Multi-Modal Foundation · SandboxAQ
AQAffinity
An open-source protein–ligand binding affinity prediction head built on OpenFold3 for drug discovery.
What AQAffinity is
AQAffinity is an open-source model developed by SandboxAQ in collaboration with the OpenFold Consortium to predict protein–ligand binding affinity. Estimating how strongly a small molecule binds to a target protein is a central challenge in computational drug discovery and structural biology.
Rather than serving as a proprietary tool, AQAffinity provides a transparent baseline for researchers. It implements a dedicated binding affinity head that functions directly on top of the OpenFold3 biomolecular structure prediction architecture.
How it works
AQAffinity acts as an affinity prediction head coupled to the OpenFold3 framework, replicating the affinity prediction module introduced in MIT and Recursion's Boltz-2 model. It can generate structural and sequence embeddings from input queries, optionally querying a Multiple Sequence Alignment (MSA) server and running sequence alignments using kalign2.
For model development and evaluation, training and validation assay datasets were prepared following the protocols established in the Boltz-2 preprint. Training can be performed directly on pre-extracted embeddings dumped during inference.
Key facts
- Base architecture
- OpenFold3
- Module architecture
- Replication of Boltz-2 binding affinity prediction head
- Key dependency
- kalign2
- Primary domain
- Drug discovery, chemistry, and structural biology
What you can use it for
- Protein–ligand affinity prediction
- Estimating binding affinities between candidate drug molecules and target proteins from sequence and structural queries.
- Structural biology baseline benchmarking
- Serving as an open baseline to benchmark novel loss functions, scoring mechanisms, and affinity prediction architectures.
- Custom affinity head training
- Extracting OpenFold3 embeddings from structural queries to train and validate custom affinity scoring heads on bespoke assay datasets.
Run it here
We list this model but cannot run it yet: gated on the Hub — never run past a gate the user did not accept.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Replicates the published Boltz-2 binding affinity module architecture within an open framework.
- Operates in tandem with OpenFold3 to leverage biomolecular structure representations.
- Accompanied by released assay datasets to allow reproducible training and validation.
Watch out for
- Requires external installation of kalign2 for sequence alignment prerequisites.
- Dependent on OpenFold3 checkpoints and embedding generation to perform predictions and training.
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · apache-2.0
This model is gated by its authors — its licence has to be accepted on the source repository before it will run.
Questions
Can I use AQAffinity commercially?
Check the licence before commercial use. This model is gated by its authors — its licence has to be accepted on the source repository before it will run.
What model architecture does AQAffinity use?
It implements the affinity prediction module architecture from Boltz-2 as a prediction head operating on top of OpenFold3.
What inputs and prerequisites are required to run predictions?
Running predictions requires query JSON inputs, runner configuration files, OpenFold3 and affinity checkpoints, and the kalign2 sequence alignment tool installed in the environment.
How are custom models trained with AQAffinity?
Users can generate and dump embeddings using the prediction pipeline on their dataset, followed by running the training pipeline on those precomputed representations.
Written from AQAffinity's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.