NOTICE

foldsteer
Copyright 2026 the authors

Licensed under the Apache License, Version 2.0. See LICENSE.

------------------------------------------------------------------------------
Attribution
------------------------------------------------------------------------------

This project is an independent, model-agnostic reimplementation of
inference-time chemical steering for AF3-style structure prediction. Its design
and default parameters derive from two prior open-source implementations, and
neither project is affiliated with or endorses this one.

1. Boltz / Boltz-1x
   https://github.com/jwohlwend/boltz
   Copyright (c) 2024 Jeremy Wohlwend, Gabriele Corso, Saro Passaro
   MIT License.

   The flat-bottom potential formulation, the Feynman-Kac steering scheme
   (including the use of stepwise energy improvement as the FK potential and
   the guidance likelihood correction), and the default term weights in
   `foldsteer.engine.default_terms` follow the released Boltz implementation
   (src/boltz/model/potentials/potentials.py and
   src/boltz/model/modules/diffusion.py).

2. Protenix / Protenix-v2
   https://github.com/bytedance/Protenix
   Copyright (c) ByteDance Ltd. and/or its affiliates
   Apache License 2.0.

   The separation of steering into a standalone, config-driven subpackage with
   a potential class registry follows the structure of `protenix/tfg/`. The
   sp2-planarity, sp3-non-planarity, and amide-planarity terms respond to the
   revised validity criteria described in the Protenix-v2 technical report.

Constraint extraction uses RDKit (BSD 3-Clause), https://www.rdkit.org/

The OpenFold3 adapter targets https://github.com/aqlaboratory/openfold-3
(Apache License 2.0) but vendors no OpenFold3 code; OpenFold3 is imported
lazily and is not a required dependency.