ReferenceAI skillsAntibody and protein designReferences

Generation versus validation

Define inputs for antibody/protein sequence or structure design and distinguish design from validation.

Skill version: 1.1.0. Resource: skill://nadoobio-protein-design/references/design-validation.md.

Protein design generates candidate sequences or structures from desired conditions. Antibodies are proteins that bind targets; their multiple chains and binding-related regions require preserving chain correspondence.

Successful candidate generation does not prove expression, stability, or actual binding. Predicting a generated sequence's structure, comparing it with the design structure, and measuring expression/binding experimentally are separate validation stages.

Sequence recovery is the fraction matching a reference sequence, not the functional success rate of new designs. Before interpreting scores, check fixed regions, output counts, and sequence lengths.

Check input–output correspondence first

MaterialsWhat to checkWhat does not establish completion
Conditioning structure and positions to changeChain roles, fixed/changeable regions, and mapping between author residue numbers/insertion codes and input positionsAssuming that positions selected on screen reached the model input
Candidate sequences and mapping tablePer-candidate chains, lengths, generation order versus score order, and preservation of positions that must not changeMapping original chains from candidate names or array order alone
Candidate reevaluation resultsStructural comparisons over the same range and scores defined by that outputPresenting reevaluation calculations as actual expression, binding or functional assays

For antibodies, verify target and antibody chain roles from the input. Do not infer roles from chain identifiers A/B alone. Check each tool’s contract to determine whether chains not directly designed remain as conditioning context and whether they appear in the output.

Teaching mapping example: author positions A:10, A:10A and A:12 may have observed sequential indices 1, 2 and 3. “Change the second position” maps to A:10A, not residue 11. Do not send that sequential index directly as a setting unless the tool’s position system is verified. Do not invent coordinates for missing residue 11.

Teaching review example: if 10 positions may change and 66 must remain fixed, check returned sequences and mappings for changes outside the allowed range. If the output contains only designed chains, compare omitted fixed chains with a separate reference. Candidate generation, preservation of fixed positions and high recovery rates do not replace functional validation.

Sources: Structure-conditioned sequence design and separate experimental validation, Author residue-number definitions in structure files. Mapping and the 10/66 positions are explanatory examples, not paper performance figures. Follow the current schema for each tool’s fixed-region representation and output format.

On this page