PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Hengyuan Cao, Shizhuo Cheng, Mingxuan Liu, Weicheng Huang, Yunhong Lu, Chenxi Cai, Yan Zhang, Min Zhang

arXiv:2607.23518v1cs.LGq-bio.BM

TL;DR

Existing binder-design methods largely assume one target and one conformational state, whereas many applications require compatibility across multiple contexts. Chamaileon unifies these settings through I3CD training and MoPS inference, and evaluation on CROSS demonstrates adaptable binders across diverse conformational and target landscapes.

  • Problem

    Existing binder-design approaches largely assume a single target and state, limiting joint design for binders compatible with multiple conformations or targets.

  • Method

    Chamaileon formulates multi-state and multi-target design as cross-context modeling, using I3CD for co-design and MoPS to optimize one sequence across structural contexts.

  • Results

    Evaluation on the CROSS benchmark demonstrates that Chamaileon generates sequences adaptable to diverse conformational landscapes and multi-target requirements.

  • Takeaways & Limitations

    Cross-context binder design provides a unified scope for designing one sequence under multiple binding constraints across conformational states and targets.

  • Takeaways & Limitations

    Explicit design of binders jointly compatible with substantially different functional states remains limited in current approaches.

Abstract

from arXiv · show

The rapid evolution of generative models has unlocked new potentials in protein binder design, a pivotal task in structural biology, by facilitating end-to-end generation via joint sequence-structure modeling or hallucination. However, existing approaches are predominantly implemented under a single-target, single-state assumption, limiting their ability to model multi-target or multi-state interactions required for advanced function-oriented protein design. Here, we introduce Chamaileon, which unifies multi-target and multi-state binder design by formulating the problem as cross-context binding landscape modeling. The framework is underpinned by a training paradigm termed In-Context Complex Co-Design (I3CD) for context-aware sequence-structure co-modeling. During inference, we employ Mixture-of-Paths Sampling (MoPS), a scalable strategy that optimizes a single sequence across contexts while alleviating the scarcity of high-quality multi-conformational paired data. Extensive evaluation on our newly constructed benchmark, CROSS, demonstrates that Chamaileon effectively generates sequences adaptable to diverse conformational landscapes and multi-target requirements. The code is available on https://github.com/caohengyuan/Chamaileon.

1. Introduction

Existing binder-design methods largely optimize against one target structure and objective, limiting function-oriented designs that must accommodate multiple conformational states or engage multiple targets. Chamaileon unifies these settings as cross-context binder design and introduces context-aware co-design, inference-time optimization, and dedicated evaluation.

  • Motivation: Most existing methods treat binder design as a one-to-one problem involving one target structure and one binding objective.This simplification mismatches function-oriented goals in biology and therapy.
  • Motivation: Multi-state design requires one binder to remain compatible with state-dependent interface geometries across substantially different functional conformations.Current approaches may tolerate modest local flexibility but remain limited for discrete, substantially different states.
  • Motivation: Multi-target design seeks a single binder satisfying multiple binding or selectivity requirements across distinct pathway components.Examples include binding proteins A and B while avoiding protein C and others.
  • Unified formulation: Cross-Context Binder Design unifies multi-state and multi-target design by requiring one sequence to satisfy multiple binding constraints with explicit trade-offs.Contexts are designated interfaces from different conformational states or distinct targets, and success requires comprehensive satisfaction across them.
  • Contributions: Chamaileon introduces I3CD for decoupled sequence-structure denoising, MoPS for iterative single-sequence optimization across structural trajectories, and CROSS for evaluation.MoPS balances energetic requirements across contexts during inference, while CROSS is curated for multi-state and multi-target binder design.

2. Related Work

Prior work has explored multi-state protein design through trajectory mixing, ensemble training, and sampling-time optimization, while validation increasingly emphasizes structural self-consistency and dynamics. However, existing optimization protocols remain limited to single rigid targets or structural objectives.

