MolGramTreeNet: A Multimodal Molecular Property Prediction Model Incorporating Syntax Tree Constraints

MolGramTreeNet is an innovative deep learning framework that explicitly encodes chemical rules and hierarchical semantics by integrating one-dimensional syntax tree structures and two-dimensional molecular graphs, enabling high-precision molecular property prediction. This method has been published in the iScience journal.

• Original Author/Maintainer: NTU-MedAILab
• Source Platform: GitHub
• Original Title: MolGramTreeNet
• Original Link: https://github.com/NTU-MedAILab/MolGramTreeNet
• Source Publication/Update Date: 2026-05-23

Molecular property prediction is a core problem in computational chemistry and drug discovery. Traditional machine learning methods face a fundamental challenge when processing molecular data: how to capture both the structural information and chemical semantics of molecules simultaneously.

Molecules can be represented in multiple ways:

• SMILES strings: One-dimensional text representation, easy to process but loses spatial structure information
• Molecular graphs: Two-dimensional graph structure, which can represent the connection relationships between atoms but is difficult to express chemical rules and hierarchical semantics
• 3D conformations: Contains spatial information, but has high computational cost and high requirements for data quality

Existing deep learning models usually focus on only one of these representations, leading to the inability to fully utilize the multimodal characteristics of molecules. For example, pure graph neural networks may ignore the types of chemical bonds and reaction rules, while pure sequence models cannot understand the topological structure of molecules.

MolGramTreeNet proposes a novel multimodal fusion method that combines one-dimensional syntax tree structures (generated via context-free grammar) with two-dimensional molecular graphs to explicitly encode chemical rules and hierarchical semantics.

Traditional molecular representation methods treat molecules as flat structures (such as SMILES strings or atom graphs), while MolGramTreeNet introduces the concept of syntax trees. Syntax trees can capture the hierarchical structure of molecules:

• Atomic layer: The most basic chemical unit
• Functional group layer: Combinations of atoms with specific chemical properties
• Substructure layer: Larger molecular fragments
• Molecular layer: The complete molecular structure

This hierarchical representation aligns with chemists' intuition. When analyzing molecules, chemists often first identify functional groups, then understand the interactions between them, and finally form a cognition of the entire molecule.

MolGramTreeNet uses Context-Free Grammar (CFG) to define the syntax rules of molecules. CFG consists of the following elements:

• Terminal symbols: Atomic types (e.g., C, N, O)
• Non-terminal symbols: Chemical structure categories (e.g., rings, chains, functional groups)
• Production rules: Describe how to build complex structures from simple ones

Through CFG, the model can learn chemically valid structure combinations and avoid generating unreasonable molecular structures. This constraint not only improves prediction accuracy but also enhances the model's interpretability.

The architecture of MolGramTreeNet includes two main branches:

1D Syntax Tree Encoder

The syntax tree encoder uses a tree-structured neural network (Tree-LSTM or similar variants) to propagate information along the hierarchical structure of the syntax tree. Each node aggregates information from its child nodes and learns the chemical semantic representation of the substructure. This bottom-up propagation ensures that the model can capture structural features at different levels of the molecule.

2D Molecular Graph Encoder

The molecular graph encoder uses graph neural networks (GNNs), such as GAT (Graph Attention Network) or MPNN (Message Passing Neural Network), to perform message passing on the atomic graph. This encoder can capture local interactions and long-range dependencies between atoms.

Fusion Layer

The outputs of the two encoders are integrated in the fusion layer. Fusion strategies may include:

• Concatenation: Concatenate the two representation vectors and feed them into a fully connected layer
• Attention mechanism: Learn weights for the two representations and perform weighted summation
• Cross-attention: Allow the two representations to attend to each other and capture their correlations

The fused representation contains both the hierarchical semantics of the syntax tree and the topological information of the molecular graph, enabling more accurate prediction of molecular properties.

MolGramTreeNet has been validated on multiple standard benchmark datasets: