The pair representation scaling change for each backend, as a diff against a
pinned upstream revision. A scalar β multiplies the pair representation,
z := (1 + β)·z, just before the Pairformer stack.
| Patch | Upstream | Revision | License |
|---|---|---|---|
boltz-2.2.1.patch |
https://github.com/jwohlwend/boltz | tag v2.2.1 (cb04aec) |
MIT |
alphafold3-97639ff.patch |
https://github.com/google-deepmind/alphafold3 | 97639ff (2026-05-06) |
CC BY-NC-SA 4.0 |
The prs CLI applies the same scaling at runtime (src/prs/boltz_hook.py),
so a stock pip install boltz==2.2.1 is enough to run a β sweep. Use the patch
to put the change in a Boltz-2 checkout instead:
git clone --branch v2.2.1 https://github.com/jwohlwend/boltz
cd boltz
git apply -p1 /path/to/pair-representation-scaling/patches/boltz-2.2.1.patchIt adds a --beta option to boltz predict and applies the scaling in
PairformerModule.forward. The confidence head builds its own Pairformer stack
from the checkpoint hyper-parameters and stays unscaled.
AlphaFold 3 is built from its own source, so the patch goes into your checkout:
prs patch-af3 /path/to/alphafold3It adds a --beta flag to run_alphafold.py, which passes the value to the
Evoformer through the PRS_BETA environment variable.