TRIDENT: Tri-Modal Molecular Representation Learning with Taxonomic Annotations and Local Correspondence

TL;DR

TRIDENT is introduced, a novel framework that integrates molecular SMILES, textual descriptions, and taxonomic functional annotations to learn rich molecular representations, and achieves state-of-the-art performance on 11 downstream tasks, demonstrating the value of combining SMILES, textual, and taxonomic functional annotations for molecular property prediction.

Abstract

Molecular property prediction aims to learn representations that map chemical structures to functional properties. While multimodal learning has emerged as a powerful paradigm to learn molecular representations, prior works have largely overlooked textual and taxonomic information of molecules for representation learning. We introduce TRIDENT, a novel framework that integrates molecular SMILES, textual descriptions, and taxonomic functional annotations to learn rich molecular representations. To achieve this, we curate a comprehensive dataset of molecule-text pairs with structured, multi-level functional annotations. Instead of relying on conventional contrastive loss, TRIDENT employs a volume-based alignment objective to jointly align tri-modal features at the global level, enabling soft, geometry-aware alignment across modalities. Additionally, TRIDENT introduces a novel local alignment objective that captures detailed relationships between molecular substructures and their corresponding sub-textual descriptions. A momentum-based mechanism dynamically balances global and local alignment, enabling the model to learn both broad functional semantics and fine-grained structure-function mappings. TRIDENT achieves state-of-the-art performance on 11 downstream tasks, demonstrating the value of combining SMILES, textual, and taxonomic functional annotations for molecular property prediction.

Figures (4)

• Figure 1: Overview of TRIDENT. TRIDENT jointly models molecular SMILES, natural language descriptions, and Hierarchical Taxonomic Annotations (HTAs) to learn rich molecular representations. The framework employs a volume-based contrastive loss for soft global tri-modal alignment and a local alignment module that links molecular substructures to sub-text spans. A momentum-based mechanism dynamically balances the contribution of global and local objectives during training. This multimodal, multi-level alignment enables precise and semantically grounded molecular understanding.
• Figure 2: Traditional molecular functional descriptions are typically obtained by inputting a molecule into PubChem, where a general functional annotation is provided, as shown in Steps 1 and 2 of the figure. To achieve more comprehensive knowledge, functional annotations of the molecule are first obtained under different classification systems, as illustrated in Step 3. Then, these annotations are summarized using GPT-4o, resulting in a higher-quality textual description, as depicted in Step 4. The blue and green highlighted sections illustrate the different perspectives between traditional text and HTA text descriptions. For detailed processing steps, please refer to the Appendix \ref{['Data_Collection_and_Processing']}.
• Figure 3: The ablation experiments are conducted on the Tox21, ToxCast, BBBP and Bace datasets. “w/o HTA” denotes that only not use hierarchical taxonomic annotation; “w/o local alignment” denotes that the local alignment is removed; and “w/o volume loss” indicates that only the volume‐based loss is changed to the standard contrastive loss.
• Figure 4: The workflow for summarizing Hierarchical Taxonomic Annotations (HTA). Using GPT-4o, detailed classification annotations are processed and summarized, resulting in high-quality HTA text descriptions for molecular data.