Research map / Probabilistic and scientific modeling / Molecules and proteins

Chain-of-Generation: Progressive Latent Diffusion for Text-Guided Molecular Design

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Structure-based molecule design with diffusion · 0 citations (OpenAlex)

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 Chain-of-Generation (CoG), a training-free multi-stage latent diffusion method that decomposes text prompts into curriculum-ordered segments (via an LLM) and progressively conditions denoising on cumulative sub-prompts, plus a post-alignment contrastive stage for a GraphLDM backbone, for text-to-molecule generation.

Problem

Text-to-molecule latent diffusion uses one-shot conditioning on the full prompt, which struggles with compositional prompts: missing substructures, poor attribution of components to text, and overloaded simultaneous constraints.

Approach

Decompose a prompt into cumulative sub-prompts ordered coarse-to-fine (scaffold, functional groups, fine modifiers) and run chained partial denoising, initializing each stage from the previous latent so earlier structure is retained; add a contrastive post-alignment between graph latents and SciBERT text embeddings.

Key results

• On ChEBI-20, GraphLDM + CoG improves BQI from 59.54 to 60.47 and Q-Cov from 43.19 to 45.58 over GraphLDM while keeping 100% validity.
• On PubChem, GraphLDM + CoG reaches BQI 58.07 and Q-Cov 43.13, versus 57.88 and 41.06 for GraphLDM; the gain is described as smaller than on ChEBI-20.
• Coarse-to-fine prompting outperforms one-shot across GPT-4o, Gemini 2.5 Pro and Grok v3 segmenters, with Grok v3 best (BQI 51.02 vs 49.64 one-shot).

Surprising findings

• Fine-to-coarse decomposition order performs markedly worse than one-shot (e.g., BQI 34-38 vs 49.64), and Q-Nov sometimes decreases with CoG on the benchmarks.
• Using only an intermediate sub-prompt without the cumulative early prompt degrades performance below the one-shot baseline, showing the value of early-stage regularization.

Limitations

• Gains on PubChem are small, attributed to noisy, loosely aligned text descriptions.
• Q-Nov occasionally decreases with CoG, attributed to reduced variability.
• Segmentation quality depends on the LLM used, and metrics are MACCS-based rather than established chemical metrics.

Experimental setup

Datasets: ChEBI-20 PubChem synthetic compositional prompt dataset

Benchmarks: ChEBI-20 PubChem

Metrics: BQI Q-Cov Q-Nov Validity MACCS Tanimoto similarity

Baselines: MolT5 small MolT5 large ChemT5 small ChemT5 base BioT5 base BioT5 plus 3M-Diffusion GraphLDM

Builds on: GraphLDM 3M-Diffusion Chain-of-Thought SciBERT

Closest papers by meaning

• LDMol: A Text-to-Molecule Diffusion Model with Structurally Informative Latent Space Surpasses AR Models (ICML 2025 · similarity 0.90)
• Text-Guided Molecule Generation with Diffusion Language Model (AAAI 2024 · similarity 0.89)
• Graph Diffusion that can Insert and Delete (NeurIPS 2025 · similarity 0.89)
• Geometric-Facilitated Denoising Diffusion Model for 3D Molecule Generation (AAAI 2024 · similarity 0.89)
• Equivariant Blurring Diffusion for Hierarchical Molecular Conformer Generation (NeurIPS 2024 · similarity 0.88)
• Graph Diffusion Transformers for Multi-Conditional Molecular Generation (NeurIPS 2024 · similarity 0.88)