Pith. sign in

REVIEW 4 major objections 7 minor 28 references

One sequence can be designed to bind multiple protein contexts by alternating co-design and folding paths at inference time.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

I3CD plus MoPS sampling produces single binder sequences that AlphaFold-Multimer scores as compatible with multiple conformational or multi-target contexts on the CROSS benchmark.

T0 review reviewed 2026-07-30 challenge →

load-bearing objection Useful framing and a real sampling trick for multi-context binders, but the headline 7 both-successes are partly AF-Multimer self-play and should be read as a methods proof-of-concept, not a solved capability. the 4 major comments →

arxiv 2607.23518 v1 pith:SESAXA4T submitted 2026-07-26 cs.LG q-bio.BM

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 This is my paper
classification cs.LG q-bio.BM
keywords protein binder designcross-context designmulti-state designmulti-target designflow matchingin-context co-designmixture-of-paths samplingCROSS benchmark
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Most computational binder design treats the problem as one sequence against one fixed target structure. Many real needs are different: a binder must stay compatible when a target switches conformation, or must engage similar epitopes on distinct proteins. Chamaileon reframes both as cross-context design—one sequence satisfying several interface constraints at once. It trains a model that denoises binder sequence and structure against a clean target context with separate noise schedules, then at inference time mixes parallel generation paths so the shared sequence is repeatedly updated against each context. On a new benchmark built for this setting, the method produces designs that pass structure-prediction success criteria on both contexts simultaneously, where sequential single-context design and two adapted baselines do not.

Core claim

Cross-context binder design—generating one sequence that binds multiple conformational states or multiple targets—can be solved without multi-state paired training data by training in-context sequence-structure co-design on ordinary complexes and, at inference, iteratively optimizing that single sequence across alternating co-design and sequence-conditioned folding trajectories (MoPS), optionally ranked by beam search on predicted interface metrics.

What carries the argument

Mixture-of-Paths Sampling (MoPS): discretize the generative trajectory and alternate which context drives joint sequence-structure co-design while the others are updated by forward folding on the shared sequence, so one sequence is refined against every context; beam search on a weighted ipAE/pLDDT/scRMSD score prunes paths.

Load-bearing premise

Predicted interface confidence and self-consistency scores from a structure predictor are treated as a faithful stand-in for real multi-context binding affinity and function, both when ranking designs and when declaring success.

What would settle it

Wet-lab binding assays (or orthogonal physics-based affinity estimates) on the reported dual-success CROSS designs that show the sequences do not bind both intended contexts, or bind an explicitly avoided negative context as strongly as the positives.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Multi-state and multi-target binder goals can share one computational formulation and one sampling procedure instead of separate pipelines.
  • Scarcity of multi-conformational complex data need not block multi-context design if single-complex training plus inference-time path mixing is enough.
  • Active negative design (bind A and B, avoid C) is available by flipping the beam-search score on unwanted contexts without retraining.
  • The same machinery is claimed to extend to more than two contexts by cycling which conformation drives co-design.
  • CROSS supplies a concrete testbed for measuring joint success across conformational and target ensembles rather than single snapshots.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If AF-based ranking and success criteria are only loosely correlated with true affinity under large interface rearrangements, MoPS may be optimizing for predictor agreement more than for biophysics—orthogonal scoring or experimental feedback loops would be the natural next filter.
  • Macro-switch adaptation cases in the qualitative taxonomy are the hardest test of whether a single sequence truly encodes fold-switching binders versus mild backbone flexibility.
  • Therapeutic multi-specifics and allosteric modulators are the direct application class if dual-success rates can be scaled beyond the small filtered benchmark set.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. The manuscript introduces Chamaileon, a framework for "cross-context" binder design: generating a single sequence compatible with multiple binding contexts, either distinct conformational states of one target (multi-state) or distinct targets (multi-target). The method has three components: (i) I3CD, a multimodal flow-matching co-design model trained on AF-Multimer-filtered PDB dimers with clean-target/noisy-binder concatenation and decoupled sequence/structure noise schedules; (ii) MoPS, an inference-time scheme that alternates co-design and sequence-conditioned forward folding across contexts, with an SDE-based stochastic translation sampler enabling beam search ranked by a composite ipAE/pLDDT/scRMSD score (Eq. 11); and (iii) CROSS, a 100-entry benchmark built from CoDNaS clusters. On CROSS, the full method yields 7 both-context successes (single sample per entry, AF-Multimer success criteria) versus 0 for two constructed baselines (RFdiffusion+ProteinMPNN fusion; alternating BindCraft) and 2 for a sequential w/o-MoPS ablation. Ablations over beam candidate count, beam frequency, and MoPS switch frequency are reported, plus single-state I3CD evaluation (Appendix F), unfiltered-pool results (Appendix H), active negative design (Appendix I), and score-weight ablations (Appendix J).

