for the demo**,

docs/SOURCES.md

@@ -36,7 +36,9 @@ record of provenance, not legal advice. Get an IP/OSS-license review before ship
 - **ChemTastesDB** — Rojas et al., curated taste dataset (sweet/bitter/umami/sour/salty/tasteless). Zenodo, DOI 10.5281/zenodo.5747393 (and the extended record). License **CC-BY-4.0** (commercially clean with attribution).
 - **cosylab BitterSweet** — Bagler lab (IIIT-Delhi). Code **AGPL-3.0** (bind-aware: train-your-own-on-data is the safe pattern; `INCLUDE_COSYLAB=False` drops it for clean licensing).
 - **FlavorDB** — Bagler lab (IIIT-Delhi); taste + odor + natural-source associations. License **CC BY-NC-SA 3.0** (NonCommercial) — incompatible with a commercial product; **not used**.
+- **FartDB** — FartLabs (HuggingFace), ~31k tastants behind the 2025 *npj Science of Food* "FART" model — the largest public tastant set. Repo labeled **MIT**, but it **composites FlavorDB (CC-BY-NC-SA) + SciFinder (CAS, proprietary) + Tas2R-DB (ACS)**; the MIT badge cannot relicense data the authors don't own. **Excluded** (NonCommercial/proprietary upstream) — the same "downloadable ≠ usable" trap as OpenPOM's GS-LF.
 - **SweetenersDB v2.0** — Bouysset et al. (2020) *Food Chem.*, building on Chéron et al. (2017); relative-to-sucrose sweetness *intensity* for the regressor. Released **MIT** by the authors' own lab (ChemSenSim, `github.com/chemosim-lab/SweetenersDB`) — the paywall is only on the journal article, not the authors' own data. **In use.**
+- **PlantMolecularTasteDB** — Gradinaru et al. (2022), *Frontiers in Pharmacology* (DOI 10.3389/fphar.2021.751712), **CC-BY**; taste-active phytochemicals (natural-product tastants). A clean candidate to extend the taste heads into the natural-product space a flavor house cares about — loader (`load_pmtdb`) wired; obtain the CC-BY supplementary and verify columns before merging.
 - **BitterDB** *(future)* — bitterness intensity, if/when added.
 - **UMP442 / BIOPEP-UWM** — umami references. BIOPEP-UWM is web-only; the GitHub repost `Shoombuatong/Dataset-Code` carries **no license** (all-rights-reserved) and is umami *peptide* data (a different class from our small-molecule head). **Not used.**
 

training/build_taste_dataset.py

@@ -9,10 +9,15 @@
 Inputs (download what you have; loaders skip cleanly if a file is absent):
   - ChemTastesDB_database.xlsx   Zenodo, CC-BY-4.0   10 classes, 4075 molecules
   - bittersweet/data/*.tsv       cosylab,  AGPL-3.0   sweet/bitter      [optional]
-  - flavordb_taste.csv           FlavorDB export      SMILES + taste     [OBTAIN]
+  - flavordb_taste.csv           FlavorDB export      SMILES + taste     [OBTAIN — NC, skip]
+  - pmtdb_taste.csv              PlantMolecularTasteDB phytochemical taste[OBTAIN — CC-BY]
   - umami_list.csv               UMP442/BIOPEP-UWM    umami SMILES       [OBTAIN]
   - sweeteners_db.csv            Cheron SweetenersDB  sweetness intensity[OBTAIN]
 
+Note: FartDB (HuggingFace, ~31k tastants) looks tempting but is EXCLUDED — its MIT
+badge sits on a composite of FlavorDB (CC-BY-NC-SA) + SciFinder (proprietary), which
+the badge cannot relicense (same trap as OpenPOM's GS-LF). See docs/DATA-SOURCES.md.
+
 Outputs:
   - taste_master.parquet     inchikey, smiles, {sweet,bitter,umami,sour,salty}=1/0/NaN,
                              multitaste flag, n_sources
@@ -42,6 +47,7 @@
 FLAVORDB_CSV = Path("flavordb_taste.csv")
 UMAMI_CSV = Path("umami_list.csv")
 SWEETENERS_CSV = Path("sweeteners_db.csv")
+PMTDB_CSV = Path("pmtdb_taste.csv")
 
 
 def canon(smiles):
@@ -227,6 +233,32 @@ def load_flavordb(path):
     return pd.DataFrame(rows)
 
 
+def load_pmtdb(path):
+    """PlantMolecularTasteDB (Gradinaru et al. 2022, Frontiers Pharmacol., CC-BY) —
+    taste-active PHYTOCHEMICALS (natural products), complementing ChemTastesDB with
+    the natural-product space a flavor house cares about. SMILES + a taste descriptor
+    column parsed onto the basic tastes; non-basic descriptors (astringent/pungent/
+    cooling) are ignored (chemesthesis, not basic taste).
+    [VERIFY columns against the exported supplementary file — see issue #25.]
+    """
+    if not path.exists():
+        print("  [skip] pmtdb")
+        return pd.DataFrame()
+    df = pd.read_csv(path)
+    sc = _find(df, ["smiles", "canonical smiles", "isomeric smiles"])
+    tc = _find(df, ["taste", "tastes", "taste class", "class", "taste_modality"])
+    rows = []
+    for _, r in df.iterrows():
+        ik, cs = canon(r[sc])
+        if ik is None:
+            continue
+        labels = parse_taste_column(r[tc])
+        if labels:
+            rows.append(_row(ik, cs, "pmtdb", labels))
+    print(f"  pmtdb: {len(rows)}")
+    return pd.DataFrame(rows)
+
+
 def reconcile(df):
     def agg(s):
         v = s.dropna().values
@@ -266,6 +298,7 @@ def build_intensity(path):
         load_chemtastes(CHEMTASTES_XLSX),
         load_cosylab(COSYLAB_DIR),
         load_flavordb(FLAVORDB_CSV),
+        load_pmtdb(PMTDB_CSV),
         load_single_taste_csv(UMAMI_CSV, "umami", "ump442"),
     ]
     parts = [p for p in parts if not p.empty]