Research map / Probabilistic and scientific modeling / Molecules and proteins
LDMol: A Text-to-Molecule Diffusion Model with Structurally Informative Latent Space Surpasses AR Models
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 LDMol, a latent diffusion model for text-to-molecule generation that trains a DiT on a contrastively pre-trained SMILES-enumeration latent space, outperforming autoregressive baselines on ChEBI-20 and PCDes.
Problem
Text-conditioned molecule generation with diffusion models struggles because discrete molecular representations (SMILES, graphs) are hard to connect to complex natural-language conditions, and prior diffusion models underperform autoregressive models.
Approach
Train a SMILES encoder with contrastive learning over SMILES enumerations (positive pairs = same molecule, different traversal order) so the latent captures whole-molecule structure, then compress it linearly and train a text-conditioned DiT latent diffusion model on that space.
Key results
- On ChEBI-20 text-to-molecule generation, LDMol reaches BLEU 0.926, Levenshtein 6.750, exact match 0.530 and FCD 0.20, outperforming MolT5-large (match 0.311, FCD 1.20) and bioT5+ (match 0.522, FCD 0.35), though validity (0.941) trails bioT5+ (1.000).
- On molecule-to-text retrieval (PCdes, 64-way, paragraph), LDMol with n=25 reaches 90.3% vs MolCA 86.4%; on the MoMu set (sentence) it reaches 78.4%.
- Ablation: removing contrastive pre-training yields validity 0.019 and match 0.000, while a beta-VAE latent gives validity 0.847 and match 0.492.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LDMol instead of the strongest reported autoregressive baseline | Text-to-molecule generation · ChEBI-20 test | BLEU: 0.872 -> 0.926 | Helped |
| Use LDMol instead of the strongest reported autoregressive baseline | Text-to-molecule generation · PCDes test | BLEU: 0.754 -> 0.857 | Helped |
| Increase retrieval noise-estimation iterations from n=10 to n=25 | LDMol molecule-to-text retrieval · PCDes test sentence retrieval | 64-way accuracy (%): 60.7 -> 62.2 | Helped |
| Increase retrieval noise-estimation iterations from n=10 to n=25 | LDMol molecule-to-text retrieval · MoMu test sentence retrieval | 64-way accuracy (%): 66.4 -> 78.4 | Helped |
| Add contrastive encoder pre-training (ablation) | LDMol · ChEBI-20 generation | Validity: 0.019 -> 0.941 | Helped |
| Add contrastive encoder pre-training (ablation) | LDMol · ChEBI-20 generation | Match: 0.000 -> 0.530 | Helped |
| Replace β-VAE latent encoding with the contrastively trained structural encoder (ablation) | LDMol · ChEBI-20 generation | Validity: 0.847 -> 0.941 | Helped |
| Replace β-VAE latent encoding with the contrastively trained structural encoder (ablation) | LDMol · ChEBI-20 generation | Match: 0.492 -> 0.530 | Helped |
| Add a linear compression layer (ablation) | LDMol · ChEBI-20 generation | Validity: 0.022 -> 0.941 | Helped |
| Replace transformer compression with linear compression (ablation) | LDMol · ChEBI-20 generation | Match: 0.084 -> 0.530 | Helped |
| Include stereoisomer hard-negative samples during contrastive encoder training (ablation) | LDMol · ChEBI-20 generation | Match: 0.278 -> 0.530 | Helped |
Surprising findings
- Without contrastive SMILES-enumeration pre-training the diffusion model essentially fails to generate valid SMILES (validity 0.019), showing latent choice dominates performance.
- A beta-VAE latent already beats several autoregressive models on match ratio, suggesting continuous latent diffusion is viable for discrete molecules.
Limitations
- Struggles to follow complex biological property conditions (acknowledged in conclusion).
- Validity on ChEBI-20 (0.941) is below bioT5+ (1.000), and the authors note lower validity in some baselines is traded against agreement.
- Inference time is comparable to AR baselines, and the authors note it could be improved.
Experimental setup
Datasets: PubChemSTM ChEBI-20 PCdes ZINC15 PubChem
Benchmarks: ChEBI-20 PCDes MoMu retrieval test set MoleculeSTM editing scenarios
Metrics: validity BLEU Levenshtein MACCS FTS RDK FTS Morgan FTS exact match FCD 64-way retrieval accuracy hit ratio
Baselines: MolT5-large bioT5 bioT5+ GIT-Mol MolXPT TGM-DLM T5-large Transformer MoleculeSTM MolCA MoMu-S SciBERT
Builds on: latent diffusion models DiT InfoNCE Delta Denoising Score MolT5
Closest papers by meaning
- Text-Guided Molecule Generation with Diffusion Language Model (AAAI 2024 · similarity 0.92)
- Chain-of-Generation: Progressive Latent Diffusion for Text-Guided Molecular Design (NeurIPS 2026 · similarity 0.90)
- GenMol: A Drug Discovery Generalist with Discrete Diffusion (ICML 2025 · similarity 0.87)
- KnowMol: Advancing Molecular Large Language Models with Multi-Level Chemical Knowledge (NeurIPS 2025 · similarity 0.87)
- How to Train Your Latent Diffusion Language Model Jointly With the Latent Space (NeurIPS 2026 · similarity 0.87)
- Derivative-Free Guidance in Continuous and Discrete Diffusion Models with Soft Value-based Decoding (NeurIPS 2025 · similarity 0.86)
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.