AQAffinity

AQAffinity is an open-source model that predicts protein-ligand binding affinity directly from a protein sequence and a small-molecule SMILES string, without requiring an experimentally determined protein structure. It was released by SandboxAQ on January 20, 2026, built on top of OpenFold3, the co-folding structure prediction model from the OpenFold Consortium. Binding affinity — how tightly a candidate compound engages its target — is a central quantity in early drug discovery, and predicting it quickly and cheaply lets teams triage compounds before committing to expensive physics-based simulations or synthesis.

AQAffinity adapts the affinity-prediction approach introduced in Boltz-2 and implements it as a dedicated head on the OpenFold3 framework. A sequence-SMILES input is passed through the OpenFold3 trunk to produce single and pair representations of the protein-ligand complex; a diffusion module folds the complex, and an affinity head reads out a predicted potency value. Because the structure is generated internally, the model is "structure-free" from the user's perspective: no crystal structure, docked pose, or prior complex is needed as input. SandboxAQ reports that this delivers predictions roughly 1,000 times faster than free-energy perturbation (FEP) methods, based on the published Boltz-2 benchmarks the architecture derives from.

The model is positioned explicitly as an open baseline for the OpenFold3 ecosystem rather than a commercial product. Its code, pretrained weights, and training and validation assay data are released to let researchers reproduce, benchmark, and fine-tune it on their own targets without vendor lock-in.

#Key Features

• Structure-free affinity prediction: AQAffinity estimates protein-ligand binding potency from a protein sequence and a ligand SMILES alone, generating the complex structure internally rather than requiring an experimental or pre-docked structure as input.
• Built on OpenFold3: The model runs as an affinity head on the OpenFold Consortium's OpenFold3 co-folding framework, extending it from structure prediction to potency estimation and giving it parity with competing co-folding-plus-affinity systems.
• Boltz-2-derived affinity head: The affinity module follows the Boltz-2 implementation, using a Pairformer stack that reads token-level distogram information from cropped pair representations focused on the binding-site region.
• Trained on the SAIR dataset: AQAffinity is trained in part on SAIR, the Structurally Augmented IC50 Repository — 1,048,857 unique protein-ligand pairs with 5.2 million cofolded 3D structures curated from ChEMBL and BindingDB.
• Fully open release: Code and weights are published on HuggingFace under Apache 2.0, alongside training and validation assay data, making it an open, reproducible baseline for affinity modeling.

#Technical Details

AQAffinity threads a sequence-SMILES pair through the OpenFold3 trunk to obtain single and pair representations of the protein-ligand complex, folds the complex with OpenFold3's diffusion module, and applies an affinity-prediction head whose Pairformer stack consumes token-level distogram features from binding-site-focused crops of the pair representation. The training data draws on SAIR — over one million unique protein-ligand pairs and 5.2 million cofolded structures assembled from ChEMBL, BindingDB, and related sources — with training and validation assay tables released as CSV files following the procedures described in the Boltz-2 preprint. In SandboxAQ's evaluations the model performs strongly on CASP16 targets while showing reduced accuracy on out-of-distribution GOSTAR assays, illustrating the generalization gap that affinity models still face on chemistry far from their training distribution. The current release does not include a binder classification head, which is noted as planned future work.

#Applications

AQAffinity is aimed at computational chemists and drug-discovery teams who need fast, early-stage potency estimates across large compound sets. Because it requires only sequence and SMILES inputs, it can score candidates before any experimental structure exists, supporting hit triage, virtual screening, and prioritization of compounds for slower, higher-fidelity FEP calculations or wet-lab assays. Its open licensing and released training data also make it a practical starting point for groups that want to fine-tune an affinity model on proprietary targets or benchmark it against in-house measurements without adopting closed commercial software.

#Impact

AQAffinity extends the open OpenFold3 stack from structure prediction into binding affinity, delivered within weeks of OpenFold3's launch as a community-oriented baseline. By pairing permissively licensed code and weights with the openly released SAIR dataset, it lowers the barrier to fast affinity prediction for academic and smaller industry teams that lack access to FEP infrastructure. Its honest reporting of weaker out-of-distribution performance on GOSTAR assays underscores that structure-free affinity prediction remains an open problem, particularly for chemistry distant from the training set. AQAffinity does not have a dedicated peer-reviewed or preprint publication; its methodology descends from the Boltz-2 affinity module, and its training data is documented in the SAIR dataset preprint from SandboxAQ.