Conformer Generation

RL-based molecular conformer generation built as a custom PufferLib environment.

The agent edits molecular torsions, proposes conformer sets, and scores accepted conformers with MMFF/Gibbs-style rewards.

Stack

• PufferLib for RL training, vectorized environments, sweeps, and rendering hooks
• RDKit for molecule parsing, torsion edits, conformer storage, and MMFF scoring
• nvMolKit for batched GPU MMFF minimization
• CUDA/PyTorch for GPU training

Key Files

• ocean/conformer_gen/: environment, RDKit wrapper, and bootstrap scripts
• config/conformer_gen.ini: training configuration
• pufferlib/: PufferLib runtime/training code

Run

./ocean/conformer_gen/bootstrap_train.sh

./ocean/conformer_gen/bootstrap_eval.sh

Checkpoints, logs, virtualenvs, build outputs, and conformer caches are intentionally not tracked.