Pith. sign in

REVIEW 4 major objections 5 minor 59 references

Protein length can be learned as a Poisson process rate and sampled jointly with structure or sequence, without fixing length first.

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 ↗

T0 review · grok-4.5

2026-07-13 00:47 UTC pith:SAUUPT7H

load-bearing objection Solid unified variable-length flow for protein modalities with real multi-task gains; novelty is tempered by TDDM/EditFlow overlap, and free-length vs contig/oracle comparisons need care but do not erase the result. the 4 major comments →

arxiv 2607.09039 v1 pith:SAUUPT7H submitted 2026-07-10 cs.LG q-bio.QM

Variable-Length Generative Protein Design via Generalized Poisson Flow

Chaoran Cheng , Zhanghan Ni , Yanru Qu , Yuxin Chen , Ruihan Guo , Jiajun Fan , Ge Liu This is my paper
classification cs.LG q-bio.QM
keywords variable-length generationprotein designgeneralized Poisson processflow matchingmotif scaffoldingpeptide co-designsimulation-free likelihood
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.

Most diffusion and flow models for proteins force you to pick a length before you generate anything. That is awkward when the best length is unknown and tightly linked to whether the design will fold. This paper claims you can treat length growth as an inhomogeneous generalized Poisson process, learn its rate by negative log-likelihood, and couple that process to ordinary within-length flow matching for coordinates, amino-acid types, or Riemannian peptide degrees of freedom. The same construction recovers the joint data distribution in the population limit and yields a generator-based upper bound on the KL divergence between data and model. Empirically, free-length sampling improves designability over fixed-length structure baselines, matches UniRef50 length and foldability statistics more closely than fixed-length sequence models, wins more unique successes on motif scaffolding, and stays competitive for peptide co-design without being told the native length. The practical payoff is that one model can explore the feasible design space without a length oracle or a grid of contig templates.

Core claim

The authors establish that an inhomogeneous generalized Poisson process on length, with rate learned by exact trajectory NLL after analytic marginalization of insertion times, can be posterior-marginalized together with continuous, discrete, or Riemannian within-length generators so that the joint multimodal data law is recovered and the training losses upper-bound the KL between data and generated laws.

What carries the argument

Generalized Poisson Flow (GPFlow): length evolves by a learned insertion rate of an inhomogeneous generalized Poisson process; existing residues are refined by a within-length generator; the rate objective is the process NLL with insertion times marginalized out, plus reconstruction and flow-matching terms.

Load-bearing premise

That the chosen insertion schedule and sampling approximations still leave free-length sampling a fair, non-oracle comparison to fixed-length baselines on the reported design metrics.

What would settle it

Train and sample GPFlow and its fixed-length backbone under identical length marginals (or oracle-matched lengths) on the same PDB or motif tasks; if free-length gains in designability and unique successes disappear once length is controlled, the central practical claim fails.

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

If this is right

  • Structure, sequence, motif, and peptide models can share one variable-length construction instead of separate length-handling tricks.
  • Motif scaffolding no longer needs predefined contig length ranges; length emerges from the learned rate.
  • Peptide co-design can drop the native-length oracle that fixed-length baselines rely on at test time.
  • The same rate-plus-within-length losses give a concrete KL upper bound that training is minimizing term by term.

Where Pith is reading between the lines

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

  • The same construction is a natural candidate for other ordered variable-length biological objects (RNA, multi-domain assemblies) once a within-length generator exists.
  • If scheduler sensitivity is the main practical bottleneck, learning or annealing the insertion schedule jointly with the rate head may be the highest-leverage follow-up.
  • Head-to-head length-matched ablations would cleanly separate the value of free length from the value of better joint modeling.

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 / 5 minor

Summary. The paper proposes Generalized Poisson Flow (GPFlow), a variable-length generative framework that couples an inhomogeneous generalized Poisson process over length with within-length continuous, discrete, or Riemannian flow matching. Theorems 3.1–3.3 construct conditional and marginal rates via a scheduler κ_t and posterior marginalization, yielding a simulation-free Poisson NLL (Eq. 4) plus reconstruction and flow-matching losses (Eq. 8). Theorem 4.1 gives a generator-based KL upper bound whose terms align with those losses. Empirically, GPFlow is instantiated on five protein tasks: unconditional structure (Proteina backbone), unconditional sequence (DiT), structure- and sequence-based motif scaffolding, and peptide structure–sequence co-design (PepFlow). Reported gains include higher designability than Proteina on PDB (96.1% vs 77.6%), closer UniRef50 pLDDT/length match than DPLM, first place on 10/16 RFDiffusion motif tasks with more unique successes, and competitive peptide metrics without a native-length oracle.

Significance. Variable length is a genuine bottleneck for diffusion/flow protein models, especially motif scaffolding and de novo peptide design where native length is unavailable. The paper’s main contribution is a unified probability-flow construction that covers Euclidean, categorical, and Riemannian modalities under one Poisson-rate objective, with population recovery and a generator KL bound that justify the training losses rather than treating the rate loss as purely empirical. Direct base-model ablations (Proteina, DPLM/EditFlow-style, PepFlow) and length-distribution recovery checks make the empirical case more informative than a pure method paper. If the comparison asymmetries are tightened, this would be a solid methods contribution for generative modeling of ordered biological sequences and structures.

major comments (4)
  1. §5.1 / Table 1 and Appendix C.1.3: The headline designability claim (96.1% GPFlow-PDB vs 77.6% Proteina-PDB) compares free-length GPFlow samples to fixed-length Proteina on the grid {50,100,150,200,250}. Appendix E.1 shows high designability across GPFlow length bins, which is supportive, but does not re-score Proteina under GPFlow’s learned length marginal (or GPFlow under the same grid). Without a length-matched or length-oracle-matched protocol, part of the gain may be comparison asymmetry rather than joint-distribution recovery alone. A matched re-evaluation (or explicit length-conditioned GPFlow ablation) is needed to make the central empirical claim load-bearing.
  2. Table 2 / Appendix C.1.3: Motif baselines are scored under predefined contig length ranges (long/med/short for 5TRV, 6E6R, 6EXZ, 7MRX), while GPFlow samples freely and aggregates those templates into single tasks. The paper correctly notes that GPFlow can produce scaffolds outside contig ranges, but unique-success rankings (first on 10/16) are not fully comparable until baselines are also allowed free length or GPFlow is constrained to the same contig ranges. Please report both free-length and contig-constrained numbers for GPFlow, or re-run baselines without hard length caps where possible.
  3. Table 3 / Appendix C.3.3: Peptide AAR, RMSD, and SSR are recomputed only on Needleman–Wunsch-matched residues and exclude |n−m| unmatched residues when lengths differ (67.9% of samples). This is disclosed, but the composite metrics then do not penalize length error, while baselines are native-length-conditioned. The claim of remaining competitive “without a native-length oracle” should be supported by (i) full-length metrics that include length mismatch and (ii) a length-matched GPFlow subset reported as primary, not only in the appendix breakdown.
  4. §3–4 vs §6 / Appendices B.5, C.1.2, C.2.5: Population recovery (Thms 3.2–3.3) and the KL bound (Thm 4.1) hold for the idealized marginal rate under a fixed κ_t. Practice uses early-completion κ_t (train min(1,t/0.6), sample min(1,t/0.3)), rate scaling 0.6–0.7, τ-leaping, localized paths, and deletion correctors. Appendix D ablations cover scheduler and γ but not rate scaling or τ-leaping against the theoretical path. Please either (a) show that these inference modifications leave the recovered length/joint marginals essentially unchanged, or (b) state clearly that the guarantees apply to the training objective and that free-length sampling quality is empirical.
minor comments (5)
  1. Many headline tables (Tables 1–3) lack error bars, multi-seed variance, or confidence intervals; even a single retrain variance for the PDB structure model would strengthen the designability claim.
  2. Figure 1 and Figure 2(e) are useful length-calibration checks; consider overlaying the fixed-length grid mass or UniRef50 histogram more explicitly so readers can judge match quality at a glance.
  3. Notation: λ_t is used both for conditional and marginal rates; a consistent λ^* vs λ_θ distinction in the main text (as in the appendix) would reduce ambiguity around Eq. 4–8.
  4. Related work correctly positions TDDM and EditFlow; a short explicit statement that GPFlow recovers their insertion-only rate objectives in the overlapping regimes (already in §2) could be moved earlier for readers who only skim the theory.
  5. Typos / polish: “a priori” spacing, occasional double spaces, and “Struct%α/β” column headers in Table 1 could be cleaned for camera-ready.

Circularity Check

0 steps flagged

No load-bearing circularity: population recovery and the KL bound follow from standard point-process NLL and posterior marginalization; empirical gains are scored against external baselines.

full rationale

Walking the derivation chain (Thm 3.1 conditional binomial path → Thm 3.2/3.3 posterior-marginalized rates and generators → Prop. A.1 Poisson NLL → simulation-free L_GP via the point-process integration formula → generator KL upper bound Thm 4.1 whose terms match the additive losses) yields standard stochastic-process arguments, not X-defined-as-Y or fitted-input-as-prediction. The rate objective is acknowledged to coincide with TDDM/EditFlow in overlapping regimes; the paper’s claimed additions (unified continuous/discrete/Riemannian construction, exact NLL interpretation, new KL bound, protein instantiations) are independent content, not a rename of a self-cited uniqueness theorem. Empirical claims (designability, UniRef50 fitness, motif unique successes, peptide metrics) are measured with external tools and baselines (ProteinMPNN, ESMFold, Proteina, DPLM, RFDiffusion, PepFlow, PDB/AFDB/UniRef50 statistics), not by re-reporting fitted constants. Scheduler/τ-leaping/localized-path choices and variable-length metric adaptations (Needleman–Wunsch matching, contig-free scoring) are evaluation or hyperparameter issues, not circular reductions of the claimed recovery theorems. Minor author-overlap reuse of PepFlow/Proteina architectures is ordinary engineering, not a self-citation that forces the central result. Score 1 only for ordinary self-containment of implementation choices, not for any identified circular step.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 2 invented entities