Significance. If the results hold under orthogonal evaluation, this is a meaningful contribution: a first unified formulation of multi-state and multi-target binder design, an inference-time method (MoPS) that avoids the need for paired multi-conformation training data, and a curated benchmark (CROSS) with public code. The ablations are genuine: the w/o MoPS conformational imbalance (17 vs 2 unique successes), the beam-search and frequency sweeps, and two engineered baselines that both fail give real evidence that path-mixing, not just the base model, drives the effect. The work is reproducible in principle (code released) and the problem is timely. Impact is currently capped by evaluation that lives entirely inside one structure predictor's metric family, small absolute success counts (7/100; 69/1867 unfiltered), and thin multi-target evidence (1 case). No experimental or physics-based validation is attempted, so claims should be understood as in silico.

major comments (4)
  1. [§4.2–§5 / Table 7] The headline result is entangled with a single evaluator at three levels: (a) the I3CD training set was filtered to dimers that AlphaFold2-Multimer already scores well (ipAE≤10, pLDDT≥80, ipTM≥0.5, §4.3); (b) MoPS beam search ranks candidates at every interval by a weighted sum of ipAE/pLDDT/scRMSD (Eq. 11); (c) success is defined by thresholds on the same three metrics from the same predictor (§5). The reported 7 both-success designs are therefore partly the output of hill-climbing the judge. Table 7 corroborates the sensitivity: reweighting Eq. 11 moves both-success between 7 and 10, so the headline count is partly a knob on the reward rather than a stable property of the designs. This does not by itself invalidate the comparison (all variants share the judge), but the claim that these are genuinely cross-context-binding sequences requires transfer to an independent assessment. Please:
  2. [§5 / Table 2] The comparison in Table 2 is asymmetric in test-time compute. Baselines produce 4 sequences per entry, while Chamaileon's single reported sample is the endpoint of beam search that screens L candidates per interval × N′ intervals, each scored by AF-Multimer (L ranges to 250 in Table 1). Reporting 'best per-context results' for baselines does not equalize the screening budget, since both-success is still counted over 4 unscreened samples. Please either (i) give baselines a matched budget of AF-scored candidates selected under Eq. 11 and re-report both-success, or (ii) state the effective number of AF-Multimer evaluations per reported Chamaileon design and discuss the comparison under that lens.
  3. [§5 / Table 1] Central quantitative claims rest on small counts from one sample per CROSS entry: 7 (full) vs 5 (w/o beam) vs 2 (w/o MoPS) both-success out of 100, with Table 1's hyperparameter sweeps showing ±3 count fluctuations across nearby settings and no seeds or intervals. A 7-vs-5 difference is plausibly within sampling noise. Please run ≥3 seeds for the module ablation and report mean±std or binomial confidence intervals. The 17-vs-2 unique-success imbalance in the w/o MoPS row is the more statistically convincing evidence for MoPS and should be foregrounded over the both-success deltas.
  4. [§1, §5, Appendix K] The abstract and §1 claim unification of multi-target (MT) and multi-state (MS) design, but CROSS contains only 15/100 MT entries and, per Appendix K, only 1 of the 7 both-success cases is MT. The MT evidence is a single proof-of-concept. Please report an MS/MT breakdown in Tables 1 and 5 and temper the MT claims accordingly. Relatedly, CROSS was curated by an AF-metric composite score (Eq. 12) and is thus enriched for AF-pleasing entries; this selection bias should be stated explicitly alongside Table 5's unfiltered-pool results (69/1867).
minor comments (7)
  1. [Throughout] Typos: 'specificically' (§1), 'numbders' (§5), 'predction' (§2.3), 'Adaption' → 'Adaptation' (Appendix G), missing spaces around 'term In-Context' (abstract/§1), 'is update by' → 'is updated by' (Algorithm 1).
  2. [§2.2, §5, Appendix F] Citations: '(Zhu)' in §2.2 is incomplete (no year/venue); Cho et al. (BoltzDesign1) lacks year/venue. The success protocol is attributed to Anonymous (2025, under review); please cite established protocols (e.g., Bennett et al. 2023; Pacesa et al. 2025) as well, and justify why the main task uses ipAE≤10 while Appendix F uses ipAE≤14.
  3. [Table 1] Table 1 formatting: the 'Beam Search Frequency 250' row shows dashes for conformation 0 and a bolded 0 for both success — clarify whether this denotes divergence/failure. Bold highlighting is scattered across non-default rows (e.g., bold ipAE 2.21 at MoPS frequency 100, which has both-success 1); apply a consistent bolding criterion.
  4. [§5] §5 states unique success equals total success 'given our single-sample generation'; the Foldseek clustering description is then vacuous for Table 1. Clarify where clustering is actually used (presumably Table 5 and Appendix F).
  5. [§4.2, Eqs. (7)–(8)] Notation: Eqs. (7)–(8) write B^m_{t,t} with a single time symbol, but §4.1 introduced decoupled schedules (t, t̃). State explicitly that co-design sets t̃=t, or carry t̃ through. Also report default values of ω's, σ_t schedule, c, N, N′, L in a single table; they are currently scattered or implicit.
  6. [Figures 5, 10] Figures 5 and 10 use a blue–red spectrum for structural deviation but give no RMSD scale/units on a colorbar; please add one.
  7. [Appendix F / Table 4] Appendix F: I3CD underperforms APM on single-state unique success (0.7 vs 1.2, Table 4). The attribution to model size (21.8M vs 199.6M parameters) is plausible but untested; please either show a capacity-matched comparison or phrase the conclusion more cautiously, since it bears on how much of Chamaileon's cross-context gain comes from MoPS versus the base model.

