Chamaileon models protein binders across targets and states

Most protein binder design tools ship with a quiet assumption baked in, one target and one conformation at a time. A paper landing this week on arXiv and accepted as a Spotlight at ICML 2026, called Chamaileon, argues that assumption is what has kept generative binder design from touching the parts of drug discovery that actually matter.

The authors, led by Hengyuan Cao, reformulate the problem as cross-context binding landscape modeling. Instead of training a separate model per target or per conformation, one model conditions on multiple contexts at once and has to produce a sequence that works across all of them. Two ideas do the heavy lifting. In-Context Complex Co-Design, which the paper calls I3CD, is the training recipe, a context-aware sequence-structure co-modeling paradigm. Mixture-of-Paths Sampling, or MoPS, is the inference trick, optimizing a single sequence across contexts as a way to sidestep the very thin pool of high-quality multi-conformational paired training data.

Why this matters if you are not in structural biology: real therapeutic targets flex, shift, and often come in families of related proteins. A binder that only works on one snapshot of one target is a research artifact, not a drug candidate. If the cross-context framing holds up, teams could compress their model zoos, reason about selectivity and off-targets in the same generation pass, and start treating conformational flexibility as part of the design objective rather than a nuisance to be handled downstream.

The honest caveat is that the arXiv abstract does not publish the specific benchmark numbers, baseline comparisons, or any wet-lab validation. The team introduces its own evaluation benchmark, CROSS, but the paper page in this form does not disclose how large the margin over prior work is, or whether the designs actually bind in a test tube. Take the framing as a strong methodological claim that still needs external replication and lab work to settle.

The direction, though, is the part worth watching. Generative protein design is quietly moving from single-shot demos to systems that try to reason about the mess biology actually presents, and cross-context modeling is a natural next step from the single-target diffusion work that has dominated the last two years.