  • Multi-state design: ProteinGenerator averages sequence logits from parallel trajectories constrained by distinct structural priors to design fold-switching proteins, but this heuristic can struggle to balance energetic trade-offs.The method represents an inference-time trajectory-mixing approach for multi-state compatibility.
  • Validation: Multi-state designs are increasingly validated through state-specific refoldability, co-folding of distinct allosteric geometries, and dynamics-oriented measures of conformational occupancy.Examples use AlphaFold Initial Guess, Chai-1 co-folding, likelihood-based bias scores, and molecular-dynamics simulation.
  • Sampling-time optimization: Recent sampling methods integrate search, reward steering, or auxiliary potentials, but existing protocols remain restricted to single rigid targets or structural objectives.Examples include Monte Carlo tree search with Feynman-Kac steering and auxiliary potentials guiding sequence diffusion.

3. Preliminary

This section introduces discrete flow models as CTMC-based flow matching on categorical state spaces and describes multimodal flow matching for protein structure–sequence co-design. It also presents independent time schedules for structural and sequence dynamics and interleaved multi-conformation sampling via forward folding.

  • Discrete Flow Models: Discrete flow models extend flow matching to discrete states by governing probability dynamics with a time-dependent CTMC rate matrix.Training uses conditional flow matching to make the rate-matrix formulation tractable.
  • Discrete Flow Models: The marginal rate matrix is parameterized as a posterior expectation of a conditional rate matrix, with categorical denoising predictions trained by cross-entropy.The conditional rate matrix is defined using the Kronecker delta, masking state, and time-dependent transition factor.
  • Discrete Flow Models: During inference, discrete sequence trajectories are generated iteratively by simulating the CTMC with Euler integration and the learned rate matrix.This procedure follows the denoising network’s learned categorical transition dynamics.
  • Multimodal Flows for Protein Co-Design: A protein structure–sequence pair combines Euclidean translations, Riemannian rotations, and discrete amino-acid types, motivating multimodal flow matching.The representation includes a mask token among the discrete amino-acid states.
  • Multimodal Flows for Protein Co-Design: Structure and sequence dynamics are decoupled through independent time schedules to support forward- and inverse-folding processes.The two schedules are denoted t and ˜t.
  • Mixture-of-Paths Sampling: MoPS interleaves sampling trajectories for different conformations through forward folding, enabling simultaneous sequence–structure co-design across binder states.The approach uses forward folding as a bridging mechanism between conformational trajectories.

4. Chamaileon

Chamaileon frames cross-context binder design as a unified problem spanning training, sampling, and evaluation. It introduces I3CD for context-aware sequence-structure co-design, MoPS for multi-conformation generation, and CROSS for benchmark evaluation.

  • Framework overview: Chamaileon unifies training, sampling, and evaluation to address the underexplored challenge of cross-context binding.The framework injects target information into the binder denoising context and supports multi-conformation sequence-structure co-design.
  • In-Context Complex Co-Design (I3CD): I3CD formulates protein design as in-context generation over multimodal data, conditioning binder denoising on clean target signals.Independent sequence and structure noise schedules enable modeling their interplay under varying noise intensities while supporting forward and inverse folding.
  • Motivation: Multi-state design is hindered by scarce structural-ensemble data and limited flexibility for variable numbers of target conformations.Only ∼11,800 NMR ensembles represent structural diversity, covering 21% of CATH superfamilies.
  • Mixture-of-Paths Sampling (MoPS): MoPS alternates co-design and sequence-conditioned generation so multiple conformations acquire distinct structures while sharing one updated sequence.The process discretizes the generative trajectory into time intervals and alternates which conformation drives co-design.
  • Mixture-of-Paths Sampling (MoPS): Beam search enhances MoPS sample quality by retaining top-performing candidates across iterative generation steps and maintaining multiple hypotheses.Stochastic sampling is introduced through structural randomization and Euler-Maruyama discretization, enabling beam search during co-design.
  • CROSS benchmark: CROSS validates Chamaileon using CoDNaS clusters to identify multi-state candidates and selecting maximally divergent conformational pairs.The interacting target chain is identified from the highest residue contact density in the original PDB entry.

5. Experiments

Chamaileon is evaluated on cross-context binder design, where one sequence must bind multiple targets through context-specific conformations. On the CROSS benchmark, it demonstrates effective cross-context design, while ablations and constructed-baseline comparisons reveal the importance of MoPS and the challenge of satisfying both contexts simultaneously.

