Pith. sign in

REVIEW 3 major objections 7 minor 54 references

ProtPainter: Draw or Drag Protein via Topology-guided Diffusion

T0 review · 3 major / 7 minor · reviewed 2026-08-16 · deepseek-v4-flash

Pith's one-line read ProtPainter claims that a 3D curve is a sufficient condition for controlling protein backbone topology: the pipeline sketches the curve into a scaffold, then steers a denoising diffusion sampler so the backbone fits the curve (scTF > 0.8)…

desk verdict A genuinely useful new conditioning modality for protein backbones, with strong visual demos, but the central guidance derivation (Eq. 13) is wrong and the evaluation is partly self-referential. read the letter →

arxiv 2504.14274 v1 pith:HFEDYIJT submitted 2025-04-19 cs.AI

classification cs.AI
keywords proteinbackbonegenerationdiffusionmodel3Dcurveconditioningtopology-guideddesignsecondarystructureannotationdrawinganddraggingRestorationTaskself-consistencyTopologyFitness
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

ProtPainter's claim is that a simple 3D curve — drawn by hand, dragged from an existing protein, or extracted from one — is a sufficient condition for controlling what fold a diffusion model generates. The paper builds a two-stage pipeline: a CurveEncoder annotates the curve with secondary-structure labels and produces a parametric sketch, and a denoising diffusion sampler is then steered, without retraining, so that the Cα coordinates of the generated backbone track the sketch through a low-pass-filter interpolation rule with a helix-percentage gating schedule. To make the claim testable, the paper introduces the Protein Restoration Task — regenerate designable backbones from curves taken from reference proteins — and a new metric, scTF, which scores topological fit by Procrustes alignment of curve-level abstractions. On that benchmark it reports backbones clearing the authors' thresholds for topology fit (scTF > 0.8) and designability (scTM > 0.5) at rates well above sequence-length, SSE, point-cloud, and latent-topology baselines, and it shows the same curve mechanism supporting drawing, dragging, jointing, and local SSE editing. The reason to care is that this moves structural control from parametric blueprints toward direct geometric manipulation: a designer sketches or edits a shape instead of configuring a list of secondary-structure parameters.

What carries the argument

The load-bearing object is the sketch-guided interpolation rule, Eq. 13: the denoised translation at each reverse step is $$z_{t-1}=\frac{\sqrt{\alpha_{t-1}\beta_t}}{1-\alpha_t}\hat z_0+\frac{\sqrt{\alpha_t(1-\alpha_{t-1})}}{1-\alpha_t}\big((1-\$\lambda$)z'_t+\$\lambda$ y_{t-1}\big),$$ a blend of the RoseTTAFold-estimated clean structure $\hat z_0$ and an interpolation between the unguided current translation $z'_t$ and the sketch coordinates $y_{t-1}$, with $\lambda$ trading diversity against guidance. The rule is derived by adapting the low-pass-filter conditioning of Choi et al. (ILVR): a linear filter $\phi$ (here the Cα-coordinate extraction scaled by $\lambda$) is assumed to pass through the reverse posterior sample, so the conditioned sample can be written as the unconditioned sample with its filtered part replaced by the filtered sketch. A second named mechanism, Helix-Gating, schedules when the sketch is trusted: while the helix percentage of $\hat z_0$ falls below the sketch's, guidance is scaled by $\gamma\,\delta(O(\hat z_0),O(y))+\eta$ (the 'confidential' phase), and full guidance is applied once it catches up. The supporting objects are the curve representation itself — helices and sheets abstracted to central axes, loops preserved, then re-sampled and SSE-annotated — and the new evaluation metric scTF, defined as $1-\text{disparsity}$ under Procrustes alignment of the two down-sampled curves.

What would settle it

Run the base DDPM's reverse transition many times from a fixed noisy backbone $z_t$, apply the frame filter $\phi$ (Cα coordinates scaled by $\lambda$), and compare the empirical distribution of $\phi(z'_{t-1})$ with the right-hand side of Eq. 12. The reverse step injects Gaussian noise of variance $\tilde\beta_t$ that the linear filter does not remove, and the posterior mean also contains a $\phi(\hat z_0)$ term, so the two sides should differ by a measurable residual of order $\lambda\sqrt{\tilde\beta_t}$; a material residual would show that Eq. 13 is a heuristic blend rather than the claimed conditional posterior.

Watch

Extended reading notes

Core claim

On its own terms, the discovery is that the three-dimensional topology of a protein backbone can be specified by a coarse 3D curve, and that a diffusion model can be made to honor that curve by a retraining-free mechanism. The condition curve is abstracted from helices and sheets as their central axes, with loops kept and the whole re-sampled and annotated with SSE labels. ProtPainter then (i) encodes the curve into a naive sketch whose helix pitch, loop spacing, and sheet placement follow parametric rules, and (ii) during DDPM reverse sampling replaces the filtered part of the unguided translation update with the filtered sketch coordinates, blending the two under a scaling factor λ while a Helix-Gating schedule limits guidance until the helix percentage of the current RoseTTAFold prediction catches up to the sketch's. The reported consequence is that generated backbones match the conditioning curve at the topology level — scTF above 0.8 — and remain designable through the standard ProteinMPNN and OmegaFold refolding cycle, with scTM above 0.5 in the majority of benchmark cases, outperforming the compared baselines in fit designability. Because the guidance is a post-hoc interpolation rather than a trained conditioner, the same base sampler also supports curve-based editing operations such as dragging, rotating, jointing, and local SSE editing while preserving designability.

Load-bearing premise

The steering rule assumes that the random noise added at each denoising step disappears once the backbone is smoothed down to its curve-level view, which is what lets the method replace the filtered unguided sample with the filtered sketch; if that assumption fails, the interpolation used for guidance is not drawing samples from the conditional distribution the paper claims.

Editorial extensions

If this is right

  • A designer can specify a protein's fold by drawing or dragging a 3D curve and still land in the standard design cycle: the generated backbones pass through ProteinMPNN sequence design and refold to structures matching the curve, with scTM above 0.5 and high pLDDT in the reported cases.
  • Curve operations — dragging, rotating, jointing, and local SSE editing — yield topology transitions between designable states, so multi-state and hinge-like scaffolds can be explored by editing a curve rather than reconfiguring a parametric blueprint.
  • Because the guidance is a retraining-free interpolation into an existing DDPM sampler, adding a new condition type only requires a curve encoder and filter that align the new condition to the frame space, while the denoiser itself is untouched.
  • The Protein Restoration Task and the scTF metric give the field a common yardstick: any future topology-conditioned generator can be scored by how well its backbones restore curves from reference proteins and survive refolding.
  • Since scTF correlates with scTM in the paper's joint analysis, topology fitness doubles as a rough designability screen: a curve that cannot be restored with high scTF is unlikely to describe a designable backbone.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The guidance acts only on the translation channel and leaves rotations to the unconditional model (Eqs. 9–10); an untested corollary is that Cα placement alone suffices to steer fold topology, which could be checked by comparing residue-orientation quality near the condition curve against the unguided base model.
  • The interpolation rule is derived from a commutation assumption about a linear filter, not from properties specific to DDPM, so the same recipe should transfer to flow-matching backbone generators with a coordinate channel; testing that transfer would show whether the mechanism is tied to the noise schedule.
  • Helix-Gating embodies a general principle — keying condition strength to a coarse geometric descriptor of the current prediction — and the paper's own deferral of β-sheet control suggests the natural extension is gating on sheet or loop content rather than helix percentage.
  • The noise-perturbation results (fit designability holds up to roughly 2 Å of point noise before degrading) imply the generator behaves as an attractor that repairs small drawing errors; if deliberate user studies confirm it, curve interfaces would not need to be precise to be useful.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. The paper proposes ProtPainter, a two-stage diffusion-based method for generating protein backbones conditioned on 3D curves. In the first stage, a CurveEncoder predicts secondary-structure labels from an input curve, and a naive sketch is generated parametrically. In the second stage, a DDPM backbone generator is guided by this sketch via a retraining-free interpolation rule, with a Helix-Gating schedule that controls how strongly the sketch is fused during sampling. The authors introduce the Protein Restoration Task and a self-consistency Topology Fitness (scTF) metric, and report that ProtPainter achieves scTF above 0.8 and scTM above 0.5 on their benchmark, along with demonstrations of drawing, dragging, jointing, and SSE editing.

Significance. If the method is sound, the curve-based conditioning modality is a genuinely useful interface for protein topology control, and the proposed benchmark and scTF metric could support future work in this direction. The scTM-based designability evaluation provides an independent signal that is not reducible to the conditioning objective, and the user-study results with curves drawn from scratch are a valuable practical demonstration. However, the central derivation of the guided-sampling update is incomplete, and parts of the evaluation are self-referential because scTF measures agreement with the same curve that the guidance explicitly targets and because key hyperparameters and thresholds are selected on the evaluation benchmark. The empirical results may still hold, but the paper currently does not establish the theoretical grounding claimed in Section 3.2.

