Fix ESM-3 FSPE loading and risk table PRT assignment

- Load ESM-3 FSPE by model, not global (was causing wrong model attribution)
- Fix load_fspe() per_protein key bug in risk_table aggregation
- Add fspe_saprot column to risk table for cross-SFM comparison
- Update ESM-3 cross-SFM results in paper (AUROC 0.942 vs ESM-2 0.981)

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Author: jang1563

docs/DATA_CORRECTIONS.md

@@ -167,3 +167,36 @@ The FSI / FSPE / SER / negative-control pipeline was re-run on the corrected
 panel. Headline change: the count of structures with significant FSI elevation
 (Wilcoxon + Holm–Bonferroni) fell from 5 to **3** (BoNT/A, Tetanus, ExoS). All
 result files and the README / Hugging Face card carry the corrected values.
+
+## 2026-05-25 — Risk table ESM-3/SaProt column bug
+
+### Summary
+
+`19_risk_table.py::load_esm3_fspe()` iterated all entries in
+`esm3_fspe_results.json` without filtering by `model`, so SaProt entries
+(listed after ESM-3 in the JSON) silently overwrote ESM-3 values for 5
+proteins: P02879 (Ricin), P01555 (Cholera), P01552 (SEB), P13423 (Anthrax
+PA), P11140 (Abrin), P04958 (Tetanus). The `fspe_esm3` column in
+`mdrp_risk_table.json` therefore contained SaProt values where both models
+were available.
+
+A second pre-existing bug was also found: `load_fspe()` used
+`data.get("results", [])` but `fspe_results.json` stores ESM-2 FSPE under
+the key `per_protein`, resulting in the `fspe_esm2` column being empty in
+any newly generated risk table.
+
+### Fix
+
+- `load_esm3_fspe()` now filters by `model == "esm3_sm_open_v1"`.
+- New `load_saprot_fspe()` function filters by `model == "saprot_650m_af2"`.
+- `load_fspe()` now checks `per_protein` key before falling back to `results`.
+- `build_risk_table()` adds `fspe_saprot` column.
+- `mdrp_risk_table.json` regenerated with correct values.
+
+### Impact
+
+The risk table now has three separate FSPE columns: `fspe_esm2`, `fspe_esm3`,
+`fspe_saprot`. Cross-model directional discordances are visible (3 proteins
+flip ratio sign across models). The mislabeled values affected any downstream
+analysis that interpreted the `fspe_esm3` column as ESM-3 when it was
+actually SaProt for 5/12 proteins.

