Making an antibody bind its target more strongly can also make it stick to things it should ignore. AbOptimizer, described in Nature Communications on 26 October, tackles that trade-off by choosing sequence edits under several objectives. The researchers then tested selected designs in the laboratory across three antibody systems.[1]
The method first decides which positions to preserve or change. A protein language model proposes plausible amino-acid substitutions, with insertions and deletions available in tasks that permit them. Predictive models score the resulting variants, and reinforcement learning adjusts the editing policy according to those scores.[1]
Three different tests of the editing process
Selected variants tested in vitro
| Measure | Value | Context |
|---|---|---|
| Emibetuzumab | Four optimised antibodies improved target binding and specificity versus the wild-type benchmark | |
| Siltuximab | Four variants improved target binding with reduced nonspecific binding; thermal effects varied | |
| Trastuzumab | Three insertion/deletion-containing variants improved HER2 binding |
Emibetuzumab tested the balance between binding its intended HGFR target and reducing nonspecific attachment. For two variants derived directly from the wild type, ELISA EC50 fell from 1.33 nM to 0.92 and 0.77 nM. Lower EC50 means less antibody was needed to reach half the maximum signal in that assay; it is a measure of apparent binding potency, not a clinical dose.[1]
Siltuximab tested a harder data situation: optimisation without antibody-specific mutation data, using a generic fitness predictor. The strongest tested variant reduced IL-6-binding EC50 from 0.22 to 0.016 nM, a reported 13.4-fold improvement. Nonspecific binding was assessed separately with ovalbumin and a cell-based assay.[1]
Its thermal results show why several properties need checking. One siltuximab variant’s melting temperature rose by about 4°C, another was comparable to the original, and two fell by about 2°C and 0.5°C. All retained high thermal stability in the authors’ assessment, but the individual directions differed. A binding gain did not guarantee an improvement in every measurement.[1]
Trastuzumab broadened the edit space to include insertions and deletions while optimising HER2 binding. The three systems therefore tested different challenges: competing properties, scarce task-specific data and changes to sequence length.[1]
Predictors guide the search; assays test selected claims
The multi-objective reward uses a Pareto approach, favouring candidates that improve the balance of properties instead of collapsing the search into one predetermined weighted score. This is useful when a gain in one property can conceal a loss in another. It also leaves the search dependent on the quality of its surrogate predictors.[1]
The laboratory assays provide evidence for the selected variants that were made and measured. Computational predictions for properties such as immunogenicity or clearance remain a different kind of evidence. The demonstrated contribution is a way to narrow the design search while keeping competing molecular requirements visible; safety, manufacturing performance and clinical efficacy require their own tests.[1]
Sources & further reading
- Multiproperty optimization of therapeutic antibodies through sequence editing via large language model and reinforcement learning Nature Communications · 26 Oct 2026 · Accessed 26 Oct 2026