xiaoyunw/force-field-guidance-checkpoints

Force-Field-Guided Protein-Ligand Generation — Checkpoints

Model checkpoints accompanying the paper "Improving protein-ligand complex generation with force field guidance" (Lai et al., 2025).

• Code: https://github.com/wangxiaoyunNV/NV-AZ-DrugDiscovery
• Paper: https://openreview.net/forum?id=oGywgZ5dR8

Contents

Path	Size	Description
edm/conditional_model_updates_487_epochs.ckpt	21 MB	EDM backbone, conditional (protein-pocket aware). 487 training epochs.
edm/model_updates_738999.ckpt	21 MB	EDM backbone, unconditional molecule generation. 738k training updates.
semlaflow/model_1743387578_299_1265.pt	151 MB	SemlaFlow flow-matching backbone weights.
semlaflow/chpt_1743387578_299_1265.pt	1 KB	SemlaFlow checkpoint metadata (hyperparameters / config).

The SemlaFlow optimizer state (~452 MB) used for resuming training is not included — it isn't needed to reproduce the paper results. Contact the authors if you need it.

Usage

From the project repo, run:

python download_checkpoints.py

Or, manually:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="xiaoyunw/force-field-guidance-checkpoints",
    local_dir=".",
    local_dir_use_symlinks=False,
)

This places files under edm/checkpoints/ and semlaflow/checkpoints/, matching the paths expected by the generation scripts.

License

Apache 2.0, matching the code repository.