The central claim rests on standard Markov/point-process math plus domain modeling choices (order-preserving insertions, modality-specific flow losses, early-completion schedulers) and many hand-chosen training/sampling knobs. No new physical entity is postulated; the invented objects are algorithmic (GPFlow process, order-preserving slots, localized paths).

free parameters (7)
  • insertion scheduler κ_t (early completion, e.g. min(1,t/0.6) train / min(1,t/0.3) sample)
    Hand-chosen absolute-continuous schedule that controls when length saturates; ablations show late insertion hurts designability.
  • loss weights w_rec, w_FM
    Hyperparameters balancing Poisson NLL, reconstruction, and within-length flow matching (Eq. 8).
  • rate scaling factors (0.7 unconditional structure; 0.6 motif)
    Inference-time multipliers on λ_θ chosen to match length distributions / AFDB lengths.
  • SDE noise scale γ (default 0.35) and score scheduler g_t
    Controls designability–diversity trade-off in structure sampling (Table 8).
  • τ-leaping / multi-place insertion and Euler step counts (400 structure, 5000 sequence, 250 peptide)
    Discretization approximations required for practical sampling; not uniquely fixed by theory.
  • localized-path λ_prop=3.0 and cubic κ_t=t^3 (sequence)
    Extra path parameters for long-sequence generation adapted from EditFlow practice.
  • CFG weights, corrector strength α, temperature/nucleus settings (sequence)
    Sampling knobs that materially affect length control and quality metrics.
axioms (6)
  • standard math Kolmogorov forward / generator identities for continuous-time Markov processes with jumps characterize the evolution of pt.
    Used throughout Theorems 3.1–3.3 and 4.1 (Appendices A.1–A.6).
  • standard math Point-process integration formula: expected sum over event times equals integral of intensity times predictable integrand.
    Justifies simulation-free NLL after marginalizing insertion times (Prop. A.1 / Brémaud).
  • domain assumption Proteins have a natural residue order, so order-preserving multi-slot insertion rates are the right specialization.
    Appendix B.1; all experiments use order-preserving GPFlow.
  • domain assumption Within-length continuous, discrete, and Riemannian flow-matching losses suffice to refine modalities between insertions.
    Section 3.2 and Appendix B.3; inherits standard FM assumptions from Proteina/PepFlow/DiT bases.
  • ad hoc to paper Mean-MSE (or geodesic-mean) regression is an adequate practical surrogate for continuous insertion kernels ρ_t.
    Appendix B.4 admits nested flow for ρ is preferred but uses mean regression for speed.
  • standard math Regularity conditions for the KL bound (positive densities, absolute continuity of jump measures, shared diffusion coefficient).
    Theorem 4.1 proof assumptions in Appendix A.6.
invented entities (2)
  • Generalized Poisson Flow (GPFlow) process on variable-length state space Y no independent evidence
    purpose: Couple length-changing Poisson insertions to within-length generators for multimodal protein generation.
    Core algorithmic object; builds on known Poisson/jump-diffusion ideas rather than a new physical entity.
  • Order-preserving multi-slot insertion rates and localized Bernoulli-propagation paths no independent evidence
    purpose: Respect residue order and improve long-sequence local context during training.
    Appendix B.1 and C.2.5 constructions specific to this framework/instantiation.

pith-pipeline@v1.1.0-grok45 · 44703 in / 3699 out tokens · 40082 ms · 2026-07-13T00:47:11.601304+00:00 · methodology

0 comments
read the original abstract

The ability to generate variable-length proteins is crucial in protein design, where the optimal length is often unknown and tightly coupled to designability. Current diffusion- and flow-based generative models typically require the protein length to be specified before sampling, limiting their flexibility in exploring the feasible design space. To address this limitation, we introduce Generalized Poisson Flow (GPFlow), a variable-length generative framework that learns the rate function of an inhomogeneous generalized Poisson process by minimizing its negative log-likelihood. We establish population-level guarantees for recovering the joint multimodal distribution and derive an upper bound on the KL divergence between the data and generated distributions. We comprehensively evaluate GPFlow across structure and sequence design, motif scaffolding, and peptide co-design, spanning Euclidean, categorical, and Riemannian modalities to fully validate its variable-length generation quality. In unconditional design, GPFlow improves structural designability and achieves the best distributional fitness for sequence design compared to their corresponding fixed-length baselines, while perfectly recovering the length distribution. In conditional motif scaffolding, GPFlow ranks first on 10 of 16 structure-based design tasks with significantly more unique successes and also achieves more passed tasks in sequence-based design. In peptide co-design, GPFlow remains competitive even without access to a native-length oracle.