  • Task: The task requires generating one binder sequence that binds multiple distinct targets while adopting different conformations for each context.
  • Benchmark & Metrics: Designs are judged successful when AlphaFold-Multimer predicts ipAE ≤10, binder pLDDT ≥70, and binder scRMSD ≤5 ˚A.The evaluation also reports averages across both conformations, unique successes for each conformation, and samples successful in both conformations.
  • Results: Chamaileon designs a single binder effective across different contexts, including both MS-type and MT-type scenarios.Qualitative cross-context results are presented in Figure 5, with detailed structural analysis provided in Appendix G.
  • Ablation Studies: Ablation results show that the w/o MoPS sequential baseline develops a strong bias in unique success numbers across conformations, failing to balance both contexts simultaneously.The baseline designs conformation 0 first and then performs sequence-conditioned generation on conformation 1.
  • Hyperparameter Sensitivity: Decreasing MoPS frequency progressively reduces discrepancies between unique success rates across conformations while increasing both success, indicating that more frequent switching better integrates context information.
  • Comparison with Constructed Baselines: Neither constructed baseline produced a binder satisfying both contexts simultaneously, whereas Chamaileon achieved non-trivial success rates.Baseline 1 fuses sequence distributions from separately generated backbones, while Baseline 2 alternates gradient updates between target contexts.

6. Conclusion … A.2. Conditional Flow Matching

Chamaileon unifies cross-context protein binder design through sequence-consistent co-modeling and scalable multi-context sampling. The appendix formalizes discrete flow models, their rate-matrix dynamics, and masking-based conditional flow matching.

  • 6. Conclusion: Chamaileon unifies cross-context protein binder design beyond the traditional single-target, single-state paradigm.The framework decouples binder sequence and structure noise schedules to maintain sequence consistency across multiple conformations.
  • 6. Conclusion: MoPS addresses scarce multi-conformational data with a scalable inference-time sampling strategy that iteratively optimizes a binder across contexts.
  • 6. Conclusion: The framework can extend from discrete states to continuous conformational landscapes and incorporate refined biophysical priors for enhanced interface complementarity.
  • 6. Conclusion: Cross-context modeling is presented as a foundation for programmable modulatory effects and multi-specific therapeutics in function-oriented protein design.
  • A. Discrete Flow Models: Discrete Flow Models generalize continuous flow matching to discrete state spaces by associating vector fields with discrete dynamics.
  • A.1. Dynamics on Discrete State Spaces: For discrete states, probability evolution follows the Kolmogorov Forward Equation, the analogue of the continuous-space Fokker–Planck or continuity equation.
  • A.1. Dynamics on Discrete State Spaces: A time-dependent rate matrix Rt governs transitions among discrete states, with probability changes determined by total inflow minus total outflow.
  • A.2. Conditional Flow Matching: Conditional flow matching constructs marginal paths from per-sample conditional paths using a masking interpolant, revealing the target token with probability t.The mask-to-target transition rate is 1/(1−t), while an already unmasked target state has zero transition rate.

A.3. Marginal Rate Parameterization and Training … D. MoPS Algorithm with Beam Search

The method parameterizes marginal rates through a denoising model, samples sequences by Euler-based unmasking, and extends this process to stochastic, cyclic multi-conformation beam search. The resulting MoPS procedure supports arbitrary conformational states while preserving the trained marginal distributions.

