DIG: A Turnkey Library for Diving into Graph Deep Learning Research

Graph deep learning has experienced significant growth, driven by advances in neural network architectures like GCN, GAT, and GraphSAGE. Early work focused on node and graph classification, with models like Kipf & Welling (2017) and Hamilton et al. (2017). Recently, research expanded into graph generation (JT-VAE, GraphAF), interpretability (GNNExplainer), and 3D modeling (SchNet, DimeNet++). Despite progress, the lack of a unified platform hampers high-level research, especially for complex tasks requiring multi-algorithm integration and benchmarking.

Existing tools are fragmented, often limited to basic tasks like node classification. Researchers face difficulties in implementing and comparing advanced methods such as molecular generation or explainability due to inconsistent interfaces and lack of standard datasets. This fragmentation slows down innovation, makes reproducibility challenging, and raises barriers for newcomers. Moreover, high computational costs and limited support for multi-task scenarios hinder progress in complex, real-world applications.

DIG introduces a unified, modular platform supporting four key directions: graph generation, self-supervised learning, explainability, and 3D graph modeling. It standardizes data interfaces, evaluation metrics, and algorithm implementations, enabling seamless task switching. The platform supports state-of-the-art algorithms like GraphAF, GraphCL, SchNet, and DimeNet++, with datasets such as QM9, ZINC, and MD17. Its design emphasizes extensibility, community contribution, and reproducibility, fostering rapid development and benchmarking of novel methods across diverse research areas.

• �� Develop unified data interfaces for multiple datasets, including QM9, ZINC, and MD17, enabling consistent data loading and preprocessing.
• �� Integrate multiple algorithms for each task, such as JT-VAE and GraphAF for graph generation, GraphCL for self-supervised learning, GNNExplainer for interpretability, and DimeNet++ for 3D modeling.
• �� Establish standardized evaluation metrics, including accuracy, MAE, and property scores, for fair comparison.
• �� Design a modular architecture allowing easy addition of new algorithms, datasets, and tasks.
• �� Leverage PyTorch and PyG for performance optimization and community support.
• �� Provide comprehensive documentation, tutorials, and example scripts to facilitate adoption.

The platform was tested on datasets like QM9, ZINC, and MD17, benchmarking algorithms against established baselines. Performance metrics included molecule generation accuracy, node classification accuracy, and MAE. Hyperparameters were tuned via grid search, and ablation studies assessed component contributions. Results showed GraphAF surpassing 85% accuracy, DimeNet++ achieving MAE below 0.01, and GraphCL improving classification by 2-3%. These experiments validated the platform’s effectiveness and reproducibility, demonstrating its suitability for high-level research.

GraphAF on ZINC achieved 85% molecule validity, outperforming previous models by over 10%. GraphCL improved node classification accuracy from 82% to 85% on TUDataset. DimeNet++ reduced MAE on QM9 to below 0.01, a 20% improvement over traditional models. These results highlight DIG’s ability to facilitate state-of-the-art performance across tasks, confirming its value as a research ecosystem.

DIG enables rapid prototyping and benchmarking for drug discovery, material design, and physical simulations. Researchers can validate new algorithms on standard datasets, accelerating innovation. Industry applications include molecular optimization, property prediction, and complex system modeling, where reproducibility and comparability are critical. The platform’s flexibility supports multi-task workflows, making it suitable for both academic research and industrial R&D.

Support is currently limited to molecular and small-scale 3D graphs; large-scale, dynamic, or multi-modal graphs are not yet fully supported. Computational costs remain high for some algorithms, restricting scalability. Future work will focus on optimizing performance, expanding task coverage, and integrating more diverse data modalities to address these limitations.