Research map / Graph learning / Graph self-supervised learning
DMol: A Highly Efficient and Chemical Motif-Preserving Molecule Generation Platform
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 DMol, a graph diffusion model for small-molecule generation with deterministic per-step perturbation counts, motif compression into supernodes, and a count-matching loss, giving ~10x fewer diffusion steps and better validity and motif preservation than DiGress and DeFoG.
Problem
Discrete graph diffusion for drug-like molecule generation needs many diffusion steps, yields low chemical validity, and fails to preserve important motifs such as carbon rings.
Approach
Perturb a deterministic number of randomly chosen nodes per step, restrict edge perturbations to the subgraph induced by those nodes, add MSE penalties on the count of changed nodes and edges, and compress frequent motifs (mostly carbon rings) into supernodes that are diffused like atoms and decoded back to motifs.
Key results
- On MOSES, DMol reaches 87.8 validity and 4.5033 ChEMBL-likeness with 38 average diffusion steps vs 500 for DiGress.
- On QM9, DMol achieves 98.3 validity, 96.0 V.U. and 75.1 V.U.N., exceeding DeFoG's 62.8 V.U.N.
- On GuacaMol, DMol reports 86.7 validity, 100 uniqueness, 100 novelty and 94.2 KL divergence, with 48 average diffusion steps and 88 h training time.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DMol instead of DiGress | molecule generation · MOSES | V: 84.8 -> 87.8 | Helped |
| Use DMol instead of DiGress | molecule generation · GUACAMOL | N: 99.8 -> 100 | Helped |
| Use DMol instead of DiGress | molecule generation · QM9 | V.U.N.: 31.8 -> 75.1 | Helped |
| Use DMol instead of DiGress | molecule generation · MOSES | QED: 0.8024 -> 0.8055 | Helped |
| Use DMol instead of DeFoG | molecule generation · GUACAMOL | SD: 0.687 -> 0.678 | Hurt |
| Add MSE penalties to the cross-entropy loss (ablation) | DMol on GUACAMOL · GUACAMOL | V: 79.6 -> 85.4 | Helped |
| Select perturbed edges from the induced subgraph rather than the whole graph (ablation) | DMol on GUACAMOL · GUACAMOL | V: 55.6 -> 85.4 | Helped |
| Set the cosine schedule parameter to c=0.008 instead of c=0.004 (ablation) | DMol on MOSES · MOSES | Validity: 86.9 -> 87.8 | Helped |
| Use a linear rather than cosine forward schedule (ablation) | DMol on MOSES · MOSES | Validity: 87.8 -> 85.2 | Hurt |
| Increase k from 1 to 2 (ablation) | DMol on MOSES · MOSES | Validity: 85.1 -> 85.6 | Helped |
| Increase k from 2 to 10 (ablation) | DMol on MOSES · MOSES | Validity: 85.6 -> 83.4 | Hurt |
Surprising findings
- Selecting edges from the whole graph instead of the induced subgraph drops MOSES-style validity from 85.4 to 55.6 in the GuacaMol ablation.
- DeFoG scores lower than DMol on ChEMBL-likeness and QED despite using comparable generation complexity.
Limitations
- Motif compression relies on a small hand-picked set of frequent motifs (3-15) that differ per dataset.
- Mismatch in reported DeFoG settings: DeFoG run with 50 steps for fair comparison rather than its tuned configuration.
- Sensitivity to the number of steps k; larger k degrades validity.
Experimental setup
Datasets: QM9 MOSES GuacaMol SBM Planar
Benchmarks: QM9 MOSES GuacaMol SBM Planar
Metrics: validity uniqueness novelty FCD ChEMBL-likeness QED shingle distance sample time training time MMD
Baselines: GraphVAE GT-VAE Set2GraphVAE SPECTRE GraphNVP GDSS GruM DiGress DisCo DeFoG
Builds on: DiGress DeFoG DisCo
Closest papers by meaning
- FlashMol: High-Quality Molecule Generation in as Few as Four Steps (NeurIPS 2026 · similarity 0.89)
- Graph Diffusion that can Insert and Delete (NeurIPS 2025 · similarity 0.89)
- Chain-of-Generation: Progressive Latent Diffusion for Text-Guided Molecular Design (NeurIPS 2026 · similarity 0.88)
- GenMol: A Drug Discovery Generalist with Discrete Diffusion (ICML 2025 · similarity 0.88)
- DiffMS: Diffusion Generation of Molecules Conditioned on Mass Spectra (ICML 2025 · similarity 0.88)
- Graph Diffusion Transformers for Multi-Conditional Molecular Generation (NeurIPS 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.