Figures

Figures reproduced from arXiv: 2607.09039 by Chaoran Cheng, Ge Liu, Jiajun Fan, Ruihan Guo, Yanru Qu, Yuxin Chen, Zhanghan Ni.

Figure 1
Figure 1. Figure 1: Length distributions and representative ex [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Performance benchmark for unconditional protein sequence design. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Representative unconditional sequence generations folded by ESMFold. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: GPFlow generates more diverse clus￾ters for the same motif (highlighted in red) [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Samples of GPFlow-generated peptides com [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Training procedure (left) and one sampling step (right) for three variants of GPFlow: Length-Only (Section 3.1), Multimodal Order-Agnostic (Section 3.2), and Multimodal Order￾Preserving (Appendix B.1). r0 = 0 and rk+1 = L + 1. The L − k unretained target indices are partitioned among the k + 1 insertion slots by: Ωi := {ℓ ∈ {1, . . . , L} \ {r1, . . . , rk} : ri < ℓ < ri+1}, 0 ≤ i ≤ k. (71) Thus, Ωi contai… view at source ↗
Figure 7
Figure 7. Figure 7: Designability per length range for generated structures. [PITH_FULL_IMAGE:figures/full_fig_p039_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Evaluation of protein length control via classifier-free guidance. Each ridge shows the [PITH_FULL_IMAGE:figures/full_fig_p040_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Aggregate protein length distribution across all 16 motif scaffolding tasks for GPFlow and [PITH_FULL_IMAGE:figures/full_fig_p040_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Protein length distributions for representative tasks for GPFlow. Raw success lengths and [PITH_FULL_IMAGE:figures/full_fig_p041_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Peptide length. Left: length distributions of GPFlow-generated and native peptides. Right: mean generated length at each native length; the dashed line marks exact recovery. F Broader Impacts This work introduces Generalized Poisson Flow (GPFlow), a framework for variable-length gen￾eration evaluated on protein design. Its connection between generalized Poisson processes and probability flows may also be … 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

59 extracted references · 1 canonical work pages

  1. [1]

    Lu, Nicolo Fusi, Ava P

    Sarah Alamdari, Nitya Thakkar, Rianne van den Berg, Alex X. Lu, Nicolo Fusi, Ava P. Amini, and Kevin K. Yang. Protein generation with evolutionary diffusion: sequence is all you need, September 2023. URL https://www.biorxiv.org/content/10.1101/2023.09. 11.556673v1. Pages: 2023.09.11.556673 Section: New Results

    doi:10.1101/2023.09 2023

  2. [2]

    Structured denoising diffusion models in discrete state-spaces.Advances in neural information processing systems, 34:17981–17993, 2021

    Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. Structured denoising diffusion models in discrete state-spaces.Advances in neural information processing systems, 34:17981–17993, 2021

    2021

  3. [3]

    Accurate prediction of protein structures and interactions using a three-track neural network.Science, 373(6557):871–876, 2021

    Minkyung Baek, Frank DiMaio, Ivan Anishchenko, Justas Dauparas, Sergey Ovchinnikov, Gyu Rie Lee, Jue Wang, Qian Cong, Lisa N Kinch, R Dustin Schaeffer, et al. Accurate prediction of protein structures and interactions using a three-track neural network.Science, 373(6557):871–876, 2021

    2021

  4. [4]

    Amin, Ruben Weitzman, Debora Marks, and Andrew Gordon Wilson

    Ethan Baron, Alan N. Amin, Ruben Weitzman, Debora Marks, and Andrew Gordon Wilson. A Diffusion Model to Shrink Proteins While Maintaining Their Function, November 2025. URL http://arxiv.org/abs/2511.07390. arXiv:2511.07390 [cs] version: 1

    arXiv 2025

  5. [5]

    The protein data bank.Nucleic acids research, 28(1):235–242, 2000

    Helen M Berman, John Westbrook, Zukang Feng, Gary Gilliland, Talapady N Bhat, Helge Weissig, Ilya N Shindyalov, and Philip E Bourne. The protein data bank.Nucleic acids research, 28(1):235–242, 2000

    2000

  6. [6]

    An extension of watanabe’s theorem of characterization of poisson processes over the positive real half line.Journal of Applied Probability, 12(2):396–399, 1975

    Pierre Brémaud. An extension of watanabe’s theorem of characterization of poisson processes over the positive real half line.Journal of Applied Probability, 12(2):396–399, 1975

    1975

  7. [7]

    Point processes and queues.Springer, 1981

    Pierre Brémaud. Point processes and queues.Springer, 1981

    1981

  8. [8]

    De novo design of all-atom biomolecular interactions with rfdiffusion3.bioRxiv, 2025

    Jasper Butcher, Rohith Krishna, Raktim Mitra, Rafael I Brent, Yanjing Li, Nathaniel Corley, Paul T Kim, Jonathan Funk, Simon Mathis, Saman Salike, et al. De novo design of all-atom biomolecular interactions with rfdiffusion3.bioRxiv, 2025

    2025

  9. [9]

    Trans-dimensional generative modeling via jump diffusion models

    Andrew Campbell, William Harvey, Christian Weilbach, Valentin De Bortoli, Thomas Rain- forth, and Arnaud Doucet. Trans-dimensional generative modeling via jump diffusion models. Advances in Neural Information Processing Systems, 36:42217–42257, 2023

    2023

  10. [10]

    Pyrosetta: a script-based interface for implementing molecular modeling algorithms using rosetta.Bioinformatics, 26(5):689–691, 2010

    Sidhartha Chaudhury, Sergey Lyskov, and Jeffrey J Gray. Pyrosetta: a script-based interface for implementing molecular modeling algorithms using rosetta.Bioinformatics, 26(5):689–691, 2010

    2010

  11. [11]

    Flow matching on general geometries

    Ricky TQ Chen and Yaron Lipman. Flow matching on general geometries. InInternational Conference on Learning Representations, volume 2024, pages 47922–47945, 2024

    2024

  12. [12]

    Categorical flow matching on statistical manifolds.Advances in Neural Information Processing Systems, 37:54787–54819, 2024

    Chaoran Cheng, Jiahan Li, Jian Peng, and Ge Liu. Categorical flow matching on statistical manifolds.Advances in Neural Information Processing Systems, 37:54787–54819, 2024

    2024

  13. [13]

    α-flow: A unified framework for continuous- state discrete flow matching models.arXiv preprint arXiv:2504.10283, 2025

    Chaoran Cheng, Jiahan Li, Jiajun Fan, and Ge Liu. α-flow: A unified framework for continuous- state discrete flow matching models.arXiv preprint arXiv:2504.10283, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    Springer, 2003

    Daryl J Daley and David Vere-Jones.An introduction to the theory of point processes: volume I: elementary theory and methods. Springer, 2003. 11

    2003

  15. [15]

    Robust deep learning–based protein sequence design using proteinmpnn.Science, 378(6615):49–56, 2022

    Justas Dauparas, Ivan Anishchenko, Nathaniel Bennett, Hua Bai, Robert J Ragotte, Lukas F Milles, Basile IM Wicky, Alexis Courbet, Rob J de Haas, Neville Bethel, et al. Robust deep learning–based protein sequence design using proteinmpnn.Science, 378(6615):49–56, 2022

    2022

  16. [16]

    Mixed continuous and categorical flow matching for 3d de novo molecule generation.ArXiv, pages arXiv–2404, 2024

    Ian Dunn and David Ryan Koes. Mixed continuous and categorical flow matching for 3d de novo molecule generation.ArXiv, pages arXiv–2404, 2024

    2024

  17. [17]

    Discrete flow matching.Advances in Neural Information Processing Systems, 37:133345–133385, 2024

    Itai Gat, Tal Remez, Neta Shaul, Felix Kreuk, Ricky TQ Chen, Gabriel Synnaeve, Yossi Adi, and Yaron Lipman. Discrete flow matching.Advances in Neural Information Processing Systems, 37:133345–133385, 2024

    2024

  18. [18]

    Proteina: Scaling flow-based protein structure generative models.arXiv preprint arXiv:2503.00710, 2025

    Tomas Geffner, Kieran Didi, Zuobai Zhang, Danny Reidenbach, Zhonglin Cao, Jason Yim, Mario Geiger, Christian Dallago, Emine Kucukbenli, Arash Vahdat, et al. Proteina: Scaling flow-based protein structure generative models.arXiv preprint arXiv:2503.00710, 2025

    Pith/arXiv arXiv 2025

  19. [19]

    Approximate accelerated stochastic simulation of chemically reacting systems.The Journal of chemical physics, 115(4):1716–1733, 2001

    Daniel T Gillespie. Approximate accelerated stochastic simulation of chemically reacting systems.The Journal of chemical physics, 115(4):1716–1733, 2001

    2001

  20. [20]

    Edit flows: Flow matching with edit operations.arXiv preprint arXiv:2506.09018, 2025

    Marton Havasi, Brian Karrer, Itai Gat, and Ricky TQ Chen. Edit flows: Flow matching with edit operations.arXiv preprint arXiv:2506.09018, 2025

    arXiv 2025

  21. [21]

    Sofroniew, Deniz Oktay, Zeming Lin, Robert Verkuil, Vincent Q

    Thomas Hayes, Roshan Rao, Halil Akin, Nicholas J. Sofroniew, Deniz Oktay, Zeming Lin, Robert Verkuil, Vincent Q. Tran, Jonathan Deaton, Marius Wiggert, Rohil Badkundri, Irhum Shafkat, Jun Gong, Alexander Derry, Raul S. Molina, Neil Thomas, Yousuf A. Khan, Chetan Mishra, Carolyn Kim, Liam J. Bartie, Matthew Nemeth, Patrick D. Hsu, Tom Sercu, Salvatore Cand...

    doi:10.1126/science.ads0018 2025

  22. [22]

    Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    2020

  23. [23]

    Generator matching: Generative modeling with arbitrary markov processes.arXiv preprint arXiv:2410.20587, 2024

    Peter Holderrieth, Marton Havasi, Jason Yim, Neta Shaul, Itai Gat, Tommi Jaakkola, Brian Karrer, Ricky TQ Chen, and Yaron Lipman. Generator matching: Generative modeling with arbitrary markov processes.arXiv preprint arXiv:2410.20587, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    Argmax flows and multinomial diffusion: Learning categorical distributions.Advances in neural information processing systems, 34:12454–12465, 2021

    Emiel Hoogeboom, Didrik Nielsen, Priyank Jaini, Patrick Forré, and Max Welling. Argmax flows and multinomial diffusion: Learning categorical distributions.Advances in neural information processing systems, 34:12454–12465, 2021

    2021

  25. [25]

    Highly accurate protein structure prediction with alphafold.nature, 596(7873):583–589, 2021

    John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ron- neberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, et al. Highly accurate protein structure prediction with alphafold.nature, 596(7873):583–589, 2021

    2021

  26. [26]

    Dictionary of protein secondary structure: Pattern recognition of hydrogen-bonded and geometrical features.Biopolymers, 22(12):2577–2637,

    Wolfgang Kabsch and Christian Sander. Dictionary of protein secondary structure: Pattern recognition of hydrogen-bonded and geometrical features.Biopolymers, 22(12):2577–2637,

  27. [27]

    doi: 10.1002/bip.360221211

    doi:10.1002/bip.360221211

  28. [28]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization, January

  29. [29]

    arXiv:1412.6980 [cs]

    URLhttp://arxiv.org/abs/1412.6980. arXiv:1412.6980 [cs]

    Pith/arXiv arXiv

  30. [30]

    Biotite: a unifying open source computational biology framework in Python.BMC Bioinformatics, 19(1):346, 2018

    Patrick Kunzmann and Kay Hamacher. Biotite: a unifying open source computational biology framework in Python.BMC Bioinformatics, 19(1):346, 2018. doi: 10.1186/s12859-018-2367-z

    work page doi:10.1186/s12859-018-2367-z 2018

  31. [31]

    Full-atom peptide design based on multi-modal flow matching.arXiv preprint arXiv:2406.00735, 2024

    Jiahan Li, Chaoran Cheng, Zuofan Wu, Ruihan Guo, Shitong Luo, Zhizhou Ren, Jian Peng, and Jianzhu Ma. Full-atom peptide design based on multi-modal flow matching.arXiv preprint arXiv:2406.00735, 2024

    Pith/arXiv arXiv 2024

  32. [32]

    Analysis of explicit tau-leaping schemes for simulating chemically reacting systems

    Tiejun Li. Analysis of explicit tau-leaping schemes for simulating chemically reacting systems. Multiscale Modeling & Simulation, 6(2):417–436, 2007

    2007

  33. [33]

    Out of many, one: Designing and scaffolding proteins at the scale of the structural universe with genie 2.arXiv preprint arXiv:2405.15489, 2024

    Yeqing Lin, Minji Lee, Zhao Zhang, and Mohammed AlQuraishi. Out of many, one: Designing and scaffolding proteins at the scale of the structural universe with genie 2.arXiv preprint arXiv:2405.15489, 2024. 12

    Pith/arXiv arXiv 2024

  34. [34]

    Evolutionary-scale prediction of atomic- level protein structure with a language model.Science, 379(6637):1123–1130, 2023

    Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Nikita Smetanin, Robert Verkuil, Ori Kabeli, Yaniv Shmueli, Allan dos Santos Costa, Maryam Fazel-Zarandi, Tom Sercu, Salvatore Candido, and Alexander Rives. Evolutionary-scale prediction of atomic- level protein structure with a language model.Science, 379(6637):1123–1130, 2023. doi:...

    doi:10.1126/science.ade2574 2023

  35. [35]

    Evolutionary-scale prediction of atomic-level protein structure with a language model.Science, 379(6637):1123–1130, 2023

    Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Nikita Smetanin, Robert Verkuil, Ori Kabeli, Yaniv Shmueli, et al. Evolutionary-scale prediction of atomic-level protein structure with a language model.Science, 379(6637):1123–1130, 2023

    2023

  36. [36]

    Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022

    Pith/arXiv arXiv 2022

  37. [37]

    Multistate and functional protein design using rosettafold sequence space diffusion.Nature biotechnology, 43(8):1288–1298, 2025

    Sidney Lyayuga Lisanza, Jacob Merle Gershon, Samuel WK Tipps, Jeremiah Nelson Sims, Lucas Arnoldt, Samuel J Hendel, Miriam K Simma, Ge Liu, Muna Yase, Hongwei Wu, et al. Multistate and functional protein design using rosettafold sequence space diffusion.Nature biotechnology, 43(8):1288–1298, 2025

    2025

  38. [38]

    Rosetta flex- pepdock web server—high resolution modeling of peptide–protein interactions.Nucleic acids research, 39(suppl_2):W249–W253, 2011

    Nir London, Barak Raveh, Eyal Cohen, Guy Fathi, and Ora Schueler-Furman. Rosetta flex- pepdock web server—high resolution modeling of peptide–protein interactions.Nucleic acids research, 39(suppl_2):W249–W253, 2011

    2011

  39. [39]

    lDDT: A local superposition-free score for comparing protein structures and models using distance difference tests.Bioinformatics, 29(21):2722–2728, 2013

    Valerio Mariani, Marco Biasini, Alessandro Barbato, and Torsten Schwede. lDDT: A local superposition-free score for comparing protein structures and models using distance difference tests.Bioinformatics, 29(21):2722–2728, 2013. doi: 10.1093/bioinformatics/btt473

    doi:10.1093/bioinformatics/btt473 2013

  40. [40]

    Scalable Diffusion Models with Transformers, March 2023

    William Peebles and Saining Xie. Scalable Diffusion Models with Transformers, March 2023. URLhttp://arxiv.org/abs/2212.09748. arXiv:2212.09748 [cs]

    Pith/arXiv arXiv 2023

  41. [41]

    Rosetta flexpepdock ab-initio: simultaneous folding, docking and refinement of peptides onto their receptors.PloS one, 6(4):e18934, 2011

    Barak Raveh, Nir London, Lior Zimmerman, and Ora Schueler-Furman. Rosetta flexpepdock ab-initio: simultaneous folding, docking and refinement of peptides onto their receptors.PloS one, 6(4):e18934, 2011

    2011

  42. [42]

    Fast solvers for discrete diffusion models: Theory and applications of high-order algorithms.arXiv preprint arXiv:2502.00234, 2025

    Yinuo Ren, Haoxuan Chen, Yuchen Zhu, Wei Guo, Yongxin Chen, Grant M Rotskoff, Molei Tao, and Lexing Ying. Fast solvers for discrete diffusion models: Theory and applications of high-order algorithms.arXiv preprint arXiv:2502.00234, 2025

    arXiv 2025

  43. [43]

    Simple and effective masked diffusion language models.Advances in Neural Information Processing Systems, 37:130136–130184, 2024

    Subham Sahoo, Marianne Arriola, Yair Schiff, Aaron Gokaslan, Edgar Marroquin, Justin Chiu, Alexander Rush, and V olodymyr Kuleshov. Simple and effective masked diffusion language models.Advances in Neural Information Processing Systems, 37:130136–130184, 2024

    2024

  44. [44]

    Generative modeling by estimating gradients of the data distribution.Advances in neural information processing systems, 32, 2019

    Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution.Advances in neural information processing systems, 32, 2019

    2019

  45. [45]

    Mmseqs2 enables sensitive protein sequence searching for the analysis of massive data sets.Nature biotechnology, 35(11):1026–1028, 2017

    Martin Steinegger and Johannes Söding. Mmseqs2 enables sensitive protein sequence searching for the analysis of massive data sets.Nature biotechnology, 35(11):1026–1028, 2017

    2017

  46. [46]

    Suzek, Hongzhan Huang, Peter McGarvey, Raja Mazumder, and Cathy H

    Baris E. Suzek, Hongzhan Huang, Peter McGarvey, Raja Mazumder, and Cathy H. Wu. UniRef: comprehensive and non-redundant UniProt reference clusters.Bioinformatics, 23(10):1282– 1288, May 2007. ISSN 1367-4803. doi: 10.1093/bioinformatics/btm098. URL https: //doi.org/10.1093/bioinformatics/btm098

    doi:10.1093/bioinformatics/btm098 2007

  47. [47]

    UniProt: the Universal Protein Knowledgebase in 2023.Nucleic Acids Research, 51(D1):D523–D531, January 2023

    The UniProt Consortium. UniProt: the Universal Protein Knowledgebase in 2023.Nucleic Acids Research, 51(D1):D523–D531, January 2023. ISSN 0305-1048. doi: 10.1093/nar/gkac1052. URLhttps://doi.org/10.1093/nar/gkac1052

    doi:10.1093/nar/gkac1052 2023

  48. [48]

    Fast and accurate protein structure search with foldseek.Nature biotechnology, 42(2):243–246, 2024

    Michel Van Kempen, Stephanie S Kim, Charlotte Tumescheit, Milot Mirdita, Jeongjae Lee, Cameron LM Gilchrist, Johannes Söding, and Martin Steinegger. Fast and accurate protein structure search with foldseek.Nature biotechnology, 42(2):243–246, 2024

