Drug Discovery · SandboxAQ
SAIR
Over one million protein-ligand pairs with cofolded 3D structures and binding affinities for drug discovery.
What SAIR is
The Structurally-Augmented IC50 Repository (SAIR) is a large-scale biochemical and structural dataset produced by SandboxAQ in collaboration with Nvidia. It pairs experimental binding potency values with three-dimensional structures of protein-ligand complexes.
A central goal of computational drug discovery is the prediction of binding affinity, which measures how strongly a potential drug molecule interacts with a protein target. Historically, structure-based deep learning methods have faced a scarcity of experimentally determined 3D structures that also possess measured affinity values (such as IC50 or Ki), forcing many computational workflows to depend solely on 1D amino acid sequences or 2D molecular graphs.
SAIR bridges this structure-activity gap by pairing experimental potency data with computationally cofolded protein-ligand complexes. It provides over one million unique pairs to serve as a resource for training, evaluating, and fine-tuning machine learning and biofoundation models.
How it was made
SAIR integrates experimental potency measurements with synthetic structural conformations. The affinity records, including IC50 and Ki values, were curated from the ChEMBL and BindingDB public databases. The corresponding 3D protein-ligand complex geometries were generated by cofolding each pair using the Boltz-1x structural prediction model.
The complete dataset spans approximately 2.5 TB of data. Tabular information and metadata are provided in an Apache Parquet file (sair.parquet), while the 5.2 million 3D structural coordinate files are hosted in a compressed directory partitioned into more than 100 individual tar.gz archives of roughly 10 GB each.
Key facts
- Unique protein-ligand pairs
- 1,048,857
- Total 3D structures
- 5.2 million
- Total data volume
- 2.5 TB
- Affinity data sources
- ChEMBL and BindingDB
- Cofolding method
- Boltz-1x
- Data formats
- Parquet and tar.gz archives
What you can use it for
- Binding affinity prediction
- Train machine learning models on 3D cofolded complexes alongside experimental IC50 and Ki values to predict potency directly from structural interactions.
- Biofoundation model benchmarking
- Benchmark the performance and structural sensitivity of biological foundation models across more than one million standardized protein-ligand pairs.
- Target-specific fine-tuning
- Fine-tune pre-trained structural algorithms on targeted complex subsets derived from curated ChEMBL and BindingDB records.
Explore the data
This dataset is gated
Its authors require their terms to be accepted before the data can be read. Accept them on the original dataset page. The platform does not accept terms on your behalf, so the rows are not shown here.Strengths and limits
Good at
- Provides 1,048,857 unique protein-ligand pairs across 5.2 million 3D structures linked to experimental binding measurements.
- Bridges the structural data gap by providing cofolded 3D coordinate data where experimental co-crystal structures are unavailable.
- Separates tabular metadata in Parquet format from partitioned structure archives to support modular downloading.
Watch out for
- All 3D complex geometries are synthetically predicted using Boltz-1x rather than experimentally determined through crystallography or cryo-EM.
- The full repository size of 2.5 TB across more than 100 individual 10 GB archive files requires significant storage and network resources.
Questions
Can I look inside SAIR without downloading it?
Not here: its authors gate it behind terms each user must accept on the original page. Its description and statistics are still shown.
What licence is SAIR under?
Its authors declare cc-by-4.0. Read the terms on the original page before reusing the data.
What model was used to generate the 3D structures?
All 3D protein-ligand complex structures in the repository were computationally cofolded using the Boltz-1x model.
Where were the experimental binding values gathered from?
Potency measurements, including IC50 and Ki values, were curated from the ChEMBL and BindingDB databases.
How is the dataset structured for downloading?
Metadata and tabular information are provided in a single sair.parquet file, while the 3D structures are packaged in over 100 individual tar.gz archives within the structures_compressed folder.
What is the total size and complex count of SAIR?
SAIR comprises 1,048,857 unique protein-ligand pairs, 5.2 million cofolded 3D structures, and roughly 2.5 TB of total data.
How to cite it
Lemos, Pablo, Zane Beckwith, Sasaank Bandi, Maarten van Damme, Jordan Crivelli-Decker, Benjamin J. Shields, Thomas Merth, Punit K. Jha, Nicola De Mitri, Tiffany J. Callahan, AJ Nish, Paul Abruzzo, Romelia Salomon-Ferrer, and Martin Ganahl. SAIR: Enabling Deep Learning for Protein-Ligand Interactions with a Synthetic Structural Dataset. 2025. doi:10.1101/2025.06.17.660168.
Written from SAIR's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.