Update app.py

app.py

@@ -731,6 +731,9 @@ def visualize_attention_and_ranges(
 
 
 # ───── Gradio callbacks ─────────────────────────────────────────
+ROOT = os.path.dirname(os.path.abspath(__file__))
+FOLDSEEK_BIN = os.path.join(ROOT, "utils", "foldseek")
+
 def extract_aa_seq_cb(structure_file, protein_text):
     """
     Extract plain amino acid sequence from uploaded PDB/mmCIF.
@@ -1156,21 +1159,26 @@ with gr.Blocks() as demo:
         <ol style="font-size:1rem;line-height:1.6;margin-left:22px;">
           <li>
             <strong>Input formats:</strong>
-            The model accepts <em>structure-aware (SA)</em> or <em>FASTA</em> protein sequences,
-            and <em>SMILES</em> or <em>SELFIES</em> representations for ligands.
-            For SA mode, <code>.pdb</code> or <code>.cif</code> files can be uploaded directly.
+            The system supports either <em>structure-aware (SA)</em> sequences derived from
+            protein structures or conventional <em>FASTA</em> sequences.
+            For structure-based analysis, users may upload <code>.pdb</code> or
+            <code>.cif</code> files to extract the corresponding sequence representation.
+            Ligands can be provided in <em>SMILES</em> or <em>SELFIES</em> format.
           </li>
     
           <li>
-            <strong>Interaction type selection:</strong>
-            Choose the desired non-covalent interaction type
-            (e.g., overall interaction or specific physicochemical channels)
-            to visualise token-level binding patterns.
+            <strong>Interaction channel selection:</strong>
+            Users may select a specific non-covalent interaction type
+            (e.g., hydrogen bonding, hydrophobic interactions) or the
+            overall interaction channel to visualise the corresponding
+            token-level binding patterns.
           </li>
     
           <li>
-            <strong>Output:</strong>
-            The demo system reports a predicted binding probability, a ranked Top-K residues table, and a token-level attention heat map.
+            <strong>Model outputs:</strong>
+            The system reports (i) a predicted binding probability for the
+            protein–ligand pair, (ii) a ranked Top-K residue table, and (iii) a token-level interaction
+            heat map illustrating spatial interaction patterns.
           </li>
         </ol>
         """)