Pith. sign in

REVIEW 3 major objections 7 minor 132 references

Context-parallel inference lets RFdiffusion 3 design full icosahedral and octahedral protein nanoparticles end-to-end without retraining.

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

2026-07-12 02:28 UTC pith:M55F2FID

load-bearing objection Solid systems port of context parallelism to RFD3 design; the scaling and commodity-GPU demos are real, but design quality is still only in-silico under a hard symmetry prior. the 3 major comments →

arxiv 2607.05439 v1 pith:M55F2FID submitted 2026-07-03 cs.LG cs.DCq-bio.QM

Design-CP: Context Parallelism for Design of Protein Nanoparticles

Lorenzo Tarricone , Helen E. Eisenach , Aiko Muraishi , Charlotte M. Deane This is my paper
classification cs.LG cs.DCq-bio.QM
keywords context parallelismprotein designRFdiffusion 3protein nanoparticlesmulti-GPU inferencepoint-group symmetryicosahedral assembliesring attention
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.

All-atom generative protein models can jointly model large multimeric complexes, but their quadratic token- and atom-pair activations quickly exceed single-GPU memory as chain and residue counts grow. Design-CP introduces two weight-preserving context-parallel inference schemes for RFdiffusion 3—1D row-sharding and 2D grid sharding with ring attention—that distribute those activations across a multi-GPU mesh. The paper shows that strong point-group symmetry collapses free coordinates to one asymmetric subunit while still modelling full inter-subunit contacts, making sampling beyond the model’s native training crop usable for closed cages. Maximum feasible subunit size grows with the expected square-root trend in GPU count, 2D sharding scales better in wall-clock time, and in-silico backbone and interface metrics for icosahedral designs track a natural nanoparticle baseline. The same approach designs octahedral cages on a small cluster of 16 GB workstation GPUs, arguing that large-assembly design need not stay limited to high-memory hardware or rigid-body docking pipelines.

Core claim

Design-CP shows that two context-parallel inference strategies for RFdiffusion 3 can shard the dominant O(I²) and O(L²) pair representations across a multi-GPU mesh while preserving pretrained weights, and that strong point-group symmetry makes this sufficient for end-to-end all-atom design of large icosahedral and octahedral nanoparticles with favourable in-silico structural and interface metrics, including on workstation-grade 16 GB GPUs.

What carries the argument

Design-CP: 1D row-sharding of pair tracks (query stripes with keys/values replicated via ALLGATHER) and 2D √P×√P grid sharding with ring attention. Both keep per-device pair memory O(I²/P) and O(L²/P) and leave model weights unchanged.

Load-bearing premise

Strong point-group symmetry is enough to keep designs protein-like when sampling far beyond the model’s training crop, and favourable clash and contact numbers are a reliable stand-in for real self-assembly.

What would settle it

Express and structurally characterise a Design-CP icosahedral or octahedral nanoparticle; if it fails to assemble as designed, or if same-size asymmetric multi-chain designs routinely match the symmetric ones in quality, the central usability claim fails.

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

If this is right

  • Maximum feasible asymmetric-subunit size grows roughly with the square root of GPU count.
  • Joint all-atom generation of closed cages becomes feasible without docking prebuilt oligomers.
  • Workstation-grade multi-GPU clusters can run designs previously limited to high-memory accelerators.
  • The same sharding patterns can, in principle, be reused by other attention-dominated protein design models.
  • Symmetry-constrained sampling can blunt quality collapse when inputs exceed the native training crop.

Where Pith is reading between the lines

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

  • If symmetry is the main quality stabiliser, large asymmetric or weakly symmetric complexes will still need longer-context training before CP alone is enough.
  • Clash and contact metrics may not predict expression or correct self-assembly, so experimental validation of designed cages is the decisive next test.
  • Composing Design-CP with single-device IO-efficient attention could further lower the per-GPU memory floor.
  • Broader access to large-assembly design could accelerate vaccine-scaffold and nanocage work outside well-resourced labs while also widening dual-use reach.

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

3 major / 7 minor

Summary. The paper introduces Design-CP, two weight-preserving context-parallel inference schemes for RFdiffusion 3 (1D row-sharding of pair tracks and 2D grid sharding with ring attention, following Fold-CP) that distribute O(I²) token-pair and O(L²) atom-pair activations across a multi-GPU mesh. The authors characterise memory ceilings and wall-clock scaling on icosahedral sampling (Table 1), arguing that max ASU length grows with the expected square-root trend in GPU count and that 2D sharding has better time scaling. They further claim that strong point-group symmetry (ASU extraction and deterministic group copies for ~90% of denoising steps) makes CP usable out of the box for end-to-end all-atom generation of icosahedral and octahedral nanoparticles, reporting favourable in-silico backbone and interface metrics (Figs. 2–3; Appendices D–F), including octahedral designs on 16×16 GB workstation GPUs, without retraining.

Significance. If the systems results hold, this is a useful and timely contribution: it removes a practical single-GPU memory barrier for AF3-class generative design models and demonstrates large symmetric assemblies on commodity hardware. Strengths include weight-preserving inference (no fine-tuning), explicit scaling measurements matching the predicted O(I²/P) memory trend, a clear 1D vs 2D wall-clock comparison, detailed engineering appendices (collectives, process_a determinism guard, distributed kNN, DTensor layout), and an honest control removing symmetry at fixed system size (Appendix G). The democratisation angle—octahedral cages on 16 GB GPUs—is practically meaningful for groups without large-memory accelerators. The work is primarily systems/methods rather than a new generative model; its lasting value depends on how carefully the design-quality claims are scoped relative to forced resymmetrisation and purely geometric metrics.

