diff --git a/configs/model/graph/dir_gnn.yaml b/configs/model/graph/dir_gnn.yaml
new file mode 100644
index 000000000..d08df0140
--- /dev/null
+++ b/configs/model/graph/dir_gnn.yaml
@@ -0,0 +1,12 @@
+# Dir-GNN: Edge Directionality Improves Learning on Heterophilic Graphs
+# Rossi et al., LoG 2024 — https://arxiv.org/abs/2305.10498
+
+defaults:
+  - /model/graph/gcn.yaml  # inherit base structure, override below
+
+backbone:
+  _target_: topobench.nn.backbones.graph.dir_gnn.DirGNN
+  hidden_channels: 64
+  num_layers: 2
+  dropout: 0.5
+  alpha: 0.5          # equal weighting of in/out aggregation
diff --git a/test/nn/backbones/graph/test_dir_gnn.py b/test/nn/backbones/graph/test_dir_gnn.py
new file mode 100644
index 000000000..7ceb33c3c
--- /dev/null
+++ b/test/nn/backbones/graph/test_dir_gnn.py
@@ -0,0 +1,98 @@
+"""Unit tests for DirGNN backbone (Rossi et al., LoG 2024)."""
+
+import pytest
+import torch
+from topobench.nn.backbones.graph.dir_gnn import DirGNN, DirSageConv
+
+
+@pytest.fixture
+def small_graph():
+    """A tiny directed graph: 10 nodes, 20 random edges."""
+    x = torch.randn(10, 8)
+    edge_index = torch.randint(0, 10, (2, 20))
+    return x, edge_index
+
+
+class TestDirSageConv:
+    def test_output_shape(self, small_graph):
+        x, edge_index = small_graph
+        conv = DirSageConv(in_channels=8, out_channels=16)
+        out = conv(x, edge_index)
+        assert out.shape == (10, 16), f"Expected (10, 16), got {out.shape}"
+
+    def test_alpha_zero(self, small_graph):
+        """alpha=0: only out-neighbour aggregation contributes."""
+        x, edge_index = small_graph
+        conv = DirSageConv(in_channels=8, out_channels=16, alpha=0.0)
+        out = conv(x, edge_index)
+        assert out.shape == (10, 16)
+
+    def test_alpha_one(self, small_graph):
+        """alpha=1: only in-neighbour aggregation contributes."""
+        x, edge_index = small_graph
+        conv = DirSageConv(in_channels=8, out_channels=16, alpha=1.0)
+        out = conv(x, edge_index)
+        assert out.shape == (10, 16)
+
+    def test_no_edges(self):
+        """Graph with no edges: should fall back to self-transform."""
+        x = torch.randn(5, 8)
+        edge_index = torch.zeros((2, 0), dtype=torch.long)
+        conv = DirSageConv(in_channels=8, out_channels=16)
+        out = conv(x, edge_index)
+        assert out.shape == (5, 16)
+
+    def test_reset_parameters(self, small_graph):
+        x, edge_index = small_graph
+        conv = DirSageConv(in_channels=8, out_channels=16)
+        conv.reset_parameters()
+        out = conv(x, edge_index)
+        assert not torch.isnan(out).any()
+
+
+class TestDirGNN:
+    def test_output_shape_2layers(self, small_graph):
+        x, edge_index = small_graph
+        model = DirGNN(in_channels=8, hidden_channels=32, out_channels=4, num_layers=2)
+        out = model(x, edge_index)
+        assert out.shape == (10, 4)
+
+    def test_output_shape_3layers(self, small_graph):
+        x, edge_index = small_graph
+        model = DirGNN(in_channels=8, hidden_channels=32, out_channels=4, num_layers=3)
+        out = model(x, edge_index)
+        assert out.shape == (10, 4)
+
+    def test_output_shape_1layer(self, small_graph):
+        x, edge_index = small_graph
+        model = DirGNN(in_channels=8, hidden_channels=32, out_channels=4, num_layers=1)
+        out = model(x, edge_index)
+        assert out.shape == (10, 4)
+
+    def test_no_nan_output(self, small_graph):
+        x, edge_index = small_graph
+        model = DirGNN(in_channels=8, hidden_channels=32, out_channels=4)
+        out = model(x, edge_index)
+        assert not torch.isnan(out).any()
+
+    def test_train_eval_mode(self, small_graph):
+        """Dropout should behave differently in train vs eval."""
+        x, edge_index = small_graph
+        model = DirGNN(in_channels=8, hidden_channels=32, out_channels=4, dropout=0.9)
+        model.train()
+        out_train = model(x, edge_index)
+        model.eval()
+        with torch.no_grad():
+            out_eval = model(x, edge_index)
+        # Shapes must match regardless of mode
+        assert out_train.shape == out_eval.shape
+
+    def test_gradient_flow(self, small_graph):
+        """Loss should produce gradients for all parameters."""
+        x, edge_index = small_graph
+        model = DirGNN(in_channels=8, hidden_channels=32, out_channels=4)
+        out = model(x, edge_index)
+        loss = out.sum()
+        loss.backward()
+        for name, param in model.named_parameters():
+            assert param.grad is not None, f"No gradient for {name}"
diff --git a/topobench/nn/backbones/graph/dir_gnn.py b/topobench/nn/backbones/graph/dir_gnn.py
new file mode 100644
index 000000000..c99499ef4
--- /dev/null
+++ b/topobench/nn/backbones/graph/dir_gnn.py
@@ -0,0 +1,283 @@
+"""Directed Graph Neural Network (Dir-GNN).
+
+Implements the Dir-GNN framework from:
+
+    Rossi et al., "Edge Directionality Improves Learning on
+    Heterophilic Graphs", Learning on Graphs Conference, 2024.
+    https://arxiv.org/abs/2305.10498
+
+The key insight is that treating a graph as *directed* increases the
+effective homophily of the graph in heterophilic settings, because
+the separate in- and out-neighbourhoods are individually more
+homophilic than the combined symmetric neighbourhood.
+
+Dir-GNN extends any Message Passing Neural Network (MPNN) to
+directed graphs by performing **separate aggregations** of incoming
+and outgoing edges (Eq. 4 in the paper)::
+
+    h_v^{(l+1)} = sigma(
+        W_self * h_v^{(l)}
+        + alpha     * W_in  * AGG_{u:(u,v)∈E}( h_u^{(l)} )
+        + (1-alpha) * W_out * AGG_{u:(v,u)∈E}( h_u^{(l)} )
+    )
+
+where alpha ∈ [0, 1] controls the balance between in- and
+out-neighbour aggregations.
+
+Theoretical result (Theorem 1, Rossi et al.): Dir-GNN matches the
+expressivity of the Directed Weisfeiler-Lehman (DWL) test, strictly
+exceeding conventional undirected MPNNs.
+"""
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch import Tensor
+from torch_geometric.nn import MessagePassing
+from torch_geometric.utils import degree
+
+
+class DirSageConv(MessagePassing):
+    """Single Dir-GNN convolutional layer (mean / degree-normalised).
+
+    Performs separate mean-aggregations over in-neighbours and
+    out-neighbours, then combines them with the node's own features
+    via learned linear maps (Eq. 4 in Rossi et al., 2024).
+
+    Parameters
+    ----------
+    in_channels : int
+        Dimensionality of input node features.
+    out_channels : int
+        Dimensionality of output node features.
+    alpha : float, optional
+        Mixing coefficient in [0, 1] for in- vs out-aggregation.
+        ``alpha = 0.5`` gives equal weight to both directions.
+        Default: ``0.5``.
+
+    Examples
+    --------
+    >>> import torch
+    >>> from topobench.nn.backbones.graph.dir_gnn import DirSageConv
+    >>> conv = DirSageConv(in_channels=8, out_channels=16)
+    >>> x = torch.randn(10, 8)
+    >>> edge_index = torch.randint(0, 10, (2, 20))
+    >>> out = conv(x, edge_index)
+    >>> out.shape
+    torch.Size([10, 16])
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        alpha: float = 0.5,
+    ) -> None:
+        # Use 'add' aggr; we apply degree normalisation manually so
+        # that in-degree and out-degree normalisation are kept separate.
+        super().__init__(aggr="add")
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.alpha = alpha
+
+        # Separate transforms for in-agg, out-agg, and identity (self).
+        # Eq. 4: W_in, W_out, W_self from the paper.
+        self.lin_in = nn.Linear(in_channels, out_channels, bias=False)
+        self.lin_out = nn.Linear(in_channels, out_channels, bias=False)
+        self.lin_self = nn.Linear(in_channels, out_channels, bias=False)
+        self.bias = nn.Parameter(torch.zeros(out_channels))
+
+        self.reset_parameters()
+
+    def reset_parameters(self) -> None:
+        """Initialise weights with Glorot uniform initialisation."""
+        nn.init.xavier_uniform_(self.lin_in.weight)
+        nn.init.xavier_uniform_(self.lin_out.weight)
+        nn.init.xavier_uniform_(self.lin_self.weight)
+        nn.init.zeros_(self.bias)
+
+    def forward(self, x: Tensor, edge_index: Tensor) -> Tensor:
+        """Forward pass for one Dir-GNN layer.
+
+        Parameters
+        ----------
+        x : Tensor, shape [N, in_channels]
+            Node feature matrix.
+        edge_index : Tensor, shape [2, E]
+            Edge indices in COO format.
+            ``edge_index[0]`` = source nodes,
+            ``edge_index[1]`` = target nodes.
+
+        Returns
+        -------
+        Tensor
+            Updated node embeddings of shape [N, out_channels].
+        """
+        num_nodes = x.size(0)
+
+        # --- In-neighbour aggregation (Eq. 4, first agg term) -------
+        # Standard propagation: messages flow from src → dst.
+        # dst receives from its *in-neighbours* (nodes that point to it).
+        deg_in = degree(
+            edge_index[1], num_nodes=num_nodes, dtype=x.dtype
+        ).clamp(min=1.0)
+        # Per-edge norm: 1 / deg_in[dst]
+        norm_in = 1.0 / deg_in[edge_index[1]]
+        agg_in = self.propagate(edge_index, x=x, norm=norm_in)
+
+        # --- Out-neighbour aggregation (Eq. 4, second agg term) ------
+        # Reverse edges so that each node now *receives* from its
+        # out-neighbours (nodes it originally pointed to).
+        edge_index_t = edge_index.flip(0)  # shape [2, E]: [dst, src]
+        deg_out = degree(
+            edge_index_t[1], num_nodes=num_nodes, dtype=x.dtype
+        ).clamp(min=1.0)
+        norm_out = 1.0 / deg_out[edge_index_t[1]]
+        agg_out = self.propagate(edge_index_t, x=x, norm=norm_out)
+
+        # --- Combine (Eq. 4) -----------------------------------------
+        out = (
+            self.lin_self(x)
+            + self.alpha * self.lin_in(agg_in)
+            + (1.0 - self.alpha) * self.lin_out(agg_out)
+            + self.bias
+        )
+        return out
+
+    def message(self, x_j: Tensor, norm: Tensor) -> Tensor:
+        """Compute degree-normalised messages.
+
+        Parameters
+        ----------
+        x_j : Tensor, shape [E, in_channels]
+            Source-node features for each edge.
+        norm : Tensor, shape [E]
+            Per-edge normalisation factor (1 / degree of target node).
+
+        Returns
+        -------
+        Tensor
+            Normalised messages of shape [E, in_channels].
