SAIR: 1.04 Million Protein-Ligand Pairs with 5.2 Million 3D Structures

SAIR provides 1,048,857 unique protein-ligand pairs and 5.2 million 3D structures curated from ChEMBL for drug discovery research. Created by SandboxAQ in collaboration with Nvidia and updated in August 2025, it pairs binding potency measurements with structural data.

Use Cases

• Training models to predict IC50 binding potency from 3D structural coordinates.
• Evaluating protein-ligand docking accuracy using the 5.2 million cofolded structures.
• Developing graph neural networks for drug discovery based on protein-ligand interaction features.

Strengths

• 1,048,857 unique protein-ligand pairs
• 5.2 million total 3D structures
• Curated from the ChEMBL database
• Openly available under CC BY 4.0

Limitations

• Structural data is cofolded rather than experimentally determined, introducing potential modeling noise.
• Coverage is restricted to the chemical space and targets available within the ChEMBL source.

Provenance

Source

SandboxAQ, Nvidia, and ChEMBL.

Collection Method

Curated from ChEMBL and augmented with computational cofolding.

Freshness

Last updated August 2025.