2.5.1

AlphaPulldownSnakemake 2.5.1

Fixed

• AF3 JSON inputs (e.g. ligands) in fold specifications are now handled correctly (#41).
  A *.json token such as ligand.json in a fold (e.g. P12345+ligand.json:80) is now
  treated as a direct AlphaFold 3 input supplied via feature_directory, instead of being
  mistaken for a protein. Previously such tokens were scheduled for download_uniprot +
  create_features and required a .pkl/_af3_input.json that never existed, so mixed
  protein + ligand folds failed. They are now symlinked and passed straight to inference,
  with no download/feature-generation jobs. Verified for both AF2-features + AF3-backend
  and pure-AF3 runs.
• symlink_features is now idempotent (skips already-correct links, repairs broken ones).

Changed

• config/config.yaml now references the published 2.4.0 AlphaFold2/AlphaFold3 backend
  container tags in its examples.
• Documented AlphaJudge reporting in config/config.yaml: a per-fold RCSB-style report.pdf
  is written next to interfaces.csv automatically, and --aggregate_report produces an
  aggregate validation PDF alongside reports/all_interfaces.csv.

Dependencies

• Requires alphapulldown-input-parser>=0.5.1 (already pinned in
  workflow/envs/alphapulldown.yaml and CI). 0.5.1 preserves .json tokens during
  fold-spec normalization, which is what makes the #41 fix possible. Users with a pinned
  older parser must upgrade.

Full changelog: 2.5.0...2.5.1

2.5.0

Highlights: SLURM --mem and (optional) GPU selection are now sized from the input sequence length, oversized folds are skipped before submission. Backward compatible — existing configs keep working.

Length-aware memory (host --mem)

Computed per stage at scheduling time:

• create_features → safety · (base + per_residue · L)
• structure_inference → safety · (base + per_token_sq · N²) (O(N²) pair representation)

Defaults differ by backend:

backend | feature base | feature/residue | inference base | inference/N² -- | -- | -- | -- | -- alphafold2 | 64000 MB | 40 MB | 16000 MB | 0.0055 alphafold3 | 40000 MB | 25 MB | 8000 MB | 0.0045

First attempt includes a safety margin (mem_safety_factor, default 1.25); OOM retries still escalate via *_ram_scaling ** (attempt-1). With unified memory, this also sizes the GPU spill ceiling.

Skip over-large folds (Closes #33)

max_total_length_alphafold2: 5000     # AF2-Multimer total complex residues
max_total_length_alphafold3: 7000
max_protein_length: 0                 # 0 = off; over-limit protein drops its folds
length_filter_fetch_uniprot: true     # false for offline runs

Lengths resolved from local FASTA → cache → UniProt. Skipped folds listed in

2.4.0

What's Changed

• Add GPU node exclusion, runtime cap, and unified memory for inference by @DimaMolod in #42

Full Changelog: 2.1.8...2.4.0

2.1.8

Use latest parser

2.1.7

Allow empty GPU model

2.1.6

Flags for creating features

2.1.5

Typo in recursive_report rule

2.1.4

Fixed a bug that homo-oligomers weren't properly modeled with alphafold3 backend

2.1.3

1. use official snakemake executor plugin
   https://snakemake.github.io/snakemake-plugin-catalog/plugins/executor/slurm.html
2. automatically bind all used disks to singularity
3. simplify README.md
4. Snakefile: change walltime to runtime everywhere
5. config.yaml
   • update alphafold_inference slurm instructions (second line was ignored before)
   • use smaller proteins for tests
   • added profiles for slurm and for local workstation into config/ so it is shipped upon snakedeploy

2.1.2

Support for AF3 data pipeline