Update app.py

app.py

@@ -32,6 +32,7 @@ TRAINING_DATA_PATH = "training_data.csv"
 
 LIME_BACKGROUND_SIZE = 1000
 LIME_NUM_SAMPLES = 3000
+LIME_RANDOM_SEED = 42
 
 
 # ============================================================
@@ -307,7 +308,28 @@ def build_lime_background():
 
 
 # ============================================================
-# CREATE LIME EXPLAINER
+# LIME EXPLAINER FACTORY
+# ------------------------------------------------------------
+# IMPORTANT FIX:
+#
+# LimeTabularExplainer(random_state=42) only seeds its internal
+# RandomState ONCE, at construction time. Every subsequent call
+# to explain_instance() draws from -- and ADVANCES -- that same
+# shared RandomState object. If you build the explainer ONCE at
+# startup and reuse it across requests (as the original script
+# did), the 1st, 2nd, 3rd, ... prediction on the SAME peptide
+# each get a different slice of the random stream, so the
+# perturbation neighborhood (and therefore the LIME weights)
+# silently drifts between runs even though the RF model and
+# MIC models are fully deterministic.
+#
+# The fix is to build a brand-new explainer, re-seeded from
+# scratch, right before every explain_instance() call. This is
+# cheap (it only recomputes discretization stats over the
+# precomputed background array -- no model inference), and it
+# guarantees the SAME peptide produces BYTE-IDENTICAL LIME
+# output every single time, regardless of how many predictions
+# came before it in the running app.
 # ============================================================
 
 lime_training_data = (
@@ -315,24 +337,29 @@ lime_training_data = (
 )
 
 
-explainer = LimeTabularExplainer(
+def build_lime_explainer():
 
-    training_data=lime_training_data,
+    return LimeTabularExplainer(
 
-    feature_names=selected_features,
+        training_data=lime_training_data,
 
-    class_names=[
-        "AMP",
-        "Non-AMP"
-    ],
+        feature_names=selected_features,
 
-    mode="classification",
+        class_names=[
+            "AMP",
+            "Non-AMP"
+        ],
 
-    discretize_continuous=True,
+        mode="classification",
+
+        discretize_continuous=True,
+
+        random_state=LIME_RANDOM_SEED
+    )
 
-    random_state=42
-)
 
+# Sanity-check that it builds correctly at startup.
+_ = build_lime_explainer()
 
 print("\nREAL LIME EXPLAINER READY.")
 
@@ -785,6 +812,13 @@ def full_prediction(sequence):
 
     try:
 
+        # Fresh, re-seeded explainer every call -- see the
+        # comment above build_lime_explainer() for why this
+        # matters. This is what makes the explanation for a
+        # given peptide reproducible run after run.
+
+        explainer = build_lime_explainer()
+
         explanation = (
             explainer.explain_instance(