Update app.py

app.py

@@ -99,7 +99,8 @@ st.markdown(base_css, unsafe_allow_html=True)
 MODEL_REPO = "ym59/velobind-models"
 MODEL_DIR = Path("output/models")
 PREP_DIR = Path("output/preprocessors")
-AD_EMB_PATH = Path("output/ad_train_embeddings.npy")
+AD_CENTROID_PATH = Path("output/models/deployment/ad_centroid.npy")
+AD_THRESHOLD_PATH = Path("output/models/deployment/ad_threshold.npy")
 
 _DESC_FNS: Optional[List[Any]] = None
 try:
@@ -176,12 +177,11 @@ def load_models() -> Tuple[Dict[str, Any], Optional[Any], Optional[Any], Optiona
                 except Exception:
                     pass
 
-        if AD_EMB_PATH.exists():
+        if AD_CENTROID_PATH.exists():
             try:
-                train_embs = np.load(str(AD_EMB_PATH))
-                at = Path("output/ad_threshold.npy")
-                if at.exists():
-                    ad_threshold = float(np.load(str(at)))
+                train_embs = np.load(str(AD_CENTROID_PATH))
+                if AD_THRESHOLD_PATH.exists():
+                    ad_threshold = float(np.load(str(AD_THRESHOLD_PATH)))
             except Exception:
                 pass
 
@@ -375,15 +375,15 @@ def predict_pkd(X: np.ndarray, fold_models: Dict[str, Any], meta: Any, iso_cal:
 
 def check_ad(esm_mean: np.ndarray, train_embs: Optional[np.ndarray], ad_threshold: float) -> Tuple[bool, float]:
     if train_embs is None:
-        return True, 0.0
+        return False, 0.0  # Fail safely to OUT OF DOMAIN if files are missing
     try:
-        from sklearn.metrics.pairwise import cosine_distances
-        q = esm_mean[-480:].reshape(1, -1)
-        d = cosine_distances(q, train_embs[:2000])[0]
-        k = float(np.sort(d)[:5].mean())
-        return k <= ad_threshold, k
-    except Exception:
-        return True, 0.0
+        q = esm_mean[-480:]
+        # Calculate Euclidean distance to the centroid
+        dist = float(np.linalg.norm(q - train_embs))
+        return dist <= ad_threshold, dist
+    except Exception as e:
+        logger.debug("check_ad error: %s", e)
+        return False, 0.0
 
 
 def clean_fasta(s: str) -> str: