timodonnell/helico-foldbench-eval-v1

Full-input FoldBench inference features for Helico exp23

This evaluation-only dataset tracks all 1,522 official FoldBench assemblies: 1482 prepared full inputs, 27 explicit resource-ceiling failures and 13 deferred assemblies awaiting external sequence alignments. No structural cropping or native pocket is used. All failures and pending entries remain in official category denominators. Pending assemblies have no inference feature payload and cannot enter a model run; a later immutable revision and completion run can add them without changing already evaluated inputs.

Features are rebuilt with the public Protenix v1 inference pipeline from full sequences and CCD chemical identities. Native complex coordinates, contacts, observation masks, pocket geometry and inter-entity covalent constraints are absent from model inputs. ref_coords are independently reconstructed CCD reference conformers. Reference structures are accessed separately for scoring from the pinned upstream source.

features/<assembly>.pt.gz contains trusted PyTorch payloads (including a Biotite AtomArray for mmCIF export); features/manifest.json and the identical root manifest.json record content hashes and explicit failures. Features are complete tensors; embedded absolute alignment paths are unused preparation metadata. To rebuild features, extract chemical_inputs.tar.gz and alignments.tar.gz, resolve the relative alignment paths against the extraction directory, and use the pinned upstream inference implementation.

MSAs prefer paired/unpaired alignments from pinned Protenix data, followed by supplied FoldBench alignments. Missing protein alignments were generated with the ColabFold MMseqs2 server on 2026-10-15; individual sequences, sources and file hashes are recorded in alignment_manifest.json. Query-only chains in ready features are peptides shorter than 15 residues. The archive also retains query-only placeholders for 13 unavailable alignments; these are used only by deferred inputs, never by ready features. Rebuilding must respect manifest status and must not substitute those placeholders into inference. All alignment files are included in the archive. Profiles use complete alignments; feature tensors retain the query and at most 511 other rows using per-assembly deterministic seeds. Current MSAs mean results should be compared to the identically configured initialization baseline, not published Protenix benchmark scores with other inputs/settings.

Maximum full-input size is 2,048 tokens. Ligands are atom tokens, so the token cap can exceed the polymer residue count. No fine-tuning training PDB ID overlaps these assemblies; preparation_summary.json records the checked training-ID file and the full input audit. This is a PDB-ID split audit, not a claim of sequence dissimilarity.

Provenance

• FoldBench source: timodonnell/helico-data@382ad3219097b764500c75475aaae3f82547f4fd. All 1,540 used source files were content-verified, including the modified DockQv2 implementation; see source_manifest.json.
• Protenix implementation: bytedance/Protenix@85767b811c40ed46e73a9b39519cf6bfca8701ba.
• Protenix data: LiteFold/protenix-data@47150f5244c15967e69315f3937d0ebcf863f317.
• Fine-tuning split: timodonnell/helico-protenix-v1-pdb@164e41cf98d7fdf4d437192a9947138ca2464011; training release-date cutoff before 2021-09-30 and FoldBench exclusions retained.
• Preparation: helico.contact_foldbench_data and experiments/exp23_contact-evaluation/prepare_foldbench_release.py in the experiment source.
• Experiment 23. Official benchmark inference is label-free; reference scoring is performed after protocol selection using validation only.

Feature-manifest SHA256: fbd9d5832930a9f62ae7e9c23e57266253bbbff9e849fb075e959f97ec00fb32.