major comments (3)
  1. The central usability claim (§1 contributions; §4.1–4.2; Abstract)—that strong point-group symmetry makes Design-CP usable out of the box for end-to-end nanoparticle design with favourable metrics—rests almost entirely on geometric in-silico checks under forced resymmetrisation (sym_step_frac default 0.9; Appendix A.2). Appendix G shows that removing the symmetry prior at identical size (60×210) yields visibly degenerate structures, so sample quality is driven by the hard ASU constraint rather than joint generative capacity beyond the native crop (384 tokens / 5000 atoms). Under that constraint free coordinates are essentially those of one ASU; reported interfaces are largely geometric consequences of group copies after a full forward pass. No sequence design, refolding oracle, or experimental expression is reported. For a journal claim of “design,” at minimum the manuscript should (i) r
  2. §4.2 and Appendix E: the claim that Design-CP “broadly preserves per-chain backbone quality on par with vanilla RFD3 monomers” is only partially supported. Chain breaks and non-loop fraction are similar, but max Cα–Cα deviation is systematically worse for icosahedral ASUs, backbone-clash outliers are more frequent, and secondary-structure composition shifts strongly (helix fraction collapses near zero; sheet fraction rises to ~0.5) relative to both single-GPU monomers and the 1NQX reference. The authors attribute this to geometric strain of the joint symmetric context, which is plausible, but then the monomer comparison is not an apples-to-apples control for sampler fidelity. The main text should foreground this β-bias and the harder problem setting, and avoid language that implies quality parity with native-crop RFD3 monomers.
  3. §4.2–4.4 and related-work framing: the paper positions Design-CP as the first end-to-end all-atom generative design of symmetric assemblies that models full inter-ASU interactions without docking. That is directionally fair as a systems advance, but there is no quantitative comparison to modern dock-and-design / RFdiffusion-oligomer + docking pipelines (e.g., Haas et al. 2025/2026) on the same geometric metrics or on any designability proxy. Without that baseline, “favourable vs 1NQX” does not establish that joint CP sampling improves over the paradigm the introduction seeks to replace. A small head-to-head on interface metrics or a clearer “systems enabler, not yet a design-quality benchmark” framing would make the contribution load-bearing claim more defensible.
minor comments (7)
  1. Acknowledgements: “Ellision Institute” appears to be a typo for “Ellison Institute.”
  2. Table 1: for P=1 the 1D and 2D max ASU lengths are both 58 with times ~1023–1025 s; a brief note that 2D degenerates to single-device behaviour (or is not used) would avoid confusion about why 2D is listed at P=1.
  3. §3.3 / Table 1: 2D requires perfect-square P; non-square rows are dashed. Stating explicitly in the table caption that 2D is undefined for non-square P would help readers scanning the table.
  4. Figure 1b and Appendix G are important controls; cross-referencing Appendix G more prominently in §4.1 (not only as a chain-length check) would strengthen the symmetry-regularisation argument.
  5. Metric thresholds (τ_clash=3.5 Å, contact band 4–10 Å, chain-break 0.75 Å) are defined in Appendix D; a one-line pointer in the main-text figure captions would improve self-contained readability.
  6. n=12 for octahedral designs (§4.4, Appendix F) is small for distributional claims; please state this limitation when interpreting secondary-structure breadth and interface tails.
  7. Code release is promised for a future RFD3 repository update (Appendix C); for journal standards, a frozen commit or archive at acceptance would aid reproducibility of the CP kernels and determinism guards.

Circularity Check

0 steps flagged

No significant circularity: empirical systems evaluation with measured memory/time scaling and geometric design metrics, not a derivation that redefines its targets.

full rationale

Design-CP is an inference-systems paper: it shards RFD3’s O(I²)/O(L²) pair activations across GPUs (1D row-sharding and 2D ring attention), measures max ASU size before OOM and wall-clock time, and reports in-silico backbone/interface metrics on symmetry-constrained samples. The square-root capacity trend is the standard consequence of fixed per-GPU memory under O(I²/P) sharding and is checked empirically against OOM, not fitted and re-labeled as a prediction. Symmetry (ASU extraction plus deterministic group copies for ~90% of steps) is an imposed sampling constraint from stock RFD3, not a parameter fit that forces clash/contact scores by construction—those scores can and do fail for some designs, and Appendix G shows quality collapses without the prior. Citations to Fold-CP and RFD3 supply the base architecture and ring algorithms; they are not uniqueness theorems or self-citation chains that make the reported scaling or metrics true by definition. There is no self-definitional loop, no fitted input called a prediction, and no renaming of a known empirical law as a first-principles result. Weaknesses (OOD sampling, in-silico-only validation, quality driven by the hard symmetry prior) are correctness/scope risks, not circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

The central claim rests on standard distributed-systems math (sharding O(n²) activations by P), pretrained RFD3 behavior under symmetrised sampling, and the domain premise that in-silico geometric metrics plus symmetry suffice to call large CP samples ‘designs.’ Free parameters are mostly stock RFD3/EDM hyperparameters and evaluation thresholds, not newly fitted constants that define the result. No new physical entities are postulated; Design-CP is an engineering method name.

