The antibody that won an AI contest and stuck to the laboratory column

Xencor’s antibody had an equilibrium dissociation constant (KD) of 2.9 picomolar, winning an AIntibody design challenge. It also failed to wash off a column testing unwanted hydrophobic interactions. Organisers judged that flaw likely to disqualify it from therapeutic development.[1]

The August paper reports 511 sequence submissions from 29 organisations, targeting the SARS-CoV-2 receptor-binding domain. Its authors describe an early-2025 snapshot of the field.[1]

A protein has to work away from its target

An antibody has surfaces that recognise a target and surfaces that encounter everything else. A promising sequence must fold into a usable protein, remain stable and avoid problematic interactions with itself or other material. Researchers call this collection of properties developability. It affects whether a molecule can be manufactured, stored and formulated successfully.[2]

Different tests ask different physical questions. Heating reveals resistance to unfolding. Self-interaction assays probe a tendency to associate with other antibody molecules. Hydrophobic interaction chromatography examines how the protein interacts with a column under specified salt conditions. A strong result in one test does not answer the others.[2]

The problem resembles designing a fastening that must hold firmly at one intended joint yet move freely through the assembly line. Grip is essential, but indiscriminate grip creates another engineering problem. For a protein, changing the sequence can alter several interactions at once; the useful design is the whole molecule.

What the winner actually passed

AIntibody scored five developability assays from zero to two and accepted totals of three or less. One failed test could coexist with a passing total, allowing the 2.9-picomolar design to win despite its column failure.[1]

Its next-best developable competitor measured 8.69 picomolar, statistically tied with the experimental comparator at 9.2 picomolar. Lower values indicate tighter equilibrium binding; they do not measure whether a treatment works in people.[1]

The distinction is between two kinds of decision. A ranking asks which candidate performs best on a chosen scale. A manufacturing decision may include a condition that cannot be traded away. If a process cannot accommodate the material, improvement elsewhere does not make that particular process feasible.

A warning flag still needs interpretation

There is a reason to avoid turning every laboratory flag into an automatic rejection. A 2017 study examined 137 antibody variable regions in a common IgG1 format, including 48 from approved medicines. Some showed unfavourable properties despite successful development. The study treated such measurements as warnings, not proof that a molecule could never become a drug.[2]

That creates a more interesting target for AI drug design than maximising one number: find a molecule whose full set of properties can be accommodated by a viable development route. Some weaknesses may be repairable. Others may make the proposed route impractical. Distinguishing them requires knowing what the assay is measuring and what the process needs.

Organisers plan to disqualify severe biophysical failures in future AIntibody challenges.[1]

Report an error or suggest a correction