Circularity Check

3 steps flagged

Moderate evaluation circularity: AF-Multimer ipAE/pLDDT/scRMSD filter the training set, rank MoPS beam candidates (Eq. 11), and define success—so headline both-success partly reflects optimizing the judge, not a by-construction derivation.

specific steps
  1. fitted input called prediction [§4.2 Eq. (11) and §5 success protocol]
    "S(i)=ωipae (max_L ipae−ipae(i))/(max_L ipae−min_L ipae)+ωplddt (plddt(i)−min_L plddt)/(max_L plddt−min_L plddt)+ωrmsd (max_L rmsd−rmsd(i))/(max_L rmsd−min_L rmsd). ... a designed binder is considered successful if it meets three distinct criteria when evaluated by AlphaFold-Multimer: an interface predicted aligned error (ipAE) ≤ 10, a binder pLDDT ≥ 70, and a binder self-consistent RMSD (scRMSD) ≤ 5 Å."

    Beam search hill-climbs a normalized weighted sum of exactly the three AF-Multimer quantities that later define success. With L parallel trajectories per interval and repeated pruning to the top-scoring candidate, reported both-success is partly the outcome of test-time optimization of the judge rather than an independent post-hoc measurement. Table 7’s weight ablations (both-success ranging 7–10) confirm the headline count is sensitive to this reward knob.

  2. fitted input called prediction [§4.3 Training Set Construction]
    "Finally, we utilized AlphaFold2-Multimer to compute predicted quality metrics, specifically ipAE, binder pLDDT, and inter-chain predicted Template Modeling score (ipTM). We retained only those pairs satisfying ipAE ≤ 10 Å, binder pLDDT ≥ 80, and ipTM ≥ 0.5. The final filtered training set comprises 60,692 pairs."

    I3CD is trained only on complexes that already pass AF-Multimer interface filters nearly identical to the evaluation thresholds (ipAE≤10, high binder pLDDT). The generator therefore learns the distribution of AF-pleasing interfaces; combined with Eq. 11 beam ranking on the same family of scores, success on AF criteria is statistically favored by construction of both the data and the search, not solely by discovering multi-context binders under an independent standard.

  3. other [§4.3 CROSS benchmark construction, Eq. (12)]
    "The selection was ranked using a composite quality score, S′_item, analogous to Equation (11): S′(i)=ω′_ipae(...)+ω′_plddt(...)+ω′_rmsd(...)+ω′_num(...), S′_item=min_{c∈C} S′(c)."

