Research map / Vision-language and multimodal / Human and embodied modeling
Atomas: Hierarchical Adaptive Alignment on Molecule-Text for Unified Molecule Understanding and Generation
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.
Atomas jointly trains a unified SMILES-text encoder with hierarchical adaptive alignment (atom, fragment, molecule levels) and a conditional decoder, improving molecule-text retrieval, property prediction, captioning and generation.
Problem
Global molecule-text contrastive alignment misses fine-grained correspondences (fragments, stereoisomers) and lacks fine-grained annotations; existing methods also are mostly designed for prediction rather than generation.
Approach
Use a unified T5 encoder for SMILES and text, and stack an adaptive polymerization module (DPC-KNN clustering plus weighted merge) with a weighted set-wise alignment module to learn atom-, fragment- and molecule-level alignments without explicit annotations, then jointly train with a conditional decoder for generation.
Key results
- Atomas-large achieves R@1 of 49.08 on text-to-molecule retrieval on PCdes versus 35.09 for MolCA.
- Atomas-large reaches Exact 0.387 and Levenshtein 12.70 on ChEBI-20 de novo generation, beating MolT5-large (0.318, 16.32).
- Atomas-base (271M) outperforms MolT5-large on captioning (BLEU-4 0.549 vs 0.508) while using about 35% of its parameters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add global alignment before conditional generation (ablation) | Atomas-base · ChEBI-20 text-based molecule generation | Exact: 0.082 -> 0.223 | Helped |
| add hierarchical alignment alongside global alignment (ablation) | Atomas-base · ChEBI-20 text-based molecule generation | Exact: 0.223 -> 0.298 | Helped |
| add global alignment alongside hierarchical alignment (ablation) | Atomas-base · ChEBI-20 text-based molecule generation | Exact: 0.266 -> 0.298 | Helped |
| replace two-stage training with joint optimization (ablation) | Atomas-base · PCdes text-to-molecule | R@1: 37.74 -> 39.08 | Helped |
| replace two-stage training with joint optimization (ablation) | Atomas-base · PCdes molecule-to-text | R@1: 36.54 -> 37.88 | Helped |
| add initial training (ablation) | Atomas-base · ChEBI-20 molecule captioning | BLEU-2: $0.6045^{\pm.003}$ -> $0.632^{\pm.005}$ | Helped |
| increase initial-training data from 0 to 15k (ablation) | Atomas-base · ChEBI-20 text-based molecule generation | Exact: 0.298 -> 0.318 | Helped |
| increase initial-training data from 15k to 51k (ablation) | Atomas-base · ChEBI-20 text-based molecule generation | Exact: 0.318 -> 0.343 | Helped |
| increase model size from 271M to 825M parameters (ablation) | Atomas · ChEBI-20 text-based molecule generation | Exact: 0.343 -> 0.387 | Helped |
| use Atomas-large instead of MolT5-large | ChEBI-20 text-based molecule generation · ChEBI-20 | Exact: 0.318 -> 0.387 | Helped |
| use Atomas-base instead of Text+Chem T5-augm | ChEBI-20 molecule captioning · ChEBI-20 | BLEU-2: 0.625 -> $0.632^{\pm.005}$ | Helped |
| use Atomas instead of MolFM | MoleculeNet scaffold split · MoleculeNet | ROC-AUC (%) average: 74.62 -> 77.01 | Helped |
Surprising findings
- Atomas-base without initial training already outperforms MolT5-large on molecule captioning, suggesting the alignment-based training matters beyond pretraining data.
Limitations
- Fine-grained alignment cannot be directly quantified due to lack of fine-grained molecule-text annotated datasets; validated only indirectly via ablations and qualitative/human evaluation.
- Only SMILES and text are used; no graph or 3D modalities.
- Human evaluation is small (five captions, three experts-style rankings).
Experimental setup
Datasets: PubChem (PubchemSTM-distill) ChEBI-20 PCdes MoleculeNet PubChem324k
Benchmarks: PCdes ChEBI-20 MoleculeNet PubChem324k
Metrics: R@1 R@5 R@10 MRR BLEU Exact match Levenshtein MACCS FTS RDK FTS Morgan FTS
Baselines: MoMu MolCA MolFM MoleculeSTM MolT5 Text+Chem T5 MolXPT MolReGPT GIT-Mol ICMA InstructMol
Builds on: MolT5 MoleculeSTM MoMu MolFM ALBEF momentum distillation density peaks clustering
Closest papers by meaning
- Learning Molecular Representation in a Cell (ICLR 2025 · similarity 0.90)
- Hierarchical Graph Tokenization for Molecule-Language Alignment (ICML 2025 · similarity 0.90)
- TRIDENT: Tri-Modal Molecular Representation Learning with Taxonomic Annotations and Local Correspondence (NeurIPS 2025 · similarity 0.89)
- Multimodal Molecular Pretraining via Modality Blending (ICLR 2024 · similarity 0.88)
- 3D-MolT5: Leveraging Discrete Structural Information for Molecule-Text Modeling (ICLR 2025 · similarity 0.88)
- Omni-Mol: Multitask Molecular Model for Any-to-any Modalities (NeurIPS 2025 · 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.