GoMS: Graph of Molecule Substructure Network for Molecule Property Prediction

While graph neural networks have shown remarkable success in molecular property prediction, current approaches like the Equivariant Subgraph Aggregation Networks (ESAN) treat molecules as bags of independent substructures, overlooking crucial relationships between these components. We present Graph of Molecule Substructures (GoMS), a novel architecture that explicitly models the interactions and spatial arrangements between molecular substructures. Unlike ESAN’s bag-based representation, GoMS constructs a graph where nodes represent subgraphs and edges capture their structural relationships, preserving critical topological information about how substructures are connected and overlap within the molecule. Through extensive experiments on public molecular datasets, we demonstrate that GoMS outperforms ESAN and other baseline methods, with particularly improvements for large molecules containing more than 100 atoms. The performance gap widens as molecular size increases, demonstrating GoMS’s effectiveness for modeling industrial-scale molecules. Our theoretical analysis demonstrates that GoMS can distinguish molecules with identical subgraph compositions but different spatial arrangements. Our approach shows particular promise for materials science applications involving complex molecules where properties emerge from the interplay between multiple functional units. By capturing substructure relationships that are lost in bag-based approaches, GoMS represents a significant advance toward scalable and interpretable molecular property prediction for real-world applications.

💡 Research Summary

The paper introduces Graph of Molecule Substructures (GoMS), a novel architecture designed to overcome the limitations of existing subgraph‑based molecular property predictors such as Equivariant Subgraph Aggregation Networks (ESAN). While ESAN treats a molecule as an unordered bag of subgraphs, discarding any information about how those fragments are arranged or interact, GoMS explicitly constructs a higher‑level graph whose nodes are chemically meaningful substructures and whose edges encode multiple relational views—topological overlap, chemical similarity, and three‑dimensional spatial arrangement.

The workflow consists of three stages. First, a domain‑informed substructure sampling policy extracts fragments using well‑established chemoinformatics rules (RECAP, BRICS, RGB). These rules preserve aromaticity, functional groups, and other chemically relevant motifs, unlike naïve random node/edge deletions. The number of fragments k is kept modest (typically one to two orders of magnitude smaller than the atom count n) to retain computational tractability for large molecules (>100 atoms).

Second, each fragment is embedded with an equivariant graph neural network (e.g., EGNN, PaiNN). The equivariance guarantees that rotated or reflected conformations map to identical latent vectors, a crucial property for 3‑D molecular tasks.

Third, a “graph‑of‑substructures” Gₛ = (Vₛ, Eₛ) is built. An edge eᵢⱼ is created only when two fragments share atoms or are directly connected by a bond in the original molecular graph. The edge feature is a concatenation of three components:

1. Topological – a normalized overlap ratio η(|Vᵢ ∩ Vⱼ| / min(|Vᵢ|,|Vⱼ|)).
2. Chemical – the Tanimoto similarity of ECFP‑4 fingerprints, capturing functional‑group likeness even at a distance.
3. Spatial – 3‑D descriptors such as centroid distance, relative orientation, and dihedral angles.

These multi‑view edges enable the model to preserve the “Arrangement‑Preserving Isomorphism” – a bijection between substructure sets that respects both subgraph isomorphism and the relational pattern R. The authors prove (Theorem 0.3) that if two molecules share identical multisets of fragments but differ in arrangement, GoMS will generate distinct graph embeddings, whereas bag‑based methods cannot.

The higher‑level graph is processed by a Graph Transformer with multi‑head attention that incorporates edge features directly into the attention computation. After L Transformer layers, node embeddings are pooled (average) and passed through a final MLP to predict the target property (regression or classification). Because the attention operates on k nodes, the computational cost scales as O(k²), which is negligible compared to O(n²) for atom‑level GNNs on large systems.

Empirical evaluation spans several public benchmarks (QM9, PCQM4M, MoleculeNet). Across all tasks, GoMS matches or exceeds state‑of‑the‑art baselines, with pronounced gains on molecules larger than 100 atoms and especially on those exceeding 200 atoms (e.g., OLED‑relevant organic semiconductors). Ablation studies confirm that each component—chemically informed fragmentation, multi‑view edge encoding, and the Transformer backbone—contributes meaningfully to performance.

In summary, GoMS offers a scalable, interpretable framework that captures both local chemical semantics and global spatial organization of functional motifs. By moving beyond bag‑of‑substructures to a relational substructure graph, it delivers superior predictive accuracy for industrial‑scale molecules where properties emerge from the interplay of multiple functional units. This makes GoMS a promising tool for materials discovery, drug design, and any domain requiring accurate property prediction on large, complex molecular systems.