dgopen model-ens v0.1.0 (weights)
Weights for dgopen, which predicts per-residue amide opening free energy (ฮG_open, kcal/mol) from one protein structure. These are 12 checkpoints: 4 training arms ร 3 seeds. The model-ens prediction is their equal-weight mean.
Use them through the package. The package downloads these files automatically and checks each
against the sha256 in weights_manifest.json:
pip install git+https://github.com/SoojungYang/dgopen.git
dgopen predict --pdb protein.pdb --out protein_dgopen.csv
Each file is a bare PyTorch state_dict (fp32, 3.49 M parameters) for dgopen.model.DistGraphTransformer
with ArchConfig() defaults. The inputs must be produced by dgopen.featurize.build_inputs plus
ESM-2 650M embeddings, so the files are not usable on their own.
Read the model card before interpreting outputs. It covers the training data (Ferrari et al. 2026 mHDX-MS on Tsuboyama et al. 2023 domains; MegaSim MD for pretraining), the evaluation, and the limitations: within-protein ranking only; absolute values compressed and offset; training domains 28โ64 residues.
License: MIT.