AlphaPulldown v2.5.0

AlphaPulldown 2.5.0 — AlphaFold3 v3.0.3 with RTX/Blackwell support

Highlights

• AlphaFold3 v3.0.3: alphafold3 submodule updated to official AF3 v3.0.3 (KosinskiLab fork main 86b9ea3), keeping gapped/discontinuous residue ID support.
• RTX PRO 6000 / Blackwell (CC 12.0) GPUs now work: the AF3 runtime moves to the JAX 0.9.1 / tokamax 0.0.11 stack, resolving the silent ptxas failures that previously crashed AF3 on RTX PRO 6000 nodes (related to google-deepmind/alphafold3#394).
• AlphaFold2 backend unchanged and verified: the alphafold submodule gets a one-line NumPy 2-safe chain-feature merge fix (5bf5b9f); the AF2 container still runs on numpy<2 / jax 0.5.3. AF2 monomer GPU inference was re-validated.
• AF3 container now reports/installs alphafold3==3.0.3; smoke-test species-identifier resolution fixed.

Validation

• Full unit (309 passed, 2 skipped) and integration (76 passed, 1 skipped) suites.
• RTX PRO 6000 GPU smoke on gpu50: alphafold3==3.0.3, jax==0.9.1, tokamax==0.0.11, real GPU matmul, rtx_smoke=ok.
• AF2 backend monomer GPU inference (TestRunModes::test__monomer) passed on gpu32.
• CI green across smoke (3.10/3.11), coverage, and both container builds.

Merged in #625 and #626.

AlphaPulldown v2.4.1

What's Changed

• mmseq precomputed enrichment by @DimaMolod in #615
• Re-pin numpy<2 after jax install in alphafold2 dockerfile by @DimaMolod in #622
• Update AlphaFold3 to Tokamax v3.0.2 by @DimaMolod in #623

Full Changelog: 2.4.0...2.4.1

AlphaPulldown v.2.4.0

What's Changed

• Fix #141 #256 #258 #293 #336 #456 #459 #463 #464 #471 by @DimaMolod in #605
• Fix #332 mmseqs2 TrueMultimer custom template support by @DimaMolod in #607
• Fix #465 reuse existing monomer features in default TrueMultimer mode by @DimaMolod in #609
• Fix #42 skip-MSA feature generation by af2/mmseqs2/af3 data pipeline by @DimaMolod in #611
• Fixing AF3 trimer pairing logic when running af3 backend with af2 features by @DimaMolod in #612

Full Changelog: 2.3.0...2.4.0

AlphaPulldown v.2.3.0

Bugs fixed

• Using Af3 backend with Af2 features occasionally led to broken pairing MSAs due to the inverse order of sequences
• Pairing never worked when mmseqs2 features were used

What's Changed

• Af3 backend supports chopped regions by @DimaMolod in #593
• Ap point af3 submodule to kosinskilab branch by @DimaMolod in #596
• Issue 588 by @DimaMolod in #597
• Reorganize tests, report coverage and 20 not tested functions in ci/cd by @DimaMolod in #598
• Tests by @DimaMolod in #600

Full Changelog: 2.2.0...2.3.0

AlphaPulldown v.2.2.0

What's Changed

• Fix #588 af3 msa order in #590
  (Using af2 features with af3 backend sometimes led to omitting paired MSAs)
• Fix #578 /af3 output dir shared root regression in #591

v2.1.9

AlphaPulldown v2.1.9

🧬 Major fixes

• Fix: normalized asym_id handling

  • always return normalized asym_id feature copy
  • fixes incorrect chain IDs in AF2 multimer outputs
  • ensures consistent chain/entity mapping in PDB and ModelCIF outputs

---

🧪 AlphaFold3 / multimer support

• Added proper support for copy numbers in AF3 homooligomers

  • correct handling of copy notation (protein:N)
  • correct mapping between input copies and internal chain indices

---

⚙️ Backend / architecture

• backend is now always initialized explicitly

• introduced lazy backend manager

  • avoids importing AF2/AF3 dependencies at import time

• enforced rule: no direct backend imports outside backend manager

👉 result: lighter CLI startup, fewer import-time crashes, cleaner backend isolation

---

📦 Dependencies & packaging

• pinned setuptools < 82
• additional pin for pdbfixer compatibility during relaxation
• updated setuptools constraints in pyproject.toml
• switched to ap-parser 0.3.0

---

🧱 CI / containers

• removed heavy JAX installation on CPU-only environments

  • avoids installing unusable GPU builds
  • reduces environment size and install failures

• disabled AlphaLink container build job

• excluded backend-specific tests from AF2/AF3/AL2/CF builds

---

📚 Documentation

• updated README:

  • version numbers
  • SLURM setup instructions
  • general cleanup and fixes

---

⚠️ user-visible changes

• multimer chain IDs may differ vs 2.1.8 due to corrected asym_id normalization
• AF3 homooligomer jobs now correctly respect copy number
• lighter installation and fewer dependency conflicts
• fewer backend-related import errors

---

🧠 Summary

v2.1.9 is a correctness + stability release with two functional improvements:

• fixes chain identity (asym_id)
• adds proper AF3 homooligomer copy support

AlphaPulldown v2.1.8

• Transition to pyproject.toml
• Fix unstable build of alphafold3 image
• Use the latest parser supporting protein:copy notation for af3 features
• Modelcif converter now works with precomputed features #575
• Do not build alphalink on release (discontinued)
• Make sure autogenerated pseudo PDB ID is parsable by AF2 (TrueMultimer)
• Fix broken links in example_<1,2,3>.md

AlphaPulldown v2.1.5

1. Reduced sized of alphafold2 and alphalink docker containers
2. Added flag validator that prevents using eg alphafold2 flags with alphafold3 backend
3. Drop redundant flags from run_multimer_jobs.py so the scirpt works with the new validator (credits to @Qrouger)
4. Use ccd instead of cached_ccd in af3 backend (again, thanks to @Qrouger !)
5. Use dense MSA representation when using alphafold2 features with alphafold3 backend
6. Move inputs parser to a separate repository
7. Renamed alphafold backend into alphafold2 (including name of the --fold-backend flag)

AlphaPulldown v2.1.4

1. Add options to save embeddings and distogram from AF3 backend.
2. Add backend-specific flag validator to run_structure_prediction.py (now you can't run backends with wrong flags!).
3. Sync alphafold2 module to the latest version (choose number of cores for MSA tools, global refactoring with pyink, etc).
4. Trim and simplify docker instructions for AlphaLink2.
5. Renamed 'fold' container into 'alphafold2'.
6. Renamed 'pulldown' dockerfile into 'alphafold2'.
7. Updated documentation.

AlphaPulldown v2.1.1

Fixed issues #524 and #528