Chamaileon: Cross-Context Binder Design with Contextualized Modeling and Mixed Sampling

Chamaileon: Cross-Context Binder Design with Contextualized Modeling and Mixed Sampling

Summary

This paper introduces Chamaileon, a unified framework for cross-context protein binder design, addressing the limitations of existing approaches that predominantly operate under a single-target, single-state assumption. The authors define cross-context binder design as a problem where one sequence must satisfy multiple binding constraints with explicit trade-offs, unifying two related challenges: multi-state (MS) binder design, which requires a binder to remain compatible with the same epitope across different conformational states of a target protein, and multi-target (MT) binder design, which requires a single binder to engage similar epitopes on distinct proteins with high binding affinity.

The paper identifies a fundamental gap in current architectures: while state-of-the-art generative models excel at single-structure objectives, they lack (i) decoupled noise schedules for sequence and structure to maintain sequence consistency across conformational states during joint target-binder modeling; (ii) inference-time optimization to navigate trade-offs between conflicting structural constraints; and (iii) comprehensive evaluation that measures success across a contextual ensemble rather than on isolated snapshots.

The framework is underpinned by two key contributions:

1. In-Context Complex Co-Design (I3CD): This is a novel training paradigm that reformulates binder generation as an in-context denoising process. Inspired by in-context generation in computer vision, I3CD concatenates clean target signals with noisy binder signals and feeds them into a model to learn joint dependencies. Crucially, the binder's sequence and structure are corrupted using decoupled noise schedules (t and t̃), enabling the model to capture the intricate sequence-structure interplay. The training objective is a hybrid discrete and continuous flow matching loss, formulated as:

E[Σ (x̂1(Bt,t̃T1,1, MH; θ) - x12 / (1-t)) - log pθ(a1Bt,t̃, T1,1, MH) + (log r1(r̂1(Bt,t̃T1,1, MH; θ)) - log r1(r1)2 / (1-t))]

2. Mixture-of-Paths Sampling (MoPS): This is a scalable inference-time sampling strategy that addresses the scarcity of multi-conformational paired data and the limited flexibility of architectures in handling variable conformations. MoPS discretizes the generative trajectory into N intervals and alternates between co-design (joint sequence-structure denoising) for one conformation and sequence-conditioned generation (forward folding) for the other conformation. This iterative process ensures the shared sequence is optimized against the structural constraints of all conformations. The authors state: "The primary advantage of this strategy is that it ensures the shared sequence is iteratively optimized against the respective structural constraints of both conformations, guaranteeing that the final designed sequence is compatible with the entire conformational ensemble."

MoPS also incorporates beam search to enhance sample quality. To enable stochasticity in structure generation, the authors extend the flow matching process from an ODE to an SDE framework, introducing a diffusion term while correcting the drift term. The beam search evaluates candidates using a composite score based on inter-chain predicted Aligned Error (ipAE), binder pLDDT, and binder self-consistent RMSD (scRMSD).

Data and Benchmark: The authors constructed a training dataset of 60,692 chain pairs from the PDB, filtered by structural quality, chain length, physical interaction, and AlphaFold2-Multimer predicted metrics (ipAE ≤ 10 Å, binder pLDDT ≥ 80, ipTM ≥ 0.5). They also introduced CROSS (Comprehensive Recognition Of Specific Surfaces), a benchmark for cross-context binder design, curated from CoDNaS clusters. The final benchmark comprises 100 entries (70 high-divergence, 30 low-divergence) selected using a composite quality score.

Experimental Results: The paper reports both qualitative and quantitative results. Qualitative results demonstrate Chamaileon's ability to design binders that bind effectively across different contexts, encompassing both MS-type and MT-type scenarios. The authors identify three structural adaptation modes: Micro Adaption, Dual-face Adaption, and Macro-switch Adaption.

Quantitative results from ablation studies show that the full version of Chamaileon achieves an ipAE of 4.32, binder pLDDT of 90.6, scRMSD of 1.96, and 8 unique successes for conformation 0; and ipAE of 4.23, binder pLDDT of 88.7, scRMSD of 2.07, and 10 unique successes for conformation 1, with 7 both success cases. The w/o MoPS baseline (sequential approach) shows significant bias in unique success numbers across conformations, demonstrating MoPS's effectiveness in synthesizing information from multiple conformations.

Comparisons with two constructed baselines (RFDiffusion + ProteinMPNN and BindCraft alternating optimization) show that neither baseline yielded any cross-context binder that simultaneously satisfies both contexts, while Chamaileon achieves non-trivial success rates.

The paper also validates I3CD on single-state binder design tasks, achieving superior novelty compared to APM (0.546 vs 0.615 TM-score) but lower success rate, attributed to the smaller model size (21.8M vs 199.6M parameters).

Additional experiments demonstrate the framework's scalability to arbitrary numbers of conformations, its support for active negative design (explicitly avoiding unwanted targets), and its robustness across all 1,867 benchmark candidates.

The authors conclude that Chamaileon provides a practical solution for designing proteins with complex functional requirements, and suggest future extensions to continuous conformational landscapes and integration of more refined biophysical priors.

Improvements to AI systems:

1. Multi-Constraint Sequence-Structure Co-Design: Implement decoupled noise schedules for sequence and structure during joint denoising, enabling AI to generate a single sequence that satisfies multiple structural contexts simultaneously (e.g., binding to different protein conformations or homologs) without retraining.

2. Inference-Time Multi-Path Optimization: Add a Mixture-of-Paths Sampling (MoPS) module that alternates between joint co-design and sequence-conditioned folding across discrete trajectory intervals, with beam search scoring on inter-chain alignment error, confidence, and self-consistency. This allows AI to iteratively refine outputs against conflicting constraints at inference, even when training data lacks paired multi-context examples.

3. Contextual Ensemble Evaluation: Replace single-snapshot success metrics with ensemble-based scoring (e.g., success across all conformations in a set, unique successes per context, and both success counts). This trains AI to optimize for worst-case or aggregate performance rather than isolated best-case, improving robustness in real-world deployment.

4. Hybrid Flow-Matching with SDE Extension: Extend deterministic flow matching to stochastic differential equations (SDEs) with corrected drift, enabling controlled exploration during generation. This improves sample diversity and avoids mode collapse when multiple valid solutions exist under conflicting constraints.

5. Active Negative Design: Incorporate explicit avoidance of unwanted targets (e.g., off-target proteins) into the training objective, allowing AI to generate binders with selective specificity—useful for therapeutic safety and reducing cross-reactivity.

6. Scalable Multi-State Generalization: Enable the framework to handle arbitrary numbers of conformations or targets (beyond pairs) by discretizing the generative trajectory and alternating across all contexts, making the AI system adaptable to complex multi-protein environments (e.g., signaling complexes or viral escape variants).

What the improved AI system can do:

• Design a single protein binder that remains functional across multiple conformational states of a target (e.g., open/closed enzyme states) or across homologous proteins with shared epitopes, with explicit trade-off control.

• Generate binders that pass ensemble-level validation (e.g., predicted alignment error < 5 Å and self-consistency RMSD < 2 Å across all contexts) rather than just one snapshot.

• Automatically balance conflicting structural constraints at inference time, producing candidates that are Pareto-optimal across contexts without requiring paired training data.

• Avoid off-target binding by actively penalizing interactions with unwanted proteins during generation, enabling safer therapeutic designs.

• Scale to multi-target or multi-state scenarios (e.g., 5+ conformations) with linear computational overhead, useful for designing broadly neutralizing antibodies or pan-variant inhibitors.

• Provide interpretable adaptation modes (micro, dual-face, macro-switch) that reveal how the AI resolves structural conflicts, aiding human understanding of design rationale.