AI Framework Designs Novel Antibodies Against Gastric Cancer Target CLDN18.2

Gastric cancer kills more people than most cancers, and the options for treating it remain thin. In China alone, nearly four in ten of the world's new gastric cancer cases appear each year. Even with chemotherapy, survival is poor. Researchers have been hunting for better targets—proteins on cancer cells that drugs can lock onto—and one has emerged with real clinical weight: a protein called CLDN18.2, which sits on the surface of stomach cancer cells and barely exists anywhere else in the body.

Zolbetuximab, an antibody drug that targets CLDN18.2, proved itself in a major trial. When added to standard chemotherapy, it stretched median overall survival from 15.5 months to 18.2 months in patients with advanced gastric cancer. That's real. It's also a proof of concept: CLDN18.2 works as a target. But making better antibodies against it has been a bottleneck. The conventional way of engineering antibodies is slow, labor-intensive, and hard to optimize for multiple properties at once—you want binding strength, but also stability, low toxicity, and low risk of triggering an immune response. Balancing all of that simultaneously, using traditional methods, is like threading a needle in the dark.

A team of researchers built an artificial intelligence system called cdrGPT to solve this problem. The core idea is elegant: antibodies have a highly variable region called CDRH3—a loop in the antibody's heavy chain that often does much of the work of recognizing and grabbing the target. Instead of engineering this loop by hand, they trained a deep-learning model to generate new sequences for it. The model learned from a database of 637,887 real antibody sequences, absorbing the statistical patterns of how nature builds these molecules. Then they fine-tuned it using reinforcement learning, steering it toward sequences that would bind well to CLDN18.2 while also having good developability—meaning they wouldn't clump up, wouldn't be cleared too fast from the bloodstream, and wouldn't trigger unwanted immune responses.

Starting with 50,000 generated sequences, they filtered them through a multi-step screen. They checked predicted binding affinity using a computational model called AlphaBind. They measured net charge and hydrophobicity to predict how the antibody would behave in solution. They predicted how likely the sequence was to be recognized as foreign by the human immune system. After all that filtering, 313 sequences survived. These weren't random—they clustered into three distinct groups based on their molecular features, with one cluster containing sequences very similar to zolbetuximab itself. The researchers then used AlphaFold3, a structure-prediction tool, to model how seven of these candidates would actually fold and bind to CLDN18.2. The predicted structures held together. The binding interfaces looked plausible. Molecular dynamics simulations—essentially watching the molecules move for 100 nanoseconds—showed the complexes remained stable.

But here's the hard truth: this is all computational. No one has yet made these antibodies in a lab, tested them against actual cancer cells, or put them in an animal. The researchers are careful about this. They note that AlphaBind was used only as a relative ranking tool, not as a measure of true binding strength. They acknowledge that their immunogenicity assessment, based on predicting MHC class II binding, doesn't capture everything that makes an antibody look foreign to the immune system. They also flag a critical gap: they couldn't confirm from their computational models that these new antibodies would specifically recognize CLDN18.2 and not its close cousin CLDN18.1, which differs in just one region. That distinction matters clinically. Cross-reactivity with CLDN18.1 could cause off-target toxicity.

What cdrGPT does offer is a tractable pipeline. It can generate diverse antibody sequences within a fixed scaffold—the zolbetuximab framework—and prioritize them based on multiple criteria at once. The model ran on four NVIDIA A100 GPUs for training and generated 20,000 candidate sequences in about four hours on a single GPU. That's fast enough to be useful. The researchers also tested the framework on other targets, like HER2, to show it could generalize. When they compared cdrGPT to two other antibody language models, AbLang and AntiBERTy, it showed trade-offs: better at predicting certain properties like hydrophobicity, weaker at predicting immunogenicity risk. No single model dominates.

The next step is the bench. The 313 candidates, or at least the most promising of them, need to be synthesized, expressed, and tested for real binding to CLDN18.2. They need to be tested against CLDN18.1 to confirm selectivity. They need to be tested in cells and in animals. Only then will anyone know whether the AI's predictions hold up in the messy reality of biology. That work hasn't happened yet. What exists now is a computational strategy—a way to narrow a vast search space down to a manageable set of candidates worth testing. For a disease where better drugs are desperately needed, that's a meaningful step forward, even if it's not the finish line.