Six years ago, predicting an antibody’s binding site from scratch was considered nearly impossible. Now, at least half a dozen biotech companies are racing to file regulatory applications for drugs designed entirely by AI systems. The first could enter human trials this year.
The Contenders
Generate:Biomedicines has the longest pipeline. Their lead candidate, GB0895, targets the same protein as tezepelumab (an existing asthma drug) but with potentially six-month dosing instead of monthly injections. They have three Phase I trials running.
Absci already has a drug in clinical development. ABS-101, designed for inflammatory bowel disease, reached clinical stage in two years - less than half the industry average of 5.5 years. The company also designed an HIV antibody targeting a deep “caldera” region on the virus that traditional methods couldn’t reach.
LabGenius, based in London, is building multi-specific antibodies that can engage T-cells against solid tumors while minimizing side effects. They expect to file their first investigational new drug application this year.
Nabla Bio successfully designed antibodies against eight different targets, including two cancer-linked membrane proteins (claudin-4 and CXCR7) that were previously difficult to target. Their hit rate runs 1-10 successful candidates per 100 designs.
Galux, a South Korean company, claims single-shot accuracy across six targets. Most notably, they designed an antibody that binds only to mutated EGFR - distinguishing it from normal EGFR by a single amino acid difference. This kind of selectivity could mean cancer drugs with far fewer side effects.
What “AI-Designed” Actually Means
The industry is still debating definitions. Some companies consider it AI-designed if the sequence comes from a model, even if scientists optimize it afterward. Others insist the antibody must be clinic-ready with no wet-lab tinkering.
The strictest definition would require what’s called de novo design - creating a binding antibody from pure computation, without starting from any existing template. Some companies, including Xaira Therapeutics (founded in 2023 with backing from Eli Lilly), are pursuing this approach.
Most practical workflows still combine AI with experimental feedback. LabGenius typically runs four optimization cycles of about six weeks each. Nabla Bio expects one to two rounds of refinement. The AI gets you close; the lab confirms it works.
What This Could Change
Antibody drugs currently take an average of 5.5 years to reach clinical trials. AI-assisted development is compressing that to two years or less. Beyond speed, the technology is opening targets that traditional methods struggled with.
G protein-coupled receptors (GPCRs) and ion channels - protein families that include many important drug targets - have been notoriously difficult to target with antibodies. Several AI companies are specifically going after these “undruggable” proteins.
The HIV work is particularly notable. Absci designed antibodies targeting a structural feature that sits deep within the virus, in a region most antibodies can’t reach. Preliminary tests suggest binding across multiple HIV subtypes.
The Fine Print
No AI-designed antibody has received regulatory approval yet. The claims about superior design speed and target engagement are promising but unproven in late-stage trials.
The industry-wide attrition rate in drug development remains around 90%. Whether AI-designed drugs perform better or worse than traditionally developed candidates is the central question these early trials will answer.
Nvidia and Eli Lilly recently announced a $1 billion partnership to build what they call a “continuous learning system” connecting computational and wet-lab research. Bets this large suggest the pharmaceutical industry is taking the technology seriously, even before definitive clinical proof.
Within the next year, we should see the first regulatory filings for antibodies where AI did most of the design work. After that, Phase I results will tell us whether the hype is warranted.