Research map / Probabilistic and scientific modeling / Molecules and proteins
Text-Guided Molecule Generation with Diffusion Language Model
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 TGM-DLM, a continuous-embedding diffusion language model that generates SMILES from text in two phases (text-guided denoising, then unconditional correction of invalid SMILES), outperforming MolT5-Base on ChEBI-20 without extra data.
Problem
Autoregressive SMILES generators struggle to incorporate global textual constraints (scaffold/attributes spread across the whole sequence) and cannot revise earlier tokens; text-guided molecule generation needs a non-autoregressive alternative.
Approach
Apply continuous embedding-space diffusion to SMILES tokens, denoising collectively and iteratively under text cross-attention, followed by a second diffusion phase trained on deliberately corrupted (unbalanced ring/parenthesis) embeddings to repair invalid SMILES.
Key results
- TGM-DLM phase one reaches BLEU 0.828 and Exact 0.242 vs MolT5-Base BLEU 0.769 and Exact 0.081 on ChEBI-20 test.
- Phase two correction raises validity from 78.9% to 87.1% (MolT5-Base 77.2%), with Text2Mol 0.581 vs 0.496 for MolT5-Base.
- Generation takes about 1.2 seconds per molecule on one NVIDIA A100 with ~180M trainable parameters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TGM-DLM instead of MolT5-Base | Text-guided molecule generation · ChEBI-20 test | Exact: 0.081 -> 0.242 | Helped |
| Use TGM-DLM instead of MolT5-Base | Text-guided molecule generation · ChEBI-20 test | MACCS FTS: 0.721 -> 0.854 | Helped |
| Use TGM-DLM instead of MolT5-Base | Text-guided molecule generation · ChEBI-20 test | Text2Mol: 0.496 -> 0.581 | Helped |
| Add phase-two correction (ablation) | TGM-DLM · ChEBI-20 test | Validity: 0.789 -> 0.871 | Helped |
| Add phase-two correction (ablation) | TGM-DLM · ChEBI-20 test | BLEU: 0.828 -> 0.826 | Hurt |
| Use 0.5× phase-two steps (ablation) | TGM-DLM · ChEBI-20 test | Validity: 0.789 -> 0.855 | Helped |
| Use twice the phase-two steps (ablation) | TGM-DLM · ChEBI-20 test | Validity: 0.871 -> 0.875 | Helped |
| Use twice the phase-two steps (ablation) | TGM-DLM · ChEBI-20 test | BLEU: 0.826 -> 0.825 | Hurt |
| Use 3× phase-two steps (ablation) | TGM-DLM · ChEBI-20 test | Validity: 0.871 -> 0.883 | Helped |
| Use 3× phase-two steps (ablation) | TGM-DLM · ChEBI-20 test | BLEU: 0.826 -> 0.825 | Hurt |
| Train one network for both phases instead of separate networks (ablation) | TGM-DLM · ChEBI-20 test | BLEU: 0.826 -> 0.819 | Hurt |
| Add text input to the correction phase (ablation) | TGM-DLM · ChEBI-20 test | Validity: 0.871 -> 0.824 | Hurt |
Surprising findings
- Adding text conditioning to the correction phase did not improve validity and capped it near 82.4%.
- Jointly training one network for both phases underperformed separate networks.
Limitations
- Validity of phase-one outputs is only 78.9%, requiring a correction phase that distorts some molecular information.
- Only evaluated on ChEBI-20, the sole public dataset for this task.
- Baselines limited to base-size T5/MolT5; compared models have slightly different parameter counts.
Experimental setup
Datasets: ChEBI-20
Benchmarks: ChEBI-20 test split
Metrics: BLEU Exact match Levenshtein MACCS FTS RDK FTS Morgan FTS FCD Text2Mol Validity
Baselines: Transformer T5-Base MolT5-Base
Builds on: Diffusion-LM DDPM MolT5 SciBERT
Closest papers by meaning
- LDMol: A Text-to-Molecule Diffusion Model with Structurally Informative Latent Space Surpasses AR Models (ICML 2025 · similarity 0.92)
- Chain-of-Generation: Progressive Latent Diffusion for Text-Guided Molecular Design (NeurIPS 2026 · similarity 0.89)
- DMol: A Highly Efficient and Chemical Motif-Preserving Molecule Generation Platform (NeurIPS 2025 · similarity 0.88)
- Data-Efficient Molecular Generation with Hierarchical Textual Inversion (ICML 2024 · similarity 0.87)
- Diffusion on Language Model Encodings for Protein Sequence Generation (ICML 2025 · similarity 0.87)
- NExT-Mol: 3D Diffusion Meets 1D Language Modeling for 3D Molecule Generation (ICLR 2025 · 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.