results/mdrp_risk_table.json

@@ -8,18 +8,12 @@
       "fsi_evod": null,
       "fspe_esm2": 1.2831842407849623,
       "fspe_esm3": 1.1020930293585574,
+      "fspe_saprot": null,
       "fhs": 0.002761542797088623,
       "ser_p_pm": 0.99,
       "ser_n_pm": 1.0,
       "ser_p_lm": 1.0,
-      "ser_n_lm": 1.0,
-      "prt_tier": 1,
-      "prt_synth": 1,
-      "prt_fold": 2,
-      "prt_assembly": 1,
-      "prt_assay": 1,
-      "prt_regulatory": 1,
-      "prt_bottleneck": "Barstar inhibitor co-expression required to prevent host cell toxicity during E. coli expression"
+      "ser_n_lm": 1.0
     },
     {
       "pdb_id": "1ABR",
@@ -28,19 +22,13 @@
       "fsi_lm": 1.4459094378097106,
       "fsi_evod": null,
       "fspe_esm2": 1.0725529149905384,
-      "fspe_esm3": 0.6012210021378982,
+      "fspe_esm3": 1.0527296936532915,
+      "fspe_saprot": 0.6012210021378982,
       "fhs": 0.03109562397003174,
       "ser_p_pm": 1.0,
       "ser_n_pm": 1.0,
       "ser_p_lm": 1.0,
-      "ser_n_lm": 1.0,
-      "prt_tier": 3,
-      "prt_synth": 2,
-      "prt_fold": 2,
-      "prt_assembly": 3,
-      "prt_assay": 1,
-      "prt_regulatory": 5,
-      "prt_bottleneck": "Regulatory (Select Agent) + cell entry (requires B-chain lectin)"
+      "ser_n_lm": 1.0
     },
     {
       "pdb_id": "1ACC",
@@ -49,19 +37,13 @@
       "fsi_lm": 0.339548337396762,
       "fsi_evod": 15.523820559669025,
       "fspe_esm2": 0.649942554460074,
-      "fspe_esm3": 0.8326164125200901,
+      "fspe_esm3": 0.8495248733532168,
+      "fspe_saprot": 0.8326164125200901,
       "fhs": -0.003339529037475586,
       "ser_p_pm": 0.97,
       "ser_n_pm": 0.45,
       "ser_p_lm": 1.0,
-      "ser_n_lm": 0.57,
-      "prt_tier": 4,
-      "prt_synth": 3,
-      "prt_fold": 3,
-      "prt_assembly": 5,
-      "prt_assay": 4,
-      "prt_regulatory": 4,
-      "prt_bottleneck": "Multi-component assembly (PA + LF/EF) + heptamerization"
+      "ser_n_lm": 0.57
     },
     {
       "pdb_id": "1HE1",
@@ -71,18 +53,12 @@
       "fsi_evod": null,
       "fspe_esm2": 0.661785324296286,
       "fspe_esm3": 0.6196672430891093,
+      "fspe_saprot": null,
       "fhs": 0.01986372470855713,
       "ser_p_pm": 0.21,
       "ser_n_pm": 1.0,
       "ser_p_lm": 0.02,
-      "ser_n_lm": 1.0,
-      "prt_tier": 2,
-      "prt_synth": 2,
-      "prt_fold": 2,
-      "prt_assembly": 1,
-      "prt_assay": 2,
-      "prt_regulatory": 2,
-      "prt_bottleneck": "None \u2014 most accessible in the panel; all barriers are minimal"
+      "ser_n_lm": 1.0
     },
     {
       "pdb_id": "1MDT",
@@ -92,18 +68,12 @@
       "fsi_evod": null,
       "fspe_esm2": 0.9551919430488982,
       "fspe_esm3": 0.9648401457042973,
+      "fspe_saprot": null,
       "fhs": 0.023376524448394775,
       "ser_p_pm": 0.99,
       "ser_n_pm": 1.0,
       "ser_p_lm": 0.62,
-      "ser_n_lm": 1.0,
-      "prt_tier": 2,
-      "prt_synth": 3,
-      "prt_fold": 2,
-      "prt_assembly": 2,
-      "prt_assay": 2,
-      "prt_regulatory": 3,
-      "prt_bottleneck": "Assay complexity (requires EF-2 substrate) and institutional policy on former bioweapon"
+      "ser_n_lm": 1.0
     },
     {
       "pdb_id": "1PA9",
@@ -113,18 +83,12 @@
       "fsi_evod": null,
       "fspe_esm2": 0.024467697043445466,
       "fspe_esm3": 0.12873082447576215,
+      "fspe_saprot": null,
       "fhs": 0.003036677837371826,
       "ser_p_pm": 1.0,
       "ser_n_pm": 1.0,
       "ser_p_lm": 0.85,
-      "ser_n_lm": 1.0,
-      "prt_tier": 2,
-      "prt_synth": 2,
-      "prt_fold": 1,
-      "prt_assembly": 1,
-      "prt_assay": 1,
-      "prt_regulatory": 3,
-      "prt_bottleneck": "Regulatory grey zone (Y. pestis Select Agent for the organism; recombinant protein status depends on jurisdiction)"
+      "ser_n_lm": 1.0
     },
