---
license: apache-2.0
extra_gated_prompt: "The AQAffinity preview model is released under Apache 2.0 license. You will automatically get access to the model after answering the following simple questions:"
extra_gated_fields:
  First Name: text
  Last Name: text
  Company Name or Affiliation: text
  Role or Job Title: text
  I want to use AQAffinity for: text
tags:
- chemistry
- biology
- protein
- ligand
- binding
- affinity
- binding affinity
- drug discovery
---

Introducing AQAffinity
============
***SandboxAQ's open protein-ligand binding affinity prediction head built on top of OpenFold3***

![](protein-ligand-binding-resized.png)

**In collaboration with the OpenFold Consortium**

## Overview ##
This repository contains an implementation of a Binding Affinity Head designed to operate on top of the OpenFold3 architecture. 
It is a direct replication of the affinity prediction module introduced in Boltz-2 (by MIT/Recursion).

The goal of this project is not to provide a final, closed commercial product, but to establish a strong, transparent baseline for the structural biology community. 
We believe that binding affinity prediction, one of the "holy grails" of drug discovery, advances fastest when training data, pipelines, and model architectures are fully open for inspection and improvement.

## Installation
### Prerequisites
This model requires `kalign2` for sequence alignment. Please install it using Mamba or Conda before installing the Python package:

```bash
mamba install kalign2 -c bioconda
```

### Option 1: Direct Install
```bash
pip install git+[https://huggingface.co/SandboxAQ/aqaffinity](https://huggingface.co/SandboxAQ/aqaffinity)
```

### Option 2: Local Install
If the repository is private or you wish to download the source code first, use the Hugging Face CLI to download the repository and then install it locally.
```bash
huggingface-cli login
hf download SandboxAQ/aqaffinity --local-dir ./aqaffinity
pip install ./aqaffinity
```
## Running AQAffinity
```bash
aqaffinity predict --query_json <of3_type_input_json> --runner_yaml <of3_runner_yaml>  --inference_ckpt_path <of3_model_weights> --use_msa_server true --output_dir <output_dir> --binding_affinity_ckpt_path <binding_head_model_weights>
```

## Training AQAffinity
Training code is included and can be done by first generating embeddings using the ```predict``` mode with the ```--dump_embeddings flag``` on your training set and then ```aqaffinity train```, documentation and examples are currently being prepared
and will be available soon but please reach out if you need helping with training in the meantime.

## Contributing to AQAffinity
We welcome contributions! This is a community effort.
Found a bug? Open an issue.
Have a better loss function? Submit a PR.
Want to benchmark on a new dataset? Share your results in the Discussions tab.

We believe "open source" means nothing without open data. Hence, we also release training and validation assay information we used to train and validate this model in csv file format.
User can use these to retrain and validate their own model. 

Both training and validation assay information was obtained following the procedures outlined in the official Boltz-2 preprint.
The corresponding SQL queries and preprocessing scripts will be added at a later point.

Together, we can build a useful transparent instrument for scientific discovery.