  • A.3. Marginal Rate Parameterization and Training: The marginal rate matrix is expressed as a posterior expectation of the conditional rate and approximated with a neural network predicting clean data from noisy inputs.This converts rate learning into denoising-model learning.
  • A.3. Marginal Rate Parameterization and Training: The network is trained with cross-entropy against ground-truth data, avoiding complex simulations or exact rate matching during optimization.The objective simplifies training by supervising predicted distributions directly.
  • A.4. Sampling via Euler Integration: Sampling starts from the fully masked state x0 = M at t = 0 and uses Euler discretization to evolve the CTMC toward t = 1.Unmasked states remain unchanged, while masked states transition according to the model’s predicted category probabilities.
  • B. Scalability of MoPS: MoPS cyclically alternates among conformations, resuming each from the previous endpoint and thereby scaling co-design to an arbitrary number of conformational states.Figure 7 states that this extension incurs no additional computational overhead.
  • C. From ODE to SDE; C.3. From ODE to SDE via Fokker-Planck Equation: An SDE is constructed whose Probability Flow ODE matches the flow-matching velocity field, so individual trajectories become stochastic while marginal distributions remain consistent with the original ODE objective.This stochasticity supplies the randomness required for beam search.
  • C. From ODE to SDE; C.1. Flow Matching and the Velocity Field: The flow-matching ODE interpolates between Gaussian noise p0(x) = N(x; 0, I) and the data distribution p1(x) = pdata(x), with dynamics governed by a learned velocity field.During inference, the neural network predicts the data sample and the corresponding estimated velocity field.
  • C.2. Connection to Score Function: The marginal score is obtained as the posterior expectation of the conditional score, enabling its computation from flow-matching model outputs for stochastic sampling.The velocity field is related to conditional expectations of x1 and x0, and the resulting score can be computed from the predicted x1.
  • C.4. Discretization; D. MoPS Algorithm with Beam Search: Euler-Maruyama discretization injects controlled noise σt into sampling, and Algorithm 1 combines this stochastic update with MoPS’s conformation switching and beam-search candidate selection.The algorithm takes timesteps, beam-search splits, candidate number L, switching frequency, and the number of conformations as inputs.

E. More Data Collection Details · F. Evaluating I3CD for Single-State Binder Design

The paper details the training and benchmark data distributions used in CROSS, then evaluates I3CD on single-state binder design across 10 target proteins. I3CD produces greater novelty than APM but a lower success rate on this task.

  • E. More Data Collection Details: The training set is characterized by the distribution of summed lengths across its two chains.This distribution is shown in Figure 8(a).
  • E. More Data Collection Details: The benchmark candidate pool contains 1,867 entries, whose target–binder combined-length distribution is reported.This distribution is shown in Figure 8(b).
  • E. More Data Collection Details: Structural differences in the candidate pool and final CROSS benchmark are analyzed using RMSD distributions.Figure 9 presents these RMSD distributions.
  • E. More Data Collection Details: Table 3 specifies the single-state benchmark’s structural information, binding specifications, and other relevant parameters.The benchmark details support evaluation of the single-state binder design task.
  • F. Evaluating I3CD for Single-State Binder Design: I3CD was evaluated for single-state binder design on 10 targets: BHRF1, SC2RBD, IL-7RA, PD-L1, TrkA, IL-17A, VEGF-A, insulin, H1, and TNF-α.The targets were selected from the main results of Zambaldi et al. (2024), with specifications given in Table 3.
  • F. Evaluating I3CD for Single-State Binder Design: Across the 10 targets, I3CD achieved superior novelty but a lower success rate than APM.Table 4 reports average unique success and novelty metrics; unique success uses Foldseek clustering, while novelty uses TM-Score.

G. Binder Structure Analysis of Cross-Context Binder Design

Chamaileon generates single sequences that satisfy multi-objective cross-context binding constraints through three structural adaptation modes: Micro Adaption, Dual-face Adaption, and Macro-switch Adaption. These modes span progressively greater backbone plasticity, from localized fluctuations to divergent conformations.

  • Adaptation-mode taxonomy: Chamaileon’s cross-context binders are hierarchically classified into three structural adaptation modes that navigate the backbone plasticity required by multi-objective constraints.The modes are Micro Adaption, Dual-face Adaption, and Macro-switch Adaption.
  • Micro Adaption: Micro Adaption preserves a highly conserved scaffold, using subtle backbone fluctuations and localized adjustments to tolerate minor interface variations.The global fold remains nearly identical across contexts.
  • Dual-face Adaption: Dual-face Adaption maintains a stable, rigid backbone while using spatially distinct protein surfaces to engage different target interfaces.Multi-specific recognition is achieved by repurposing different regions of the same fold with minimal structural deformation.
  • Macro-switch Adaption: Macro-switch Adaption uses large-scale backbone rearrangements or partial fold-switching when interface geometries differ drastically.The single sequence adopts divergent conformations to optimize binding in each context.