major comments (3)
  1. [Section 3.2, Eq. (12)] Equation (12) does not follow from Equation (5). Applying any linear operation φ to a sample from the posterior in Eq. (5) gives φ(z'_{t-1}) = [√(α_{t-1}β_t)/(1−α_t)] φ(z0) + [√(α_t(1−α_{t-1}))/(1−α_t)] φ(z_t) + φ(σ_t ε), where ε is the posterior sampling noise. The derivation keeps only the middle term, silently dropping the z0 term and the filtered noise; near t=1 the z0 term is O(1) and the noise is not negligible. Since Eq. (13) is the core interpolation rule of the guidance, this omission is load-bearing: the manuscript does not establish that the procedure samples the conditional distribution announced in Eq. (7). The authors should either provide a correct derivation under explicit assumptions or explicitly present Eqs. (11)–(13) as a heuristic guidance rule supported by empirical validation.
  2. [Section 3.2, Eqs. (7), (11), (14)] The conditional distribution q(y_{t-1}|y, ẑ_t^0) in Eq. (7) is never defined, and the filter φ is introduced as a "linear low-pass filtering" operation but then defined as φ_λ(X_t) = λ z(X_t), which is a scalar rescaling of the Cα coordinates, not an idempotent low-pass projection. Because scalar rescaling satisfies φ(φ(z)) = λ² z rather than φ(φ(z)) = φ(z), the ILVR replacement identity in Eq. (11) does not enforce φ(x_{t-1}) = φ(y_{t-1}) as claimed. Additionally, Eq. (14) defines y_{t-1} deterministically as a scaled version of y, rather than as a draw from q. The authors should either supply a correct derivation with a clearly defined idempotent filter and conditional distribution, or reframe the whole guidance mechanism as a heuristic and move the theoretical claim to the empirical evaluation.
  3. [Section 4.1 and Appendices G.5/G.6] The headline evaluation is partly self-referential. scTF measures Procrustes agreement between the generated or refolded backbone curve and the same condition curve that the guidance explicitly tries to match, so high scTF is close to a construction check. The threshold scTF > 0.8 and the hyperparameters λ = 2/3, γ = 0.2, η = 0.7 are selected using the Protein Restoration Task benchmark on which the reported numbers are computed (Table G.5, Figure 13). The scTM and pLDDT results provide an independent designability signal, but the "state-of-the-art control" claim rests on scTF. Please report results with hyperparameters fixed before evaluation or on a held-out split, and show that the reported margins are insensitive to the scTF threshold and to reasonable variations of λ, γ, and η.
minor comments (7)
  1. [Section 3.2, Eq. (9)] The condition "if pθ(r_{t−1}|x_t) = pθ(r_{t−1}|x_t,c_T)" makes the preceding equality tautological; the intended independence assumption between rotation and translation updates should be stated explicitly and justified.
  2. [Section 3.2, Eq. (13)] The notation z′_t in Eq. (13) is inconsistent with z′_{t−1} in Eq. (11); please clarify whether the update uses the unconditional sample at time t−1 or the current noisy coordinate z_t.
  3. [Section 3.3 and Figure 2] The phrase "confidential phase" appears to be a typo for "confidence phase" (and "confphase" in Figure 2); please correct it throughout.
  4. [Table 1] The header of Table 1 lists six CATH IDs but seven numeric entries; the column alignment should be fixed so that each topology family has exactly one count and one set of scTF values.
  5. [References] The citation "Zhang et al." is ambiguous because it is used for multiple distinct works (TopoDiff and the GPCR review); please disambiguate with different citation keys or full author lists.
  6. [Section 3.2, first paragraph] The sentence "The translation part is generated from the 3D Gaussian noise by DDPM" should clarify that translations use the DDPM posterior in Eq. (4) while rotations are sampled by Brownian motion on SO(3); as written it could be read as applying DDPM to the full frame.
  7. [Figure 12.d] The axis label "Number of Inversion steps t" is unclear; please specify whether t increases from noise to clean signal or vice versa, and whether "inversion" refers to the reverse diffusion process.

Circularity Check

1 steps flagged · score 5.0 of 10

scTF>0.8 is largely a construction check: Eq. 13 injects the condition curve into the output and scTF measures agreement with that same curve; scTM designability is independently evaluated.

  1. self definitional [Section 3.2, Eq. 13; Section 4.1, Metrics; Appendix A.3]
    "zt−1 = sqrt(α_{t−1}β_t)/(1−α_t) · zˆt_0 + sqrt(α_t(1−α_{t−1}))/(1−α_t) · (1−λ)·z′_t + λ·y_{t−1}. (13) ... scTF is calculated as the Procrustes similarity between the refolded (or un-refolded) backbone curve and the curve guidance."

