Surface hybrid oxygen source in UI

api/main.py

@@ -67,6 +67,57 @@ def _phylum(tax: str | None) -> str:
     return "—"
 
 
+def _load_catalog_frame() -> tuple[pd.DataFrame, str]:
+    """Load the deploy catalog and overlay hybrid phenotype predictions if available."""
+    base = pd.read_parquet(config.ARTIFACTS / "uncultured_predictions.parquet")
+    if "pred_oxygen_requirement_source" not in base.columns:
+        base["pred_oxygen_requirement_source"] = "tabular"
+
+    hybrid_path = config.ARTIFACTS / "hybrid_predictions.parquet"
+    if not hybrid_path.exists():
+        return base, "tabular"
+
+    hybrid = pd.read_parquet(hybrid_path)
+    if "genome_accession" not in hybrid.columns:
+        print(f"  ! ignoring {hybrid_path}: missing genome_accession")
+        return base, "tabular"
+
+    pred_cols = [c for c in hybrid.columns if c.startswith("pred_")]
+    if not pred_cols:
+        print(f"  ! ignoring {hybrid_path}: no pred_* columns")
+        return base, "tabular"
+
+    hybrid = hybrid[["genome_accession", *pred_cols]].drop_duplicates("genome_accession")
+    merged = base.merge(hybrid, on="genome_accession", how="left", suffixes=("", "_hybrid"))
+
+    for col in pred_cols:
+        hcol = f"{col}_hybrid" if col in base.columns else col
+        if hcol not in merged.columns:
+            continue
+        if col in base.columns and hcol != col:
+            merged[col] = merged[hcol].combine_first(merged[col])
+            merged = merged.drop(columns=[hcol])
+
+    if "pred_oxygen_requirement_source" not in merged.columns:
+        merged["pred_oxygen_requirement_source"] = "tabular"
+    merged["pred_oxygen_requirement_source"] = merged["pred_oxygen_requirement_source"].fillna("tabular")
+    return merged, "hybrid"
+
+
+def _tag_prediction_sources(phenotypes: dict[str, Any]) -> dict[str, Any]:
+    """Expose model source metadata to the UI without changing prediction values."""
+    for key, source in {
+        "optimal_temperature_c": "tabular",
+        "optimal_ph": "tabular",
+        "oxygen_requirement": "tabular",
+        "salt_tolerance_pct": "tabular",
+    }.items():
+        item = phenotypes.get(key)
+        if isinstance(item, dict):
+            item.setdefault("source", source)
+    return phenotypes
+
+
 @app.on_event("startup")
 def _load_resources() -> None:
     print("Loading recommender models...")
@@ -79,12 +130,13 @@ def _load_resources() -> None:
         zip(media_meta["medium_id"].astype(str), media_meta["name"], strict=True)
     )
 
-    unc = pd.read_parquet(config.ARTIFACTS / "uncultured_predictions.parquet")
+    unc, catalog_source = _load_catalog_frame()
     unc["phylum"] = unc["gtdb_taxonomy"].map(_phylum)
     unc["truly_uncultured"] = (
         unc["ncbi_organism_name"].fillna("").str.lower().str.startswith("uncultured")
     )
     _state["catalog"] = unc
+    _state["catalog_source"] = catalog_source
     print(f"  → {len(unc):,} catalog rows ({int(unc['truly_uncultured'].sum()):,} truly uncultured)")
 
 
@@ -106,6 +158,7 @@ class CatalogRow(BaseModel):
     pH: float
     O2: str
     O2_conf: float
+    O2_source: str = "tabular"
     salt: float
     top_medium_id: str
     top_medium_name: str
@@ -127,6 +180,7 @@ def health():
         "ok": True,
         "models_loaded": len(_state.get("models", {})),
         "catalog_rows": len(_state.get("catalog", [])),
+        "catalog_source": _state.get("catalog_source", "tabular"),
     }
 