H. Additional Quantitative Results across All Benchmark Candidates · I. Active Negative Design

Evaluation across all 1,867 benchmark candidates showed lower success than the primary benchmark, supporting quality-based filtering for CROSS. Active negative design weakened binding to unwanted contexts while preserving the framework’s two-context design capability without architectural changes.

  • H. Additional Quantitative Results across All Benchmark Candidates: 1,867 benchmark candidates were evaluated to further assess Chamaileon’s robustness across the full benchmark.The results were reported in Table 5.
  • H. Additional Quantitative Results across All Benchmark Candidates: The all-candidate evaluation achieved a lower success rate than Table 1, indicating greater difficulty on lower-quality targets.This result further supports the quality-based filtering strategy used to construct CROSS.
  • I. Active Negative Design: Active negative design modifies beam-search ranking by negating scores on the unwanted target while aggregating scores normally.This mechanism requires no architectural change.
  • I. Active Negative Design: During MoPS, sampling alternates only between the two desired targets to support active negative design.The evaluation used four cases where Chamaileon had successfully designed binders for both desired contexts, each paired with a third context.
  • I. Active Negative Design: Chamaileon can be extended to explicit negative design without architectural modifications.The conclusion follows from the score-function change and the observed weakening on the unwanted context.
  • I. Active Negative Design: The active/negative sampling strategy produced binders with weaker binding to the third, unwanted context.This finding directly evaluates the intended repulsion of the unwanted context.
  • I. Active Negative Design: ipAE, pLDDT, and scRMSD on target 2 all degraded relative to the 2-context design baseline, confirming repulsion of the unwanted conformation.The degradation occurred under active negative design compared with the baseline.

J. Ablation Study on Hyperparameters of the Score Function

The ablation study examines how weighting coefficients in MoPS’s beam-search score function affect three binding-quality metrics. Increasing emphasis on one metric generally improves it at the expense of others, while the chosen setting balances overall quality.

  • Ablation Study: The study ablates individual weighting coefficients in the MoPS beam-search score function, with results reported in Table 7.The ablations assess the effect of each coefficient on the final outcomes.
  • Ablation Study: Shifting weight toward a specific metric generally improves that metric while potentially degrading the others.The evaluated metrics are ipAE, binder pLDDT, and binder scRMSD.
  • Ablation Study: The extreme setting (1.0, 0.0, 0.0) achieves a higher both success count but noticeably weaker scRMSD and pLDDT than most alternatives.This illustrates the trade-off produced by strongly prioritizing one component of the score.
  • Ablation Study: The chosen weighting setting balances high quality across ipAE, binder pLDDT, and binder scRMSD.Rather than maximizing a single metric, the selected configuration targets balanced performance across all three.

K. Detailed Construction of the CROSS Benchmark

CROSS is constructed through multi-stage similarity balancing and quality filtering to produce a compact, reliable benchmark. It distinguishes multi-target from multi-state cases using a 95% target sequence-identity threshold and includes a proof-of-concept multi-target success while acknowledging substantial limitations.

  • Benchmark construction: CROSS starts from 1,867 candidates, applies target similarity-based category balancing and quality-based filtering, and reduces evaluation cost while retaining reliable entries.The construction pipeline is designed to produce a compact, high-quality benchmark.
  • Benchmark construction: The 95% sequence-similarity threshold is applied to binders within CoDNAS-derived clusters, treating highly similar proteins as the same binder.The threshold is not applied to targets.
  • Scenario coverage: Using a 95% target sequence-identity threshold, CROSS contains 15 multi-target entries and 85 multi-state entries.Multi-target cases involve distinct targets, whereas multi-state cases involve different conformations of the same or nearly identical target.
  • Multi-target results: One of Chamaileon’s 7 both-success outcomes is a genuine multi-target case, simultaneously binding 94%-identity targets with ipAE values of 3.32 and 3.16.The case also has pLDDT values of 93.6 and 94.1 and scRMSD values of 0.90 and 1.17, respectively.
  • Limitations: Multi-target binder design remains challenging because no existing end-to-end approach is available, while adapted pipelines have fundamental limitations.Chamaileon provides a feasible solution and a successful multi-target proof of concept, but considerable room for improvement remains.
Loading 2607.23518v1…