FuzzyBench-NOE: A Curated NMR Distance Restraint Dataset for Fuzzy IDP–Target Protein Complexes

Overview This dataset provides the first comprehensive collection of resources for benchmarking AI structure prediction methods on fuzzy protein complexes — biomolecular assemblies where an intrinsically disordered protein (IDP) retains conformational disorder upon binding its target. Crystal structures capture only the most ordered snapshot of these ensembles, making standard structure prediction benchmarks misleading. This dataset addresses this gap through three complementary ground truth sources: experimental NOE distance restraints from solution NMR, interface hotspot annotations from crystal structures, and per-residue thermodynamic helicity profiles from the Hadži statistical mechanical model. Dataset Contents Fuzzy_Benchmarking.zip/ │ ├── all_pdbs.csv # Full system registry (105 systems) │ ├── restraints # NMR distance restraint files (75 systems) │ ├── systems_metadata.csv │ ├── NOE_restraints_files/ │ │ ├── merged_15072_2jmx.str │ │ ├── merged_15357_2khs.str │ │ └── ... (75 NMR-STAR files total) │ └── restraint_violation_analysis.py │ ├── predictions # Predicted structures (105 systems × 4 predictors) │ └── <pdb_id>/ │ ├── AF3/ │ │ ├── model_0.cif │ │ ├── model_1.cif │ │ └── model_2.cif │ ├── AF2MM/ │ │ ├── model_1.pdb │ │ ├── model_2.pdb │ │ └── model_3.pdb │ ├── Boltz2/ │ │ ├── model_0.cif │ │ ├── model_1.cif │ │ └── model_2.cif │ └── Chai1/ │ ├── model_0.cif │ ├── model_1.cif │ └── model_2.cif │ └── hadzi_dssp # Hadži–DSSP analysis outputs ├── hotspots/ # Interface hotspot annotations (105 systems) │ ├── 2KGB_hotspots.json │ ├── 6E4H_hotspots.json │ └── ... (105 JSON files total) └── analysis/ # Per-residue Hadži–DSSP profiles (105 × 4 predictors) ├── 2KGB_AF3.json ├── 2KGB_AF2MM.json └── ... (420 JSON files total) Dataset Statistics Property Value Total fuzzy complexes 105 Systems with NOE restraints (group_b) 75 Systems without NOE restraints (group_a) 30 Predictors evaluated 4 (AF3, AF2MM, Boltz-2, Chai-1) Models per system per predictor 3 Total predicted structures 1,257 Total hotspot JSON files 105 Total analysis JSON files 420 IDP length range 8–172 residues (median 45) Interface size range 2–51 residues (median 10) Hotspot count range 0–9 (median 4) DockQ summary (mean ± SD across 105 systems, 3 models each): Predictor Version Mean DockQ Median NOE violation rate AF3 AlphaFold3 (2024) 0.381 ± 0.216 0.387 30% AF2MM AlphaFold2-Multimer v3 0.364 ± 0.202 0.376 30% Boltz-2 Boltz-2 (2025) 0.313 ± 0.190 0.310 30% Chai-1 Chai-1 (2024) 0.264 ± 0.188 0.242 32% All predictions used sequence-only input with no structural templates. Predictor versions: AlphaFold3 (Abramson et al. 2024, accessed via AF3 server), AlphaFold2-Multimer v3 (Evans et al. 2021), Boltz-2 (Passaro et al. 2025, v0.4.2), Chai-1 (Chai Discovery 2024, v0.5.0). Curation Criteria Systems were selected to meet all of the following criteria: FuzDB inclusion — the complex is catalogued in FuzDB v4.0, which requires experimental evidence of bound-state disorder. Disorder-to-order transitions are excluded by FuzDB's curation criteria. PDB structure available — a deposited structure exists in the RCSB Protein Data Bank. IDP binds primarily via α-helical motifs — required for compatibility with the Hadži helix–coil thermodynamic framework. A subset of 75 systems additionally satisfies: BMRB NOE restraints available — a corresponding BMRB entry with distance restraint data in NMR-STAR format, manually verified to contain usable NOE restraints. Systems meeting criteria 1–3 but lacking BMRB data (30 systems) are included in systems_metadata.csv and the predictions/hotspots/analysis directories but are not represented in the restraint files. File Formats systems_metadata.csv Column Description pdb_id 4-character RCSB PDB identifier bmrb_id BMRB entry number (restraint source; blank if unavailable) str_filename Filename of the NMR-STAR restraint file has_noe Whether the NOE restraint file is included (True/False) source Dataset tier: group_a (no NOE) or group_b (with NOE) fuzdb_entry FuzDB entry URL pdb_url RCSB PDB entry URL bmrb_url BMRB entry URL (where applicable) BMRB accession numbers for all 75 group_b systems are listed in systems_metadata.csv. Restraint files follow the naming convention merged_<bmrb_id>_<pdb_id>.str. Prediction structures Predicted structures follow the naming convention <pdb_id>/<predictor>/<model_id>.