LeFlur Benchmarks

Pre-tokenized protein structure benchmarks used to evaluate the LeFlur protein and protein-ligand generative model. Each subdirectory contains one .pt file per target, ready to be consumed by the lobster_generate forward/inverse folding pipelines via the ${paths.benchmarks.<name>} Hydra interpolation in paths/public.yaml.

Quickstart

pip install lbster
lobster_leflur_benchmarks fetch cameo
lobster_leflur_benchmarks fetch multiflow_test

# Inverse folding on the CAMEO 2022 monomer benchmark
lobster_generate --config-name experiment/generate_inverse_folding \
    paths=public model.ckpt_path=leflur-ted \
    'generation.input_structures=${paths.benchmarks.cameo}/*.pt'

Benchmarks

cameo (cameo-2022/)

CAMEO 2022 monomer benchmark, pre-tokenized into one .pt per target. Drives the CAMEO 2022 rows of LeFlur Table 1 (inverse folding) and Table 3 (forward folding).

• License: academic; per-target PDB licenses apply
• Citation: Haas et al., 'Continuous Automated Model EvaluatiOn (CAMEO) complementing the critical assessment of structure prediction in CASP12', Proteins (2018). Per-target PDB structures are released under their original PDB / RCSB licenses.
• File glob: cameo-2022/*.pt
• Per-record schema (top-level keys of each .pt):
  • pdb_path
  • sequence
  • sequence_str
  • coords_res
  • chains_ids
  • indices
  • mask
  • real_chains

multiflow_test (multiflow-test/)

Filtered MultiFlow test set, pre-tokenized into one .pt per target. Drives the MultiFlow rows of LeFlur Table 1 (inverse folding) and Table 3 (forward folding).

• License: MIT (matches upstream MultiFlow data release)
• Citation: Campbell et al., 'Generative Flows on Discrete State-Spaces: Enabling Multimodal Flows with Applications to Protein Co-Design', ICML 2024.
• File glob: multiflow-test/*.pt
• Per-record schema (top-level keys of each .pt):
  • sequence
  • coords_res
  • mask
  • indices
  • chains

posebusters_benchmark_no_overlap (posebusters-benchmark-no-overlap/)

PoseBusters benchmark, deduplicated against the LeFlur training set (the 'no-overlap' subset). Pre-tokenized as paired {protein,ligand}.pt files. Drives the publication PoseBusters rows of LeFlur Table 2 (PL inverse folding) and Table 4 (PL forward folding) — this is the canonical leflur-pl evaluation set.

• License: CC-BY-4.0 (matches upstream PoseBusters release)
• Citation: Buttenschoen et al., 'PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences', Chem. Sci. (2024). 'no-overlap' filtering removes targets that share clusters with the LeFlur protein-ligand training set, isolating true held-out generalisation.
• File glob: posebusters-benchmark-no-overlap/*.pt
• Per-record schema (top-level keys of each .pt):
  • pdb_path
  • sequence
  • sequence_str
  • coords_res
  • chains_ids
  • indices
  • mask
  • real_chains
  • atom_names
  • atom_coords
  • atom_indices
  • element_indices
  • bond_matrix

posebusters_benchmark (posebusters-benchmark/)

Full PoseBusters benchmark (no overlap filtering applied), pre-tokenized as paired {protein,ligand}.pt files. Reported in supplementary tables alongside the deduplicated 'no_overlap' subset — most publication numbers use the 'no_overlap' variant.

• License: CC-BY-4.0 (matches upstream PoseBusters release)
• Citation: Buttenschoen et al., 'PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences', Chem. Sci. (2024).
• File glob: posebusters-benchmark/*.pt
• Per-record schema (top-level keys of each .pt):
  • pdb_path
  • sequence
  • sequence_str
  • coords_res
  • chains_ids
  • indices
  • mask
  • real_chains
  • atom_names
  • atom_coords
  • atom_indices
  • element_indices
  • bond_matrix

Notes on the CAMEO pdb_path field

The CAMEO records expose a pdb_path string that names the source PDB file (e.g. 7dz2.C.pdb). Only the basename is published here — the absolute path inside the original Genentech file tree has been stripped on upload. Downstream code that wants to load the raw PDB should look it up against an RCSB / CAMEO mirror by basename.

See also

• Sidney-Lisanza/leflur — the LeFlur model repo (3 canonical checkpoints).
• LeFlur README — paper tables + one-line reproduction commands.