free parameters (5)
  • sym_step_frac (default 0.9)
    Fraction of the denoising trajectory during which ASU resymmetrisation is enforced; chosen by configuration, not derived, and load-bearing for the claim that symmetry makes OOD CP usable.
  • RFD3 native crop limits (384 tokens, 5000 atoms)
    Training-regime boundary taken from RFD3; all large-assembly claims are relative to exceeding this crop without fine-tuning.
  • Interface metric thresholds (τ_clash=3.5 Å, contact band 4–10 Å, τ_prox=15 Å, chain-break 0.75 Å)
    Hand-chosen geometric cutoffs that define ‘favourable’ interface and backbone metrics in Appendix D; changing them would reclassify designs.
  • EDM / Karras schedule defaults (T=200, σ_data=16, s_min/s_max, p=7, γ0=0.6, step scale 1.5)
    Stock sampling hyperparameters that control trajectory quality; not re-tuned here but required for the reported samples.
  • Atom kNN budget k (default 128) and inter-chain quota
    Sparse atom-attention neighbour budget that trades accuracy for memory; affects whether inter-ASU interactions are actually seen under CP.
axioms (5)
  • standard math Steady-state pair memory per device scales as O(I²/P) (and O(L²/P) or O(Lk/P) with sparsification) under both 1D and 2D sharding.
    Standard context-parallel memory accounting used to predict square-root growth of max ASU with GPU count (§3.2–3.3, §4.3).
  • domain assumption Weight-preserving redistribution of activations yields numerically equivalent (or bit-identical under determinism guards) inference to single-GPU RFD3 for the same seed and trajectory.
    Assumed via ALLGATHER/broadcast and DTensor replication; 1D needs an explicit process_a broadcast because index_reduce is nondeterministic (Appendix B.3).
  • ad hoc to paper Strong point-group symmetry sufficiently regularizes sampling outside the training crop so that joint all-atom generation remains protein-like without retraining.
    Core usability claim of §4.1–4.2; supported by qualitative contrast to unsymmetric large systems (Fig. 1b, Appendix G) but not proven generally.
  • domain assumption In-silico backbone and Cα interface metrics are adequate primary quality criteria when structure-prediction oracles degrade with chain count.
    Explicit methodological choice in §4.2 citing AF-Multimer degradation; substitutes for experimental or refolding validation.
  • domain assumption RFD3’s pretrained pair-biased attention and EDM loop remain valid when pair tensors are only ever stored as shards/quadrants.
    Required for ‘no retraining’ claim; relies on exact reformulation of attention as cross-attention or ring attention.
invented entities (1)
  • Design-CP (1D row-sharding + 2D grid CP for RFD3 design) no independent evidence
    purpose: Name the inference-time parallelisation package that makes large symmetric generative design feasible on multi-GPU meshes.
    Method branding rather than a new physical object; independent evidence is the scaling tables and design metrics in the paper itself, not an external measurement.

pith-pipeline@v1.1.0-grok45 · 36360 in / 3985 out tokens · 34284 ms · 2026-07-12T02:28:41.804911+00:00 · methodology

0 comments
read the original abstract

Many all-atom generative protein models can in principle design large multimeric complexes by jointly modelling all chains, but their quadratic token- and atom-pair representations quickly exceed single-GPU memory as the number of chains and residues modelled grows. We introduce Design-CP, two context-parallel (CP) inference strategies for RFdiffusion 3 (1D row-sharding and 2D grid sharding with ring attention) that distribute the quadratic activations across a multi-GPU mesh while preserving pretrained weights. We characterise their scaling when sampling icosahedral assemblies, showing that the maximum feasible asymmetric subunit (ASU) size grows with the expected square-root trend in GPU count and that 2D sharding achieves better wall-clock scaling. Moreover, we show how strong point-group symmetry constraints make CP usable out of the box for end-to-end, all-atom design of icosahedral nanoparticles, yielding favourable in silico structural and interface metrics. Finally, we demonstrate octahedral nanoparticle design on a small cluster of workstation-grade 16GB GPUs, illustrating how Design-CP can be a practical path towards democratising large-assembly protein design.

Figures

Figures reproduced from arXiv: 2607.05439 by Aiko Muraishi, Charlotte M. Deane, Helen E. Eisenach, Lorenzo Tarricone.

Figure 1
Figure 1. Figure 1: Symmetric design with Design-CP.a, Schematic depiction of the sharding techniques implemented in Design-CP. Every square represents a sub-tensor of the self-attention matrix, and its colour represents its assigned device. 1D sharding partitions the queries across different GPUs, where they are used to calculate cross-attention against all keys. 2D sharding partitions both queries and keys and uses ring att… view at source ↗
Figure 2
Figure 2. Figure 2: Designing icosahedral nanoparticles with Design-CP. a, Depiction of the ASU (red) and its eight nearest neighbours in the icosahedral assembly (cyan), shown schematically (left) and annotated on a naturally occurring icosahedral protein nanoparticle: Lumazine Synthase (right, PDB: 1NQX), which has been used as a scaffold for vaccine development. b–e, Per-chain comparison of Design-CP-generated icosahedral … view at source ↗
Figure 3
Figure 3. Figure 3: Designing octahedral nanoparticles on workstation-grade GPUs. a, A representative Design-CP octahedral assembly with 176 residues per chain (24 chains, 4,224 residues total) generated on 16 NVIDIA RTX A4000 GPUs (16 GB per device). b–e, Backbone-sanity metrics over n = 12 such designs: b chain breaks, c backbone clashes, d non-loop fraction, e max CA deviation. f–g, Symmetry-interface metrics: f minimum in… view at source ↗
Figure 4
Figure 4. Figure 4: Per-chain comparison of Design-CP icosahedral designs against vanilla RFD3 monomers (full eight panels). Distributions of standard backbone-sanity and composition metrics, computed chain-by-chain on n = 40 Design-CP icosahedral assemblies (blue, 60 chains × 210 residues per chain) and n = 40 vanilla single-GPU RFD3 monomers of length 210 (green). The dotted line in each panel reports the corresponding valu… view at source ↗
Figure 5
Figure 5. Figure 5: Full distributions of in silico metrics for octahedral designs (n = 12). Headline metrics from Figure 3b–g are reproduced here together with the additional panels discussed in this appendix. (a) Backbone-sanity metrics: chain breaks, backbone clashes, non-loop fraction, max CA deviation, helix fraction, sheet fraction, radius of gyration, alanine content, glycine content. (b) Symmetry-interface metrics: AS… view at source ↗
Figure 6
Figure 6. Figure 6: Effect of removing the symmetry constraint at the same system size. Six Design-CP samples generated with 60 chains of 210 residues each (12,600 residues total) under no point-group symmetry constraint, with each colour denoting a distinct chain. Compare with the well-formed icosahedral nanoparticles of Figure 2f. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_6.png] 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

132 extracted references · 68 canonical work pages · 8 internal anchors

  1. [1]

    and Steinegger, Martin and Kucukbenli, Emine and Vahdat, Arash and Kreis, Karsten , month = mar, year =

