What it does

Chainer Chemistry is a Python toolkit for training graph neural networks on molecular data. It wraps implementations of NFP, GGNN, SchNet, GAT, GIN, MPNN, and several other GNN variants, plus loaders for standard chemistry benchmarks like QM9, Tox21, and MoleculeNet. The pitch: feed it a molecule, get a property prediction.

The interesting bit

The library sits at an awkward historical intersection. It was built on Chainer—once PyTorch’s Japanese rival, now itself in maintenance mode—and the README explicitly warns that “further development will be limited to only serious bug-fixes and maintenance.” The footnotes are a miniature archaeology of a framework’s decline: Python 2 support dropped, ChainerX migration breaking compatibility, pinned RDKit versions from 2017.

Key highlights

• Bundles 13+ GNN architectures including newer additions like MEGNet and CGCNN for materials science
• Supports the Graph Warp Module (GWM), an auxiliary module for boosting GNN power
• Includes Weisfeiler-Lehman Embedding preprocessing for molecular graphs
• Has been used in published research projects including molecular graph generation with normalizing flows
• Requires manual RDKit installation; dependency matrix shows tight version coupling with Chainer releases

Caveats

• Active development has ceased; the project is in maintenance-only mode
• Chainer itself is no longer actively developed, making this a dependency on a deprecated framework
• The README states it “cannot guarantee the reproducibility of any results published in papers”

Verdict

Worth studying if you’re surveying historical GNN implementations or maintaining legacy Chainer-based research code. Everyone else should look to PyTorch Geometric, DGL, or Hugging Face’s chemistry tools for active development and modern performance.