diff --git a/jraph/_src/models.py b/jraph/_src/models.py
index e6a2562..dd31752 100644
--- a/jraph/_src/models.py
+++ b/jraph/_src/models.py
@@ -547,19 +547,26 @@ def _ApplyGCN(graph):
     nodes = update_node_fn(nodes)
     # Equivalent to jnp.sum(n_node), but jittable
     total_num_nodes = tree.tree_leaves(nodes)[0].shape[0]
+    
+    # Handle None senders and receivers by initializing empty arrays
+    if senders is None:
+        senders = jnp.array([], dtype=jnp.int32)
+    if receivers is None:
+        receivers = jnp.array([], dtype=jnp.int32)
+        
     if add_self_edges:
-      # We add self edges to the senders and receivers so that each node
-      # includes itself in aggregation.
-      # In principle, a `GraphsTuple` should partition by n_edge, but in
-      # this case it is not required since a GCN is agnostic to whether
-      # the `GraphsTuple` is a batch of graphs or a single large graph.
-      conv_receivers = jnp.concatenate((receivers, jnp.arange(total_num_nodes)),
+        # We add self edges to the senders and receivers so that each node
+        # includes itself in aggregation.
+        # In principle, a `GraphsTuple` should partition by n_edge, but in
+        # this case it is not required since a GCN is agnostic to whether
+        # the `GraphsTuple` is a batch of graphs or a single large graph.
+        conv_receivers = jnp.concatenate((receivers, jnp.arange(total_num_nodes)),
                                        axis=0)
-      conv_senders = jnp.concatenate((senders, jnp.arange(total_num_nodes)),
+        conv_senders = jnp.concatenate((senders, jnp.arange(total_num_nodes)),
                                      axis=0)
     else:
-      conv_senders = senders
-      conv_receivers = receivers
+        conv_senders = senders
+        conv_receivers = receivers
 
     # pylint: disable=g-long-lambda
     if symmetric_normalization: