DimeNet

The paper introduced directional message passing for molecular graphs, using distances and angles rather than only pairwise distances.

Metadata

Field	Value
Paper	Directional Message Passing for Molecular Graphs
Authors	Johannes Gasteiger, Janek Gross, Stephan Gunnemann
Year	2020
Venue	ICLR 2020
arXiv	2003.03123
OpenReview PDF	B1eWbxStPH
Project	DimeNet project page
Status	verified

Question

Many molecular GNNs use distances between atoms:

$$d_{ij}=\Vert r_i-r_j\Vert .$$

Distances are rotation and translation invariant, but they do not fully describe directional local geometry. Molecular interactions often depend on bond angles and angular potentials. DimeNet asks:

$$\text{Can molecular message passing use directional information without breaking geometric symmetry?}$$

The key move is to represent messages, not only atoms, and associate each message with a direction in coordinate space.

Main Claim

DimeNet proposes directional message passing. Messages are attached to directed edges:

$$m_{j\to i}$$

rather than only node states:

$$h_i.$$

The update uses both distances and angles:

$$d_{ij}=\Vert r_i-r_j\Vert ,{\theta }_{kji}=\angle (r_k-r_j,r_i-r_j).$$

The durable claim is:

$$\text{distance-only molecular GNN}\to \text{directional message passing with angular information}.$$

This makes DimeNet a key bridge between SchNet-style distance models and later geometric molecular architectures.

Architecture Contract

Item	Contract
Input	atoms, coordinates, and molecular graph/radius graph
Node unit	atom
Message unit	directed edge/message $j\to i$
Geometric inputs	interatomic distances and bond angles
Main operator	directional message passing
Basis functions	spherical Bessel functions and spherical harmonics
Output	molecular property, energy, force-related targets
Symmetry target	invariant scalar prediction with direction-aware intermediate messages
Main domain	molecular property prediction and quantum chemistry

Why Messages Instead of Only Atoms?

In a standard MPNN, a node aggregates messages from neighbors:

$$m_i=\sum _{j\in N(i)}M(h_i,h_j,e_{ij}).$$

DimeNet treats the message $m_{j\to i}$ as the carrier of geometry. A message has an associated direction:

$$e_{j\to i}=\frac{r_i-r_j}{\Vert r_i-r_j\Vert }.$$

When the molecule rotates by $R$, the direction rotates:

$$e_{j\to i}^{\prime }=R{e}_{j\to i}.$$

The architecture can use angles between directions, which are invariant to global rotation:

$$\cos {\theta }_{kji}=\frac{(r_k-r_j{)}^{\top }(r_i-r_j)}{\Vert r_k-r_j\Vert \Vert r_i-r_j\Vert }.$$

This lets the model encode directional geometry while keeping final scalar predictions invariant.

Directional Message Passing

Consider two directed messages:

$$m_{k\to j}\text{and}{m}_{j\to i}.$$

DimeNet transforms messages using the angle between their directions:

$${\theta }_{kji}=\angle (k\to j,j\to i).$$

A simplified update is:

$$m_{j\to i}^{(t+1)}=\sum _{k\in N(j)\setminus \{i\}}\Phi \left(m_{k\to j}^{(t)},d_{ji},d_{kj},{\theta }_{kji}\right).$$

The exclusion $k\ne i$ avoids immediate backtracking in the message update.

This resembles belief propagation:

$$\text{incoming messages to }j\to \text{outgoing message }j\to i.$$

The architecture is edge/message-centric, not only node-centric.

Distances and Angles

Distance features encode pairwise separation:

$$d_{ij}=\Vert r_i-r_j\Vert .$$

Angular features encode triplet geometry:

$${\theta }_{kji}=\arccos \left(\frac{(r_k-r_j{)}^{\top }(r_i-r_j)}{\Vert r_k-r_j\Vert \Vert r_i-r_j\Vert }\right).$$

The triplet $(k,j,i)$ is important because it describes how two interactions meet at atom $j$.

Feature	Geometry
$d_{ij}$	pairwise distance
${\theta }_{kji}$	local angle at atom $j$
directed edge $j\to i$	orientation of message flow
triplet $k\to j\to i$	angular dependency between messages

Basis Functions

DimeNet uses physically motivated basis representations instead of only Gaussian radial basis functions.