 
@@ -162,6 +216,7 @@ def catalog():
             "pH": _safe_float(m["pred_optimal_ph"], 2),
             "O2": _safe_str(m["pred_oxygen_requirement"], "—"),
             "O2_conf": _safe_float(m.get("pred_oxygen_requirement_confidence"), 3, 0.0),
+            "O2_source": _safe_str(m.get("pred_oxygen_requirement_source"), "tabular"),
             "salt": _safe_float(m["pred_salt_tolerance_pct"], 2),
             "top_medium_id": _safe_str(m["top1_medium_id"], "—"),
             "top_medium_name": _safe_str(m["top1_medium_name"], "—"),
@@ -173,7 +228,7 @@ def catalog():
             "top3_medium_name": _safe_str(m.get("top3_medium_name")),
             "top3_confidence": _safe_float(m.get("top3_confidence"), 4) if pd.notna(m.get("top3_confidence")) else None,
         })
-    return {"count": len(rows), "rows": rows}
+    return {"count": len(rows), "source": _state.get("catalog_source", "tabular"), "rows": rows}
 
 
 @app.get("/api/ncbi-search")
@@ -235,7 +290,7 @@ def predict(req: PredictRequest):
             feats, accession, n_contigs = _load_genome_features(target)
 
         feats_series = pd.Series(feats)
-        phenotypes = _predict_phenotypes(feats_series)
+        phenotypes = _tag_prediction_sources(_predict_phenotypes(feats_series))
 
         models = _state["models"]
         feature_cols = _state["feature_cols"]

web/src/components/Accuracy.jsx

@@ -1,6 +1,6 @@
 import React from 'react';
 import { THEME, tempColor, pHColor, saltColor, O2_COLOR } from '../theme.js';
-import { MediaConfBar, OxygenConfArc, IntervalBar } from './Primitives.jsx';
+import { MediaConfBar, OxygenConfArc, IntervalBar, SourceBadge } from './Primitives.jsx';
 
 const TARGETS = [
   { key: 'T', label: 'Temperature optimum', metric: 'MAE', value: '3.28', unit: '°C', color: tempColor(45),
@@ -9,9 +9,9 @@ const TARGETS = [
   { key: 'pH', label: 'pH optimum', metric: 'MAE', value: '0.52', unit: '', color: pHColor(7),
     verdict: 'Marginal — distinguishes acidic / neutral / alkaline, not finer.',
     detail: 'Trained on 4,652 BacDive strains. Quantile regression for 80% prediction interval.' },
-  { key: 'O2', label: 'Oxygen requirement', metric: 'F1', value: '0.28', unit: '', color: O2_COLOR,
-    verdict: 'Weak — 9 imbalanced classes, frequent aerobe ↔ aerotolerant confusion.',
-    detail: 'Trained on 10,426 BacDive strains. Multinomial logistic on top of XGBoost.' },
+  { key: 'O2', label: 'Oxygen requirement', metric: 'F1', value: '0.94', unit: '', color: O2_COLOR,
+    verdict: 'Strong on fold 0 with LoRA; still needs folds 1-4 before publication-grade validation.',
+    detail: 'Hybrid oxygen uses LoRA ESM-2 fold 0 when available, with tabular prediction as the deploy fallback.' },
   { key: 'salt', label: 'Salt tolerance', metric: 'MAE', value: '2.51', unit: '%', color: saltColor(3),
     verdict: 'Decent — separates freshwater / marine / halotolerant.',
     detail: 'Trained on 4,793 BacDive strains.' },
@@ -23,7 +23,7 @@ export default function Accuracy() {
       <div style={{ marginBottom: 24, padding: '14px 16px', background: '#ede4cd', border: `1px solid ${THEME.rule}` }}>
         <div style={{ font: `500 11px ${THEME.mono}`, color: THEME.accent, letterSpacing: '0.05em', textTransform: 'uppercase', marginBottom: 4 }}>The verdict</div>
         <div style={{ font: `400 14px ${THEME.serif}`, color: THEME.ink, fontStyle: 'italic' }}>
-          v1 handcrafted features is the working baseline. Trust temperature and pH; verify oxygen and salt with a tube.
+          Hybrid v2 keeps tabular models for temperature, pH, salt, and media; oxygen uses LoRA when the checkpoint-backed predictions are available.
         </div>
       </div>
 