    Didi, Kieran and Zhang, Zuobai and Zhou, Guoqing and Reidenbach, Danny and Cao, Zhonglin and Cha, Sooyoung and Geffner, Tomas and Dallago, Christian and Tang, Jian and Bronstein, Michael M. and Steinegger, Martin and Kucukbenli, Emine and Vahdat, Arash and Kreis, Karsten , month = mar, year =. Scaling. doi:10.48550/arXiv.2603.27950 , abstract =

    doi:10.48550/arxiv.2603.27950

  2. [2]

    Boltz-2:

    Passaro, Saro and Corso, Gabriele and Wohlwend, Jeremy , month = jun, year =. Boltz-2:

  3. [3]

    Science , author =

    Scalable emulation of protein equilibrium ensembles with generative deep learning , volume =. Science , author =. 2025 , note =. doi:10.1126/science.adv9817 , abstract =

    doi:10.1126/science.adv9817 2025

  4. [4]

    and Dutordoir, Vincent and Mathieu, Emile and Komorowska, Urszula J

    Didi, Kieran and Vargas, Francisco and Mathis, Simon V. and Dutordoir, Vincent and Mathieu, Emile and Komorowska, Urszula J. and Lio, Pietro , month = mar, year =. A framework for conditional diffusion modelling with applications in motif scaffolding for protein design , url =. doi:10.48550/arXiv.2312.09236 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2312.09236

  5. [5]

    Fast and

    Lin, Yeqing and Lee, Minji and Vermani, Aakarsh and Jiang, Ellena and Cooman, Sebastiaan De and Špeťko, Matej and AlQuraishi, Mohammed , month = may, year =. Fast and. doi:10.64898/2026.05.01.722168 , abstract =

    work page doi:10.64898/2026.05.01.722168 2026

  6. [6]

    and Guo, Daniel and Viggiano, Ben and Brennan-Almaraz, Lucas and Hur, Evelyn and Mai, Tina and Yin, Peter and King, Samuel H

    Merchant, Aditi T. and Guo, Daniel and Viggiano, Ben and Brennan-Almaraz, Lucas and Hur, Evelyn and Mai, Tina and Yin, Peter and King, Samuel H. and Ashley, Euan A. and Hie, Brian L. , month = jun, year =. A high-level programming language for generative biology with. doi:10.64898/2026.06.22.733870 , abstract =

    work page doi:10.64898/2026.06.22.733870 2026

  7. [7]

    and Schaeffer, R

    Baek, Minkyung and DiMaio, Frank and Anishchenko, Ivan and Dauparas, Justas and Ovchinnikov, Sergey and Lee, Gyu Rie and Wang, Jue and Cong, Qian and Kinch, Lisa N. and Schaeffer, R. Dustin and Millán, Claudia and Park, Hahnbeom and Adams, Carson and Glassman, Caleb R. and DeGiovanni, Andy and Pereira, Jose H. and Rodrigues, Andria V. and van Dijk, Alberd...

    doi:10.1126/science.abj8754

  8. [8]

    and Baranov, Max and Costello, Zak and Barber, Karl W

    Ingraham, John B. and Baranov, Max and Costello, Zak and Barber, Karl W. and Wang, Wujie and Ismail, Ahmed and Frappier, Vincent and Lord, Dana M. and Ng-Thow-Hing, Christopher and Van Vlack, Erik R. and Tie, Shan and Xue, Vincent and Cowles, Sarah C. and Leung, Alan and Rodrigues, João V. and Morales-Perez, Claudio L. and Ayoub, Alex M. and Green, Robin ...

    doi:10.1038/s41586-023-06728-8

  9. [9]

    and Anishchenko, I

    Dauparas, J. and Anishchenko, I. and Bennett, N. and Bai, H. and Ragotte, R. J. and Milles, L. F. and Wicky, B. I. M. and Courbet, A. and de Haas, R. J. and Bethel, N. and Leung, P. J. Y. and Huddy, T. F. and Pellock, S. and Tischer, D. and Chan, F. and Koepnick, B. and Nguyen, H. and Kang, A. and Sankaran, B. and Bera, A. K. and King, N. P. and Baker, D....

    doi:10.1126/science.add2187

  10. [10]

    2024 , pages =

    bioRxiv , author =. 2024 , pages =. doi:10.1101/2024.07.09.602403 , abstract =

    work page doi:10.1101/2024.07.09.602403 2024

  11. [11]

    and Bambrick, Joshua and Bodenstein, Sebastian W

    Abramson, Josh and Adler, Jonas and Dunger, Jack and Evans, Richard and Green, Tim and Pritzel, Alexander and Ronneberger, Olaf and Willmore, Lindsay and Ballard, Andrew J. and Bambrick, Joshua and Bodenstein, Sebastian W. and Evans, David A. and Hung, Chia-Chun and O’Neill, Michael and Reiman, David and Tunyasuvunakool, Kathryn and Wu, Zachary and Žemgul...

    doi:10.1038/s41586-024-07487-w

  12. [12]

    and Lee, Jin Sub and Bruguera, Elise S

    Candido, Salvatore and Hayes, Thomas and Derry, Alexander and Rao, Roshan and Lin, Zeming and Verkuil, Robert and Wu, Bryan Z. and Lee, Jin Sub and Bruguera, Elise S. and Keval, Jehan A. and Kopylov, Mykhailo and Pak, John E. and Wu, Wesley and Thomas, Neil and Mataraso, Samson and Hsu, Alvin and Trotman-Grant, Ashton C. and Fatras, Kilian and Costa, Alla...

    work page doi:10.64898/2026.06.03.729735 2026

  13. [13]

    , month = dec, year =

    Raghu, Rishwanth and Levy, Axel and Wetzstein, Gordon and Zhong, Ellen D. , month = dec, year =. Multiscale guidance of protein structure prediction with heterogeneous cryo-. doi:10.48550/arXiv.2506.04490 , abstract =

    doi:10.48550/arxiv.2506.04490

  14. [14]

    arXiv.org , author =

  15. [16]

    doi:10.48550/arXiv.2506.19820 , abstract =

    Faltings, Felix and Stark, Hannes and Barzilay, Regina and Jaakkola, Tommi , month = jun, year =. doi:10.48550/arXiv.2506.19820 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2506.19820

  16. [17]

    Yang, Wei and Wang, Shunzhi and Lee, Gyu Rie and Zhang, Jason Z. and Courbet, Alexis and Juergens, David and Wang, Xinru and Schlichthaerle, Thomas and Abedi, Mohamad and Ragotte, Robert and An, Linna and Kalvet, Indrek and Pellock, Sam and Mihaljevic, Ljubica and Glasscock, Cameron and Pillai, Arvind and Broerman, Adam and Ennist, Nathan and Haefner, Ell...

    work page doi:10.1038/s41586-026-10328-7

  17. [18]

    Arithmetic in the

    Feucht, Sheridan and Haklay, Tal and Bhalla, Usha and Wurgaft, Daniel and Rager, Can and Sarfati, Raphaël and Merullo, Jack and McGrath, Thomas and Lewis, Owen and Lubana, Ekdeep Singh and Fel, Thomas and Geiger, Atticus , month = may, year =. Arithmetic in the. doi:10.48550/arXiv.2605.01148 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2605.01148

  18. [19]

    doi:10.1101/2025.02.27.640585 , abstract =

    Feldman, Jonathan and Skolnick, Jeffrey , month = jul, year =. doi:10.1101/2025.02.27.640585 , abstract =

    work page doi:10.1101/2025.02.27.640585 2025

  19. [20]

    Adams, Etowah and Bai, Liam and Lee, Minji and Yu, Yiyang and AlQuraishi, Mohammed , month = jun, year =. From. doi:10.1101/2025.02.06.636901 , abstract =

    doi:10.1101/2025.02.06.636901 2025

  20. [21]

    Induction Meets Biology: Mechanisms of Repeat Detection in Protein Language Models

    Kesten-Pomeranz, Gal and Nikankin, Yaniv and Reusch, Anja and Tsaban, Tomer and Schueler-Furman, Ora and Belinkov, Yonatan , year =. Induction. doi:10.48550/ARXIV.2602.23179 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2602.23179

  21. [22]

    doi:10.64898/2026.04.22.720175 , abstract =

    Feldman, Jonathan and Skolnick, Jeffrey , month = apr, year =. doi:10.64898/2026.04.22.720175 , abstract =

    work page doi:10.64898/2026.04.22.720175 2026

  22. [23]

    doi:10.48550/arXiv.2503.05025 , abstract =

    Stark, Hannes and Jing, Bowen and Geffner, Tomas and Yim, Jason and Jaakkola, Tommi and Vahdat, Arash and Kreis, Karsten , month = mar, year =. doi:10.48550/arXiv.2503.05025 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.05025

  23. [24]

    Generative

    Campbell, Andrew and Yim, Jason and Barzilay, Regina and Rainforth, Tom and Jaakkola, Tommi , month = jun, year =. Generative. doi:10.48550/arXiv.2402.04997 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.04997

  24. [25]

    Liu, Xingchao and Gong, Chengyue and Liu, Qiang , month = sep, year =. Flow. doi:10.48550/arXiv.2209.03003 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2209.03003

  25. [26]

    Han, Xiaolong and Wang, Zehong and Zhao, Bo and Zhang, Binchi and Li, Jundong and Borth, Damian and Yu, Rose and Maron, Haggai and Ye, Yanfang and Yin, Lu and Neri, Ferrante , month = mar, year =. A. doi:10.48550/arXiv.2603.10090 , abstract =

    doi:10.48550/arxiv.2603.10090

  26. [27]

    Liang, Stephen D , year =

  27. [28]

    Towards the explainability of protein language models , copyright =

    Hunklinger, Andrea and Ferruz, Noelia , month = may, year =. Towards the explainability of protein language models , copyright =. Nature Machine Intelligence , publisher =. doi:10.1038/s42256-026-01232-w , abstract =

    work page doi:10.1038/s42256-026-01232-w

  28. [29]

    Song, Jiaming and Meng, Chenlin and Ermon, Stefano , year =

  29. [30]

    Denoising

    Ho, Jonathan and Jain, Ajay and Abbeel, Pieter , month = dec, year =. Denoising. doi:10.48550/arXiv.2006.11239 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2006.11239 2006

  30. [31]

    and Kumar, Abhishek and Ermon, Stefano and Poole, Ben , month = feb, year =

    Song, Yang and Sohl-Dickstein, Jascha and Kingma, Diederik P. and Kumar, Abhishek and Ermon, Stefano and Poole, Ben , month = feb, year =. Score-. doi:10.48550/arXiv.2011.13456 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2011.13456 2011

  31. [32]

    and Yang, John J

    Parsan, Nithin and Yang, David J. and Yang, John J. , month = mar, year =. Towards. doi:10.48550/arXiv.2503.08764 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.08764

  32. [33]

    doi:10.48550/arXiv.2002.05202 , abstract =

    Shazeer, Noam , month = feb, year =. doi:10.48550/arXiv.2002.05202 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2002.05202 2002

  33. [34]

    and Bale, Jacob B

    King, Neil P. and Bale, Jacob B. and Sheffler, William and McNamara, Dan E. and Gonen, Shane and Gonen, Tamir and Yeates, Todd O. and Baker, David , month = jun, year =. Accurate design of co-assembling multi-component protein nanomaterials , volume =. Nature , publisher =. doi:10.1038/nature13404 , abstract =

    work page doi:10.1038/nature13404

  34. [37]

    Generating

    Lin, Yeqing and AlQuraishi, Mohammed , month = jun, year =. Generating. doi:10.48550/arXiv.2301.12485 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2301.12485

  35. [38]

    and Dowling, Quinton and Hsia, Yang and Fries, Chelsea N

    Sheffler, William and Yang, Erin C. and Dowling, Quinton and Hsia, Yang and Fries, Chelsea N. and Stanislaw, Jenna and Langowski, Mark D. and Brandys, Marisa and Li, Zhe and Skotheim, Rebecca and Borst, Andrew J. and Khmelinskaia, Alena and King, Neil P. and Baker, David , month = may, year =. Fast and versatile sequence-independent protein docking for na...

    work page doi:10.1371/journal.pcbi.1010680

  36. [39]

    Sequence

    Li, Shenggui and Xue, Fuzhao and Baranwal, Chaitanya and Li, Yongbin and You, Yang , month = may, year =. Sequence. doi:10.48550/arXiv.2105.13120 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2105.13120

  37. [40]

    Proteina:

    Geffner, Tomas and Didi, Kieran and Zhang, Zuobai and Reidenbach, Danny and Cao, Zhonglin and Yim, Jason and Geiger, Mario and Dallago, Christian and Kucukbenli, Emine and Vahdat, Arash and Kreis, Karsten , month = mar, year =. Proteina:. doi:10.48550/arXiv.2503.00710 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.00710

  38. [41]

    and Liang, Huazhu and Tam, John and Miletic, Sean and Berman, Jacob M

    Ragotte, Robert J. and Liang, Huazhu and Tam, John and Miletic, Sean and Berman, Jacob M. and Palou, Roger and Weidle, Connor and Li, Zhijie and Glögl, Matthias and Beilhartz, Greg L. and Carr, Kenneth D. and Borst, Andrew J. and Coventry, Brian and Wang, Xinru and Rubinstein, John L. and Tyers, Mike and Schramek, Daniel and Melnyk, Roman A. and Baker, Da...

    work page doi:10.1073/pnas.2509329122

  39. [42]

    Vázquez Torres, Susana and Leung, Philip J. Y. and Venkatesh, Preetham and Lutz, Isaac D. and Hink, Fabian and Huynh, Huu-Hien and Becker, Jessica and Yeh, Andy Hsien-Wei and Juergens, David and Bennett, Nathaniel R. and Hoofnagle, Andrew N. and Huang, Eric and MacCoss, Michael J. and Expòsit, Marc and Lee, Gyu Rie and Bera, Asim K. and Kang, Alex and De ...

    work page doi:10.1038/s41586-023-06953-1

  40. [43]

    Nature Methods , publisher =

    Simon, Elana and Zou, James , month = oct, year =. Nature Methods , publisher =. doi:10.1038/s41592-025-02836-7 , abstract =

    doi:10.1038/s41592-025-02836-7

  41. [44]

    Two Stages of Folding: Convergent Mechanisms in AI Protein Folding Trunks

    Lu, Kevin and Brinkmann, Jannik and Huber, Stefan and Mueller, Aaron and Belinkov, Yonatan and Bau, David and Wendler, Chris , year =. Mechanisms of. doi:10.48550/ARXIV.2602.06020 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2602.06020

  42. [45]

    Improving diffusion-based protein backbone generation with global-geometry-aware latent encoding , volume =

    Zhang, Yuyang and Liu, Yuhang and Ma, Zinnia and Li, Min and Xu, Chunfu and Gong, Haipeng , month = jul, year =. Improving diffusion-based protein backbone generation with global-geometry-aware latent encoding , volume =. Nature Machine Intelligence , publisher =. doi:10.1038/s42256-025-01059-x , abstract =

    work page doi:10.1038/s42256-025-01059-x

  43. [46]

    and Divine, Robby and Miranda, Marcos C

    Yang, Erin C. and Divine, Robby and Miranda, Marcos C. and Borst, Andrew J. and Sheffler, Will and Zhang, Jason Z. and Decarreau, Justin and Saragovi, Amijai and Abedi, Mohamad and Goldbach, Nicolas and Ahlrichs, Maggie and Dobbins, Craig and Hand, Alexis and Cheng, Suna and Lamb, Mila and Levine, Paul M. and Chan, Sidney and Skotheim, Rebecca and Fallas,...

    work page doi:10.1038/s41594-024-01288-5

  44. [47]

    and Ueda, George and Fallas, Jorge A

    Divine, Robby and Dang, Ha V. and Ueda, George and Fallas, Jorge A. and Vulovic, Ivan and Sheffler, William and Saini, Shally and Zhao, Yan Ting and Raj, Infencia Xavier and Morawski, Peter A. and Jennewein, Madeleine F. and Homad, Leah J. and Wan, Yu-Hsin and Tooley, Marti R. and Seeger, Franziska and Etemadi, Ali and Fahning, Mitchell L. and Lazarovits,...

    work page doi:10.1126/science.abd9994

  45. [48]

    Lu, Chixiang and Zhong, Yunhua and Liang, Shikang and Qi, Xiaojuan and Jiang, Haibo , year =

  46. [49]

    Lin, Yeqing and Lee, Minji and Zhang, Zhao and AlQuraishi, Mohammed , month = may, year =. Out of. doi:10.48550/arXiv.2405.15489 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2405.15489

  47. [50]

    doi:10.48550/arXiv.2310.02391 , abstract =

    Bose, Avishek Joey and Akhound-Sadegh, Tara and Huguet, Guillaume and Fatras, Kilian and Rector-Brooks, Jarrid and Liu, Cheng-Hao and Nica, Andrei Cristian and Korablyov, Maksym and Bronstein, Michael and Tong, Alexander , month = apr, year =. doi:10.48550/arXiv.2310.02391 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2310.02391

  48. [51]

    Yim, Jason and Campbell, Andrew and Foong, Andrew Y. K. and Gastegger, Michael and Jiménez-Luna, José and Lewis, Sarah and Satorras, Victor Garcia and Veeling, Bastiaan S. and Barzilay, Regina and Jaakkola, Tommi and Noé, Frank , month = oct, year =. Fast protein backbone generation with. doi:10.48550/arXiv.2310.05297 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2310.05297

  49. [52]

    and Grigoryan, Lilit and Navarro, Mary Jane and Catanzaro, Nicholas J

    Hendricks, Grace G. and Grigoryan, Lilit and Navarro, Mary Jane and Catanzaro, Nicholas J. and Hubbard, Miranda L. and Powers, John M. and Mattocks, Melissa and Treichel, Catherine and Walls, Alexandra C. and Lee, Jimin and Ellis, Daniel and Wang, Jing Yang (John) and Cheng, Suna and Miranda, Marcos C. and Valdez, Adian and Chao, Cara W. and Chan, Sidney ...

    work page doi:10.1126/scitranslmed.adu2085

  50. [53]

    Biotechnology Reports , author =

    Second career of a biosynthetic enzyme:. Biotechnology Reports , author =. 2020 , keywords =. doi:10.1016/j.btre.2020.e00494 , abstract =

    work page doi:10.1016/j.btre.2020.e00494 2020

  51. [54]

    ACS nano , author =

    Lumazine. ACS nano , author =. 2025 , keywords =. doi:10.1021/acsnano.5c06081 , abstract =

    work page doi:10.1021/acsnano.5c06081 2025

  52. [55]

    Computationally designed

  53. [57]

    bioRxiv: The Preprint Server for Biology , author =

    De novo design of protein nanoparticles with integrated functional motifs , issn =. bioRxiv: The Preprint Server for Biology , author =. 2026 , pages =. doi:10.64898/2025.12.19.695620 , abstract =

    work page doi:10.64898/2025.12.19.695620 2026

  54. [58]

    , month = dec, year =

    Haas, Cyrus M. , month = dec, year =. De novo design of protein nanoparticles with integrated functional motifs , doi =

  55. [60]

    Proceedings of the National Academy of Sciences , author =

    From sequence to scaffold:. Proceedings of the National Academy of Sciences , author =. 2025 , pages =. doi:10.1073/pnas.2409566122 , abstract =

    work page doi:10.1073/pnas.2409566122 2025

  56. [61]

    Proceedings of the National Academy of Sciences , author =

    Rapid and automated design of two-component protein nanomaterials using. Proceedings of the National Academy of Sciences , author =. 2024 , pages =. doi:10.1073/pnas.2314646121 , abstract =

    work page doi:10.1073/pnas.2314646121 2024

  57. [62]

    Nature , author =

    Design of a hyperstable 60-subunit protein icosahedron , volume =. Nature , author =. 2016 , pages =. doi:10.1038/nature18010 , language =

    work page doi:10.1038/nature18010 2016

  58. [63]

    Science , author =

    Accurate design of megadalton-scale two-component icosahedral protein complexes , volume =. Science , author =. 2016 , pages =. doi:10.1126/science.aaf8818 , abstract =

    work page doi:10.1126/science.aaf8818 2016

  59. [64]

    Science , author =

    Computational. Science , author =. 2012 , pages =. doi:10.1126/science.1219364 , abstract =

    work page doi:10.1126/science.1219364 2012

  60. [65]

    Proceedings of the National Academy of Sciences , author =

    Nanohedra:. Proceedings of the National Academy of Sciences , author =. 2001 , pages =. doi:10.1073/pnas.041614998 , abstract =

    work page doi:10.1073/pnas.041614998 2001

  61. [66]

    Bioinformatics , author =

    P-. Bioinformatics , author =. 1997 , pages =. doi:10.1093/bioinformatics/13.3.291 , abstract =

    work page doi:10.1093/bioinformatics/13.3.291 1997

  62. [67]

    Hallucinating symmetric protein assemblies

  63. [68]

    Predicting the structure of large protein complexes using

    Bryant, Patrick and Pozzati, Gabriele and Zhu, Wensi and Shenoy, Aditi and Kundrotas, Petras and Elofsson, Arne , month = oct, year =. Predicting the structure of large protein complexes using. Nature Communications , publisher =. doi:10.1038/s41467-022-33729-4 , abstract =

    work page doi:10.1038/s41467-022-33729-4

  64. [69]

    Lin, Dejun and Chu, Simon and Iyer, Vishanth and Lee, Youhan and John, John St and Boyd, Kevin and Roland, Brian and Ren, Xiaowei and Zhou, Guoqing and Cao, Zhonglin and Binder, Polina and Zhautouskaya, Yuliya and Zakrzewski, Jakub and Stadler, Maximilian and Gion, Kyle and Peng, Yuxing and Chen, Xi and Zhang, Tianjing and Junk, Philipp and Dimon, Michell...

    work page doi:10.48550/arxiv.2603.14806

  65. [70]

    Efficient

    Duan, Jiangfei and Zhang, Shuo and Wang, Zerui and Jiang, Lijuan and Qu, Wenwen and Hu, Qinghao and Wang, Guoteng and Weng, Qizhen and Yan, Hang and Zhang, Xingcheng and Qiu, Xipeng and Lin, Dahua and Wen, Yonggang and Jin, Xin and Zhang, Tianwei and Sun, Peng , month = jul, year =. Efficient. doi:10.48550/arXiv.2407.20018 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2407.20018

  66. [71]

    doi:10.48550/arXiv.2203.00854 , abstract =

    Cheng, Shenggan and Zhao, Xuanlei and Lu, Guangyang and Fang, Jiarui and Yu, Zhongming and Zheng, Tian and Wu, Ruidong and Zhang, Xiwen and Peng, Jian and You, Yang , month = feb, year =. doi:10.48550/arXiv.2203.00854 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2203.00854

  67. [72]

    and Staats, Charles , month = oct, year =

    Rabe, Markus N. and Staats, Charles , month = oct, year =. Self-attention. doi:10.48550/arXiv.2112.05682 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2112.05682