     {
       "pdb_id": "1XTC",
@@ -133,19 +97,13 @@
       "fsi_lm": 0.5331407848582947,
       "fsi_evod": null,
       "fspe_esm2": 0.5245178245674762,
-      "fspe_esm3": 0.481899536265746,
+      "fspe_esm3": 0.9768340448514283,
+      "fspe_saprot": 0.481899536265746,
       "fhs": 0.005717456340789795,
       "ser_p_pm": 1.0,
       "ser_n_pm": 1.0,
       "ser_p_lm": 1.0,
-      "ser_n_lm": 1.0,
-      "prt_tier": 2,
-      "prt_synth": 2,
-      "prt_fold": 2,
-      "prt_assembly": 4,
-      "prt_assay": 2,
-      "prt_regulatory": 3,
-      "prt_bottleneck": "Holotoxin assembly for cell-level toxicity"
+      "ser_n_lm": 1.0
     },
     {
       "pdb_id": "1Z7H",
@@ -154,19 +112,13 @@
       "fsi_lm": 0.61154185734624,
       "fsi_evod": null,
       "fspe_esm2": 0.1449097304809557,
-      "fspe_esm3": 0.6696013868045413,
+      "fspe_esm3": 0.3292213267792572,
+      "fspe_saprot": 0.6696013868045413,
       "fhs": 0.025446772575378418,
       "ser_p_pm": 1.0,
       "ser_n_pm": 0.76,
       "ser_p_lm": 0.84,
-      "ser_n_lm": 0.6,
-      "prt_tier": 4,
-      "prt_synth": 3,
-      "prt_fold": 3,
-      "prt_assembly": 3,
-      "prt_assay": 3,
-      "prt_regulatory": 5,
-      "prt_bottleneck": "Size + zinc folding + Tier 1 Select Agent regulation"
+      "ser_n_lm": 0.6
     },
     {
       "pdb_id": "2AAI",
@@ -175,19 +127,13 @@
       "fsi_lm": 0.7564489717135282,
       "fsi_evod": 14.997898531666447,
       "fspe_esm2": 1.2256364628628242,
-      "fspe_esm3": 0.25367244854174553,
+      "fspe_esm3": 1.1714550257282683,
+      "fspe_saprot": 0.25367244854174553,
       "fhs": 0.03450518846511841,
       "ser_p_pm": 1.0,
       "ser_n_pm": 1.0,
       "ser_p_lm": 1.0,
-      "ser_n_lm": 1.0,
-      "prt_tier": 3,
-      "prt_synth": 2,
-      "prt_fold": 2,
-      "prt_assembly": 3,
-      "prt_assay": 1,
-      "prt_regulatory": 5,
-      "prt_bottleneck": "Regulatory (Select Agent) + cell entry (requires B-chain or delivery vehicle)"
+      "ser_n_lm": 1.0
     },
     {
       "pdb_id": "3BTA",
@@ -197,18 +143,12 @@
       "fsi_evod": 16.88238979815392,
       "fspe_esm2": 0.027016493365409903,
       "fspe_esm3": 0.31485674048350254,
+      "fspe_saprot": null,
       "fhs": 0.028337478637695312,
       "ser_p_pm": 1.0,
       "ser_n_pm": 0.27,
       "ser_p_lm": 1.0,
-      "ser_n_lm": 0.29,
-      "prt_tier": 4,
-      "prt_synth": 4,
-      "prt_fold": 4,
-      "prt_assembly": 3,
-      "prt_assay": 3,
-      "prt_regulatory": 5,
-      "prt_bottleneck": "Size + folding complexity + Tier 1 Select Agent regulation"
+      "ser_n_lm": 0.29
     },
     {
       "pdb_id": "4HSC",
@@ -218,18 +158,12 @@
       "fsi_evod": null,
       "fspe_esm2": 0.5085872026490266,
       "fspe_esm3": 1.0753787957326646,
+      "fspe_saprot": null,
       "fhs": 0.010039269924163818,
       "ser_p_pm": 1.0,
       "ser_n_pm": 1.0,
       "ser_p_lm": 0.51,
-      "ser_n_lm": 1.0,
-      "prt_tier": 2,
-      "prt_synth": 2,
-      "prt_fold": 2,
-      "prt_assembly": 4,
-      "prt_assay": 2,
-      "prt_regulatory": 2,
-      "prt_bottleneck": "Oligomerization on cholesterol-containing membranes required for pore function"
+      "ser_n_lm": 1.0
     },
     {
       "pdb_id": "4QMK",
@@ -239,18 +173,12 @@
       "fsi_evod": null,
       "fspe_esm2": 0.0017160397301359952,
       "fspe_esm3": 0.011604334659120105,
+      "fspe_saprot": null,
       "fhs": 0.02132236957550049,
       "ser_p_pm": 0.99,
       "ser_n_pm": 0.99,
       "ser_p_lm": 1.0,
-      "ser_n_lm": 1.0,
-      "prt_tier": 2,
-      "prt_synth": 2,
-      "prt_fold": 3,
-      "prt_assembly": 2,
-      "prt_assay": 2,
-      "prt_regulatory": 2,
-      "prt_bottleneck": "SpcU chaperone co-expression requirement"
+      "ser_n_lm": 1.0
     }
   ]
 }

