DIG

DIG — Dive into Graphs — is an open-source library with 2,000+ stars on GitHub for graph deep learning research, covering graph generation, explainable ML, self-supervised learning and 3D graphs in one research codebase.

Overview

Graph-structured data — molecules, social networks, knowledge graphs — needs models that respect edges as well as nodes, and DIG exists to make research on those models faster to run and easier to compare. Developed by the DIVE lab, the library implements methods across several graph-learning frontiers: generating new graphs (for example candidate molecules), explaining what a graph neural network based a prediction on, self-supervised pretraining for graphs, and learning on 3D graph structures. Each area ships with the benchmarks and evaluation protocol its subfield expects, so a researcher can reproduce a published result and then test a new idea against it on equal footing. The most recent releases landed in 2024, so treat DIG as a solid, citable research platform for those topic areas rather than a fast-moving product. Documentation is hosted on Read the Docs, and the code is GPL-3.0 licensed.

Key Features

• Graph generation – implementations for generating graph structures such as molecules.
• Explainable ML for graphs – methods that explain graph neural network predictions.
• Self-supervised learning – pretraining approaches for graph data.
• 3D graph support – methods that use spatial structure, not just topology.
• Benchmarks included – standard datasets and evaluation for fair comparisons.
• Documented library – guides and API docs on Read the Docs.

💳 Pricing

DIG is free and open source under the GPL-3.0 license. Note the copyleft license if you plan to redistribute software built on it; research and internal use are unrestricted.

Who Is It For

Graph machine-learning researchers — especially in drug discovery, explainability and self-supervised learning — who want benchmarked implementations to build on instead of re-deriving every baseline from its paper.