    CROSS entries themselves are quality-ranked with a composite built from the same AF-derived ipAE/pLDDT/scRMSD family used in training filters and MoPS beam search. The benchmark is therefore not an external oracle fully independent of the metric stack the method optimizes, tightening the evaluation loop even though labels are not literally fitted parameters renamed as predictions.

full rationale

Chamaileon is an empirical ML methods paper, not a first-principles derivation. I3CD training is standard multimodal flow matching on PDB dimers; MoPS is an inference procedure that alternates co-design and forward-folding across contexts. There is no self-definitional equation, no uniqueness theorem imported from the authors, and no ansatz smuggled in via self-citation that forces the central claim. The load-bearing issue is evaluation entanglement with one predictor family: (i) the I3CD training set retains only complexes AF-Multimer already scores well; (ii) MoPS beam search explicitly ranks trajectories by a weighted composite of ipAE, binder pLDDT, and binder scRMSD; (iii) success on CROSS is defined by thresholds on those same three quantities from AlphaFold-Multimer. Table 7 further shows both-success counts move when Eq. 11 weights change. That is real Goodhart-style circularity between search objective and reported metric—closer to “fitted/optimized input called prediction” than to a mathematical identity—but it does not collapse the method to its inputs by construction: simultaneous both-context success is still nontrivial (w/o MoPS is badly imbalanced; constructed baselines report 0 both-success), and the algorithmic claim that MoPS balances multi-context constraints has independent content. Score 4 reflects partial evaluation circularity with a still-substantive central contribution.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 3 invented entities

The central empirical claim rests on standard generative-modeling and structural-biology assumptions plus several paper-specific modeling and evaluation choices (decoupled noise, path alternation, AF-metric success, quality-filtered CROSS). Free parameters are the usual ML/sampler knobs and the composite ranking weights; invented entities are named procedures and the benchmark, not new physical objects.

free parameters (6)
  • Beam score weights ω_ipae, ω_plddt, ω_rmsd = 0.5, 0.3, 0.2 (default)
    Hand-chosen mixture that ranks MoPS candidates; ablated in Table 7; default (0.5, 0.3, 0.2) affects both-success.
  • MoPS switch frequency and beam frequency / candidate count L = e.g. MoPS freq 10, beam candidates 4, beam freq 50
    Discrete inference hyperparameters swept in Table 1; primary drivers of reported both-success.
  • SDE noise scale σ_t and rotation constant c
    Control stochasticity for beam search and sample quality; taken from flow-matching practice, not derived.
  • CROSS composite selection weights ω' and 70/30 high/low RMSD split = top-70 RMSD>1.59Å + top-30 ≤1.59Å
    Determine which 100 of 1867 candidates enter the main table; quality filter directly shapes reported success rates.
  • AF success thresholds (ipAE≤10, pLDDT≥70, scRMSD≤5Å) = 10 / 70 / 5Å
    Binary success definition borrowed from Anonymous (2025) protocol; changing them would change Unique/Both Success counts.
  • Training filters (resolution, length, contact, ipAE/pLDDT/ipTM gates) = res≤5Å, len constraints, ipAE≤10, pLDDT≥80, ipTM≥0.5
    Define the 60,692-pair I3CD corpus; ad hoc quality cuts.
axioms (5)
  • domain assumption Discrete and multimodal flow matching (masking CTMC + Euclidean/SO(3) flows) are valid generative models for joint sequence-structure co-design.
    Preliminary §3 and loss Eqs. 5–6 build directly on Campbell et al. 2024 and related flow-matching work.
  • ad hoc to paper A single shared sequence can be made jointly compatible with multiple discrete binder conformations by alternating co-design and forward-folding paths without multi-state paired supervision.
    Core MoPS premise in §4.2; motivated by data scarcity, not proved.
  • domain assumption AlphaFold-Multimer confidence and self-consistency metrics proxy true binding competence across contexts.
    Success definition and beam ranking in §5 and Eq. 11; standard in recent binder papers but unvalidated here experimentally.
  • ad hoc to paper CoDNaS clusters at ≥95% binder sequence identity plus PDB partner chains yield meaningful multi-state and multi-target design tests.
    CROSS construction §4.3 and Appendix K; 95% threshold equates near-identical binders and assigns their partners as contexts.
  • ad hoc to paper Clean-target / noisy-binder concatenation with hotspot mask suffices for context-conditional co-design (visual in-context analogy).
    I3CD §4.1; transfers a vision-style conditioning pattern to complexes.
invented entities (3)
  • I3CD (In-Context Complex Co-Design) no independent evidence
    purpose: Training paradigm that conditions binder denoising on concatenated clean target multimodal tokens with decoupled sequence/structure noise.
    Named method contribution; implemented as GNN flow model on PDB dimers.
  • MoPS (Mixture-of-Paths Sampling) no independent evidence
    purpose: Inference algorithm that cyclically mixes co-design and sequence-conditioned trajectories so one sequence is optimized against multiple contexts.
    Primary mechanism claimed to solve multi-context design without multi-state training pairs.
  • CROSS benchmark independent evidence
    purpose: Curated 100-entry (plus 1867-pool) eval set for multi-state/multi-target binder success under AF metrics.
    New dataset construction from CoDNaS/PDB; labels are structural, not wet-lab affinities.

reviewed 2026-07-30 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Chamaileon: Cross-Context Binder Design with Contextualized Modeling and Mixed Sampling." pith.science (2026). https://pith.science/paper/SESAXA4T

@misc{pith2026260723518,
  author       = {Pith},
  title        = {Pith review of: Chamaileon: Cross-Context Binder Design with Contextualized Modeling and Mixed Sampling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SESAXA4T}},
  note         = {Machine review of arXiv:2607.23518}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

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.

Figures

Figures reproduced from arXiv: 2607.23518 by Chenxi Cai, Hengyuan Cao, Mingxuan Liu, Min Zhang, Shizhuo Cheng, Weicheng Huang, Yan Zhang, Yunhong Lu.

Figure 1
Figure 1. Figure 1: Illustration of Cross-Context Binder Design. Multi￾state (MS) design (left) focuses on maintaining interface compati￾bility across the target’s conformational landscape, preventing the loss of affinity when the target reshapes during functional switch￾ing. Multi-target (MT) design (right) focuses on multi-specificity, where a single sequence is optimized to engage similar epitopes on distinct proteins with… view at source ↗
Figure 2
Figure 2. Figure 2: Training pipeline of I3CD. We concatenate clean target signals (blue tokens) with noisy binder signals (orange tokens) and feed them into the proposed I3CD paradigm to learn the joint dependencies of target and binder via discrete and continuous flow matching. Crucially, the binder’s sequence and structure are corrupted using decoupled noise schedules (t,t˜), enabling the model to effectively capture the i… view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of MoPS. By leveraging forward folding as a bridging mechanism, MoPS interleaves the sampling trajectories of different conformations. This enables simultaneous sequence-structure co-design across multiple binder states. The name Chamaileon embodies our vision of AI-designed binders as chameleons, capable of adaptively adjusting their conformations in response to the context. the model inputs … view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of beam search. In each time interval, the strategy retains only the highest-scoring candidate (solid line) while discarding less likely paths (dashed lines), ensuring efficient navigation toward the global optimum. conformations. Consequently, at tτ1 , both conformations possess distinct structures but share the same updated se￾quence. In the subsequent interval (tτ1 to tτ2 ), we alternate th… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative results of Chamaileon on cross-context binder design. For each case, the target protein is shown in deep blue-red spectrum with warmer color (red) indicating higher structural deviations, while the designed binder is shown in Binder State I (blue) and Binder State II (pink). The “Aligned” column displays the superposition of both states, with structures aligned based on the target backbones to … view at source ↗
Figure 6
Figure 6. Figure 6: Data collection pipeline. (a) Construction of the I3CD training set. (b) Construction of the CROSS benchmark. based on four distinct criteria. First, to ensure high struc￾tural quality, we excluded entries with a crystallographic resolution greater than 5 A. Second, we imposed constraints ˚ on chain length: each chain must contain at least 16 residues, with a combined total length not exceeding 512 residue… view at source ↗
Figure 7
Figure 7. Figure 7: Scalability of MoPS with respect to the number of conformations. By cyclically alternating among target states, MoPS extends binder co-design to an arbitrary number of conformations without incurring additional computational overhead. This formulation reveals that learning the rate matrix is equivalent to learning a denoising model. Therefore, the network p θ 1|t is trained to minimize the cross-entropy lo… view at source ↗
Figure 8
Figure 8. Figure 8: Dimer length distributions for the training set and the benchmark candidate pool. a b [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Distributions of structural differences (RMSD) in the candidate pool and the final CROSS benchmark. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Taxonomy of structural adaptation modes in cross-context binder design. The figure illustrates three distinct modes of backbone plasticity: Micro Adaption (top), Dual-face Adaption (middle), and Macro-switch Adaption (bottom). Binder conformations for Context I (blue) and Context II (pink) are superimposed. Deep blue-red spectrum regions indicate structural deviations (RMSD) between the two binder backbon… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

28 extracted references · 10 canonical work pages

  1. [1]

    More Data Collection Details In this section, we provide further details regarding our data collection process

    E. More Data Collection Details In this section, we provide further details regarding our data collection process. Figure 8 (a) illustrates the distribution of the sum of lengths for the two chains in the training set. Figure 8 (b) displays the distribution of the combined lengths of the target and binder within the candidate pool of 1,867 entries collect...

