This repository contains the code for Nesso-1, a coarse-grained cofolding model that can predict binding affinity. For more details about the method and evaluations, refer to this Technical Report. This project is licensed under a permissive Apache License 2.0.
PyPI release coming soon!
pip install "git+https://github.com/recursionpharma/nesso.git"GPU / CUDA : The nesso package relies on PyTorch. For GPU support, install a CUDA-enabled PyTorch build first, then install nesso. To additionally enable cuEquivariance kernels for extra speedups (CUDA 12 only; Apache License 2.0):
pip install "nesso[kernels] @ git+https://github.com/recursionpharma/nesso.git"If the kernels extra fails to install or you are running on CPU, use the default install and/or pass --no_kernels at runtime.
We recommend using uv to install the project in development mode.
git clone https://github.com/recursionpharma/nesso.git && cd nesso
uv sync # runtime + dev tools (pytest, ruff, pre-commit)
pytest # run the CPU-only test suite
pre-commit run --all-files # lint + format(Plain pip install -e . installs the runtime only; the dev tools live in the
dev dependency group, which uv sync installs.)
1. Create an input YAML (e.g. complex.yaml):
sequences:
- protein:
id: A
sequence: MKTAYIAKQ... # amino acid sequence
- ligand:
id: B
smiles: "Fc1ccc(cc1)C(=O)Nc1ccc(cc1)S(=O)(=O)N"
properties:
- affinity:
binder: B2. Run prediction:
nesso predict complex.yaml --out_dir ./outputNote: For input formats, caching, and all CLI options, see docs/prediction.md.
- Pocket Conditioning
- Structural Templating
@article{valencelabs2026nesso1,
author = {Valence Labs, Recursion, Nikhil Shenoy, David Errington, Emmanuel Bengio, Kacper Kapu\'sniak, Kerstin Klaeser, Yui Tik Pang, Vladimir Radenkovic, Prudencio Tossou, Therence Bois, Andrew Wedlake, Francesco Di Giovanni},
title = {Nesso-1: Accelerating Open-Source Binding Affinity Predictions},
year = {2026}
}This project is licensed under Apache License 2.0.
Third-party attributions: licenses/THIRD_PARTY_NOTICES.md
We thank the following open-source codebases for enabling the development of this project,