Source-linked AI summary
Molecular Mechanics-Driven Graph Neural Network with Multiplex Graph for Molecular Structures
Shuo Zhang, Yang Liu, Lei Xie
TL;DR
Predicting molecular properties from structure is important for AI-assisted molecular design, but adding information to GNNs can increase computational cost. MXMNet uses a molecular-mechanics-driven multiplex graph with distinct local and global message passing; it outperforms baselines on QM9 and PDBBind, with a 13% lower QM9 mean std. MAE than DimeNet.
Problem
GNNs can gain expressive power by incorporating molecular information, but this increases computational complexity and can limit their use under restricted memory.
Source: §1 · pp. 1–2Method
MXMNet represents molecules as a local/global two-layer multiplex graph and uses corresponding message-passing modules, with angular information handled locally.
Source: §1 · p. 2, §4.1 · p. 4, §4.4 · p. 7Results
On QM9 and PDBBind, MXMNet outperforms baselines; its QM9 mean std. MAE is 13% lower than DimeNet's.
Source: §5.2 · p. 8, §5.3 · p. 9Takeaways & Limitations
Experiments on QM9 and PDBBind demonstrate MXMNet's power and efficiency relative to state-of-the-art baselines.
Source: §6 · p. 9Dihedral angles remain future work, and the effects of multiple molecular conformations remain unclear.
Source: §6 · p. 9
Abstract
from arXiv · showhide
The prediction of physicochemical properties from molecular structures is a crucial task for artificial intelligence aided molecular design. A growing number of Graph Neural Networks (GNNs) have been proposed to address this challenge. These models improve their expressive power by incorporating auxiliary information in molecules while inevitably increase their computational complexity. In this work, we aim to design a GNN which is both powerful and efficient for molecule structures. To achieve such goal, we propose a molecular mechanics-driven approach by first representing each molecule as a two-layer multiplex graph, where one layer contains only local connections that mainly capture the covalent interactions and another layer contains global connections that can simulate non-covalent interactions. Then for each layer, a corresponding message passing module is proposed to balance the trade-off of expression power and computational complexity. Based on these two modules, we build Multiplex Molecular Graph Neural Network (MXMNet). When validated by the QM9 dataset for small molecules and PDBBind dataset for large protein-ligand complexes, MXMNet achieves superior results to the existing state-of-the-art models under restricted resources.
1 Introduction
Angle-aware GNNs can be computationally expensive, limiting their use under restricted memory. MXMNet separates local and global interactions and reports strong performance and efficiency on QM9 and PDBBind.
1 Introduction
- Prior GNNs add chemical, distance, and angular information to improve representation, but angle-aware message passing can require O(Nk^2) or O(N^3) messages and strain memory on macromolecules.The paper notes that such models may have limited expressive power or fail on proteins or RNAs under restricted memory.Source: §1 · pp. 1–2
- On QM9 and PDBBind, MXMNet outperforms baseline models, uses significantly less memory than the previous state-of-the-art model, and achieves 260% training speedup.The QM9 comparison in Figure 1 includes SchNet, PhysNet, and DimeNet.Source: Figure 1 · p. 2, §1 · p. 2
- MXMNet uses a two-layer multiplex graph: local connections mainly capture covalent interactions, while global connections cover non-covalent interactions.Its MXM module uses angle-aware message passing locally and efficient message passing globally, avoiding angular computation for nonlocal interactions.Source: §1 · p. 2
2 Related Work
Molecular GNNs progressed from chemical-bond graphs to distance- and angle-aware models, with increased expressiveness accompanied by scalability costs. MXMNet explicitly organizes geometric information into multiplex layers and uses different message-passing schemes for them.
2 Related Work
- Early molecular GNNs used chemical bonds as edges and atoms as nodes, while later models incorporated 3D positions and interatomic distances.Distance-based models commonly use a cutoff to limit complexity and overfitting, but cutoff choices can leave some molecules indistinguishable.Source: §2 · pp. 2–3
- Angular information can improve expressive power, but angle-aware GNNs have higher time and space complexity and do not scale to macromolecules or large-batch learning.The cited related work describes cutoff-based models as sometimes failing to distinguish certain molecules.Source: §2 · p. 3
- Unlike prior work that implicitly represents molecular graphs as multiplex graphs, MXMNet explicitly builds layers from molecular geometry and assigns them different message-passing schemes.A multiplex graph has multiple edge types among the same nodes, organized as separate layers.Source: §2 · p. 3
3 Preliminaries
The preliminaries define molecular-property prediction, graph and message-passing notation, and multiplex graphs. They also establish the angle-count complexity that motivates efficient treatment of molecular interactions.
3 Preliminaries
- A multiplex graph consists of a shared node set and multiple edge sets, each forming a graph layer.The paper denotes the layers as plexes and represents the multiplex graph as a set of graphs.Source: §3 · p. 3
- Message passing iteratively updates node embeddings using node features and edge features.The notation defines h_i as a node embedding, e_ji as an edge embedding, and the message and update functions as learnable.Source: §3 · p. 3
- A graph with N nodes and average degree k has O(Nk^2) angles, requiring at least O(Nk^2) messages per graph operation in recent angle-aware approaches.Each node's average k incident edges define k(k − 1)/2 angles.Source: §3 · p. 3
- Molecular-property prediction maps atomic numbers and 3D coordinates to a real-valued target, optionally using auxiliary chemical information.The stated regression function is f({Z, r}) → R, or f({Z, r, Θ}) → R when auxiliary information Θ is included.Source: §3 · p. 4
4 Approach
MXMNet represents molecules with local and global graph layers inspired by molecular mechanics, then uses distinct message-passing operations to model their interactions. The design aims to balance expressive power with computational efficiency.
