Research map / Vision-language and multimodal / Human and embodied modeling
3D-MolT5: Leveraging Discrete Structural Information for Molecule-Text Modeling
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 3D-MolT5, a T5-based molecule-text model that tokenizes 3D molecular substructures with E3FP into discrete 3D tokens, sums them with 1D SELFIES embeddings, and jointly pre-trains with denoising and translation objectives.
Problem
Existing molecule-language models either ignore 3D molecular structure or add external 3D encoders that are pre-trained separately, causing weak cross-modal interaction, costly modality alignment, and dependence on external encoders.
Approach
Map 3D substructures around each atom to discrete tokens via the E3FP fingerprint, align them atom-by-atom with 1D SELFIES tokens, and sum the embeddings so a single T5 model handles 1D, 3D, and text tokens with no external structure encoder.
Key results
- On PubChemQC computed property prediction (Specialist), 3D-MolT5 reaches HOMO/LUMO/H-L gap MAE 0.08 eV versus 0.26/0.25/0.28 eV for 3D-MoLM.
- On PubChem 3D molecule captioning, 3D-MolT5 Specialist achieves BLEU-2 42.05 and ROUGE-L 42.33 versus 30.32 and 31.23 for 3D-MoLM.
- On ChEBI-20 text-to-molecule generation, 3D-MolT5 reaches exact match 0.487, the highest among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use unified 1D/3D tokenization and multi-task pre-training | 3D-MolT5 · PubChemQC Specialist | H-L Gap MAE (eV): 0.28 -> 0.08 | Helped |
| Use unified 1D/3D tokenization and multi-task pre-training | 3D-MolT5 · PubChem descriptive property prediction Specialist | BLEU-2: 32.00 -> 51.24 | Helped |
| Use unified 1D/3D tokenization and multi-task pre-training | 3D-MolT5 · PubChem 3D molecule captioning Specialist | ROUGE-L: 38.76 -> 42.33 | Helped |
| Use unified 1D/3D tokenization and multi-task pre-training | 3D-MolT5 · ChEBI-20 molecule generation Specialist | Exact match: 0.413 -> 0.487 | Helped |
| Remove all 3D structure information (ablation) | 3D-MolT5 · PubChemQC H-L Gap prediction | MAE: 0.0791 -> 0.0968 | Hurt |
| Remove the 1D + 3D joint denoising task (ablation) | 3D-MolT5 · PubChemQC 3D-related property prediction | MAE | Hurt |
| Remove the translation tasks (ablation) | 3D-MolT5 · PubChemQC 3D-related property prediction | MAE | Hurt |
| Concatenate 3D tokens with 1D SELFIES tokens instead of summing embeddings (ablation) | 3D-MolT5 · PubChem 3D molecule-to-text translation | BLEU-2: 37.78 vs 37.65 | No clear effect |
| Use sequential concatenation rather than embedding summation (ablation) | 3D-MolT5 · PubChem 3D molecule-to-text translation | training time to converge: more than 1.5 times | Hurt |
Surprising findings
- A variant using only 3D tokens (no 1D SELFIES) achieved H-L gap MAE 0.15 on QM9, beating Uni-Mol's 0.21, suggesting discrete 3D tokens lose little information.
- Sequential concatenation of 1D and 3D tokens performed comparably to embedding summation but needed over 1.5x training time.
Limitations
- Discrete E3FP tokenization is sparser than continuous coordinates and may lose information; the authors argue the loss is marginal without a full quantification.
- Model focuses on 3D molecular understanding; 3D structure generation is not supported.
- Generalist multi-task tuning slightly underperforms Specialist models, attributed to task conflicts.
Experimental setup
Datasets: PubChem C4 PubMed Central PCQM4Mv2 3D-MoLM PubChem pairs QM9 PubChemQC ChEBI-20 Mol-Instructions MoleculeNet USPTO-50k
Benchmarks: PubChem PubChemQC QM9 PubChem 3D captioning ChEBI-20 MoleculeNet Mol-Instructions USPTO-50k
Metrics: MAE valid answer rate BLEU ROUGE METEOR exact match Levenshtein FTS FCD Text2Mol
Baselines: Llama2-7B 2D-MoLM 3D-MoLM Uni-Mol BioT5+ Mol-Instructions InstructMol MolT5 MoMu UniMoT MolX BioT5
Builds on: T5 E3FP 3D-MoLM MolT5 nanoT5
Closest papers by meaning
- Towards 3D Molecule-Text Interpretation in Language Models (ICLR 2024 · similarity 0.94)
- Omni-Mol: Multitask Molecular Model for Any-to-any Modalities (NeurIPS 2025 · similarity 0.89)
- Improving Large Molecular Language Model via Relation-aware Multimodal Collaboration (AAAI 2026 · similarity 0.88)
- Atomas: Hierarchical Adaptive Alignment on Molecule-Text for Unified Molecule Understanding and Generation (ICLR 2025 · similarity 0.88)
- KnowMol: Advancing Molecular Large Language Models with Multi-Level Chemical Knowledge (NeurIPS 2025 · similarity 0.88)
- Mol-Instructions: A Large-Scale Biomolecular Instruction Dataset for Large Language Models (ICLR 2024 · similarity 0.88)
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.