    The headline claim “generate topology-fit (scTF > 0.8)” is close to a construction check. Equation 13 explicitly blends the condition curve y into every denoising-translation update through the λ·y_{t−1} term, and scTF is defined as the Procrustes similarity between the generated backbone's abstracted curve and exactly that same condition curve y. Thus the metric reports whether the injected condition survived generation and refolding; a high value is partly enforced by the guidance rule itself rather than being an independent property of the generator. The 0.8 threshold is also calibrated on the authors' own scTF–scTM relationship (Figure 13), so the specific reported cutoff is self-referential. The designability component (scTM, via ProteinMPNN/OmegaFold) is external and not circular.

full rationale

The central circularity is confined to the topology-fitness claim: condition y is inserted directly into the sample path (Eq. 13), and scTF compares the output to that same y, so “scTF > 0.8” is substantially a check that the guidance mechanism does what it was constructed to do. The designability result (scTM > 0.5) is evaluated through independent refolding with ProteinMPNN and OmegaFold, so it provides non-circular evidence. I did not score the Eq. 12–13 derivation gap as circularity: dropping the z0 term and filtered noise, and identifying scalar scaling with the idempotent low-pass filter of ILVR, are correctness/derivation weaknesses rather than cases where a prediction is equivalent to its input by definition. No load-bearing self-citation or author-imported uniqueness theorem appears; the method is benchmarked against external tools and external baselines. Overall, the paper has one partially circular headline metric but retains independent designability content, hence a score of 5 rather than 0 or 8.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The method depends on several domain assumptions inherited from protein structure prediction and on hyperparameters fitted to the evaluation benchmark. No new physical entities are posited; CurveEncoder, Helix-Gating, and the naive sketch are trained or constructed algorithmic components, not postulated phenomena.

free parameters (5)
  • lambda (sketch fusion factor) = 2/3
    Interpolation strength between unconditional sample and sketch guidance. Selected via ablation on HHH_ems in Appendix G.5 and Figure 12.a, then used in all reported results.
  • gamma (Helix-Gating scaling factor) = 0.2
    Grid searched over 0 to 0.3 with eta in Appendix G.5 on the HHH_ems dataset to maximize an average of scTF1, scTM, and scTF2.
  • eta (Helix-Gating offset) = 0.7
    Same grid search as gamma; the pair (gamma=0.2, eta=0.7) is selected from the HHH_ems benchmark and applied to all datasets.
  • scTF threshold for topology fitness = 0.8
    The threshold for declaring a backbone 'topology-fit' is chosen from the proportional relationship between scTF and scTM in Figure 13, i.e., selected on the evaluation data.
  • Curve sampling rate = 40% of C-alpha count
    Chosen from a plateau observed in Appendix G.3 using a single 54-residue protein; this controls how much topology information the condition curve carries.
assumptions (5)
  • domain assumption RoseTTAFold provides sufficiently accurate estimates of z_0 to guide translation updates in Equation 13.
    Adopted from RFDiffusion; no independent verification is provided in this paper beyond the reported scTM results.
  • domain assumption The pretrained unconditional diffusion model remains on the manifold of designable backbones under sketch-guided interpolation.
    The method is retraining-free; no proof is given that guided samples do not drift toward non-designable regions, which is why designability is measured post hoc.
  • domain assumption A downsampled C-alpha curve, consisting of SSE central axes plus loop coordinates, contains enough information to reconstruct a designable protein backbone.
    The entire restoration task rests on this premise; the paper provides empirical scTF versus scTM correlation but no guarantee for novel topologies.
  • domain assumption Procrustes disparity after uniform resampling is a valid monotone proxy for topological similarity and designability.
    scTF is proposed in this paper, and its cutoff is fit to the data rather than justified by an external benchmark.
  • standard math Standard DDPM and SO(3) diffusion mathematics, Equations 1 through 6, are correct.
    Background from Ho et al. and Yim et al.; assumed as established mathematical machinery.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ProtPainter: Draw or Drag Protein via Topology-guided Diffusion." pith.science (2026). https://pith.science/paper/HFEDYIJT

@misc{pith2026250414274,
  author       = {Pith},
  title        = {Pith review of: ProtPainter: Draw or Drag Protein via Topology-guided Diffusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HFEDYIJT}},
  note         = {Machine review of arXiv:2504.14274}
}
read the original abstract

Recent advances in protein backbone generation have achieved promising results under structural, functional, or physical constraints. However, existing methods lack the flexibility for precise topology control, limiting navigation of the backbone space. We present ProtPainter, a diffusion-based approach for generating protein backbones conditioned on 3D curves. ProtPainter follows a two-stage process: curve-based sketching and sketch-guided backbone generation. For the first stage, we propose CurveEncoder, which predicts secondary structure annotations from a curve to parametrize sketch generation. For the second stage, the sketch guides the generative process in Denoising Diffusion Probabilistic Modeling (DDPM) to generate backbones. During this process, we further introduce a fusion scheduling scheme, Helix-Gating, to control the scaling factors. To evaluate, we propose the first benchmark for topology-conditioned protein generation, introducing Protein Restoration Task and a new metric, self-consistency Topology Fitness (scTF). Experiments demonstrate ProtPainter's ability to generate topology-fit (scTF > 0.8) and designable (scTM > 0.5) backbones, with drawing and dragging tasks showcasing its flexibility and versatility.