  68. [73]

    Thorben and Chmiela, Stefan and Müller, Klaus-Robert and Unke, Oliver T

    Frank, J. Thorben and Chmiela, Stefan and Müller, Klaus-Robert and Unke, Oliver T. , month = mar, year =. Machine learning global atomic representations with. Nature Machine Intelligence , publisher =. doi:10.1038/s42256-026-01195-y , abstract =

    work page doi:10.1038/s42256-026-01195-y

  69. [74]

    Efficiently

    Pope, Reiner and Douglas, Sholto and Chowdhery, Aakanksha and Devlin, Jacob and Bradbury, James and Levskaya, Anselm and Heek, Jonathan and Xiao, Kefan and Agrawal, Shivani and Dean, Jeff , month = nov, year =. Efficiently. doi:10.48550/arXiv.2211.05102 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2211.05102

  70. [75]

    Efficient

    Narayanan, Deepak and Shoeybi, Mohammad and Casper, Jared and LeGresley, Patrick and Patwary, Mostofa and Korthikanti, Vijay Anand and Vainbrand, Dmitri and Kashinkunti, Prethvi and Bernauer, Julie and Catanzaro, Bryan and Phanishayee, Amar and Zaharia, Matei , month = aug, year =. Efficient. doi:10.48550/arXiv.2104.04473 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2104.04473

  71. [76]

    Wohlwend, Jeremy and Corso, Gabriele and Passaro, Saro and Reveiz, Mateo and Leidal, Ken and Swiderski, Wojtek and Portnoi, Tally and Chinn, Itamar and Silterra, Jacob and Jaakkola, Tommi and Barzilay, Regina , month = nov, year =. Boltz-1. doi:10.1101/2024.11.19.624167 , abstract =

    doi:10.1101/2024.11.19.624167 2024

  72. [77]

    and Fatehi, Farzad and Mallik, Bhoomika Basu and Antanasijevic, Aleksandar and Borst, Andrew J

    Khmelinskaia, Alena and Bethel, Neville P. and Fatehi, Farzad and Mallik, Bhoomika Basu and Antanasijevic, Aleksandar and Borst, Andrew J. and Lai, Szu-Hsueh and Chim, Ho Yeung and Wang, Jing Yang ‘John’ and Miranda, Marcos C. and Watkins, Andrew M. and Ogohara, Cassandra and Caldwell, Shane and Wu, Mengyu and Heck, Albert J. R. and Veesler, David and War...

    work page doi:10.1038/s41594-025-01490-z

  73. [78]

    and Aziznejad, Shayan and Bixby, Eli , month = apr, year =

    Deutschmann, Nicolas and Ferragu, Constance and Ziegler, Jonathan D. and Aziznejad, Shayan and Bixby, Eli , month = apr, year =. doi:10.48550/arXiv.2603.11703 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2603.11703

  74. [79]

    and Liu, Cheng-Hao , month = apr, year =

    Rector-Brooks, Jarrid and Lambert, Théophile and Skreta, Marta and Roth, Daniel and Long, Yueming and Li, Zi-Qi and Zhang, Xi and Cretu, Miruna and Li, Francesca-Zhoufan and Ganapathy, Tanvi and Jin, Emily and Bose, Avishek Joey and Yang, Jason and Neklyudov, Kirill and Bengio, Yoshua and Tong, Alexander and Arnold, Frances H. and Liu, Cheng-Hao , month =...

    Pith/arXiv arXiv doi:10.48550/arxiv.2604.05181

  75. [80]

    and Kim, Jinho and Cheng, Lucy and El Nesr, Gina and Xu, Minkai and Shuai, Richard W

    Chu, Alexander E. and Kim, Jinho and Cheng, Lucy and El Nesr, Gina and Xu, Minkai and Shuai, Richard W. and Huang, Po-Ssu , month = jul, year =. An all-atom protein generative model , volume =. Proceedings of the National Academy of Sciences , publisher =. doi:10.1073/pnas.2311500121 , abstract =

    doi:10.1073/pnas.2311500121

  76. [81]

    Pillai, Arvind and Idris, Abbas and Philomin, Annika and Weidle, Connor and Skotheim, Rebecca and Leung, Philip J. Y. and Broerman, Adam and Demakis, Cullen and Borst, Andrew J. and Praetorius, Florian and Baker, David , month = aug, year =. De novo design of allosterically switchable protein assemblies , volume =. Nature , publisher =. doi:10.1038/s41586...

    work page doi:10.1038/s41586-024-07813-2

  77. [82]

    and Fu, Xiang and Liao, Yi-Lun and Gharakhanyan, Vahe and Miller, Benjamin Kurt and Sriram, Anuroop and Ulissi, Zachary W

    Joshi, Chaitanya K. and Fu, Xiang and Liao, Yi-Lun and Gharakhanyan, Vahe and Miller, Benjamin Kurt and Sriram, Anuroop and Ulissi, Zachary W. , month = may, year =. All-atom. doi:10.48550/arXiv.2503.03965 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.03965

  78. [83]

    P(all-atom)

    Qu, Wei and Guan, Jiawei and Ma, Rui and Zhai, Ke and Wu, Weikun and Wang, Haobo , month = aug, year =. P(all-atom). doi:10.1101/2024.08.16.608235 , abstract =

    doi:10.1101/2024.08.16.608235 2024

  79. [84]

    Journal of Structural Biology , author =

    Protein tentacles , volume =. Journal of Structural Biology , author =. 2017 , pages =. doi:10.1016/j.jsb.2017.05.012 , abstract =

    work page doi:10.1016/j.jsb.2017.05.012 2017

  80. [85]

    Nature Communications , author =

    Structural puzzles in virology solved with an overarching icosahedral design principle , volume =. Nature Communications , author =. 2019 , pages =. doi:10.1038/s41467-019-12367-3 , abstract =

    work page doi:10.1038/s41467-019-12367-3 2019

Showing first 80 references.