Designed by AI: the future of antibody drugs

Artificial intelligence (AI) is set to transform the way antibody drugs are designed — and not just by improving or refining existing processes, but designing antibodies from the ground up. 

The hope is that AI can bring significant benefits for patients. It could design cancer drugs that don’t carry harsh side effects, create drugs for conditions that were previously thought ‘undruggable’ and cut years off the pharmaceutical pipeline, meaning patients can access drugs significantly faster. 

Monoclonal antibodies (mAbs) are a powerful class of drug, able to selectively bind to drug targets. An estimated 20% of all new drug approvals now fall into the category, including well-known blockbuster drugs such as adalimumab to treat autoimmune diseases and trastuzumab for breast cancer^​1​. 

However, it has not been possible to create antibodies for all targets using current methods, which are laborious and do not always yield results. 

Debbie Law, chief scientific officer of Xaira Therapeutics, a start-up founded in California in 2023, is convinced we’re at a turning point for drug design. She says AI is going to “fundamentally change” how we design new antibody drugs.

Designing drugs from scratch

Xaira is one of a few start-ups with a platform capable of de novo antibody design. This is a ‘from scratch’ design process that can discover an antibody that not only binds to a biological protein target, but crucially also has the right physical and chemical properties to be developed as a drug.

Traditionally, antibody drugs are designed in one of two ways. The first — the hybridoma method — uses the immune system of an animal, usually a mouse, to produce antibodies that are then engineered to be better tolerated by the human immune system. The second generates billions of different synthetic antibodies: laboratory-created proteins that mimic the function of natural antibodies. These are filtered down through multiple high-throughput tests to find those that bind to the target, which are developed further to improve their drug-like properties. 

“At the end of it, you typically only get a few antibodies that satisfy those criteria,” says Surge Biswas, co-founder and chief executive at AI antibody design start-up Nabla Bio. “And you don’t really have a lot of control over where the antibody is binding on the target — you’re just throwing darts at a dart board and hoping you get a bull’s eye.”

The figure below shows a timeline of the progress being made by the pharmaceutical industry using generative AI^​2–7​.

Designed to reduce side effects

Companies such as LabGenius, Nabla and Xaira are also using AI to optimise their offering. London-based LabGenius is using generative AI, which creates new content based on patterns from existing data, to create complex multi-specific antibodies for cancer therapies, with fewer side effects. Multi-specifics are antibodies engineered to bind to two or more different targets. 

Such antibodies are typically designed in sequence: first, finding the parts that bind to the targets and concentrating on linking them together to form a molecule that has the other properties needed to make a drug. 

“The challenge with that approach is by optimising for one thing, you can then find yourself de-optimising for another,” says James Field, chief executive of LabGenius. Its platform can rapidly co-optimise for all desired properties through a combination of automated high-throughput experimentation and AI. Field explains that this opens up the design space to a myriad of different variants.

LabGenius has been using its platform to design T-cell engager antibodies that target solid tumours and mark them for T-cells to destroy. The problem with such immunotherapies is their inability to target and kill cancer cells selectively, without causing off-target toxicity to healthy cells. Even with targeted antibodies, off-target effects can cause serious side effects, including neurological issues (e.g. confusion and seizures), infections, and cytopenias (e.g. low blood cell counts). 

The AI platform suggests a variety of molecules within a broad design space that are experimentally tested, creating an iterative closed loop cycle of generative AI and automated lab experiments until they have the best antibody candidate to take forward — typically after four cycles, with each cycle taking six weeks. 

“We go all the way from having a computational design to having a molecule that’s been fully characterised across functional and developability assays,” explains Field.

LabGenius has demonstrated that its platform can produce highly selective T-cell engagers and expect to have an investigational new drug (IND) filed in 2026. Field says designing molecules as complex as these T-cell engagers “would be impossible [without machine learning]… these are such rare molecules that you would never have found them unless you deployed these methods”. 

South Korean AI pioneer Galux is also using its generative AI engine to create cancer therapeutics that minimise off-target effects. The company has successfully generated an antibody targeting the epidermal growth factor receptor protein, found on the surface of some cancer cells. They used a de novo approach to design the molecule to bind only to mutated estimated glomerular filtration rate — or EGFR — which is found exclusively on cancer cells and would therefore minimise off-target effects in cancer immunotherapies. The mutated protein is different by only one amino acid, which illustrates the level of selectivity their AI designs can achieve.

Drugging the undruggable 

The other high hope for AI designed antibodies is to tackle currently ‘hard to drug’ targets. These include cross-membrane proteins such as G protein-coupled receptors (GPCRs) and ion channels, which are involved in cell-signalling and make up about 60% of drug targets^​8​. They are not soluble, which makes it difficult to use high-throughput in vitro testing to find antibodies able to target them. AI design can bypass this stage and even go one step further, creating antibodies with properties that go beyond our own immune systems.

“We will be able to make molecules that, for example, recognise very small ‘real estate’ on a protein,” says Law. 