Figures

Figures reproduced from arXiv: 2504.14274 by the authors.

Figure 1
Figure 1. Architecture. Sketching: given a 3D curve, SSEcurve is predicted by CurveEncoder. Then a naive sketch is generated parametrically. Guided Sampling: the sketch is fused into a diffusion sampling process with the guidance of RoseTTAFold and Helix-Gating interpolation. Zoom in Gaussian noise Condition Curve Unconditional Confidential phase Controllable phase b Helix-Gating 0ˆ t z t 1 y − y Helix%(⋅) Helix%(⋅) Fusion Sc… view at source ↗
Figure 2
Figure 2. Sketch Fusion Scheduling with Helix-Gating. a. Helix-Gating splits the sampling process [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Examples of ProtPainter on de novo protein design, binder design, and motif scaffolding. From left to right, the original structures are 6s9l, 1tqg, 7f4d MR, 7f4d GB, and 103l. Curves are visualized in 3D space. Other examples are shown in [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Draw and edit process. (a) Structures are visualized in the MDS topology space, with [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: scTF vs scTM and scRMSD. Figures from left to right show the test results on datasets [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]
Figure 6
Figure 6. Figure 6: Topology similarity example between two different structures (green and orange). [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: Protein Restoration Task and designability test. Using ProtPainter, we sample [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: a. Datasets visualization on protein space. We calculate scTM between structures and plot [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: Conditional Generation Visualization. We sample 1500 backbones conditioned on 1tqg [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]
Figure 10
Figure 10. Figure 10: CurveEncoder Ablation Result. D.4 TRAIN We processed 15000 data samples consisting of PDB (Burley et al., 2023; Berman et al., 2003) and scaffolds, generating three different granularity curve datasets with sampling rates of 40%, 80%, and 120%. Then the curves are smo…
Figure 11
Figure 11. Figure 11: Designability and scTM compared to RFDiffusion. [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]
Figure 12
Figure 12. Figure 12: Ablation. a. λ control as a tradeoff between diversity and similarity. b. Example samples of different λ. c. Selection method ablation; Score is computed as (scTF + scTM)/2. d. Two-Phase timing ablation. Users can control the generation between diversity and similarit…
Figure 13
Figure 13. Figure 13: Selection Method Ablation. The score is computed as (scTF+scTM)/2. [PITH_FULL_IMAGE:figures/full_fig_p022_13.png]
Figure 14
Figure 14. Figure 14: Gallery. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_14.png]
Figure 15
Figure 15. Figure 15: User cases gallery. K BETA SHEET EXAMPLES We have achieved controlling beta sheet percentage by increasing diversity, inspired by RFDiffusion. The cases are presented in [PITH_FULL_IMAGE:figures/full_fig_p024_15.png]
Figure 16
Figure 16. Figure 16: Beta sheet examples. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_16.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 33 canonical work pages

  1. [1]

    Announcing the worldwide protein data bank

    Helen Berman, Kim Henrick, and Haruki Nakamura. Announcing the worldwide protein data bank. Nature structural & molecular biology, 10 0 (12): 0 980--980, 2003

  2. [2]

    Se (3)-stochastic flow matching for protein backbone generation

    Avishek Joey Bose, Tara Akhound-Sadegh, Kilian Fatras, Guillaume Huguet, Jarrid Rector-Brooks, Cheng-Hao Liu, Andrei Cristian Nica, Maksym Korablyov, Michael Bronstein, and Alexander Tong. Se (3)-stochastic flow matching for protein backbone generation. arXiv preprint arXiv:2310.02391, 2023

  3. [3]

    Rcsb protein data bank (rcsb

    Stephen K Burley, Charmi Bhikadiya, Chunxiao Bi, Sebastian Bittrich, Henry Chao, Li Chen, Paul A Craig, Gregg V Crichlow, Kenneth Dalenberg, Jose M Duarte, et al. Rcsb protein data bank (rcsb. org): delivery of experimentally-determined pdb structures alongside one million computed structure models of proteins from artificial intelligence/machine learning...