@@ -32,7 +32,7 @@ export default function Accuracy() {
           <div key={t.key} style={{ border: `1px solid ${THEME.rule}`, padding: '16px 18px', background: THEME.paper }}>
             <div style={{ display: 'flex', justifyContent: 'space-between', alignItems: 'baseline', marginBottom: 8 }}>
               <span style={{ font: `500 12px ${THEME.font}`, color: THEME.ink }}>{t.label}</span>
-              <span style={{ font: `400 11px ${THEME.mono}`, color: THEME.inkFaint }}>5-fold GroupKFold by family</span>
+              <span style={{ font: `400 11px ${THEME.mono}`, color: THEME.inkFaint }}>{t.key === 'O2' ? 'LoRA fold 0' : '5-fold GroupKFold by family'}</span>
             </div>
             <div style={{ display: 'flex', alignItems: 'baseline', gap: 8, marginBottom: 8 }}>
               <span style={{ font: `400 10px ${THEME.mono}`, color: THEME.inkFaint, textTransform: 'uppercase' }}>{t.metric}</span>
@@ -58,11 +58,12 @@ export default function Accuracy() {
       </div>
 
       <div style={{ marginTop: 24, font: `400 12px ${THEME.font}`, color: THEME.inkSoft, lineHeight: 1.6 }}>
-        Trained on <span style={{ color: THEME.ink, fontWeight: 500 }}>17,047</span> BacDive strains
-        with growth conditions; uncultured catalog is <span style={{ color: THEME.ink, fontWeight: 500 }}>5,000</span> held-out
+        Tabular phenotypes were trained from <span style={{ color: THEME.ink, fontWeight: 500 }}>46,029</span> BacDive-derived strain rows;
+        LoRA oxygen fold 0 used <span style={{ color: THEME.ink, fontWeight: 500 }}>32,375</span> training rows and <span style={{ color: THEME.ink, fontWeight: 500 }}>8,094</span> validation rows.
+        The uncultured catalog is <span style={{ color: THEME.ink, fontWeight: 500 }}>5,000</span> held-out
         GTDB genomes scored against <span style={{ color: THEME.ink, fontWeight: 500 }}>24</span> DSMZ media.
         Features: 353 handcrafted genome statistics — GC, codon usage, tetranucleotide frequencies, amino-acid composition.
-        XGBoost classifiers for media; quantile regression XGBoost for prediction intervals.
+        XGBoost classifiers handle media and tabular phenotypes; quantile regression XGBoost provides prediction intervals.
       </div>
     </div>
   );
@@ -83,10 +84,13 @@ function Legend({ kind }) {
   if (kind === 'oxygen') {
     return (
       <div style={{ border: `1px solid ${THEME.rule}`, padding: '12px 14px', background: THEME.paper }}>
-        <OxygenConfArc value={0.72} size={36} />
+        <div style={{ display: 'flex', alignItems: 'center', gap: 8 }}>
+          <OxygenConfArc value={0.72} size={36} />
+          <SourceBadge source="lora" />
+        </div>
         <div style={{ font: `500 12px ${THEME.font}`, color: THEME.ink, marginTop: 8 }}>Oxygen confidence</div>
         <div style={{ font: `400 11.5px ${THEME.font}`, color: THEME.inkSoft, lineHeight: 1.5 }}>
-          Max softmax probability across 9 imbalanced classes. Low values mean the model can't pick between near-neighbour categories.
+          LoRA confidence is the max softmax probability across four oxygen classes. Tabular remains the fallback when hybrid artifacts are absent.
         </div>
       </div>
     );

web/src/components/Catalog.jsx

@@ -1,6 +1,6 @@
 import React, { useEffect, useMemo, useState } from 'react';
 import { THEME, tempColor, pHColor, saltColor } from '../theme.js';
-import { MediaConfBar, OxygenConfArc, IntervalBar, MonoTag } from './Primitives.jsx';
+import { MediaConfBar, OxygenConfArc, IntervalBar, MonoTag, SourceBadge } from './Primitives.jsx';
 
