Distributional Priors Guided Diffusion for Generating 3D Molecules in Low Data Regimes

AAAI 2026 (Oral)
Official Appendix and Code Release for Distributional Priors Guided Diffusion for Generating 3D Molecules in Low Data Regimes

---

🌟 Overview

This repository provides the appendix and official implementation of our diffusion-based framework for generating chemically valid 3D molecular conformations in extremely low-data regimes. The method integrates distributional priors with guided diffusion to improve robustness, sample efficiency, and generalization.

---

📄 Appendix

The full appendix is available in this repository:
https://github.com/HaokaiHong/GODD/blob/main/Appendix.pdf

---

🛠️ Environment Setup

Install required packages via:

pip install -r requirements.txt

A simplified version of the dependency list can also be found here.

Note: If you prefer an RDKit-based environment, the easiest setup is:

conda create -c conda-forge -n my-rdkit-env rdkit

Then install the remaining packages inside this environment. The code should still run without RDKit, although some functionalities may be limited.

---

🔧 Dataset Split (QM9)

To generate the QM9 splits:

sh sh/split_qm.sh

---

🚀 Training

Scaffold Domain

sh sh/train_gadm_scaffold.sh

Ring Domain

sh sh/train_gadm_ring.sh

Trained models will be saved to:

./Models

---

📊 Evaluation

Scaffold Evaluation

Example (Class III domain):

python eval_analyze.py \
    --model_path ./Models/qm9_scaffold_outputs/da_qm9_scaffold_masked \
    --n_samples 10_000 \
    --save_to_xyz True \
    --target_domain ClassIII \
    --dataset qm9_scaffold_ClassIII

Ring Evaluation

Example (generate 8-ring structures):

python eval_analyze.py \
    --model_path ./Models/qm9_ring_outputs/da_qm9_ring_masked \
    --n_samples 10_000 \
    --save_to_xyz True \
    --target_domain 8 \
    --dataset qm9_ring_n_8

---

📫 Contact

For questions, issues, or feature requests:

• Email: haokai.hong@connect.polyu.hk
• Issues: Please open a GitHub issue.

---

🔒 License

This project is released under the MIT License.