  4. [4]

    Analog bits: Generating discrete data using diffusion models with self-conditioning

    Ting Chen, Ruixiang Zhang, and Geoffrey Hinton. Analog bits: Generating discrete data using diffusion models with self-conditioning. arXiv preprint arXiv:2208.04202, 2022

  5. [5]

    Ilvr: Conditioning method for denoising diffusion probabilistic models

    Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 14367--14376, 2021

  6. [6]

    Difftopo: Fold exploration using coarse grained protein topology representations

    Bruno Correia. Difftopo: Fold exploration using coarse grained protein topology representations. bioRxiv, pp.\ 2024--02, 2024

  7. [7]

    Robust deep learning--based protein sequence design using proteinmpnn

    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 0 (6615): 0 49--56, 2022

  8. [8]

    A fully automatic evolutionary classification of protein folds: Dali domain dictionary version 3

    Sabine Dietmann, Jong Park, Cedric Notredame, Andreas Heger, Michael Lappe, and Liisa Holm. A fully automatic evolutionary classification of protein folds: Dali domain dictionary version 3. Nucleic acids research, 29 0 (1): 0 55--57, 2001

Show all 54 references
  1. [9]

    Subcellular organization of gpcr signaling

    Kelsie Eichel and Mark von Zastrow. Subcellular organization of gpcr signaling. Trends in Pharmacological Sciences, 39 0 (2): 0 200--208, 2018

  2. [10]

    Fast protein structure searching using structure graph embeddings

    Joe G Greener and Kiarash Jamali. Fast protein structure searching using structure graph embeddings. bioRxiv, pp.\ 2022--11, 2022

  3. [11]

    Geometrically equivariant graph neural networks: A survey

    Jiaqi Han, Yu Rong, Tingyang Xu, and Wenbing Huang. Geometrically equivariant graph neural networks: A survey. arXiv preprint arXiv:2202.07230, 2022

  4. [12]

    A generic framework for hierarchical de novo protein design

    Zander Harteveld, Jaume Bonet, St \'e phane Rosset, Che Yang, Fabian Sesterhenn, and Bruno E Correia. A generic framework for hierarchical de novo protein design. Proceedings of the National Academy of Sciences, 119 0 (43): 0 e2206111119, 2022

  5. [13]

    e al Defferrard, Andreas Loukas, Pierre Vandergheynst, et al. Exploring “dark matter

    Zander Harteveld, Alexandra Van Hall-Beauvais, Irina Morozova, Joshua Southern, Casper Goverde, Sandrine Georgeon, St \'e phane Rosset, Mich \"e al Defferrard, Andreas Loukas, Pierre Vandergheynst, et al. Exploring “dark matter” protein folds using deep learning. bioRxiv, pp.\...

  6. [14]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33: 0 6840--6851, 2020

  7. [15]

    Generative models for protein structures and sequences

    Chloe Hsu, Clara Fannjiang, and Jennifer Listgarten. Generative models for protein structures and sequences. nature biotechnology, 42 0 (2): 0 196--199, 2024

  8. [16]

    A backbone-centred energy function of neural networks for protein design

    Bin Huang, Yang Xu, Xiuhong Hu, Yongrui Liu, Shanhui Liao, Jiahai Zhang, Chengdong Huang, Jingjun Hong, Quan Chen, and Haiyan Liu. A backbone-centred energy function of neural networks for protein design. Nature, 602 0 (7897): 0 523--528, February 2022. ISSN 0028-0836, 1476-46...

  9. [17]

    Pei, Brent L

    Po-Ssu Huang, Gustav Oberdorfer, Chunfu Xu, Xue Y. Pei, Brent L. Nannenga, Joseph M. Rogers, Frank DiMaio, Tamir Gonen, Ben Luisi, and David Baker. High thermodynamic stability of parametrically designed helical bundles. Science, 346 0 (6208): 0 481--485, October 2014. ISSN 00...

  10. [18]

    Huddy, Yang Hsia, Ryan D

    Timothy F. Huddy, Yang Hsia, Ryan D. Kibler, Jinwei Xu, Neville Bethel, Deepesh Nagarajan, Rachel Redler, Philip J. Y. Leung, Connor Weidle, Alexis Courbet, Erin C. Yang, Asim K. Bera, Nicolas Coudray, S. John Calise, Fatima A. Davila-Hernandez, Hannah L. Han, Kenneth D. Carr,...