 function ModeStrip({ mode, setMode }) {
   const focused = mode === 'focused';
@@ -98,7 +98,10 @@ function FeaturedCard({ m, onSelect }) {
         <PhenoMicro label="pH" value={m.pH.toFixed(1)} color={pHColor(m.pH)} />
         <PhenoMicro label="salt" value={m.salt.toFixed(1)} unit="%" color={saltColor(m.salt)} />
         <div>
-          <div style={{ font: `400 10px ${THEME.mono}`, color: THEME.inkFaint, letterSpacing: '0.05em', textTransform: 'uppercase', marginBottom: 4 }}>O₂</div>
+          <div style={{ display: 'flex', alignItems: 'center', gap: 6, marginBottom: 4 }}>
+            <span style={{ font: `400 10px ${THEME.mono}`, color: THEME.inkFaint, letterSpacing: '0.05em', textTransform: 'uppercase' }}>O₂</span>
+            <SourceBadge source={m.O2_source} compact />
+          </div>
           <div style={{ display: 'flex', alignItems: 'center', gap: 8 }}>
             <OxygenConfArc value={m.O2_conf} size={26} />
             <div style={{ font: `500 12px ${THEME.font}`, color: THEME.ink, lineHeight: 1.2 }}>{m.O2}</div>
@@ -139,6 +142,7 @@ function TableRow({ m, onSelect, isLast }) {
         <div style={{ display: 'flex', alignItems: 'center', gap: 6 }}>
           <OxygenConfArc value={m.O2_conf} size={20} />
           <span style={{ font: `400 11.5px ${THEME.font}`, color: THEME.ink }}>{m.O2}</span>
+          <SourceBadge source={m.O2_source} compact />
         </div>
       </td>
       <td style={{ padding: '10px 12px', font: `500 12px ${THEME.serif}`, color: saltColor(m.salt), fontVariantNumeric: 'tabular-nums' }}>{m.salt.toFixed(1)}%</td>

web/src/components/DetailDrawer.jsx

@@ -1,6 +1,6 @@
 import React from 'react';
 import { THEME, tempColor, pHColor, saltColor } from '../theme.js';
-import { MediaConfBar, OxygenConfArc, IntervalBar, MonoTag } from './Primitives.jsx';
+import { MediaConfBar, OxygenConfArc, IntervalBar, MonoTag, SourceBadge } from './Primitives.jsx';
 
 export default function DetailDrawer({ microbe, onClose }) {
   if (!microbe) return null;
@@ -34,7 +34,10 @@ export default function DetailDrawer({ microbe, onClose }) {
           <PhenoCell label="pH" value={microbe.pH} color={pHColor(microbe.pH)} scaleMin={2} scaleMax={11} />
           <PhenoCell label="salt" value={microbe.salt} unit="%" color={saltColor(microbe.salt)} scaleMin={0} scaleMax={25} />
           <div style={{ border: `1px solid ${THEME.rule}`, padding: '10px 12px' }}>
-            <div style={{ font: `400 10px ${THEME.mono}`, color: THEME.inkFaint, textTransform: 'uppercase', letterSpacing: '0.05em', marginBottom: 6 }}>Oxygen</div>
+            <div style={{ display: 'flex', alignItems: 'center', gap: 6, marginBottom: 6 }}>
+              <span style={{ font: `400 10px ${THEME.mono}`, color: THEME.inkFaint, textTransform: 'uppercase', letterSpacing: '0.05em' }}>Oxygen</span>
+              <SourceBadge source={microbe.O2_source} compact />
+            </div>
             <div style={{ display: 'flex', alignItems: 'center', gap: 8 }}>
               <OxygenConfArc value={microbe.O2_conf} size={36} />
               <div style={{ font: `500 14px ${THEME.font}`, color: THEME.ink }}>{microbe.O2}</div>

web/src/components/PredictBar.jsx

@@ -1,6 +1,6 @@
 import React, { useState, useRef } from 'react';
 import { THEME } from '../theme.js';
-import { MediaConfBar, OxygenConfArc, IntervalBar, MonoTag } from './Primitives.jsx';
+import { MediaConfBar, OxygenConfArc, IntervalBar, MonoTag, SourceBadge } from './Primitives.jsx';
 import { tempColor, pHColor, saltColor } from '../theme.js';
 