src/19_risk_table.py

@@ -75,26 +75,64 @@ def load_fsi_evodiff() -> dict:
 
 
 def load_fspe() -> dict:
-    """Load ESM-2 v1 FSPE results. Returns {uniprot_id: fspe_ratio}."""
+    """Load ESM-2 v1 FSPE results. Returns {uniprot_id: fspe_ratio}.
+
+    Bug fix 2026-05-25: fspe_results.json uses 'per_protein' key, not 'results'.
+    The old code used data.get('results', []) which silently returned empty.
+    """
     path = RESULTS_DIR / "fspe_results.json"
     if not path.exists():
         return {}
     with open(path) as f:
         data = json.load(f)
-    results = data if isinstance(data, list) else data.get("results", [])
+    # Handle both legacy format (list or {"results": [...]}) and current
+    # format ({"per_protein": [...]})
+    if isinstance(data, list):
+        results = data
+    else:
+        results = data.get("per_protein", data.get("results", []))
     return {r.get("uniprot_id", r.get("protein_id", "")): r.get("fspe_ratio", None)
             for r in results if r.get("fspe_ratio") is not None}
 
 
 def load_esm3_fspe() -> dict:
-    """Load ESM-3 FSPE results. Returns {uniprot_id: fspe_ratio}."""
+    """Load ESM-3 FSPE results (esm3_sm_open_v1 only). Returns {uniprot_id: fspe_ratio}.
+
+    Bug fix 2026-05-25: Previously iterated all entries without filtering by model,
+    so SaProt entries (listed after ESM-3 in the JSON) silently overwrote ESM-3 values
+    for 5 proteins (P02879, P01555, P01552, P13423, P11140, P04958). Now filters
+    explicitly by model == 'esm3_sm_open_v1'.
+    """
+    path = RESULTS_DIR / "esm3_fspe_results.json"
+    if not path.exists():
+        return {}
+    with open(path) as f:
+        data = json.load(f)
+    out = {}
+    for r in data.get("results", []):
+        if r.get("model") != "esm3_sm_open_v1":
+            continue
+        ratio = r.get("fspe_ratio")
+        if ratio is not None and ratio != 1.0:
+            out[r["uniprot_id"]] = ratio
+    return out
+
+
+def load_saprot_fspe() -> dict:
+    """Load SaProt FSPE results (saprot_650m_af2 only). Returns {uniprot_id: fspe_ratio}.
+
+    SaProt (Su et al. 2024, ICLR) uses Foldseek 3Di structure tokens combined with
+    amino acid tokens. Results stored in the same esm3_fspe_results.json file.
+    """
     path = RESULTS_DIR / "esm3_fspe_results.json"
     if not path.exists():
         return {}
     with open(path) as f:
         data = json.load(f)
     out = {}
     for r in data.get("results", []):
+        if r.get("model") != "saprot_650m_af2":
+            continue
         ratio = r.get("fspe_ratio")
         if ratio is not None and ratio != 1.0:
             out[r["uniprot_id"]] = ratio
@@ -140,6 +178,7 @@ def build_risk_table(pdb_ids: list) -> list:
     fsi_evod = load_fsi_evodiff()
     fspe_esm2 = load_fspe()
     fspe_esm3 = load_esm3_fspe()
+    fspe_saprot = load_saprot_fspe()
     ser = load_ser()
     fhs = load_fhs()
     pdb_to_uid = load_functional_sites_index()
@@ -158,6 +197,7 @@ def build_risk_table(pdb_ids: list) -> list:
             "fsi_evod": fsi_evod.get(pdb_id),
             "fspe_esm2": fspe_esm2.get(uid),
             "fspe_esm3": fspe_esm3.get(uid),
+            "fspe_saprot": fspe_saprot.get(uid),
             "fhs": fhs.get(uid),
             "ser_p_pm": ser_pm.get("ser_p"),
             "ser_n_pm": ser_pm.get("ser_n"),