Artificial intelligence (AI) is beginning to change how new drugs are designed. One of its most promising uses is improving therapeutic antibodies, proteins that recognize specific targets in the body and are already used to treat cancer, autoimmune diseases and infections.

Finding a useful antibody, however, remains largely an experimental process. An antibody must bind strongly to its target, but it must also have the physical properties needed to become a medicine. AI could potentially make this process faster by identifying the most promising antibody sequences before they are made and tested in the laboratory.

A new study tested how well current AI methods can design antibodies. In a blinded competition called AIntibody, 29 organizations submitted 511 AI-designed or AI-selected antibodies. The antibodies were then made and tested experimentally, with the results showing that AI was most successful when it was asked to improve an antibody that already had useful properties. It was much less reliable when asked to identify the best antibody or design one from scratch.

Putting AI Predictions to the Test

The AIntibody competition tested three different stages of antibody discovery. In the first, AI models were asked to improve upon an existing antibody. In the second, they had to identify the strongest antibodies from groups of related candidates. In the third, they had to design new antibody sequences that had not appeared in the original experimental dataset. All of the antibodies were tested against the same target, a protein from SARS-CoV-2.

The teams did not receive information on how their new antibodies performed until the competition ended. The antibodies were evaluated for two key properties. The first was affinity, or how strongly an antibody binds to its target. The second was developability, which describes whether an antibody has the physical properties needed to potentially become a useful medicine. An antibody can bind extremely well and still be a poor drug candidate if it is unstable, sticks to other molecules or tends to clump together.

AI Can Improve Existing Antibodies

The clearest success came when AI was asked to improve an existing antibody. About 13% of submissions produced antibodies with at least a 20-fold improvement in binding strength while remaining developable.

One AI-designed antibody bound its target with an affinity comparable to the best antibody produced through conventional laboratory methods in the study. This suggests AI can be useful when scientists already have a promising antibody and want to improve it.

The potential advantage is speed. A successful computational approach could replace several weeks of experimental screening with a much smaller number of AI-generated candidates. Instead of testing thousands of possible combinations, scientists might be able to make and test only a handful of computationally selected antibodies.

Predicting the Best Antibody Is Harder

The results became less consistent when AI had to identify the best antibody from groups of related candidates. Most AI approaches performed no better than simply choosing antibodies at random. Only one participating group consistently beat the random-selection baseline in this challenge.

This highlights a fundamental difficulty in antibody design. There are many different ways to change an antibody’s sequence. A change that improves binding may also make the antibody unstable or cause it to interact with unintended targets. AI therefore has to optimize several properties at the same time, which is much harder than simply predicting whether an antibody looks like one that worked in the past.

Designing Antibodies from Scratch Remains Difficult

The third challenge asked AI to create antibodies outside the sequences it had been given. Here, the results were mixed. Some AI-designed antibodies bound their targets extremely strongly. About one-third of the submissions were able to bind at concentrations below one billionth of a mole per liter.

But many other designs did not bind. And among the strongest binders, developability often became a problem, showing why experimental testing remains essential. An AI model may identify a sequence that looks promising on a computer, but only laboratory experiments can determine whether the resulting antibody actually works.

The Next Generation of AI Antibodies

AI showed its strongest performance when improving existing antibodies using experimental information. Its performance was much less reliable when asked to predict the best candidate from a large group or design completely new sequences. That distinction matters as AI becomes more common in drug discovery. A model can produce an impressive prediction without necessarily producing an effective medicine.

The AIntibody challenge provides a way to measure that difference. By requiring computational predictions to be tested experimentally under the same conditions, future studies can determine which approaches genuinely improve antibody discovery.

AI may eventually reduce the enormous amount of laboratory testing required to develop new antibodies. But for now, the most powerful approach may be using AI to make those experiments smaller, faster and more targeted.