  11. [19]

    Illuminating protein space with a programmable generative model

    John B Ingraham, Max Baranov, Zak Costello, Karl W Barber, Wujie Wang, Ahmed Ismail, Vincent Frappier, Dana M Lord, Christopher Ng-Thow-Hing, Erik R Van Vlack, et al. Illuminating protein space with a programmable generative model. Nature, 623 0 (7989): 0 1070--1078, 2023

  12. [20]

    Diversity and modularity of g protein-coupled receptor structures

    Vsevolod Katritch, Vadim Cherezov, and Raymond C Stevens. Diversity and modularity of g protein-coupled receptor structures. Trends in pharmacological sciences, 33 0 (1): 0 17--27, 2012

  13. [21]

    Dynamics-informed protein design with structure conditioning

    Urszula Julia Komorowska, Simon V Mathis, Kieran Didi, Francisco Vargas, Pietro Lio, and Mateja Jamnik. Dynamics-informed protein design with structure conditioning. In The Twelfth International Conference on Learning Representations, 2024

  14. [22]

    De novo protein design—from new structures to programmable functions

    Tanja Kortemme. De novo protein design—from new structures to programmable functions. Cell, 187 0 (3): 0 526--544, 2024

  15. [23]

    Improving antibody design with force-guided sampling in diffusion models

    Paulina Kulyt \.e , Francisco Vargas, Simon Valentin Mathis, Yu Guang Wang, Jos \'e Miguel Hern \'a ndez-Lobato, and Pietro Li \`o . Improving antibody design with force-guided sampling in diffusion models. arXiv preprint arXiv:2406.05832, 2024

  16. [24]

    Protein topology determines binding mechanism

    Yaakov Levy, Peter G Wolynes, and Jos \'e N Onuchic. Protein topology determines binding mechanism. Proceedings of the National Academy of Sciences, 101 0 (2): 0 511--516, 2004

  17. [25]

    Generating novel, designable, and diverse protein structures by equivariantly diffusing oriented residue clouds

    Yeqing Lin and Mohammed Alquraishi. Generating novel, designable, and diverse protein structures by equivariantly diffusing oriented residue clouds. In International Conference on Machine Learning, pp.\ 20978--21002. PMLR, 2023

  18. [26]

    Out of many, one: Designing and scaffolding proteins at the scale of the structural universe with genie 2

    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

  19. [27]

    Zero-shot 3d drug design by sketching and generating

    Siyu Long, Yi Zhou, Xinyu Dai, and Hao Zhou. Zero-shot 3d drug design by sketching and generating. Advances in Neural Information Processing Systems, 35: 0 23894--23907, 2022

  20. [28]

    Lutz, Shunzhi Wang, Christoffer Norn, Alexis Courbet, Andrew J

    Isaac D. Lutz, Shunzhi Wang, Christoffer Norn, Alexis Courbet, Andrew J. Borst, Yan Ting Zhao, Annie Dosey, Longxing Cao, Jinwei Xu, Elizabeth M. Leaf, Catherine Treichel, Patrisia Litvicov, Zhe Li, Alexander D. Goodson, Paula Rivera-Sánchez, Ana-Maria Bratovianu, Minkyung Bae...

  21. [29]

    Cath--a hierarchic classification of protein domain structures

    Christine A Orengo, Alex D Michie, Susan Jones, David T Jones, Mark B Swindells, and Janet M Thornton. Cath--a hierarchic classification of protein domain structures. Structure, 5 0 (8): 0 1093--1109, 1997

  22. [30]

    How well do multivariate data sets match? the advantages of a procrustean superimposition approach over the mantel test

    Pedro R Peres-Neto and Donald A Jackson. How well do multivariate data sets match? the advantages of a procrustean superimposition approach over the mantel test. Oecologia, 129: 0 169--178, 2001

  23. [31]

    Arvind Pillai, Abbas Idris, Annika Philomin, Connor Weidle, Rebecca Skotheim, Philip J. Y. Leung, Adam Broerman, Cullen Demakis, Andrew J. Borst, Florian Praetorius, and David Baker. De novo design of allosterically switchable protein assemblies. ISSN 0028-0836, 1476-4687. doi...

  24. [32]

    Design of stimulus-responsive two-state hinge proteins

    Florian Praetorius, Philip JY Leung, Maxx H Tessmer, Adam Broerman, Cullen Demakis, Acacia F Dishman, Arvind Pillai, Abbas Idris, David Juergens, Justas Dauparas, et al. Design of stimulus-responsive two-state hinge proteins. Science, 381 0 (6659): 0 754--760, 2023

  25. [33]

    The structure and function of g-protein-coupled receptors

    Daniel M Rosenbaum, S ren GF Rasmussen, and Brian K Kobilka. The structure and function of g-protein-coupled receptors. Nature, 459 0 (7245): 0 356--363, 2009

  26. [34]

    E (n) equivariant graph neural networks

    V ctor Garcia Satorras, Emiel Hoogeboom, and Max Welling. E (n) equivariant graph neural networks. In International conference on machine learning, pp.\ 9323--9332. PMLR, 2021

