Research map / Vision-language and multimodal / Human and embodied modeling
TRIDENT: Tri-Modal Molecular Representation Learning with Taxonomic Annotations and Local Correspondence
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TRIDENT, a tri-modal molecular representation framework aligning SMILES, text, and taxonomic annotations (HTA) with a volume-based global loss, a functional-group local loss, and momentum-weighted combination.
Problem
Molecular property prediction methods largely ignore taxonomic functional annotations and substructure-level correspondence, and pairwise alignment struggles with multi-level textual information.
Approach
Extends GRAM's volume-based tri-modal contrastive loss to SMILES, text and hierarchical taxonomic annotations, adds a local functional-group-to-subtext InfoNCE loss, and balances the two with an exponential-moving-average weight alpha.
Key results
- TRIDENT (M-M) reaches 78.46% average ROC-AUC on 8 MoleculeNet classification tasks, versus 77.01% for Atomas.
- TRIDENT (M-S) reaches 95.08% AUC on DILI, above the best baseline Atomas at 90.17%.
- On Caco-2 regression, TRIDENT (M-M) reports RMSE 0.41, tied with MolT5 at the best value.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add traditional Text input alongside SMILES and HTA (ablation) | TRIDENT (M-M) · BBBP | ROC-AUC (%): 72.02±0.36 -> 73.95±1.01 | Helped |
| add traditional Text input alongside SMILES and HTA (ablation) | TRIDENT (M-M) · Tox21 | ROC-AUC (%): 78.21±0.19 -> 79.36±0.13 | Helped |
| add traditional Text input alongside SMILES and HTA (ablation) | TRIDENT (M-M) · ToxCast | ROC-AUC (%): 67.04±0.38 -> 67.80±0.37 | Helped |
| add traditional Text input alongside SMILES and HTA (ablation) | TRIDENT (M-M) · Bace | ROC-AUC (%): 81.28±0.92 -> 82.39±0.56 | Helped |
| replace momentum-based loss weighting with direct addition (ablation) | TRIDENT · Tox21, ToxCast, BBBP, Bace | ROC-AUC (%): 79.36±0.13 -> 77.79±0.81; 67.80±0.37 -> 66.73±0.65; 73.95±1.01 -> 72.15±0.81; 82.39±0.56 -> 81.42±0.69 | Hurt |
| replace momentum-based weighting with sigmoid-weighted increasing local-loss weight (ablation) | TRIDENT · Tox21, ToxCast, BBBP, Bace | ROC-AUC (%): 79.36±0.13 -> 76.68±0.79; 67.80±0.37 -> 65.49±0.82; 73.95±1.01 -> 71.68±0.79; 82.39±0.56 -> 80.91±0.83 | Hurt |
| remove HTA (ablation) | TRIDENT · Tox21, ToxCast, BBBP, Bace | ROC-AUC (%) | Hurt |
| remove local alignment (ablation) | TRIDENT · Tox21, ToxCast, BBBP, Bace | ROC-AUC (%) | Hurt |
| replace volume loss with standard contrastive loss (ablation) | TRIDENT · Tox21, ToxCast, BBBP | ROC-AUC (%) | Hurt |
Surprising findings
- Replacing the volume loss with standard contrastive loss causes instability on several datasets, per the authors' ablation.
- HTA text alone as the text input already outperforms most baselines but falls short of the full tri-modal input.
Limitations
- Molecular properties such as toxicity also depend on targets and metabolites, which the model does not capture.
- Reported results are averaged over three random seeds, and the HTA annotations are validated only by domain experts without a detailed protocol.
Experimental setup
Datasets: PubChem molecule-text pairs MoleculeNet Therapeutics Data Commons
Benchmarks: MoleculeNet TDC DILI TDC Carcinogens TDC Skin Reaction TDC AMES TDC hERG TDC CYP P450 2C19 TDC Caco-2
Metrics: ROC-AUC accuracy RMSE
Baselines: MOLFORMER KV-PLM MegaMolBART MoleculeSTM MolFM MoMu Atomas MolCA-SMILES MolXPT BioT5 BioT5+ MolT5
Builds on: GRAM CLIP MoleculeSTM MolCA
Closest papers by meaning
- Atomas: Hierarchical Adaptive Alignment on Molecule-Text for Unified Molecule Understanding and Generation (ICLR 2025 · similarity 0.89)
- Multimodal Molecular Pretraining via Modality Blending (ICLR 2024 · similarity 0.88)
- FGBench: A Dataset and Benchmark for Molecular Property Reasoning at Functional Group-Level in Large Language Models (NeurIPS 2025 · similarity 0.87)
- MolVision: Molecular Property Prediction with Vision Language Models (NeurIPS 2025 · similarity 0.87)
- 3D-MolT5: Leveraging Discrete Structural Information for Molecule-Text Modeling (ICLR 2025 · similarity 0.87)
- From Artificially Real to Real: Leveraging Pseudo Data from Large Language Models for Low-Resource Molecule Discovery (AAAI 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.