Research map / Graph learning / Graph self-supervised learning
GraphAF: a Flow-based Autoregressive Model for Molecular Graph Generation
Chence Shi*, Minkai Xu*, Zhaocheng Zhu, Weinan Zhang, Ming Zhang, Jian Tang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply valency checking during sequential generation (ablation) | GraphAF · ZINC250k | validity: 68% -> 100% | Improved |
| Use GraphAF rather than GCPN for unconstrained validity | Molecular graph generation on ZINC250k · ZINC250k | validity without valency check: 20% -> 68% | Improved |
| Use GraphAF rather than GraphNVP for molecular generation | Molecular graph generation · ZINC250k | validity: 42.60% -> 100% | Improved |
| Fine-tune the pretrained GraphAF policy with reinforcement learning for penalized logP | GraphAF · Penalized logP optimization | top-1 penalized logP: 8.63 -> 12.23 | Improved |
| Fine-tune GraphAF with reinforcement learning for QED | GraphAF · QED optimization | top-1 QED: 0.948 -> 0.948 | No clear effect |
| Use GraphAF rather than GCPN for constrained penalized logP optimization | Constrained molecular optimization at δ=0.0 · ZINC250k low-score molecules | improvement: $4.20\pm 1.28$ -> $13.13\pm 6.89$ | Improved |
| Use GraphAF rather than GCPN for constrained penalized logP optimization | Constrained molecular optimization at δ=0.4 · ZINC250k low-score molecules | improvement: $2.49\pm 1.30$ -> $8.21\pm 6.51$ | Improved |
| Train GraphAF instead of GCPN | Molecular graph generation · ZINC250k training | training time: 8 hours -> 4 hours | Improved |
| Use GraphAF rather than GraphRNN on COMMUNITY-SMALL | Generic graph generation · COMMUNITY-SMALL | Orbit MMD: 0.04 -> 0.02 | Improved |
Later papers that compared with this method
3 later accepted papers report GraphAF as a baseline in their results tables; 3 of them report a better number on most of the results they share with it.
- GGBall: Graph Generative Model on Poincaré Ball (ICLR 2026 · better on 3 of 3 shared results)
- Modular Flows: Differential Molecular Generation (NeurIPS 2022 · better on 2 of 2 shared results)
- GraphDF: A Discrete Flow Model for Molecular Graph Generation (ICML 2021 · better on 3 of 3 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DMol: A Highly Efficient and Chemical Motif-Preserving Molecule Generation Platform (NeurIPS 2025 · similarity 0.89)
- Pretraining Generative Flow Networks with Inexpensive Rewards for Molecular Graph Generation (ICML 2025 · similarity 0.88)
- NEAT: Neighborhood-Guided, Efficient, Autoregressive Set Transformer for 3D Molecular Generation (NeurIPS 2026 · similarity 0.88)
- Graph Diffusion that can Insert and Delete (NeurIPS 2025 · similarity 0.87)
- Aligning Transformers with Continuous Feedback via Energy Rank Alignment (NeurIPS 2025 · similarity 0.87)
- Generative Flows on Synthetic Pathway for Drug Design (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.
