[sokrypton]

Moves both notebooks to alphafold3-colabfold 3.1.18 (released today) and keeps RGI working on it.

RGI

RGI currently reaches the diffusion sampler by SHA256-checking alphafold3.model.network.diffusion_head.sample's source and rewriting four lines of it in memory. That pins it to 3.1.11. 3.1.17's sampler already hashes differently, and from 3.1.18 the sampler loop is refactored, so none of the four lines exist.

3.1.18 exposes the same point as a supported hook, diffusion_head.DENOISED_HOOK:

• The signature is hook(positions_denoised, sigma, step).
• It's applied after the denoiser, and after chai's second-order call.
• sigma is the step's pre-churn noise level and step its zero-based index, exactly what the source patch passed.

Changes:

• colabfold/rgi/sampler.py: guided(diffusion_head, minimize) uses the hook when the installed release has it, and otherwise instruments the 3.1.11 sampler exactly as before.
• colabfold/rgi/runtime.py: uses guided(). The 3.1.11 pin applies only when the hook is absent.
• tests/test_rgi_sampler.py: the sampler tests go through guided(), so they hold the same expectations against either release.
• SUPPORTED_MODELS gains protenix1, helixfold3 and esmfold2_fast. RGI was tested on each; see below.

Notebooks

• Both notebooks: VERSION = '3.1.18' in ColabFold2_preview.ipynb and AlphaFold3_of3.ipynb.
• ColabFold2_preview models: adds protenix1, helixfold3, esmfold2_fast and bioemu2, each with a row in the Instructions table.
• Licence: helixfold3's weights are non-commercial (NC-SA). Its table row and the Licence section say so.
• No confidence head: a model without one (bioemu2) now shows "Mean pLDDT: n/a" instead of averaging its all-zero pLDDTs to "0.000".

Testing

On a real Colab T4 (Python 3.13, JAX 0.11.1), with 3.1.18 installed from PyPI and colabfold.rgi from this branch, every cell ran except the download cell:

run	result
helixfold3, MMseqs2 server	mean pLDDT 96.0, pTM 0.93
openbind0 with an 8 Å RGI distance restraint	8.000 Å in all 5 samples
bioemu2	completes; pLDDT and pTM shown as n/a
AlphaFold3_of3.ipynb, openfold3	completes; mean pLDDT 56.3

tests/test_rgi_sampler.py passes against both alphafold3 3.1.11 and 3.1.18. These tests skip in this repo's CI, which doesn't install alphafold3.

Locally (A10, 3.1.18), the distance test pulled residues 1–5 of a 47-residue bundle to 8 Å from residues 40–46. RGI's own measurement gave 8.0 Å in every sample on 14 models: alphafold3, chai1, boltz2, protenix2, protenix1, openbind0, openfold3, opendde, rosettafold3, helixfold3, esmfold2, esmfold2_fast, esmfold2_lm600m and esmfold2_lm300m. Without RGI these were 11–15 Å. bioemu2 is unaffected because its own sampler isn't the one the hook reaches, so it stays vanilla only.

Ligand conformer restraints on a SMILES with three stereocentres gave correct stereochemistry in 5/5 samples on alphafold3 and boltz2, against 2/5 and 1/5 without RGI.

Unrelated, and present on main before this change: tests/test_rgi_notebook.py::test_native_form_runs_without_widgets_and_reads_every_edit[ColabFold2_preview.ipynb-openbind0] fails wherever rgi_toolkit is installed (CI skips it), because the notebook's form text no longer matches the test's expected copy.

cc @YoshitakaMo: this changes how RGI reaches the sampler.