<ext>: AF3 — model_0.cif, model_1.cif, model_2.cif (AlphaFold3, top 3 by confidence) AF2MM — model_1.pdb, model_2.pdb, model_3.pdb (AlphaFold2-Multimer v3) Boltz2 — model_0.cif, model_1.cif, model_2.cif (Boltz-2) Chai1 — model_0.cif, model_1.cif, model_2.cif (Chai-1) All predictions used sequence-only input with no structural templates. Hotspot JSON files (hotspots/<pdb_id>_hotspots.json) One file per system. Fields: Field Description hotspot_positions Zero-indexed positions of hotspot residues in the IDP sequence hotspot_residues List of hotspot residues with resnum, resname, aa, index, esri, nesri, neec interface_residues All interface residues within 7Å of partner chain (same fields) n_interface_residues Total interface residue count n_hotspots Number of hotspot residues idp_sequence One-letter IDP amino acid sequence Hotspots are defined as interface residues with normalized spatial residue interaction score (NESRI > 1.0) and normalized energy contribution (NEEC > 1.0), following the PPCheck protocol (Sukhwal & Bhardwaj, 2015). Analysis JSON files (analysis/<pdb_id>_<predictor>.json) One file per system per predictor (420 total). Fields: Field Description system_id PDB accession predictor One of: AF2MM, AF3, Boltz2, Chai1 idp_sequence One-letter IDP amino acid sequence hotspot_positions Zero-indexed hotspot positions (from hotspot JSON) delta_g_int Interaction free energy per hotspot in kcal/mol (default −1.2) scale_factor w_i pre-scaling factor (2.4× applied internally) n_residues IDP length n_models Number of predicted models used hadzi_pw Per-residue Hadži helical probability p(w_i), length = n_residues dssp_avg Per-residue DSSP helicity averaged across models, length = n_residues plddt_avg Per-residue mean pLDDT across models (null if not extracted) metrics Summary: pearson_r, pearson_p, spearman_rho, spearman_p, rmse, mae, hadzi_fractional_helicity, dssp_fractional_helicity, helicity_bias, n_residues Chain and Residue Mapping The restraint violation analysis script automatically detects chain IDs and residue numbering offsets between NMR-STAR files and predicted structures using residue-name matching. Of the 75 systems, 74 were handled fully automatically. One system (2MLZ) required a manual chain assignment override due to non-standard numbering in the deposited restraint file; the override is documented in the script's STRUCTURE_REMAP configuration block. Usage Running the NOE violation analysis # Score a single predicted structure against its BMRB restraints python restraint_violation_analysis.py \ --pdb predictions/6E4H/AF3/model_0.cif \ --bmrb restraints/NOE_restraints_files/merged_xxxxx_6e4h.str \ --output results/6E4H_AF3_violations.csv Loading analysis JSONs import json # Load hotspot annotations for a system with open("hotspots/6E4H_hotspots.json") as f: hotspots = json.load(f) print(hotspots["n_hotspots"], hotspots["hotspot_positions"]) # Load Hadži–DSSP analysis for AF3 predictions with open("analysis/6E4H_AF3.json") as f: data = json.load(f) print(data["metrics"]["pearson_r"], data["metrics"]["helicity_bias"]) License This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. You are free to share and adapt the material for any purpose, provided appropriate credit is given. Citation If you use this dataset, please cite: Do AI Structure Predictors Capture Bound-State Disorder? A Benchmark on Fuzzy Protein Complexes Juan Velasquez, Sebastien Ghent, Vladimir N. Uversky, Taseef Rahman bioRxiv 2026.05.30.729023; doi: https://doi.org/10.64898/2026.05.30.729023 Contact Taseef Rahman — University of South FloridaBellini College of AI, Cybersecurity, and [e-mail]