OnixBind: Aligning co-folding models to experimental fitness

| OnixBind | OnixBind-Flash |

1. OnixBind Overview

OnixBind takes protein sequences with their MSA and a small molecule as direct inputs, and predicts the binding affinity of the complex as a single pX value per record, where a higher value means stronger binding.

This repository is inference only: it contains no training code, no optimiser state, and no dataset tooling. The released model uses one shared structural trunk and diffusion pass with two checkpoint-specific affinity heads. Both heads are evaluated on the same structural representations, and their equally weighted mean is reported as the final pX prediction.

An A100 80GB or higher-memory GPU is recommended.

2. Installation

2.1 OnixBind

1. Prepare Input File: Create an AlphaFold 3 input JSON following our input format specification. A complete record is provided at src/examples/5S8I_A.json.

2. Download Cache Data and Model: You can download from google drive. Place onixbind.pt in src/weights/ and ccd.pickle in runtime/alphafold3/constants/converters/ before running inference.

3. Installation and demo:

To more complete installation instructions and usage, please refer to the Installation Guide.

conda env create --file src/environment.yaml
conda activate onixbind
export CUTLASS_PATH=/path/to/cutlass

cd src && bash predict.sh

4. Output: Predictions will be saved to: ./output/predictions

2.2 Lightweight model: OnixBind-Flash

We provide a Jupyter Notebook demo for using OnixBind-Flash to perform inference on 10 targets.

1. Download Required Files: To run the demo, Please download the demo trunk here, and OnixBind-Flash model checkpoints here.

2. File Placement: Unzip the downloaded files and place them in the src/onixbind-flash folder.

3. Run Demo: Open and execute

inference_onixbind_flash.ipynb

This will generate predictions for the input file demo_input.csv, and the results will be saved in inference_result.csv.

3. Acknowledgements

• AuroBind is the V1 version of OnixBind.
• This repository uses the AlphaFold 3 inference data pipeline (feature processing and MSA handling), vendored under runtime/.
• The implementation of fast layernorm operators is inspired by OneFlow and FastFold, following Protenix's usage.

4. License

Unless otherwise stated, this code repository and the released OnixBind and OnixBind-Flash model weights are licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0). Third-party components retain their original licenses and copyright notices.