alegendaryfish/CDDB-PDB-Protein-50-512

CDDB–PDB Protein Structures, 50–512 Residues

Curated sequences, observed atomic coordinates, physical side-chain torsions, observation masks, and separately filtered intrinsic-backbone labels for protein generation and conditional modeling.

Training PDB cutoff: 31 December 2023, inclusive, using the entry's initial public release date. The source snapshot was collected on 11 September 2026. These dates serve different purposes: historical entries use their audited coordinates from the current snapshot, including later revisions. This is not a reconstruction of the archive's 2023 coordinate versions.

The training collection contains 591,399 records: 430,069 CDDB predictions and 161,330 experimental PDB chain segments. These represent 518,383 distinct amino-acid sequences, 465,082 retained verified sequence clusters, and 56,311 distinct PDB entries. Record counts include alternative structural observations; they are not counts of independent sequence families. The joint_intrinsic training split contains 524,200 records. All post-cutoff PDB entries are excluded from this collection, including its internal development splits. A separate temporal evaluation configuration contains 249 validation and 467 test targets.

Loading

from datasets import load_dataset

train = load_dataset("alegendaryfish/CDDB-PDB-Protein-50-512", "joint_native",
                     split="train", streaming=True)
test = load_dataset("alegendaryfish/CDDB-PDB-Protein-50-512", "temporal_evaluation",
                    split="test", streaming=True)

The default joint_native configuration preserves observed geometry. joint_intrinsic also requires acceptable fitted backbone labels. The all_native archive includes declared quality exceptions and is not an instruction to train on every row. Source-specific configurations have cddb_ or pdb_ prefixes. The temporal configuration contains only PDB targets eligible in both joint views and has no training split.

Configuration	All internal splits	CDDB	PDB
all_native	591,399	430,069	161,330
joint_native	286,351	186,690	99,661
joint_intrinsic	529,276	429,576	99,700

Training and evaluation separation

Fixed protected inputs include the published MultiFlow conditional test set, La-Proteina motif source entries, previously used project evaluation panels, and public CASP13–15 targets. Full parent sequences and observed backbones supplement the benchmark inputs. Both CDDB and PDB are screened, since a predicted-structure collection can contain sequences from experimental targets.

The additional temporal validation window is 1 January–31 December 2024; the test window is 1 January 2025–11 September 2026. Candidate selection uses single-protein-chain entries, eligibility in both joint views, one representative per connected sequence/source group, deterministic selection and length strata. Here single-chain means one protein chain in the entry; it does not prove a biologically monomeric assembly. Up to 256 validation and 512 test candidates were fixed before similarity searches and model scoring. Candidates matching CASP, and test candidates matching temporal validation, are removed without replacement. Every original candidate remains protected from training.

Detected sequence matches at 30% identity and 80% shorter-sequence coverage, and structural near copies at shorter-normalized TM score 0.9, 80% coverage and 50 aligned residues, are excluded from the training collection. Exclusion extends to connected groups of verified sequence clusters and source parents. Direct matches, date exclusions, and group closure are reported separately. Exclusion reasons can overlap; their counts must not be summed. Group closure includes relatives without a direct above-threshold match.

The independent audit found no retained protected identities or saved matches meeting these thresholds. Searches are sensitive heuristics, not a guarantee against every remote homologue or shared fold. Only the registered temporal panel has this relative-exclusion claim; all other future PDB records form an unscored reserve. CDDB has no equivalent experimental release date and is described as predicted data screened against the protected inputs, not as a historically reconstructed 2023 collection. Its upstream predictor's training history is not controlled by this release.

Internal splits and sampling

The following development splits are separate from temporal_evaluation. Retained connected groups keep their source-collection assignment. No retained source parent or verified sequence cluster crosses these internal splits.

Internal split	Joint native	Joint intrinsic
train	283,742	524,200
validation	1,258	2,528
test	1,351	2,548

Choose the CDDB/PDB mixture explicitly. The sample_weight_* columns contain recomputed inverse eligible source-cluster sizes; uniform row sampling does not apply them automatically. Alternative structures remain available within a cluster. Cluster sizes describe the present release. If a historical clustering anchor is absent from retained sequences, cluster_representative_sequence is null; original cluster identifiers are preserved for provenance.

Geometry and quality

Coordinates are in angstroms and torsions in radians. Native coordinates, observation masks, torsions and eligibility flags are unchanged from the fully audited source bank. Missing atoms are not imputed. Both observed and valid chi masks are provided. The chi convention is the negative of the conventional MDTraj dihedral; use the included decoder.

Intrinsic labels require fitted N/CA/C RMSD at most 0.5 Å and the documented geometry checks. Their explicit proline closure exception does not certify native proline ring closure. Report native and fitted representations separately. See Methods and Source curation.

License

CDDB-derived records retain CC BY 4.0 attribution; PDB observations follow wwPDB's CC0 policy. See License. CASP coordinates used for screening are not redistributed as training examples. The benchmarks/ directory records source URLs and target membership.