← Back to Home

AI Can Design Antibodies That Bind Tightly. Why Does It Still Lose to Random Selection When Choosing the Best Sequences?

Blind testing of 511 antibody submissions from 29 organizations demonstrates AI’s ability to optimize antibodies while exposing weaknesses in fine-grained ranking: producing standout candidates and reliably selecting better antibodies are two different challenges.

By SURL BioNews

Getting an antibody to bind tightly to its target is an important starting point in drug development. Identifying promising candidates from large numbers of sequences in advance could reduce synthesis and testing costs. However, the international AIntibody challenge shows that AI can produce antibodies with excellent binding capabilities without necessarily distinguishing accurately between closely related sequences. This experimental benchmark puts the gap between the ability to design and the ability to select on the same scorecard.

GENEWIZ, which was responsible for preparing the antibodies, described the testing process on November 11; the full study was published in *Nature Biotechnology* on August 19. The challenge tested 511 antibody submissions from 29 organizations. Participants had to finalize their sequences before experiments, after which the antibodies were evaluated using a standardized preparation process and experimental staff who did not know the participants’ identities. The official website also emphasizes that success in individual tasks did not consistently carry over to other tasks.

The testing target was the receptor-binding domain of SARS-CoV-2. According to participating company Cradle, teams received sequencing data from yeast display screening, along with measured affinities for 142 antibodies. The three tasks required participants to improve existing antibodies, select the strongest-binding antibodies from groups of sequences, and design new antibody-binding regions outside the database, respectively. The final task still used existing data as a starting point and kept the antibody framework fixed, so it cannot be considered design from scratch without any prior information.

The results did include highlights. GENEWIZ reported that multiple teams produced antibodies with affinities below 100 picomolar (pM) that also met the developability criteria used in this challenge. Affinity here is measured by the dissociation constant: a lower value generally indicates tighter binding. Cradle reported that its best design achieved 8.69 pM and passed the relevant thresholds. However, the company also acknowledged that its model assigned very similar predicted scores to candidates within the same sequence group, making it difficult to resolve differences that were present experimentally.

This is precisely the problem exposed by the ranking task. The study asked teams to identify better candidates from three sequence groups defined by the third complementarity-determining region of the antibody heavy chain (HCDR3). In the overall comparison, only one team’s method outperformed the random-selection baseline, and its advantage did not extend across all sequence groups. This does not mean that the other models were entirely unable to identify antibodies that bind to the target. Rather, when the starting data already contained many binding sequences, the models could not reliably select the better ones.

The scoring approach also affects how the results should be understood. Cradle argued for clearer definitions of ranking metrics and improvements to developability assessment, to prevent the performance of a single best molecule from obscuring differences across an entire batch of designs. Beyond tight binding, antibodies must also be assessed for properties such as stability, aggregation, and nonspecific binding. Passing a set of early screening criteria does not establish suitability for manufacture as a drug, nor does it support conclusions about efficacy and safety in humans.

The challenge’s stronger message is that AI already has practical potential in antibody engineering tasks with abundant data and a clearly defined scope, but designing new sequences, predicting subtle differences, and accounting for multiple properties remain distinct capabilities. Because testing focused on a single antigen under fixed data conditions, the results are also insufficient to generalize to other disease targets or development workflows that repeatedly incorporate new experimental data. For teams preparing to adopt these tools, the key remains whether they can consistently outperform simple, reliable experimental benchmarks in actual tasks.

References

  1. GENEWIZ
  2. Nature Biotechnology
  3. Cradle
  4. AIntibody / Specifica