SandboxAQ SAIR Dataset Release

SandboxAQ has released the Structurally Augmented IC50 Repository (SAIR), the largest dataset of co-folded 3D protein-ligand structures paired with experimentally measured IC₅₀ labels. This release bridges a critical data gap in pharmaceutical R&D by directly linking molecular 3D structures to drug potency, enabling researchers to move more drug design and optimization from wet labs to in silico environments.

Scaling Protein-Ligand Structural Data

SAIR provides open access to over 5 million AI-generated, high-accuracy protein-ligand 3D structures. The dataset was created by compiling over 1 million unique computationally co-folded protein–ligand pairs, resulting in 5.24 million distinct 3D complexes (five co-folded structures per pair). Each structure is paired with a curated IC₅₀ measurement sourced from ChEMBL or BindingDB.

This scale of data addresses the limitations of traditional structure-based discovery, which relies on time-intensive experimental methods like X-ray crystallography and cryo-EM. Unlike earlier algorithms such as AlphaFold or Vina, which provide static snapshots, SAIR's co-folded approach provides a more scalable link between high-quality 3D structures and drug potency.

High-Performance Computing and Generation

Generating the SAIR dataset required significant computational resources and infrastructure optimization:

  • Compute Power: The dataset was generated using the Boltz1 cofolding AI model on a cluster of 760 NVIDIA H100 processors via NVIDIA DGX Cloud on Google Cloud Platform.
  • Resource Utilization: Through collaboration between NVIDIA AI Accelerator and SandboxAQ, the teams achieved over 95% GPU compute utilization.
  • Efficiency: The total compute time exceeded 130,000 GPU hours. The optimized workflow reduced the generation timeline from an estimated three months to three weeks, representing a 4X speed-up.

Validation and Data Quality

To ensure the physical plausibility and chemical sanity of the AI-generated complexes, SandboxAQ utilized PoseBusters, an industry-standard open-source benchmarking tool. Validation results show that 97% of the structures in SAIR passed all PoseBusters checks.

Additionally, SandboxAQ benchmarked leading affinity prediction methods—including graph neural networks, 3D CNNs, and empirical scoring functions—across the synthetic structures and experimental IC₅₀ values. Detailed results are available in a scientific manuscript hosted on bioRxiv.

Addressing the "Dark Proteome"

SAIR provides structural hypotheses for proteins that lack experimental data, specifically targeting the "dark proteome" (disease-relevant proteins without existing experimental structures).

  • Expanding Target Horizons: More than 40% of the proteins in the SAIR dataset have no available structures in the Protein Data Bank (PDB), regardless of whether they are bound to a ligand.
  • Polypharmacology Insights: The dataset's cross-target breadth allows researchers to identify how a single molecule interacts with multiple proteins, facilitating the prediction of off-target effects and the identification of drug repurposing opportunities.

Access and Implementation

SAIR is released under a permissive CC BY 4.0 license and is hosted on Hugging Face. The dataset consists of a main table (sair.parquet) and multiple compressed structure archives (.tar.gz) located in the structures_compressed/ directory.

Researchers can access the data using the huggingface_hub, pandas, and pyarrow libraries. The structure archives are large (approximately 10 GB each), and the documentation recommends downloading only the necessary archives for local extraction.

Sources