  27. [35]

    Cath: increased structural coverage of functional space

    Ian Sillitoe, Nicola Bordin, Natalie Dawson, Vaishali P Waman, Paul Ashford, Harry M Scholes, Camilla SM Pang, Laurel Woodridge, Clemens Rauer, Neeladri Sen, et al. Cath: increased structural coverage of functional space. Nucleic acids research, 49 0 (D1): 0 D266--D273, 2021

  28. [36]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pp.\ 2256--2265. PMLR, 2015

  29. [37]

    Score-based generative modeling through stochastic differential equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020

  30. [38]

    Fast and accurate protein structure search with foldseek

    Michel Van Kempen, Stephanie S Kim, Charlotte Tumescheit, Milot Mirdita, Jeongjae Lee, Cameron LM Gilchrist, Johannes S \"o ding, and Martin Steinegger. Fast and accurate protein structure search with foldseek. Nature biotechnology, 42 0 (2): 0 243--246, 2024

  31. [39]

    Membrane-protein topology

    Gunnar von Heijne. Membrane-protein topology. Nature reviews Molecular cell biology, 7 0 (12): 0 909--918, 2006

  32. [40]

    Protein conformation generation via force-guided se (3) diffusion models

    Yan Wang, Lihao Wang, Yuning Shen, Yiqun Wang, Huizhuo Yuan, Yue Wu, and Quanquan Gu. Protein conformation generation via force-guided se (3) diffusion models. arXiv preprint arXiv:2403.14088, 2024

  33. [41]

    De novo design of protein structure and function with rfdiffusion

    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 0 (7976): 0 1089--1100, 2023

  34. [42]

    Protein structural topology: Automated analysis and diagrammatic representation

    David R Westhead, Timothy WF Slidel, Tomas PJ Flores, and Janet M Thornton. Protein structural topology: Automated analysis and diagrammatic representation. Protein Science, 8 0 (4): 0 897--904, 1999

  35. [43]

    High-resolution de novo structure prediction from primary sequence

    Ruidong Wu, Fan Ding, Rui Wang, Rui Shen, Xiwen Zhang, Shitong Luo, Chenpeng Su, Zuofan Wu, Qi Xie, Bonnie Berger, et al. High-resolution de novo structure prediction from primary sequence. BioRxiv, pp.\ 2022--07, 2022

  36. [44]

    Topology: a unique dimension in protein engineering

    Lianjie Xu and Wen-Bin Zhang. Topology: a unique dimension in protein engineering. Science China Chemistry, 61: 0 3--16, 2018

  37. [45]

    Fast protein backbone generation with se (3) flow matching

    Jason Yim, Andrew Campbell, Andrew YK Foong, Michael Gastegger, Jos \'e Jim \'e nez-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:2310.0529...

  38. [46]

    Se (3) diffusion model with application to protein backbone generation

    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 b

  39. [47]

    G protein-coupled receptors (gpcrs): advances in structures, mechanisms, and drug discovery

    Mingyang Zhang, Ting Chen, Xun Lu, Xiaobing Lan, Ziqiang Chen, and Shaoyong Lu. G protein-coupled receptors (gpcrs): advances in structures, mechanisms, and drug discovery. Signal Transduction and Targeted Therapy, 9 0 (1): 0 88, 2024

  40. [48]

    Topodiff: Improving protein backbone generation with topology-aware latent encoding

    Yuyang Zhang, Zinnia Ma, and Haipeng Gong. Topodiff: Improving protein backbone generation with topology-aware latent encoding. In NeurIPS 2023 Generative AI and Biology (GenBio) Workshop

  41. [49]

    Deciphering of microbial community and antibiotic resistance genes in activated sludge reactors under high selective pressure of different antibiotics

    Renxin Zhao, Jie Feng, Jie Liu, Wenjie Fu, Xiaoyan Li, and Bing Li. Deciphering of microbial community and antibiotic resistance genes in activated sludge reactors under high selective pressure of different antibiotics. Water Research, 151: 0 388--402, 2019

  42. [50]

    Scaffold-lab: Critical evaluation and ranking of protein backbone generation methods in a unified framework

    Zhuoqi Zheng, Bo Zhang, Bozitao Zhong, Kexin Liu, Zhengxin Li, Junjie Zhu, Jinyu Yu, Ting Wei, and Hai-Feng Chen. Scaffold-lab: Critical evaluation and ranking of protein backbone generation methods in a unified framework. bioRxiv, pp.\ 2024--02, 2024

  43. [51]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

  44. [52]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  45. [53]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  46. [54]

    " id="W5M0MpCehiHzreSzNTczkc9d

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.