    2025

  2. [5]

    Dimension-reduction attack! video generative models are experts on controllable image synthesis.arXiv preprint arXiv:2505.23325, 2025a

    Cao, H., Feng, Y ., Gong, B., Tian, Y ., Lu, Y ., Liu, C., and Wang, B. Dimension-reduction attack! video generative models are experts on controllable image synthesis.arXiv preprint arXiv:2505.23325, 2025a. Cao, H., Lu, Y ., Wang, Q., Li, T., Xu, X., and Zhang, M. Ad- versarial self flow matching: Few-steps image generation with straight flows. 2025b. Ca...

    Pith/arXiv arXiv

  3. [6]

    doi: 10.1038/s41586-022-04654-9

    ISSN 1476-4687. doi: 10.1038/s41586-022-04654-9. Cavanagh, P. E., Xue, A. G., Dai, S. A., Qiang, A., Matsui, T., and Ting, A. Y . Computational design of conformation- biasing mutations to alter protein functions.Science, pp. eadv7953, January

    work page doi:10.1038/s41586-022-04654-9

  4. [7]

    It first undergoes sequence-conditioned generation to align with the timestep of the preceding conformation, and subsequently advances one time interval via co-design

    Specifically, each conformation resumes from the endpoint of its previous co-design step. It first undergoes sequence-conditioned generation to align with the timestep of the preceding conformation, and subsequently advances one time interval via co-design. All conformations are initialized at t= 0 . The process terminates when the first conformation reac...

    2021

  5. [8]

    doi: 10.1016/j.cell.2024.02.034

    ISSN 0092-8674, 1097-4172. doi: 10.1016/j.cell.2024.02.034. Cho, Y ., Pacesa, M., Zhang, Z., Correia, B. E., and Ovchin- nikov, S. BoltzDesign1: Inverting All-Atom Structure Prediction Model for Generalized Biomolecular Binder Design. Chu, A. E., Lu, T., and Huang, P.-S. Sparks of function by de novo protein design.Nature Biotechnology, 42 (2):203–215, February

    work page doi:10.1016/j.cell.2024.02.034 2024

  6. [9]

    doi: 10.1038/s41587-024-02133-2

    ISSN 1546-1696. doi: 10.1038/s41587-024-02133-2. Feng, Y ., Zhang, L., Cao, H., Chen, Y ., Feng, X., Cao, J., Wu, Y ., and Wang, B. Omnitry: Virtual try-on anything without masks.Advances in Neural Information Process- ing Systems, 38:132640–132667,

    work page doi:10.1038/s41587-024-02133-2

  7. [10]

    URL https://arxiv.org/abs/2503.00710. Guo, A. B., Akpinaroglu, D., Kelly, M. J., and Kortemme, T. Deep learning guided design of dynamic proteins, July

    Pith/arXiv arXiv

  8. [11]

    URL https://www.biorxiv.org/content/ early/2025/09/04/2025.09.03.672144

    doi: 10.1101/2025.09.03.672144. URL https://www.biorxiv.org/content/ early/2025/09/04/2025.09.03.672144. Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., Tunyasuvunakool, K., Bates, R., ˇZ´ıdek, A., Potapenko, A., Bridgland, A., Meyer, C., Kohl, S. A. A., Ballard, A. J., Cowie, A., Romera-Paredes, B., Nikolov, S., Jain, R., Ad...

    work page doi:10.1101/2025.09.03.672144 2025

  9. [12]

    doi: 10.1016/S0140-6736(24)00101-6

    ISSN 1474-547X. doi: 10.1016/S0140-6736(24)00101-6. Kortemme, T. De novo protein design—From new struc- tures to programmable functions.Cell, 187(3):526–544, February

    work page doi:10.1016/s0140-6736(24)00101-6

  10. [13]

    doi: 10.1016/j.cell.2023.12.028

    ISSN 0092-8674, 1097-4172. doi: 10.1016/j.cell.2023.12.028. Krishna, R., Wang, J., Ahern, W., Sturmfels, P., Venkatesh, P., Kalvet, I., Lee, G. R., Morey-Burrows, F. S., An- ishchenko, I., Humphreys, I. R., McHugh, R., Vafeados, D., Li, X., Sutherland, G. A., Hitchcock, A., Hunter, C. N., Kang, A., Brackenbrough, E., Bera, A. K., Baek, M., DiMaio, F., and...

    work page doi:10.1016/j.cell.2023.12.028 2023

  11. [14]

    doi: 10.1126/science.adl2528

    ISSN 0036-8075, 1095-9203. doi: 10.1126/science.adl2528. Lipman, Y ., Chen, R. T. Q., Ben-Hamu, H., Nickel, M., and Le, M. Flow matching for generative modeling,

    work page doi:10.1126/science.adl2528

  12. [15]

    Lisanza, S

    URLhttps://arxiv.org/abs/2210.02747. Lisanza, S. L., Gershon, J. M., Tipps, S. W. K., Sims, J. N., Arnoldt, L., Hendel, S. J., Simma, M. K., Liu, G., Yase, M., Wu, H., Tharp, C. D., Li, X., Kang, A., Bracken- brough, E., Bera, A. K., Gerben, S., Wittmann, B. J., McShan, A. C., and Baker, D. Multistate and functional protein design using RoseTTAFold sequen...

    Pith/arXiv arXiv

  13. [16]

    doi: 10.1038/s41587-024-02395-w

    ISSN 1546-1696. doi: 10.1038/s41587-024-02395-w. Liu, J., Liu, G., Liang, J., Li, Y ., Liu, J., Wang, X., Wan, P., Zhang, D., and Ouyang, W. Flow-grpo: Training flow matching models via online rl,

    doi:10.1038/s41587-024-02395-w

  14. [17]

    Lu, Y ., Wang, Q., Cao, H., Wang, X., Xu, X., and Zhang, M

    URL https: //arxiv.org/abs/2505.05470. Lu, Y ., Wang, Q., Cao, H., Wang, X., Xu, X., and Zhang, M. Inpo: Inversion preference optimization with reparametrized ddim for efficient diffusion model align- ment. InProceedings of the Computer Vision and Pattern Recognition Conference, pp. 28629–28639, 2025a. Lu, Y ., Wang, Q., Cao, H., Xu, X., and Zhang, M. Smo...

    Pith/arXiv arXiv 2016

  15. [19]

    doi: 10.1038/s41586-025-09429-6

    ISSN 0028-0836, 1476-4687. doi: 10.1038/s41586-025-09429-6. Passaro, S., Corso, G., Wohlwend, J., Reveiz, M., Thaler, S., Somnath, V . R., Getz, N., Portnoi, T., Roy, J., Stark, H., Kwabi-Addo, D., Beaini, D., Jaakkola, T., and Barzilay, R. Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction. pp. 2025.06.14.659707, June

    doi:10.1038/s41586-025-09429-6 2025

  16. [20]

    Ravussin, E., Sanchez-Delgado, G., Martin, C

    doi: 10.1101/2025.06.14.659707. Ravussin, E., Sanchez-Delgado, G., Martin, C. K., Beyl, R. A., Greenway, F. L., O’Farrell, L. S., Roell, W. C., Qian, H.-R., Li, J., Nishiyama, H., Haupt, A., Pratt, E. J., Urva, S., Milicevic, Z., and Coskun, T. Tirzepatide did not impact metabolic adaptation in people with obesity, but increased fat oxidation.Cell Metabol...

    doi:10.1101/2025.06.14.659707 2025

  17. [21]

    doi: 10.1016/j

    ISSN 1550-4131. doi: 10.1016/j. cmet.2025.03.011. Singhal, R., Horvitz, Z., Teehan, R., Ren, M., Yu, Z., McK- eown, K., and Ranganath, R. A General Framework for Inference-time Scaling and Steering of Diffusion Models, July

    doi:10.1016/j 2025

  18. [23]

    doi: 10.1016/j.sbi.2024.102794

    ISSN 0959-440X. doi: 10.1016/j.sbi.2024.102794. Wu, S., Huang, M., Wu, W., Cheng, Y ., Ding, F., and He, Q. Less-to-more generalization: Unlocking more controllability by in-context generation.arXiv preprint arXiv:2504.02160,

    arXiv 2024

  19. [24]

