graph-data

TorchDrug Graph Data

Use this sub-skill when the task is about TorchDrug data objects rather than model layers, training loops, or domain-specific task recipes.

Read First

• For constructors, attribute contexts, batching, masking, split helpers, and dataset loading patterns, read references/data-objects.md.
• For common failures around RDKit input, attribute lengths, device movement, dataset downloads, split lengths, and lazy loading, read references/troubleshooting.md.
• For a no-network sanity check that creates tiny in-memory graphs and molecules, run python scripts/smoke_graph_data.py from this sub-skill directory. Add --skip-rdkit when RDKit or molecule parsing is unavailable.

Use This For

• Creating data.Graph, data.Molecule, data.Protein, and their packed variants from in-memory edge lists, SMILES strings, or protein sequences.
• Adding node, edge, graph, atom, bond, molecule, or residue attributes with the right context managers.
• Batching samples with Graph.pack, PackedMolecule.from_smiles, PackedProtein.from_sequence, data.graph_collate, or data.DataLoader.
• Applying subgraph, node_mask, edge_mask, graph_mask, subbatch, compact, unpack, and dataset split helpers.
• Loading TorchDrug datasets or building small local MoleculeDataset / ProteinDataset objects with transform and lazy-loading options.

Route Elsewhere

• Use layers-and-extensions for graph neural network layers, graph construction modules, and extension compilation.
• Use training-engine for core.Engine, optimizers, training/evaluation loops, checkpoint save/load, and Configurable workflows.
• Use molecular-workflows for molecule property prediction, generation, reaction, or chemistry task recipes built on these data objects.
• Use protein-workflows for protein prediction, sequence/structure recipes, and protein-specific task pipelines.
• Use knowledge-graphs for knowledge graph datasets, entity/relation indexing, and KG tasks.