---

name: three-d-graphs description: "Use DIG for 3D graph learning and 3D molecular geometry generation: QM93D, MD17, ECdataset, FOLDdataset, SchNet, DimeNet++, SphereNet, ComENet, ProNet, and G-SphereNet." disable-model-invocation: true metadata: disco-role: operating license: GPL 3.0

Three-D Graphs

Use this sub-skill for DIG workflows that model 3D graph geometry, molecular properties, or protein structures.

Include

• dig.threedgraph.dataset: QM93D, MD17, ECdataset, FOLDdataset.
• dig.threedgraph.method: run, SchNet, DimeNetPP, SphereNet, ComENet, ProNet.
• dig.threedgraph.evaluation: ThreeDEvaluator.
• dig.threedgraph.utils: xyz_to_dat.
• dig.ggraph3D: QM93DGEN, collate_fn, G_SphereNet, random generation, target-property optimization, and bond-length/geometry evaluation.

Exclude

• 2D molecule generation, RDKit property optimization, or SMILES workflows: use ../molecular-graph-generation/SKILL.md.
• SSL, explainability, OOD, augmentation, fairness, or large-scale graph workflows: route to sibling sub-skills.

Start Here

• Read references/workflows.md for the 3D property-prediction and 3D-generation flows.
• Read references/data-formats.md for QM93D, MD17, EC/FOLD, and QM93DGEN input layouts.
• Read references/troubleshooting.md when hdf5 layouts are wrong, PySCF is slow, or a model config expects GPU.
• Run scripts/three_d_smoke.py for a safe import-and-evaluator smoke check.

Core Workflows

• 3D property prediction: load QM93D or MD17, split with the provided helper, and train SchNet, DimeNetPP, SphereNet, ComENet, or ProNet using the run wrapper and ThreeDEvaluator.
• Protein graph classification: use ECdataset or FOLDdataset with ProNet and the 3D runner.
• 3D molecule generation: use QM93DGEN, collate_fn, and G_SphereNet to train or generate 3D molecular geometries.
• 3D generation evaluation: use dig.ggraph3D.evaluation.RandGenEvaluator and PropOptEvaluator for validity and property summaries.

Quick Validation

python scripts/three_d_smoke.py --help
python scripts/three_d_smoke.py

The smoke script only imports public APIs and evaluates a tiny MAE example plus a tiny 3D validity example.