AI antibody validation:
from prediction to evidence.

What a model prediction does—and does not—answer

Sequence and structure models help researchers explore molecular geometry and prioritize candidates. A high-confidence predicted fold does not, by itself, establish target affinity, specificity, expression yield, solution stability, or manufacturability. Binding predictions are influenced by target state, conformation, glycosylation, assay format and physical conditions.

For AI-first teams, the practical challenge is not to reject modeling; it is to use models to choose the smallest experiment set that can falsify the most important assumptions quickly.

A staged validation funnel

1. Define the desired molecule and target form. Record sequence format (VHH, IgG, scFv, Fab or multispecific), antigen construct, expected epitope, desired application and minimum decision-relevant binding behavior.
2. Express and assess recoverability. Check soluble expression, purification recovery and major purity features with appropriate analytical methods. No visible product or severe heterogeneity may be more important than a predicted affinity score.
3. Establish binding with controls. Use BLI or SPR when suitable, with concentration series, reference subtraction and appropriate negative and positive controls. Consider an orthogonal format when immobilization, avidity or nonspecific interactions may bias an apparent binding result.
4. Check homogeneity and stability. Assess monomer content and soluble aggregates by SEC, then add thermal or short stress readouts such as nanoDSF when sufficient material exists.
5. Review the candidate as a whole. Integrate quality, binding evidence and sample supply into a go / refine / stop recommendation. Do not use an arbitrary universal Kd cutoff.

Why testing “Kd only” can mislead

An impressive apparent equilibrium dissociation constant can coexist with low expression, sample heterogeneity or fast aggregation. Apparent binding can also be driven by multivalent avidity, assay-surface effects or a mischaracterized antigen. Kinetic fits require inspection of raw sensograms, residuals, reference controls and fitting assumptions—not simply a single reported number.

For very tight binding or slow kinetics, study design, contact times and assay sensitivity matter. A negative measurement is not automatically proof of no binding when target presentation or sample quality is unsuitable.

Suggested minimum decision package

• Sequence-format and antigen-construct definition; a documented experimental hypothesis.
• Expression and purified material yield, with practical sample-handling notes.
• Purity or fragment profile plus SEC monomer/aggregate assessment where feasible.
• Concentration-dependent binding evidence, including reference and control materials.
• Thermal or short stress stability screen when appropriate; clear assay limitations.
• One concise candidate comparison that separates measurements, interpretation and unresolved risk.

The optimal panel depends on the construct and study objective. Some designs need potency or cell-binding assays; others require cross-reactivity, colloidal stability or a more detailed developability review.

Where Seq2Study fits

Seq2Study is an independent, pre-launch research-use service concept focused on coordinating fit-for-purpose experimental evidence through external laboratory partners and presenting the results as a preclinical candidate-decision package. Exact assays, availability, cost, timeline, material requirements and international logistics are subject to feasibility and written scope.

Review the assay library, see the illustrative decision report, or read the developability study-design guide.