Absci has demonstrated in its collaboration with researchers at California Institute of Technology, by designing an antibody that targets the “caldera” region of the HIV virus, which is conserved across HIV strains, so it could lead to a multi-variant vaccine. This is an area where traditional approaches have previously failed. “It was in a very deep crevice and the natural immune system couldn’t generate the antibodies that could bind to this particular region,” says Sean McClain, founder and chief executive of Absci. 

The company eventually achieved this using AI: preliminary screening shows binding to multiple HIV sub-types via the region previously considered out of reach^​9​. 

Nabla has so far successfully designed antibodies against eight targets, including the first binders of any kind to engage with two cancer-linked membrane-bound targets, claudin-4 (CLDN4) and CXCR7^​10​. 

“You can generate your antibody to bind a specific epitope, or even a specific side chain, or eventually down to a specific atom on the target that you want to bind,” says Biswas. 

Speeding up the pharmaceutical pipeline

Other companies are focusing on how generative AI can improve the drug-like properties of some current antibody drugs. 

Founded in 2018, Generate:Biomedicines has the longest pipeline of all the generative AI antibody design companies. The Massachusetts-based organisation has three ongoing phase I trials, two of which are investigating an antibody, GB0895, to treat severe asthma and COPD. It has based its design on the existing antibody drug, tezepelumab, which is administered monthly. “We have generated a monoclonal antibody with such high affinity and half-life extension that we can actually give this potentially every six months,” says Dinesh De Alwis, the company’s head of clinical drug development.

It’s not just finding an antibody that binds to a target that will speed up discovery, says McClain. “What these [AI] models are great at is being able to look at this whole entire space, and then hone in on the sequences that are going to give us the drug-like attributes that we ultimately want.” That means they can design for developability, low immunogenicity, high stability and ease of manufacture, in a fraction of the time — so patients can ultimately access these drugs sooner.

The company currently has its first drug in a clinical trial. ABS-101 treats inflammatory bowel disease by targeting the immune-regulating receptor TL1A. McClain says getting this far in only two years, rather than the industry average of five and a half, shows the boost generative AI can provide. 

The real benefits to both speed of development and the type of targets that can be tackled will come with de novo design — a novel antibody drug in one shot, without the need for multiple rounds of optimisation. In 2023, this became controversial when Absci was criticised for its claim to have designed an antibody targeting (HER2) de novo. However, similarities to the existing breast cancer antibody drug trastuzumab suggested they had not started from scratch. “The dominant thinking about what de novo means is you should be able to just take the target and then, without using information from any known antibodies, can you design a new antibody from scratch?” asks Biswas. “Otherwise, it almost feels like cheating.”

McClain says Absci is proud of its work, which included a completely novel design of one of the most highly variable regions of the antibody. In the following two and a half years, the company has progressed to “being able to de novo design antibodies from scratch with challenging targets, where there are no known binders”, he says, and these antibodies will still need some optimisation to make sure they have fully achieved the desired target product profile. 

Although its de novo capabilities are improving, Xaira Therapeutics is also using what it calls co-evolution or the ‘lab in the loop’ method. This combines machine learning design with experimental data generation, which feeds more data into the model. “As they evolve and get better, you will need to do less and less iteration,” says Law. 

Nabla is also not yet able to just push a button and have an antibody ready for clinical trials, according to Biswas. They still need one or two additional rounds of optimisation. “But the hit rate of antibodies that satisfy all of those key therapeutic criteria is just way higher,” he adds. The latest tests show that out of 100 antibodies the model designs, 1 to 10 will hit the target, which is “orders of magnitude higher than conventional approaches”. 

Chaok Seok, chief executive of Galux, says the company is more bullish about its approach — the company can already produce accurate de novo antibodies in one shot and has successfully generated antibodies for six different targets this way^​11​. 

Designing the future

The promise of one-shot de novo design may well still be in the future for the complex antibodies that LabGenius designs, Field anticipates. “If you even try to model the structures of these complex multi-specifics, often, you get something that looks like meatballs and spaghetti,” he jokes. 

Biswas agrees it’s not a trivial challenge, but says Nabla has already tackled multi-specifics.

Some companies also have their eyes on the longer-term goal of using generative AI to model biology itself, to better predict how an antibody will translate in the clinic. Xaira has huge ambitions in what it calls ‘biology AI’, explains Law. In June 2025, it released a pre-print with single cell RNA sequencing data from 8 million cells of different types with different genes turned on or off, to start to better understand the networks that regulate cells^​12​. This is the first step to creating a virtual cell to hunt for new targets as well as test antibody design. “These are not new ideas — it’s just that now we’re getting to the stage that the implementation is in sight,” says Law.

In the shorter term, could generative AI methods start to replace the current mature antibody design technologies developed over the past 30 to 40 years? Seok thinks that, in some cases, this could happen within as little as five years. She sees a lot of enthusiasm in the field. “This industry is very early stage, but also very fast moving with a lot of tech giants investing a lot,” says Seok.

While no AI-designed antibody has yet reached regulatory approval, these pioneering companies are convinced their technologies can provide drugs with fewer side effects as well as new treatments — all at an accelerated pace. 

Seok continues: “Personally, I can’t wait to see where we can go and the difference we can make to patients through AI. It’s an exciting time to be in this area, in part because things that we almost dreamed about are becoming reality.”