diff --git a/tfold/deploy/psp_featurizer.py b/tfold/deploy/psp_featurizer.py
index 1809187..91f91e3 100644
--- a/tfold/deploy/psp_featurizer.py
+++ b/tfold/deploy/psp_featurizer.py
@@ -53,20 +53,70 @@ def restore(cls, path):
         logging.info('model weights restored from %s', path)
         return model
 
+    def _normalize_msa(self, msa):
+        """
+        Normalizes the MSA so that all sequences have the same length as the query (first sequence).
+        If a sequence is shorter, it will be padded with gap characters ('-').
+        If it is longer, it will be truncated.
+        """
+        target_length = len(msa[0].strip())
+        normalized = []
+        for seq in msa:
+            s = seq.strip()
+            if len(s) < target_length:
+                s = s + "-" * (target_length - len(s))
+            elif len(s) > target_length:
+                s = s[:target_length]
+            normalized.append(s)
+        return normalized
+
+    def _normalize_deletion_matrix(self, deletion_matrix, msa):
+        """
+        Normalizes the deletion matrix so that each row has the same length as the corresponding MSA sequence.
+        Each row is converted to a list of integers (using 0 for non-digit characters)
+        and padded or trimmed to match the length of the query sequence (assumed equal to len(msa[0])).
+        """
+        normalized = []
+        target_length = len(msa[0])
+        for d, seq in zip(deletion_matrix, msa):
+            if isinstance(d, str):
+                # Convert each character: if digit, cast to int; otherwise 0.
+                row = [int(ch) if ch.isdigit() else 0 for ch in d]
+            elif isinstance(d, (list, tuple)):
+                try:
+                    row = [int(x) for x in d]
+                except Exception:
+                    row = [0] * len(seq)
+            else:
+                row = [0] * len(seq)
+            # Pad or truncate to target_length.
+            if len(row) < target_length:
+                row.extend([0] * (target_length - len(row)))
+            elif len(row) > target_length:
+                row = row[:target_length]
+            normalized.append(row)
+        return normalized
+
     def _get_feature_dict(self, msa_path, idx_resd_beg=None, idx_resd_end=None):
         if os.path.exists(msa_path):
             with open(msa_path) as f:
                 MSA, deletion_matrix = parse_a3m(f.read())
         else:
-            raise ValueError(f'<msa_path> {msa_path} is not existed')
+            raise ValueError(f'<msa_path> {msa_path} does not exist')
 
         feature_dict = {}
-        query_seq = MSA[0]
+        # Assume the first sequence is the query.
+        query_seq = MSA[0].strip()
+
         if idx_resd_beg is not None and idx_resd_end is not None:
             query_seq = query_seq[idx_resd_beg:idx_resd_end]
             MSA = [s[idx_resd_beg:idx_resd_end] for s in MSA]
             deletion_matrix = [d[idx_resd_beg:idx_resd_end] for d in deletion_matrix]
 
+        # Normalize both the MSA and the deletion matrix.
+        MSA = self._normalize_msa(MSA)
+        deletion_matrix = self._normalize_deletion_matrix(deletion_matrix, MSA)
+
         feature_dict.update(data_pipeline.make_sequence_features(query_seq, 'test', len(query_seq)))
         feature_dict.update(data_pipeline.make_msa_features([MSA], [deletion_matrix]))
         feature_dict.update(template_feats_placeholder())
@@ -85,7 +135,8 @@ def forward(self,
         try:
             process_feature_dict = feature_processor.process_features(feature_dict, mode='predict')
         except IndexError as e:
-            logging.error(f'Fail to parse idx_resd_beg: {idx_resd_beg}, idx_resd_end: {idx_resd_end}...')
+            logging.error(
+                f'Failed to parse features with idx_resd_beg: {idx_resd_beg}, idx_resd_end: {idx_resd_end}...')
             raise e
         process_feature_dict = {
             k: torch.as_tensor(v, device=self.device)
@@ -127,8 +178,9 @@ def _forward_impl(self, batch):
             # Enable grad iff we're training and it's the final recycling layer
             is_final_iter = cycle_no == (num_iters - 1)
             with torch.no_grad():
-                # Run the next iteration of the model
-                outputs, m_1_prev, z_prev, x_prev = self.model.iteration(feats, prevs, _recycle=(num_iters > 1))
+                outputs, m_1_prev, z_prev, x_prev = self.model.iteration(
+                    feats, prevs, _recycle=(num_iters > 1)
+                )
                 feat_tns = [
                     outputs['msa'].clone().cpu(),
                     outputs['pair'].clone().cpu(),