Does an antibody that binds well make a good development candidate?

When reading an announcement that AI has designed an antibody, the first question is what counted as success. Was it strong binding to the target? The ability to choose good candidates from a large pool? Or the prospect of advancing a drug development programme?

An antibody-design benchmark discussed by Cradle on October 12 provides a useful way to distinguish these questions. The original paper appeared in August and evaluated technologies as they stood in the first quarter of 2025. It is not a contest among new models launched this week, but it reveals strengths and weaknesses that are hard to see from companies' own evaluations alone. Cradle's discussion ↗

The researchers evaluated 511 antibody sequences submitted by 29 organisations. They produced the antibodies under common conditions and measured their performance without knowing which organisation had proposed them. The tasks were separated into improving the binding affinity of an existing antibody, selecting strong binders from an existing sequence set, and designing new sequences for the binding regions.

The results are difficult to reduce to a single overall ranking. Methods that performed well at optimisation or design were not consistently strong at selection. In the task of selecting high-affinity binders from closely related sequence groups, all but one method failed to outperform random selection. The target was a SARS-CoV-2 protein for which extensive experimental data were available, so this finding should not be generalised to all antibody AI. Nevertheless, the ability to produce good candidates and the ability to identify them need to be evaluated separately. Original paper ↗

Another important question is what it means to have passed a developability assessment.

Cradle's best candidate showed high affinity, with a dissociation constant of 8.69 picomolar. Yet an assay of its tendency to react with material other than the target raised a concern, and its hydrophobicity assessment also warranted attention. It nevertheless passed the study's aggregate scoring threshold across five assessments. The company disclosed these individual results and explained that the candidate had inherited properties of its starting sequence. Cradle's disclosure of the results ↗

This does not mean that describing it as a pass was wrong. Meeting a combined criterion across several properties is different from having no concerns on any individual property.

Conversely, an unfavourable result in one assay does not by itself mean that an antibody cannot become a drug. A comparison of clinical-stage antibodies that informed this assessment found variation in individual properties, including among approved drugs. Assay results inform development risk; they do not alone determine a candidate's future. Study of clinical-stage antibodies ↗

Developers need to know which weaknesses they can accept for their intended product and which they must address. Strong binding does not move a programme forward if subsequent optimisation damages another essential property. Conversely, a candidate that is not the strongest binder may be easier to develop if it consistently meets the required conditions.

In our article on the business of research AI ↗, we considered how experimental work following candidate generation could earn revenue. Here, we want to go one step further and ask what should count as success in those experiments.

For a company adopting AI, a useful comparison is not simply how many sequences a model proposed. How many candidates had to be made, how many measurements were needed, and how much time and money were spent before obtaining a candidate that met the company's non-negotiable requirements? Does that performance carry over to other targets?

The study did not measure cost-effectiveness. It does, however, show specifically which evaluations are still needed to translate strong model performance into development value.