Chamaileon

Chamaileon is a generative model for protein binder design that drops an assumption almost every predecessor makes: that there is one target, in one structural state. Real goals often violate it: a binder meant to modulate an allosteric switch has to stay compatible with two conformations of the same protein, and a cross-reactive binder has to grip similar epitopes on several related proteins. Sequences optimized against one static context tend to fail for its neighbors.

Developed by Hengyuan Cao, Min Zhang, and colleagues at Zhejiang University, Chamaileon reframes the task as cross-context binding landscape modeling: one binder sequence is optimized to fold and dock acceptably across a supplied set of contexts. In-Context Complex Co-Design (I3CD) is the training paradigm, corrupting only the binder while the target structure is held clean as conditioning context, with decoupled noise schedules for sequence and structure. Mixture-of-Paths Sampling (MoPS) is the inference strategy, alternating denoising between contexts around a shared sequence and running a beam search over candidate trajectories — which sidesteps the scarcity of high-quality multi-conformational paired data rather than requiring more of it. The work was accepted as an ICML 2026 Spotlight.

The paper also contributes CROSS, a benchmark for this setting, since existing binder benchmarks score one target at a time. Chamaileon sits beside single-context binder generators such as RFdiffusion, BoltzGen, and PXDesign, and complements conformational-ensemble predictors like AlphaFlow by designing against ensembles rather than producing them.

#Key Features

• Cross-context objective: One binder sequence is optimized jointly against multiple target contexts — several conformational states, or several related targets — instead of one fixed structure.
• In-Context Complex Co-Design: Training corrupts only binder sequence and structure, on decoupled schedules, while the concatenated target signal stays clean, making the target an explicit context for co-design.
• Mixture-of-Paths Sampling: Inference alternates between co-designing one context and forward-folding the other under the shared sequence, with a beam search scoring parallel trajectories by interface error, binder pLDDT, and self-consistent RMSD.
• CROSS benchmark: A 100-entry benchmark of paired conformational states curated from CoDNaS, split into 70 high-divergence and 30 low-divergence pairs, distributed with the code repository.
• Open code and weights: Both the implementation and the released checkpoint are Apache-2.0 licensed.

#Technical Details

Chamaileon is a multimodal flow-matching model built on the MultiFlow codebase, combining continuous flows over backbone frames with discrete flow matching over amino-acid identities so that sequence and structure are generated together. Training used 60,692 protein complex pairs from the PDB, filtered to resolution ≤5 Å, combined chain lengths of 16–512 residues, inter-chain contacts under 8 Å, and AlphaFold-Multimer quality thresholds (interface predicted aligned error ≤10, binder pLDDT ≥80, ipTM ≥0.5), for 88 epochs on eight 80 GB NVIDIA A100 GPUs with AdamW and dynamic batching.

CROSS was assembled from CoDNaS clusters at ≥95% sequence identity by taking each cluster's maximally divergent conformation pair, reducing 1,867 candidates to 100 scored entries. Quantitative results are reported for the multi-state setting: interface predicted aligned error of 4.32, binder pLDDT of 90.6, and self-consistent RMSD of 1.96 Å on the first conformation, and 4.23 / 88.7 / 2.07 Å on the second, with seven cases yielding a binder that succeeds against both conformations at once (thresholds ipAE ≤10, binder pLDDT ≥70, scRMSD ≤5 Å). Two constructed baselines succeed on both in zero cases: an RFdiffusion plus ProteinMPNN pipeline averaging per-position residue distributions across contexts, and a BindCraft-style hallucination loop with gradient updates alternating between contexts. Evaluation runs AlphaFold2-Multimer via ColabDesign, with ProteinMPNN for self-consistency and Foldseek for uniqueness.

#Applications

The intended users are protein engineers who need a binder that works across a landscape rather than a snapshot: allosteric modulators that must engage a target in more than one state, broad-spectrum neutralizing agents against a family of drifting antigens, and reagents for flexible targets where a single crystal structure is a poor design template. Because contexts are supplied at inference, one checkpoint covers both the multi-state and multi-target framings without retraining.

#Impact

Chamaileon's contribution is conceptual as much as architectural: it unifies multi-target and multi-state binder design under a single cross-context formulation and supplies a benchmark that makes the setting measurable, where prior evaluations scored one context at a time. The Apache-2.0 release of code and weights makes the method unusually reusable for an academic paper. The evidence is entirely in silico: success is defined by AlphaFold2-Multimer confidence and self-consistency metrics, no experimental characterization of designs is reported, and dual-context success stays rare in absolute terms across the 100 CROSS entries. The authors note that CROSS treats conformational variation as discrete states rather than continuous ensembles, and that scarce multi-conformational data remains the binding constraint on training rather than on sampling.