+        """
+        return norm.view(-1, 1) * x_j
+
+
+class DirGNN(nn.Module):
+    """Directed Graph Neural Network (Dir-GNN).
+
+    A multi-layer stack of :class:`DirSageConv` layers with ReLU
+    activations and optional dropout.  Implements the framework from
+    Rossi et al. (2024) which proves that separate in/out aggregations
+    match the expressive power of the Directed Weisfeiler-Lehman test.
+
+    The model is particularly effective on heterophilic graphs, where
+    using directed edges provably increases effective homophily compared
+    to symmetrised (undirected) variants.
+
+    The forward signature accepts ``**kwargs`` so that the standard
+    :class:`~topobench.nn.wrappers.GNNWrapper` (which passes
+    ``batch`` and ``edge_weight``) can wrap this backbone without
+    modification.
+
+    Parameters
+    ----------
+    in_channels : int
+        Dimensionality of input node features.
+    hidden_channels : int
+        Dimensionality of hidden layer representations.
+    out_channels : int
+        Dimensionality of the final output representations.
+    num_layers : int, optional
+        Total number of Dir-GNN convolutional layers.  Default: ``2``.
+    dropout : float, optional
+        Dropout probability applied after each hidden layer.
+        Default: ``0.5``.
+    alpha : float, optional
+        In/out mixing coefficient passed to every
+        :class:`DirSageConv`.  Default: ``0.5``.
+
+    Examples
+    --------
+    >>> import torch
+    >>> from topobench.nn.backbones.graph.dir_gnn import DirGNN
+    >>> model = DirGNN(
+    ...     in_channels=16, hidden_channels=64,
+    ...     out_channels=7, num_layers=2,
+    ... )
+    >>> x = torch.randn(100, 16)
+    >>> edge_index = torch.randint(0, 100, (2, 300))
+    >>> out = model(x, edge_index)
+    >>> out.shape
+    torch.Size([100, 7])
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        hidden_channels: int,
+        out_channels: int,
+        num_layers: int = 2,
+        dropout: float = 0.5,
+        alpha: float = 0.5,
+    ) -> None:
+        super().__init__()
+        self.dropout = dropout
+        self.num_layers = num_layers
+        self.out_channels = out_channels
+
+        if num_layers < 1:
+            msg = f"num_layers must be >= 1, got {num_layers}"
+            raise ValueError(msg)
+
+        self.convs = nn.ModuleList()
+
+        if num_layers == 1:
+            self.convs.append(
+                DirSageConv(in_channels, out_channels, alpha=alpha)
+            )
+        else:
+            # Input layer
+            self.convs.append(
+                DirSageConv(in_channels, hidden_channels, alpha=alpha)
+            )
+            # Intermediate hidden layers
+            for _ in range(num_layers - 2):
+                self.convs.append(
+                    DirSageConv(hidden_channels, hidden_channels, alpha=alpha)
+                )
+            # Output layer
+            self.convs.append(
+                DirSageConv(hidden_channels, out_channels, alpha=alpha)
+            )
+
+    def forward(
+        self,
+        x: Tensor,
+        edge_index: Tensor,
+        **kwargs,
+    ) -> Tensor:
+        """Forward pass through all Dir-GNN layers.
+
+        Parameters
+        ----------
+        x : Tensor, shape [N, in_channels]
+            Node feature matrix.
+        edge_index : Tensor, shape [2, E]
+            Edge connectivity in COO format.
+        **kwargs : dict
+            Ignored extra keyword arguments (e.g. ``batch``,
+            ``edge_weight``) passed by
+            :class:`~topobench.nn.wrappers.GNNWrapper`.
+
+        Returns
+        -------
+        Tensor
+            Node-level output representations of shape
+            [N, out_channels].
+        """
+        for i, conv in enumerate(self.convs):
+            x = conv(x, edge_index)
+            if i < self.num_layers - 1:
+                x = F.relu(x)
+                x = F.dropout(x, p=self.dropout, training=self.training)
+        return x