A discovery funnel can become efficient at removing antibodies before it becomes good at identifying the ones a program needs. Strong affinity and clean early assay results make a candidate easy to advance, but they may say little about its ability to internalize, assemble receptors, or function in the intended therapeutic format. Once a different epitope or engagement geometry has been discarded, later engineering may have fewer useful options.

The challenge is deciding which differences deserve to survive long enough to be tested. Sequence diversity, functional diversity, and developability each contribute to that decision, but they answer different questions. A strong selection strategy narrows the panel as the evidence becomes more relevant to the intended mechanism, while keeping enough alternatives to challenge the current ranking.

Preserve diversity that could change the outcome

Complementary B-cell populations can contribute different germline usage, CDR3 diversity, and functional binders. That breadth creates opportunities at the start of discovery, but its practical value depends on what passes through the selection funnel. A large panel can still be functionally narrow if most of its members occupy the same epitope class or behave similarly in the assays that matter.

The shortlist should preserve differences that could change therapeutic activity. An internalizing modality needs room for alternative uptake and trafficking behavior. A receptor agonist needs candidates that support different engagement geometries. Retaining several near-identical binders may offer less useful optionality than carrying a smaller set of representatives across those functional classes.

Epitope information can help teams make that distinction earlier. Project-specific competition models trained on a subset of measured pairs offer a way to prioritize experimental work without completing every possible pairwise comparison. Their value lies in helping preserve meaningful alternatives while the panel is still broad.

The important competition relationships still need experimental confirmation before they determine a format choice. A predicted bin can guide the next experiment, but uncertainty about whether two antibodies compete should remain visible when that relationship is central to the proposed mechanism.

Bring forward the assay that could change the ranking

An early functional assay earns its place when it separates candidates that affinity measurements cannot. For an antibody-drug conjugate program, the relevant differences may include internalization, receptor fate, and the intracellular response to the payload. A candidate that binds tightly but takes an unproductive route may be less useful than one with modest affinity and more favorable trafficking.

Screens that couple antibody-mediated uptake to a downstream delivery readout can bring some of that discrimination into discovery. They allow selection to reflect transport behavior as well as target recognition. Even then, uptake or activity in a screening system does not establish that the final conjugate will deliver the required intracellular effect.

Representative format testing is especially valuable when it can overturn the ranking from unconjugated antibodies. If conjugation changes uptake, trafficking, or payload response, that information belongs in the selection decision while alternatives remain available. Otherwise, the program risks optimizing a parent antibody whose apparent advantage does not survive conversion.

Agonist programs face a similar problem through a different mechanism. Antibodies that bring receptors together depend on architecture, orientation, valency, and domain size, so affinity alone can be a poor predictor of activity. Removing a binder before testing relevant configurations may eliminate a useful receptor-assembly solution.

The decisive experiment should therefore be defined before the shortlist becomes fixed. Receptor blockade, internalization, and agonism require different evidence, even when the initial discovery platform is the same. That distinction should determine which candidates advance and which functional alternatives remain worth testing.

Make developability comparisons interpretable

Developability results are most useful when the reference molecules, format, and assay conditions support a fair comparison. Profiling single-domain antibodies in a common Fc fusion architecture, for example, can reveal differences across nonspecific binding, self-association, thermal stability, hydrophobicity, and aggregation propensity. Holding the architecture constant makes the comparison easier to interpret.

It also defines the boundary of the conclusion. A domain assessed as an Fc fusion may behave differently as a free domain or within a multispecific molecule. Favorable behavior in the reference format is useful evidence, but the intended architecture still needs its own assessment.

Nonspecific binding deserves particular attention because it can flag exposure risk. Associations between higher nonspecific binding and faster clearance in mouse studies support early screening, without establishing the pharmacokinetics of every format carrying the same domain. The result should guide investigation rather than substitute for the relevant exposure data.

Individual readouts should remain visible as candidates are ranked. Strong thermal stability does not automatically offset substantial nonspecific binding, and a self-interaction signal needs to be considered against the intended concentration and formulation requirements. Combining everything too quickly into one score can conceal the liability that engineering actually needs to address.

Keep predictions tied to what they predict

Computational assessment can expand the number of candidates a team evaluates and help focus experimental resources. Its usefulness depends on keeping the predicted endpoint explicit. A model that predicts hydrophobic interaction chromatography retention is predicting behavior in that assay, even when surface hydrophobic patches and aromatic exposure provide a plausible explanation.

That result can support triage or suggest mutations, but it does not directly measure every downstream consequence associated with hydrophobicity. The same distinction applies when sequence-based models reproduce structural descriptors used to assess viscosity risk. Improving a descriptor is encouraging only to the extent that the change improves the experimental property relevant to the molecule.

This is particularly important when predictions become design objectives. Optimization can move a proxy in the desired direction without delivering the intended benefit. Teams need experimental checks that connect the model output to the property they are trying to improve, while monitoring whether the redesign compromises function or creates another liability.

The training data need similar scrutiny. Mutations that change folding or stability can produce apparent binding effects that resemble changes at the interface. Without controls for protein quality, a model may learn a relationship that the assay does not actually establish. That ambiguity becomes consequential when the prediction determines which variants are never made.

Retain candidates for a specific next decision

A useful selection review distinguishes candidates ready to advance from those worth targeted engineering and those retained to resolve an open mechanism or format question. The latter groups need explicit experiments attached. Keeping an antibody because it is different is less persuasive than explaining what its difference could reveal or enable.

A distinctive functional profile may justify fixing a localized developability problem. A near-duplicate with several unfavorable properties may offer little reason for further investment. Those judgments become clearer when the team can state what would change the ranking and how the next experiment will answer it.

Product requirements should anchor that discussion. Compare activity in a representative format, examine the individual developability findings, and identify where predictions remain uncertain. An experiment that distinguishes two competing mechanisms may be more valuable than producing another variant close to the current leader.

The aim is a funnel that becomes narrower because the evidence becomes more decisive. Preserving relevant biological and engineering diversity gives teams a meaningful choice; timely functional testing and interpretable developability data give them a reason to commit. Together, those practices improve the chance that the selected candidate carries the biology the program needs into a format it can develop.

Continue the conversation at PEGS AI

Carry the conversations from PEGS Europe into PEGS AI in January. Join the PEGS community to explore how AI and experimental evidence can work together to guide antibody discovery, candidate selection, and developability engineering.

Register for PEGS AI