TorchDrug

Use TorchDrug as a modular PyTorch graph-learning stack:

1. load a datasets.* dataset,
2. choose a models.* representation model,
3. wrap it in a tasks.* objective,
4. train and evaluate it with core.Engine.

The current official documentation and latest published release are both 0.2.1 (released July 2023; rechecked October 1, 2026). Treat newer Python or PyTorch combinations as unverified rather than silently assuming compatibility.

Start with the version guard

Before generating or debugging code, inspect the environment:

python --version
python -c "import torch; print(torch.__version__)"
python -c "import torchdrug; print(torchdrug.__version__)"

The supported matrix for TorchDrug 0.2.1 is:

• Python 3.7 through 3.10
• PyTorch 1.8 through 2.0
• Linux, Windows, or macOS
• Apple Silicon: PyTorch 1.13 or later, CPU only; no MPS support

If the project uses Python 3.11+ or PyTorch 2.1+, create a compatible environment or explicitly test a source build. Do not present such combinations as supported.

Installation

Prefer a dedicated Python 3.10 environment and pin the TorchDrug release:

uv venv --python 3.10
source .venv/bin/activate
uv pip install "torch==2.0.0" "numpy==1.26.4" "setuptools<81" wheel

Install torch-scatter and torch-cluster wheels matched to the exact PyTorch and CUDA pair, following the official installation page. For a CPU-only PyTorch 2.0 environment on a platform listed in that wheel index, use:

uv pip install --only-binary :all: "torch-scatter==2.1.2" "torch-cluster==1.6.3" \
  --find-links "https://data.pyg.org/whl/torch-2.0.0+cpu.html"
uv pip install "torchdrug==0.2.1" "numpy==1.26.4" "scipy==1.13.1" \
  "rdkit-pypi==2022.9.5" "fair-esm==2.0.0" "decorator==5.1.1"

Do not copy a CUDA wheel URL between environments. Match the PyTorch version, CUDA build, Python ABI, and platform. On Apple Silicon, the official docs require building torch-scatter and torch-cluster from source; the wheel index above has no macOS ARM64 wheels. Install PyTorch before building with --no-build-isolation. A working compiler/SDK is also required; having PyTorch installed alone does not guarantee a successful native build. See review and environment evidence for the exact audit stack. Use the fair-esm distribution, which imports as esm; the newer distribution named esm is a different SDK. Do not install both RDKit distributions (rdkit and rdkit-pypi) into one environment. NumPy 1.x avoids the old binary stack's NumPy 2 ABI incompatibility; setuptools<81 retains pkg_resources for PyTorch 2.0.

Canonical property-prediction workflow

Use the documented ClinTox → GIN → PropertyPrediction → Engine pattern. The random split below is a tutorial baseline. For generalization to new molecular scaffolds, use data.scaffold_split or the benchmark's specified split, keep duplicate molecules in one partition, and record the actual split sizes and class counts. Scaffold-group allocation may not match the requested lengths exactly.

First run the ClinTox cache preparation. The release's old HTTP download URL fails; the current official HTTPS asset has the identical release MD5. The full training examples are illustrative and were not run to convergence during this review.

import torch
from torchdrug import core, datasets, models, tasks

dataset = datasets.ClinTox("~/molecule-datasets/")
lengths = [int(0.8 * len(dataset)), int(0.1 * len(dataset))]
lengths.append(len(dataset) - sum(lengths))
train_set, valid_set, test_set = torch.utils.data.random_split(
    dataset, lengths, generator=torch.Generator().manual_seed(1)
)

model = models.GIN(
    input_dim=dataset.node_feature_dim,
    hidden_dims=[256, 256, 256, 256],
    short_cut=True,
    batch_norm=True,
    concat_hidden=True,
)
task = tasks.PropertyPrediction(
    model,
    task=dataset.tasks,
    criterion="bce",
    metric=("auprc", "auroc"),
)

optimizer = torch.optim.Adam(task.parameters(), lr=1e-3)
solver = core.Engine(
    task,
    train_set,
    valid_set,
    test_set,
    optimizer,
    batch_size=1024,
)
solver.train(num_epoch=100)
solver.evaluate("valid")

Add gpus=[0] only when a supported CUDA device is available. Omit gpus for CPU execution.

For binary classification, task.predict(batch) returns logits; apply

…