    2024

  49. [49]

    A comprehensive review of protein language models.arXiv preprint arXiv:2502.06881, 2025

    Lei Wang, Xudong Li, Han Zhang, Jinyi Wang, Dingkang Jiang, Zhidong Xue, and Yan Wang. A comprehensive review of protein language models.arXiv preprint arXiv:2502.06881, 2025. 13

    Pith/arXiv arXiv 2025

  50. [50]

    Diffusion Language Models Are Versatile Protein Learners, October 2024

    Xinyou Wang, Zaixiang Zheng, Fei Ye, Dongyu Xue, Shujian Huang, and Quanquan Gu. Diffusion Language Models Are Versatile Protein Learners, October 2024. URL http:// arxiv.org/abs/2402.18567. arXiv:2402.18567 [cs]

    Pith/arXiv arXiv 2024

  51. [51]

    De novo design of protein structure and function with rfdiffusion.Nature, 620(7976):1089–1100, 2023

    Joseph L Watson, David Juergens, Nathaniel R Bennett, Brian L Trippe, Jason Yim, Helen E Eisenach, Woody Ahern, Andrew J Borst, Robert J Ragotte, Lukas F Milles, et al. De novo design of protein structure and function with rfdiffusion.Nature, 620(7976):1089–1100, 2023

    2023

  52. [52]

    Fast protein backbone generation with se (3) flow matching.arXiv preprint arXiv:2310.05297, 2023

    Jason Yim, Andrew Campbell, Andrew YK Foong, Michael Gastegger, José Jiménez-Luna, Sarah Lewis, Victor Garcia Satorras, Bastiaan S Veeling, Regina Barzilay, Tommi Jaakkola, et al. Fast protein backbone generation with se (3) flow matching.arXiv preprint arXiv:2310.05297, 2023

    Pith/arXiv arXiv 2023

  53. [53]

    Se (3) diffusion model with application to protein backbone generation.arXiv preprint arXiv:2302.02277, 2023

    Jason Yim, Brian L Trippe, Valentin De Bortoli, Emile Mathieu, Arnaud Doucet, Regina Barzilay, and Tommi Jaakkola. Se (3) diffusion model with application to protein backbone generation.arXiv preprint arXiv:2302.02277, 2023

    Pith/arXiv arXiv 2023

  54. [54]

    Scoring function for automated assessment of protein structure template quality.Proteins: Structure, Function, and Bioinformatics, 57(4):702–710,

    Yang Zhang and Jeffrey Skolnick. Scoring function for automated assessment of protein structure template quality.Proteins: Structure, Function, and Bioinformatics, 57(4):702–710,

  55. [55]

    doi: 10.1002/prot.20264

    doi:10.1002/prot.20264

  56. [56]

    ut(x)· ∇logf(x) + 1 2 σ2 t ∇ · ∇logf(x) + Z [logf(y)−logf(x)]Q(dy|x) −˜ut(x)· ∇f(x) f(x) − 1 2 σ2 t ∇ · ∇f(x) f(x) − Z f(y) f(x) −1 ˜Q(dy|x) # (55) =E x∼pt

    Yang Zhang and Jeffrey Skolnick. Tm-align: a protein structure alignment algorithm based on the tm-score.Nucleic acids research, 33(7):2302–2309, 2005. 14 Supplementary Material Contents A Theoretical Details 16 A.1 Proof for Theorem 3.1 (Conditional Rate) . . . . . . . . . . . . . . . . . . . . . . 16 A.2 Proof for Theorem 3.2 (Marginal Rate) . . . . . ....

    2005

  57. [57]

    All training was carried out on 4 NVIDIA RTX PRO 6000 GPUs, taking approximately 3 days. Sampling.Generation proceeds by simulating the learned CTMC from the empty sequence ( t= 0 ) to the data distribution (t= 1 ) over a uniform time grid with step size τ= 1/T (Euler solver). Each step independently samples the number of insertions at each position i fro...

  58. [58]

    Inserted token identities are drawn from Qins tk,i(·|Xtk)

    Insertion sub-step.For each position i of the current sequence Xtk, we sample insertions from a Poisson distribution with intensity (1 +α)λ ins tk,i(Xtk)τ. Inserted token identities are drawn from Qins tk,i(·|Xtk). This produces an intermediate stateYthat overshoots byατin time

  59. [59]

    XtX i=0 λi t − XtX i=0 X k∈Ωi λeff k,t logλ i t # ,(98) and the modified reconstruction loss is: Lloc rec =E Y1,Yt

    Deletion sub-step.For each position i of Y , we sample whether to delete that position with probability α˜λdel t′ k,i(Y)τ , where t′ k =t k + (1 +α)τ and ˜λdel is the learned reverse (deletion) rate. Deleted positions are removed, yielding Xtk+1. This pulls back by ατ, so the net time advancement isτ. We further apply classifier-free guidance (CFG) indepe...