The AI That Learns From Three Experiments How to Find the Next Blockbuster Antibody
In 2023, the world spent over $200 billion on antibody drugs — treatments like Keytruda for cancer and Humira for autoimmune disease. But finding a single antibody that binds well enough to become a drug is like searching for a specific grain of sand on a beach. A typical drug discovery campaign screens thousands of candidates, each one requiring expensive lab tests to measure how tightly it grabs its target. Most fail.
A team at Henlius, a Shanghai-based biotech company, has built a new AI system that changes how these candidates are ranked. Instead of judging each antibody in isolation, their model called AbICL learns from a handful of known comparisons and uses those reference points to sort through the rest. Think of it this way: if you know that antibody A binds better than B, and B binds better than C, those data points help you figure out where D fits without testing it from scratch. The paper was published on arXiv on July 7.
Background & Context
Antibody drugs work by grabbing onto a specific target, usually a protein on cancer cells, a virus, or an overactive immune signal, and either blocking it or flagging it for destruction. The tighter the grip (called binding affinity), the better the drug tends to work.
For decades, the standard approach was to predict affinity as a single number, like “this antibody binds at 5 nanomolar.” But those numbers are noisy. They vary depending on which lab ran the test, what temperature the experiment used, even which batch of reagents was in stock. A more reliable approach, called affinity ranking, simply asks: “Out of these two antibodies, which one binds better?” No numbers, just order.
Existing ranking systems treat each comparison as an independent puzzle. They build a single scoring function during training and apply it identically to every new antibody, regardless of the target. The problem is that binding is not universal. An antibody that grips the COVID spike protein beautifully might completely ignore a cancer marker. Each target has its own biochemical personality.
What the Researchers Did
The Henlius team reframed the problem as in-context learning, or ICL: the same mechanism that lets large language models adapt to new tasks after seeing just a few examples. Their model, AbICL, takes a handful of known antibody comparisons for a specific target and uses them as demonstrations to calibrate its ranking for that same target. No retraining needed.
The architecture has two pieces. First, a structural encoder turns each antibody-antigen pair into a numerical representation, capturing the shape and chemical features of the interface where the two molecules meet. Second, a Context Ranking Head looks at the known demonstrations alongside the new pair being evaluated and produces a ranking score that accounts for the local affinity landscape.
The training method matters as much as the architecture. AbICL uses episodic meta-training — it learns not from individual comparisons but from entire mini-scenarios. Each training episode mimics what happens at test time: the model sees a small support set of labeled comparisons, reasons over them, and then ranks new query pairs from the same antigen. Over thousands of episodes spanning different antibodies and targets, the model learns how to extract useful ranking patterns from context rather than memorizing specific antibodies.
What They Found
The team tested AbICL on the AbRank benchmark, which includes three increasingly difficult data splits and two evaluation benchmarks. The results were striking, especially in the hardest cases.
| Setting | WALLE-Affinity | AbICL (No-context) | AbICL (Test-context) |
|---|---|---|---|
| Balanced / Unrelated Complex | 0.880 | 0.849 | 0.888 |
| Balanced / Local Perturbation | 0.668 | 0.647 | 0.872 |
| Hard Ab / Unrelated Complex | 0.838 | 0.760 | 0.904 |
| Hard Ab / Local Perturbation | 0.588 | 0.490 | 0.749 |
| Hard Ag / Unrelated Complex | 0.711 | 0.700 | 0.787 |
| Hard Ag / Local Perturbation | 0.581 | 0.572 | 0.778 |
AUROC scores: higher is better. Test-context uses a small number of measured reference comparisons at inference time.
The most dramatic jump happened on the Local Perturbation benchmark, which tests whether the model can tell apart nearly identical antibody variants — the kind of subtle discrimination that determines whether a drug candidate moves forward or gets shelved. With just one or two reference comparisons per target, AbICL jumped from 0.647 to 0.872 AUROC, a gain of 22.5 percentage points. For context, the previous best method (WALLE-Affinity) scored 0.668 on the same task.
The model also shined under distribution shift. When tested on antibodies and antigens it had never seen during training (the Hard Ab and Hard Ag splits), contextual demonstrations delivered larger relative gains than on easier splits. The harder the generalization task, the more the model benefited from having a few reference points.
One counterintuitive finding: the demonstrations do not need to come from the exact same antigen to be useful. Providing comparisons from a biologically related target still improved ranking, just not as much as same-antigen examples. This means the model picks up on transferable patterns: certain structural motifs that tend to mean better binding, regardless of the specific protein.
Why It Matters
The economics of antibody discovery are brutal. Screening a million candidates to find one drug costs tens of millions of dollars and takes years. Any technology that reduces the number of wet-lab experiments needed directly shortens timelines and cuts costs.
AbICL’s approach is pragmatic. It does not require retraining for each new drug target, a process that would itself require generating expensive training data. Instead, it asks for a tiny number of experimental measurements (sometimes just one or two) and bootstraps from there. A lab could run a handful of binding assays on Monday, feed those results to AbICL, and have a ranked shortlist of top candidates by Tuesday morning.
The ICL design also means the model can improve as more data arrives, without any code changes. Run five experiments, rank the next batch. Run ten more experiments, refine the ranking again. Each new measurement becomes another demonstration in the context window, making subsequent predictions sharper.
How It Could Change Our Lives
Imagine a future pandemic. A new virus emerges, and within weeks researchers have identified its spike protein structure. Pharma companies generate thousands of candidate antibodies and need to narrow them to the top 50 for urgent testing. Instead of running thousands of binding assays — each taking days and costing thousands of dollars — they run 20, feed those to AbICL, and get a ranked list in hours. The difference between six months and six weeks to identify a lead candidate could translate to lives saved.
Closer to home, this matters for diseases that do not attract blockbuster drug investment. Rare cancers, neglected tropical diseases, antibiotic-resistant infections. These areas struggle to attract the funding needed for brute-force screening. A method that squeezes more information out of fewer experiments makes it economically viable to search for antibodies against targets that would otherwise be ignored.
The Bigger Picture
AbICL sits at the intersection of two accelerating trends: the maturation of protein structure prediction (led by AlphaFold and its successors) and the spread of in-context learning beyond language models into scientific domains. The structural encoder already benefits from the revolution in predicting how proteins fold and interact. As those structural models improve, the raw material fed into AbICL becomes richer, and the few-shot ranking should improve in tandem.
The paper also hints at something broader: in-context learning may be a natural fit for scientific problems where “universal” models fail. Biology is full of context-dependent phenomena — a mutation that is deadly in one genetic background is harmless in another, a drug that works in one patient fails in the next. Frameworks that adapt to local evidence rather than imposing global rules could become the default in computational biology.
Limitations & What’s Next
The study only addresses ranking — ordering antibodies from best to worst. It does not predict actual binding strength in physical units, which matters for dosing decisions. And the researchers found that the rules governing effective context construction are not yet fully understood. Sometimes demonstrations from less similar antigens actually worked better than same-antigen examples, a result the authors describe as unexpected and worth deeper study.
The model was tested on computational benchmarks, not in a real drug discovery campaign. The acid test will be whether AbICL’s rankings hold up when the top candidates go into actual lab experiments — and whether the savings in screening costs materialize outside of a benchmark dataset.
📄 Zhiyuan Chen, Jing Hu, Junzhe Wang, Yueyang Huang, Xinyi Yang, Zhaoyang Wang, Feng Zhu. “AbICL: In-Context Learning for Antigen-Specific Antibody Affinity Ranking.” arXiv:2607.05846, July 2026. All authors affiliated with Henlius, Shanghai.