feat(demo): structures + names on substitution; capability notes

training/app.py

@@ -18,6 +18,7 @@
 prediction core doesn't change.
 """
 
+from functools import lru_cache
 from pathlib import Path
 
 from fastapi import FastAPI
@@ -45,6 +46,37 @@ def _resolve(text: str):
     return None
 
 
+def _svg(smi, w=320, h=220):
+    """2D structure SVG; None if drawing is unavailable (headless box w/o libXrender)."""
+    mol = Chem.MolFromSmiles(smi) if smi else None
+    if mol is None:
+        return None
+    try:
+        from rdkit.Chem.Draw import rdMolDraw2D
+        d = rdMolDraw2D.MolDraw2DSVG(w, h)
+        d.DrawMolecule(mol)
+        d.FinishDrawing()
+        return d.GetDrawingText()
+    except Exception:  # noqa: BLE001 — missing X11 libs etc.; degrade gracefully
+        return None
+
+
+@lru_cache(maxsize=8192)
+def _name(smi):
+    """PubChem common name (Title) for a SMILES — cached, best-effort, short timeout."""
+    import json
+    import urllib.parse
+    import urllib.request
+    url = ("https://pubchem.ncbi.nlm.nih.gov/rest/pug/compound/smiles/"
+           f"{urllib.parse.quote(smi)}/property/Title/JSON")
+    try:
+        with urllib.request.urlopen(url, timeout=4) as r:
+            d = json.load(r)
+        return d["PropertyTable"]["Properties"][0].get("Title")
+    except Exception:  # noqa: BLE001 — not found / timeout / throttled
+        return None
+
+
 class Query(BaseModel):
     smiles: str
     k: int = 8
@@ -64,25 +96,17 @@ def api_neighbors(q: Query):
     smi = _resolve(q.smiles)
     if not smi:
         return {"neighbors": []}
-    return P.substitute(smi, k=q.k)
+    res = P.substitute(smi, k=q.k)
+    for n in res.get("neighbors", []):  # enrich each candidate with a structure + a name
+        n["svg"] = _svg(n["smiles"], 132, 96)
+        n["name"] = _name(n["smiles"])
+    return res
 
 
 @app.post("/api/structure")
 def api_structure(q: Query):
-    """2D structure depiction (SVG). Degrades to {svg: None} if RDKit's drawing module
-    can't load (e.g. a headless box missing libXrender) — never 500s the demo."""
-    smi = _resolve(q.smiles)
-    mol = Chem.MolFromSmiles(smi) if smi else None
-    if mol is None:
-        return {"svg": None}
-    try:
-        from rdkit.Chem.Draw import rdMolDraw2D
-        d = rdMolDraw2D.MolDraw2DSVG(320, 220)
-        d.DrawMolecule(mol)
-        d.FinishDrawing()
-        return {"svg": d.GetDrawingText()}
-    except Exception:  # noqa: BLE001 — missing X11 libs etc.; degrade gracefully
-        return {"svg": None}
+    """2D structure depiction (SVG); {svg: None} if drawing is unavailable."""
+    return {"svg": _svg(_resolve(q.smiles))}
 
 
 @app.get("/", response_class=HTMLResponse)

training/workbench.html

@@ -49,10 +49,14 @@
   .chip{font-family:var(--mono);font-size:12px;padding:5px 11px;border-radius:20px;
     border:1px solid var(--line);background:var(--surface)}
   .chip.on{color:#fff;border-color:transparent}
-  .neighbor{display:flex;justify-content:space-between;align-items:center;gap:12px;
+  .neighbor{display:flex;align-items:center;gap:12px;
     padding:11px 0;border-bottom:1px solid var(--line)}
   .neighbor:last-child{border-bottom:none}
-  .neighbor .s{font-family:var(--mono);font-size:12.5px;word-break:break-all}
+  .neighbor .nb-struct{flex:0 0 auto;width:64px;height:48px;background:#fff;border:1px solid var(--line);border-radius:6px}
+  .neighbor .nb-struct svg{width:64px;height:48px}
+  .neighbor .nb-info{flex:1;min-width:0}
+  .neighbor .nb-name{font-weight:600;font-size:13px;text-transform:capitalize;color:var(--ink)}
+  .neighbor .s{font-family:var(--mono);font-size:12px;color:var(--muted);word-break:break-all}
   .neighbor .sim{font-family:var(--mono);font-size:12px;color:var(--accent);white-space:nowrap}
   .neighbor .kt{font-family:var(--mono);font-size:11px;color:var(--muted)}
   .empty{color:var(--muted);font-size:14px;padding:8px 0}
@@ -127,10 +131,13 @@ <h2>Aroma</h2>
 
   <p class="note" id="footnote" style="display:none">
     Bars are model predictions; a <b>measured</b> tag means the molecule is in the
-    curated data (ground truth, not a guess). Salty is lookup-only and sour is a
-    rule. Substitutions rank by structural similarity (Tanimoto over molecular
-    fingerprints). Aroma isn't in the public demo — it trains on your licensed or
-    in-house odor data, on-prem. Results are leads for the bench, not verdicts.
+    curated data (ground truth, not a guess). The trained taste/tox heads are fit on
+    <b>organic</b> molecules — carbon-free inputs are flagged out-of-domain. Salty is
+    lookup-only, sour is a rule. Tox-assay flags are <b>indicative, caution-only</b>
+    (Tox21 in-vitro models) — never a determination. Quantitative dosing / odor-activity
+    needs odor-threshold tables (licensed or customer data), so it stays qualitative here.
+    Substitutions rank by structural similarity (Tanimoto). Aroma isn't in the public
+    demo — it trains on your data, on-prem. Results are leads for the bench, not verdicts.
   </p>
 </main>
 
@@ -215,8 +222,12 @@ <h2>Aroma</h2>
   }else{
     $('neighbors').innerHTML = neighbors.map(x=>`
       <div class="neighbor">
-        <div><div class="s">${x.smiles}</div>
-        <div class="kt">${x.known_tastes.length?x.known_tastes.join(' · '):'—'}</div></div>
+        <div class="nb-struct">${x.svg||''}</div>
+        <div class="nb-info">
+          <div class="nb-name">${x.name||'—'}</div>
+          <div class="s">${x.smiles}</div>
+          <div class="kt">${x.known_tastes.length?x.known_tastes.join(' · '):'—'}</div>
+        </div>
         <div class="sim">${Math.round(x.similarity*100)}% match</div>
       </div>`).join('');
   }