AI-Edited Antibody Bound IL-6 13x Tighter in Lab

An antibody rewritten by an AI system latched onto its target around 13 times more tightly than the original in laboratory tests, and held up slightly better to heat, according to a study published on Monday in Nature Communications.

The system behind it, AbOptimizer, is the work of a team at Huazhong Agricultural University in Wuhan with the company Wuhan Huamei Biotech. It is a tool for the unglamorous middle stage of drug development: not inventing an antibody from nothing, but taking one that already exists and tuning it. Every result in the paper comes from a test tube or a plate of cells. Nothing has been tested in an animal, let alone a person.

Therapeutic antibodies are large proteins that stick to a chosen target in the body, and the first version a lab finds is rarely good enough. It may bind too weakly, stick to things it should not, or fall apart when warmed. Fixing one of those faults often spoils another, which is the trade-off the researchers set out to attack.

The model marks which letters to change before it changes them

AbOptimizer treats an antibody’s sequence as something to be edited rather than redrawn. A neural network looks at the sequence and labels each position with an action: keep it, swap it, and, optionally, insert a new position or delete one. Only then does an antibody language model, AbLang, propose what the swapped positions should become. Scoring models rate the result on several properties at once, and reinforcement learning pushes the editor towards edits that improve the whole balance rather than one number.

The scoring step is where this kind of work usually stands or falls, and the authors are candid that their scorers are stand-ins. For one of the three antibodies they had no mutation data at all, so they used a general-purpose fitness score from a structure-informed language model instead.

Three antibodies, nine to eleven variants, one assay bench

The team tried it on three real medicines or near-medicines. Emibetuzumab, which targets the HGFR receptor and has reached phase II trials in non-small-cell lung cancer, binds well but sticks to too much else. Nine candidate variants were produced and tested; four showed what the paper repeatedly calls “slight” improvements in binding alongside genuinely reduced non-specific binding, with half-maximal binding concentrations falling from 1.33 nanomolar for the original to 0.92 and 0.77 nanomolar for two of them.

Siltuximab, approved for Castleman disease, gave the headline number. Four variants improved on both counts, and the best of them, Ab3-4, reached a half-maximal binding concentration of 0.016 nanomolar against interleukin-6 compared with 0.22 for the original, a 13.4-fold improvement by the authors’ own calculation. Its melting temperature rose by about 4C. That part is not uniform: two sibling variants lost thermal stability, by roughly 2C and 0.5C, though all four stayed within what the paper describes as a high range.

Trastuzumab, the HER2 antibody familiar from breast cancer treatment, was used to test insertions and deletions rather than straight swaps. Three of five variants bound HER2 more tightly than the original’s 2.61 nanomolar, the best reaching 0.82 nanomolar, a 3.2-fold gain.

The authors say their own validation is limited

The caveats are in the paper, not around it. Binding here is measured as an EC50 from an ELISA plate, an apparent measure of how much antibody is needed for half-maximal binding, rather than a full binding constant. Non-specific binding was checked against ovalbumin and hamster ovary cells. Thermal stability came from a melting-point instrument. None of that says anything about whether an edited antibody works in a living body, survives in blood, or is safe.

The discussion says so plainly: “the validation scope remains limited relative to the breadth of the computational evaluations,” and the authors call for testing across more antibody scaffolds, antigen classes and developability endpoints. They also warn that language-model fitness scores can drift away from the property you actually want, and that insertions and deletions risk larger structural disruption than swaps.

Six of the thirteen authors are listed as inventors on a patent application covering the framework, filed by Wuhan Huamei Biotech, and five are employees of that company. The compiled wrapper code is published under a CC BY-NC 4.0 licence, which permits academic use but not commercial use, so a rival drug firm cannot simply pick it up. The archived deposit on Zenodo carries the same code.

Antibody tuning is a quieter target than designing proteins from scratch

The AI-for-biology stories that travel furthest are usually about creation: a model producing a binder that never existed, as in Anthropic’s claim that Claude designed protein binders against scorpion neurotoxins, or generative systems such as Westlake’s Pinal and the open RFD4-Proteina model that build proteins to order. Optimisation is less cinematic and closer to where money is lost. Most antibody programmes fail on the dull properties, and a method that can nudge several of them at once, using a handful of edits rather than a library of thousands, is aimed squarely at the cost of the design-build-test cycle that investors keep funding, including in rounds such as Pathos AI’s $237m Series E.

What should temper the excitement is the sample size. Across three antibodies, roughly twenty variants reached the bench and about eleven improved on something. That is a reasonable hit rate for a first demonstration and far too small a base for claiming a general method, which is close to what the authors themselves say. The commercial framing matters too: a tool whose code is non-commercial, whose patent sits with one company and whose validation stops at the plate is a research result, not a product. The number worth watching is not 13.4-fold binding but whether any edited variant is ever put into an animal and reported.

The paper does not say whether Wuhan Huamei Biotech intends to take any of these variants further, and no follow-up study has been announced.