SandboxAQ Releases SAIR: 5.24M Co-Folded Protein-Ligand Structures With IC50 Labels
Hugging Face
- Data
- Research
- Benchmarks
- #sair
- #sandboxaq
- #protein-ligand
- #ic50
SandboxAQ announced SAIR, an open dataset of 5.24 million AI-generated protein-ligand 3D complexes paired with curated IC50 measurements, derived from over 1 million computationally co-folded pairs. The company reports generating it in three weeks on 760 NVIDIA H100 GPUs via DGX Cloud, with 97% of structures passing PoseBusters checks. It is published on Hugging Face under a CC BY 4.0 license.
What SandboxAQ says it released
SandboxAQ announced SAIR, a dataset linking predicted 3D protein-ligand structure to measured potency. Per the post, over 1 million unique protein-ligand pairs were co-folded with the Boltz1 model and expanded into 5.24 million distinct complexes, five conformations per pair. Each complex is paired with a curated IC50 value drawn from ChEMBL or BindingDB, connecting geometry to empirical binding data. SandboxAQ reports that 97% of structures passed PoseBusters checks for chemical sanity and physical plausibility, and that more than 40% of the proteins have no PDB structure. The company states the dataset was computed in about 130,000 GPU hours on 760 NVIDIA H100 processors via DGX Cloud on Google Cloud, finishing in three weeks against a three-month estimate, and is hosted on Hugging Face under CC BY 4.0. These figures are SandboxAQ's own claims, not independent findings.
A realistic trial for a computational chemist
A medicinal chemist or computational chemist working on a target with no experimental structure could try SAIR as a hypothesis source rather than ground truth. The plausible use, given the text, is to pull predicted complexes for related targets, inspect proposed binding poses and potency associations, and shortlist compounds for virtual screening or later wet-lab assays. Because the structures are AI-predicted, such a team could treat any pose as a starting hypothesis to be checked against chemistry intuition and, where affordable, experimental data. A modeler training an affinity predictor could also try SAIR as a training or benchmarking resource, keeping a held-out split to avoid self-confirmation on the same predicted geometries.
How to run a bounded evaluation
A team could try a small, time-boxed pilot: select one target family with known actives, load the SAIR parquet table via the Hugging Face Hub as the post demonstrates, and compare model rankings against compounds with existing measured potency. Measure success against a pre-pilot baseline, such as the team's current docking or scoring workflow, tracking rank correlation and hit rate on a fixed test set. Failure checks should include spot-checking PoseBusters pass rates on the downloaded subset, watching for duplicate or inconsistent IC50 entries, and verifying that structures correspond to the intended protein. Keep a human reviewer in the loop to reject implausible poses before any synthesis decision, and cap disk use, since the post notes structure archives can be around 10 GB each.
Trade-offs, prerequisites and when to skip it
Adopting SAIR requires infrastructure for large downloads, Parquet tooling and, ideally, GPU capacity for downstream modeling, plus expertise to interpret predicted poses. The main trade-off is scale and openness against provenance: potency labels come from public databases with varying assay conditions, and structures are model outputs, so apparent signal may reflect Boltz1 biases rather than biology. It is not a replacement for experimental structural validation. A team should avoid relying on SAIR alone for go/no-go decisions, for regulatory submissions, or where licensing reuse of underlying public data is uncertain despite the stated CC BY 4.0 terms. Because the evidence here is a publisher's summary, treat reported speed-ups and pass rates as claims awaiting independent reproduction.
Context and limitations
- Based on a short source excerpt; the full article and bioRxiv manuscript were not analyzed.
- Quality, coverage and benchmark figures are SandboxAQ's self-reported claims, not independently verified.
- Structures are AI-predicted, not experimental; potency labels are curated IC50 values of varying assay context.
- Based on bounded source text, not independent testing. Practical suggestions are editorial proposals.