id stringclasses 10
values | protein_context stringclasses 10
values | alphafold_signals stringclasses 10
values | proposed_inference stringclasses 10
values | gold_misinterpretation stringclasses 2
values | gold_error_type stringclasses 9
values | gold_correction stringclasses 10
values |
|---|---|---|---|---|---|---|
AFMC-001 | enzyme active site loop | pLDDT:loop=42 core=92; PAE:loop-high; disorder:moderate | Loop geometry is precise so we can design a tight-binding inhibitor around it | yes | low_confidence_region_overtrust | Do not use low pLDDT loop geometry for docking. Focus on high-confidence core or validate loop experimentally |
AFMC-002 | soluble monomeric enzyme | pLDDT:overall=90; PAE:low; experimental:none | Structure is reliable for fold-level annotation | no | no_error | Use for fold annotation and cautious hypothesis generation |
AFMC-003 | multi-domain protein with hinge | pLDDT:domains=88 hinge=55; PAE:interdomain-high | Interdomain angle is fixed so domain arrangement is biologically locked | yes | interdomain_orientation_overclaim | Treat domain orientation as uncertain when PAE is high. Consider flexibility and validate with experiments |
AFMC-004 | intrinsically disordered region candidate | pLDDT:IDR=28 core=85; disorder:high | IDR forms a stable helix in vivo because AlphaFold shows a helix fragment | yes | disorder_as_structure | High disorder and low pLDDT suggest no stable fold. Use disorder tools and experiments |
AFMC-005 | membrane protein | pLDDT:TM=84 loops=40; context:membrane; experimental:none | All loop positions are accurate for antibody epitope mapping | yes | loop_position_overtrust | Use TM core as guide. Do not assume loop placement. Validate with epitope mapping experiments |
AFMC-006 | protein complex suspected | pLDDT:monomer=87; PAE:low; oligomer_context:unknown | Monomer model proves it does not dimerize | yes | complex_negation_from_monomer | Monomer prediction does not rule out oligomerization. Use AlphaFold-Multimer or experiments |
AFMC-007 | binding interface hypothesis | pLDDT:interface=60; PAE:interface-high | Interface residues define a stable binding pocket | yes | interface_overconfidence | High PAE and mid pLDDT make interface uncertain. Validate binding site with experiments or co-folding |
AFMC-008 | metal-binding enzyme | pLDDT:core=92; cofactors:none | Predicted pocket confirms metal coordination geometry | yes | cofactor_absence_misread | AlphaFold often omits cofactors. Do not infer coordination geometry without ligand or experimental data |
AFMC-009 | low complexity region | pLDDT:low_complex=25; disorder:high | Low complexity region forms a stable domain | yes | low_complexity_as_domain | Treat as disordered or context-dependent. Avoid rigid structural claims |
AFMC-010 | well-studied protein | pLDDT:88; PAE:low; experimental:known_xray_match | AlphaFold supports existing crystal structure alignment | no | no_error | Use as supportive evidence and note agreement with experiment |
AlphaFold Misinterpretation Classifier
Purpose
Detect common ways people overclaim from AlphaFold outputs.
Input fields
- protein_context
- alphafold_signals
- proposed_inference
Required output
Return one JSON object
- misinterpretation
yes or no - error_type
must match allowed list - correction
one sentence
Allowed error_type values
- no_error
- low_confidence_region_overtrust
- interdomain_orientation_overclaim
- disorder_as_structure
- loop_position_overtrust
- complex_negation_from_monomer
- interface_overconfidence
- cofactor_absence_misread
- low_complexity_as_domain
Scoring
- misinterpretation_accuracy
- error_type_accuracy
- correction_similarity
- format_pass_rate
Run scoring
python scorer.py --predictions predictions.jsonl --test_csv data/test.csv
- Downloads last month
- 12