For distances, it uses radial basis functions related to spherical Bessel functions:

$$R_n(d)$$

For angles, it uses spherical harmonics:

$$Y_l^m(\theta ,\phi ).$$

In molecular graphs, the combined basis represents distance-angle structure:

$$B_{n,l}(d,\theta )\approx R_n(d)Y_l(\theta ).$$

The reading point is not to memorize every basis detail. The important architecture idea is:

$$\text{geometric feature}\to \text{orthogonal basis expansion}\to \text{message update}.$$

This gives the network a structured way to model smooth molecular geometry.

Relation to SchNet

SchNet uses distance-conditioned filters:

$$m_i=\sum _j{h}_j\odot W(d_{ij}).$$

DimeNet adds directional message embeddings and angles:

$$m_{j\to i}\leftarrow \Phi (m_{k\to j},d_{ji},d_{kj},{\theta }_{kji}).$$

Axis	SchNet	DimeNet
primary geometry	distance	distance + angle
representation center	atom states	directed messages
interaction unit	pair $i,j$	triplet $k,j,i$
basis	radial distance features	radial + angular basis
physical motivation	continuous filter over atom distances	angular molecular interactions

DimeNet should be read as a direct refinement of distance-only atomistic GNNs.

Relation to MPNN

MPNN defines:

$$m_i=\sum _{j\in N(i)}M(h_i,h_j,e_{ij}).$$

DimeNet changes the unit being updated:

$$h_i\to m_{j\to i}.$$

This is a deeper change than adding one extra edge feature. It changes the computational graph from atom updates to directed-message updates.

Relation to Equivariant GNNs

DimeNet uses directional information, but it is not the same as a full tensor-field equivariant network.

Model	Geometry Handling
SchNet	invariant distance-based filters
DimeNet	directional messages and angular basis
EGNN	equivariant coordinate updates
SE(3)-Transformer	equivariant tensor features and attention

DimeNet’s final molecular property prediction is invariant, while the message directions rotate with the molecule. Later equivariant architectures make transformation behavior more explicit in feature types.

Evidence to Read

DimeNet is evaluated on molecular property and force/energy style benchmarks.

Evidence	What It Supports
QM9	angular message passing improves equilibrium molecular property prediction
MD17	direction-aware modeling helps molecular dynamics trajectories
comparison to SchNet-like baselines	angles and directed messages matter
parameter efficiency	orthogonal basis can improve performance with fewer parameters
ablations over geometric features	distance-only vs distance-angle reasoning

The central evidence question is:

$$\text{Does adding angular information improve molecular predictions beyond distance-only message passing?}$$

Evaluation Risks

For DimeNet-style models, check:

Risk	Check
conformer availability	true 3D coordinates may not be available in real screening tasks
conformer leakage	near-identical geometries across splits can inflate results
computational cost	triplet interactions can be more expensive than pairwise messages
cutoff sensitivity	angular triplets depend on radius graph construction
target type	some targets need long-range electrostatics beyond local angles
coordinate quality	noisy conformers can degrade angle-based features

Adding angles is not automatically better if the geometry is unreliable or unavailable.

Failure Modes and Caveats

• Directional message passing requires 3D coordinates or plausible synthetic coordinates.
• Triplet construction can increase cost significantly.
• Angular features may overfit conformer-specific artifacts if splits are weak.
• Distance and angle features are invariant summaries; they do not expose the full vector/tensor representation used by later equivariant networks.
• Molecular long-range effects may require explicit global or electrostatic modeling.

Why This Matters for Architecture Reading

DimeNet teaches a key geometric modeling lesson:

$$\text{pairwise distances}\ne \text{full local molecular geometry}.$$

Angles are part of molecular structure. For protein-ligand and molecular modeling work, this paper is a useful checkpoint before reading more complex equivariant architectures.

The reading question is:

$$\text{Does the architecture use enough geometry for the target, without assuming unavailable coordinates?}$$

Links

• Graph neural networks
• Equivariance
• Invariance
• Spherical harmonics
• Conformer
• Molecular graph
• Molecular property prediction
• Geometry
• Neural Message Passing
• SchNet
• E(n) Equivariant GNN
• SE(3)-Transformer

One-Line Memory

DimeNet is the molecular GNN paper that moves from distance-only atom interactions to directed messages transformed by distances and angles.