    Zhang, Y

    URL https://arxiv.org/ abs/2409.08022. Zhang, Y . and Skolnick, J. Scoring function for automated assessment of protein structure template quality.Proteins: Structure, Function, and Bioinformatics, 57(4):702–710,

    Pith/arXiv arXiv

  20. [28]

    We compared our method against APM (Chen et al., 2025), and the results are presented in Table

    2025

  21. [2004]

    Extending Conformational Ensemble Prediction to Multidomain Proteins and Protein Complex

    Zhu, J. Extending Conformational Ensemble Prediction to Multidomain Proteins and Protein Complex. 12 Chamaileon: Cross-Context Binder Design with Contextualized Modeling and Mixed Sampling A. Discrete Flow Models Diffusion models and flow matching have recently achieved substantial progress (Wang et al., 2026; Lu et al., 2026a; 2025a;b; 2026b; Feng et al....

    2026

  22. [2016]

    doi: 10.1093/database/baw038

    ISSN 1758-0463. doi: 10.1093/database/baw038. Pacesa, M., Nickel, L., Schellhaas, C., Schmidt, J., Py- atova, E., Kissling, L., Barendse, P., Choudhury, J., Kapoor, S., Alcaraz-Serna, A., Cho, Y ., Ghamary, K. H., Vinu´e, L., Yachnin, B. J., Wollacott, A. M., Buckley, S., Westphal, A. H., Lindhoud, S., Georgeon, S., Goverde, C. A., Hatzopoulos, G. N., G ¨...

    work page doi:10.1093/database/baw038

  23. [2021]

    van Kempen, M., Kim, S

    URL https://arxiv.org/abs/2011.13456. van Kempen, M., Kim, S. S., Tumescheit, C., Mirdita, M., Lee, J., Gilchrist, C. L. M., S ¨oding, J., and Steineg- ger, M. Fast and accurate protein structure search with Foldseek.Nature Biotechnology, 42(2):243–246, February

    Pith/arXiv arXiv 2011

  24. [2022]

    doi: 10.1016/S1470-2045(22)00335-7

    ISSN 1474-5488. doi: 10.1016/S1470-2045(22)00335-7. Campbell, A., Yim, J., Barzilay, R., Rainforth, T., and Jaakkola, T. Generative flows on discrete state-spaces: enabling multimodal flows with applications to protein co- design. InProceedings of the 41st International Confer- ence on Machine Learning, ICML’24. JMLR.org,

    work page doi:10.1016/s1470-2045(22)00335-7 2045

  25. [2023]

    doi: 10.1038/ s41467-023-38328-5

    ISSN 2041-1723. doi: 10.1038/ s41467-023-38328-5. Budde, L. E., Sehn, L. H., Matasar, M., Schuster, S. J., Assouline, S., Giri, P., Kuruvilla, J., Canales, M., Diet- rich, S., Fay, K., Ku, M., Nastoupil, L., Cheah, C. Y ., Wei, M. C., Yin, S., Li, C.-C., Huang, H., Kwan, A., Penuel, E., and Bartlett, N. L. Safety and efficacy of mosunetuzumab, a bispecifi...

    2041

  26. [2024]

    doi: 10.1038/s41586-024-07487-w

    ISSN 1476-4687. doi: 10.1038/s41586-024-07487-w. Abrudan, A., Ojeda, S. P., Joshi, C. K., Greenig, M., En- gelberger, F., Khmelinskaia, A., Meiler, J., Vendruscolo, M., and Knowles, T. P. J. Multi-state protein design with dynamicmpnn,

    doi:10.1038/s41586-024-07487-w

  27. [2025]

    Anonymous

    URL https://arxiv.org/ abs/2507.21938. Anonymous. Scaling atomistic protein binder design with generative pretraining and test-time compute. In Submitted to The Fourteenth International Conference on Learning Representations,

    arXiv

  28. [2026]

    doi: 10.1126/science.adv7953

    ISSN 0036-8075, 1095-9203. doi: 10.1126/science.adv7953. Chen, R., Xue, D., Zhou, X., Zheng, Z., Zeng, X., and Gu, Q. An all-atom generative model for designing protein complexes, September

    work page doi:10.1126/science.adv7953

This paper was first reviewed by grok-4.5 on July 30, 2026.