4.1 Multiplex Molecular Graphs
- The multiplex graph combines a local layer for local geometric information with a global layer, using chemical bonds or small-cutoff neighbors locally and larger-cutoff neighbors globally.The local geometric information includes angles and local pairwise distances; the global information includes local and nonlocal pairwise distances.Source: §4.1 · p. 4
4.2 Multiplex Molecular (MXM) Module
- Global message passing captures pairwise distances and up to two-hop neighbors through two one-hop operations, requiring O(2Nk) rather than O(Nk^2) messages per iteration.A residual update function is used between the two operations.Source: §4.2 · p. 5
- Local message passing uses two edge-level steps to capture two-hop and one-hop angles with their pairwise distances, then aggregates edge messages to update node embeddings.The module also uses cross-layer mapping to transfer node embeddings between the local and global layers.Source: §4.2 · p. 6
4.3 Multiplex Molecular Graph Neural Network (MXMNet)
- MXMNet embeds atomic numbers, derives distance and angle features from coordinates, stacks MXM modules, and sums node-level outputs across nodes and layers.The distance and angle representations use basis functions proposed in.Source: §4.3 · pp. 6–7
4.4 Expressive Power and Complexity of MXMNet
- MXMNet uses global pairwise distances and local angular information, making it more expressive than distance-only GNNs, but theoretically less expressive than GNNs capturing both globally due to uncaptured nonlocal angles.Source: §4.4 · p. 7
5 Experiments
The experiments evaluate MXMNet on QM9 molecular-property prediction and PDBBind protein–ligand binding-affinity prediction against established baselines. Results report strong QM9 performance and memory-efficient PDBBind training, alongside ablations and efficiency comparisons.
5.1 Experimental Setup
- QM9 covers around 130,000 molecules with up to 9 heavy atoms; PDBBind uses roughly 4,000 refined complexes averaging around 200 atoms, with different graph cutoffs and evaluation metrics.QM9 uses MAE; PDBBind reports binding-free-energy MAE and logKi Pearson correlation. Baselines include SchNet, PhysNet, MEGNet-full, Cormorant, MGCN, and DimeNet.Source: §5.1 · p. 7
5.2 Results on QM9
- MXMNet variants outperform baselines on 9 QM9 targets, and the BS=128, dg=10Å configuration lowers mean std. MAE by 13% versus DimeNet.The reported results identify this configuration as having the lowest std. MAE among the models.Source: §5.2 · p. 8, Table 1 · p. 8
- Ablations removing either message-passing module or parts of a module reduce performance; local-only messaging is inadequate, and both one-hop and two-hop angles matter.With global-only messaging, two message passings outperform one.Source: §5.2 · pp. 8–9
- MXMNet variants use much less QM9 training memory than DimeNet, and BS=128 achieves 2.6× training speedup against DimeNet at BS=32.SchNet and PhysNet use less memory than MXMNet but have higher mean std. MAEs.Source: §5.2 · p. 9
5.3 Results on PDBBind
- On PDBBind, MXMNet trains on the GPU while DimeNet runs out of memory, and MXMNet has higher Pearson R and lower MAE than SchNet and PhysNet.Table 3 reports Pearson correlation and MAE for the models; “-” denotes an out-of-memory issue.Source: §5.3 · p. 9, Table 3 · p. 9
6 Conclusion
The conclusion presents MXMNet as a molecular-property GNN built on a molecular-mechanics-rooted multiplex graph. Experiments on QM9 and PDBBind demonstrate its reported power and efficiency, while conformational effects remain unclear.
6 Conclusion
- MXMNet represents molecules as a molecular-mechanics-rooted multiplex graph, and QM9 and PDBBind experiments demonstrate its power and efficiency against state-of-the-art baselines.The authors identify dihedral angles and the effect of multiple conformations as directions for future work.Source: §6 · p. 9
7 Appendix
The appendix records dataset processing and target choices, baseline implementation sources, and training and model settings for the experiments.
7.1 Dataset Sources
- QM9 processing removes about 3,000 molecules that fail geometric consistency checks or are difficult to converge; PDBBind uses logKi as the predicted target.For QM9, Δϵ is predicted from ϵLUMO − ϵHOMO, and U0, U, H, and G use atomization energies adjusted by atomic reference energies.Source: §7.1 · p. 12
7.2 Baseline Sources
- For PDBBind comparisons, the authors use PyTorch Geometric implementations of SchNet and DimeNet and adapt a PhysNet implementation for framework consistency.MXMNet is also built on the referenced implementation framework.Source: §7.2 · p. 12
7.3 Implementation Details
- The appendix specifies shared model settings and training procedures, including two-layer MLPs, Swish activations, separate single-target QM9 models, and three weight-sharing PDBBind replica networks.QM9 runs are repeated three times; PDBBind experiments use 10-fold cross-validation and are repeated five times. Table 4 lists experiment hyperparameters.Source: §7.3 · p. 12, Table 4 · p. 13