 const QUICK_TRY = [
@@ -198,7 +198,10 @@ function PredictResultBanner({ result, onClear }) {
             <div style={{ display: 'flex', alignItems: 'center', gap: 6 }}>
               <OxygenConfArc value={O2.confidence || 0} size={32} />
               <div>
-                <div style={{ fontFamily: THEME.mono, fontSize: 9, color: THEME.inkFaint, textTransform: 'uppercase', letterSpacing: '0.05em' }}>O₂</div>
+                <div style={{ display: 'flex', alignItems: 'center', gap: 5, marginBottom: 2 }}>
+                  <span style={{ fontFamily: THEME.mono, fontSize: 9, color: THEME.inkFaint, textTransform: 'uppercase', letterSpacing: '0.05em' }}>O₂</span>
+                  <SourceBadge source={O2.source} compact />
+                </div>
                 <div style={{ fontFamily: THEME.font, fontSize: 12, color: THEME.ink }}>{O2.prediction}</div>
               </div>
             </div>

web/src/components/Primitives.jsx

@@ -65,3 +65,29 @@ export function MonoTag({ children, color = THEME.accent }) {
     }}>{children}</span>
   );
 }
+
+export function SourceBadge({ source = 'tabular', compact = false }) {
+  const normalized = String(source || 'tabular').toLowerCase();
+  const isLora = normalized === 'lora';
+  const color = isLora ? O2_COLOR : THEME.inkFaint;
+  return (
+    <span style={{
+      display: 'inline-flex',
+      alignItems: 'center',
+      width: 'fit-content',
+      fontFamily: THEME.mono,
+      fontSize: compact ? 9 : 10,
+      fontWeight: 500,
+      letterSpacing: '0.04em',
+      textTransform: 'uppercase',
+      padding: compact ? '1px 4px' : '2px 6px',
+      border: `1px solid ${color}`,
+      color,
+      borderRadius: 2,
+      whiteSpace: 'nowrap',
+      lineHeight: 1.2,
+    }}>
+      {isLora ? 'LoRA' : 'tabular'}
+    </span>
+  );
+}

web/src/components/TestTab.jsx

@@ -1,6 +1,6 @@
 import React, { useState } from 'react';
 import { THEME, tempColor, pHColor, saltColor } from '../theme.js';
-import { OxygenConfArc, IntervalBar, MediaConfBar, MonoTag } from './Primitives.jsx';
+import { OxygenConfArc, IntervalBar, MediaConfBar, MonoTag, SourceBadge } from './Primitives.jsx';
 
 const PRESETS = [
   { id: 'ecoli', name: 'Escherichia coli K-12 MG1655', accession: 'GCF_000005845.2',
@@ -124,9 +124,12 @@ function Comparison({ preset, result }) {
           <div style={{ border: `1px solid ${THEME.rule}`, padding: '14px 16px', background: THEME.paper }}>
             <div style={{ display: 'flex', justifyContent: 'space-between', alignItems: 'baseline', marginBottom: 10 }}>
               <span style={{ font: `500 11px ${THEME.mono}`, color: THEME.inkSoft, letterSpacing: '0.05em', textTransform: 'uppercase' }}>Oxygen requirement</span>
-              <span style={{ font: `500 11px ${THEME.mono}`, color: O2.prediction === pub.O2 ? THEME.pos : THEME.warn }}>
-                {O2.prediction === pub.O2 ? '✓ match' : '△ mismatch'}
-              </span>
+              <div style={{ display: 'flex', alignItems: 'center', gap: 8 }}>
+                <SourceBadge source={O2.source} compact />
+                <span style={{ font: `500 11px ${THEME.mono}`, color: O2.prediction === pub.O2 ? THEME.pos : THEME.warn }}>
+                  {O2.prediction === pub.O2 ? '✓ match' : '△ mismatch'}
+                </span>
+              </div>
             </div>
             <div style={{ display: 'grid', gridTemplateColumns: '1fr 1fr', gap: 18, alignItems: 'center' }}>
               <div>