Pith. sign in

REVIEW 6 major objections 6 minor 27 references

Protenix-Mini: Efficient Structure Predictor via Compact Architecture, Few-Step Diffusion and Switchable pLM

T0 review · 6 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Two diffusion steps match 200 for protein structure prediction.

desk verdict Useful few-step ODE and pruning findings, but the central claim is muddied by a contradiction about which model was used. read the letter →

arxiv 2507.11839 v1 pith:HQSY3I4K submitted 2025-07-16 cs.LG q-bio.QM

classification cs.LGq-bio.QM
keywords proteinstructurepredictionfew-stepODEsamplingdiffusionmodelsmodelpruningESMlanguageAlphaFold3-stylearchitecturelightweightinferenceflowmatching
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

The paper claims that AlphaFold3-style diffusion structure predictors do not actually need their default hundreds of sampling steps: a deterministic ODE sampler with two specific settings, no added noise and unit step scale, reproduces the accuracy of 200-step sampling without any retraining. On top of that, it claims that many pairformer and diffusion transformer blocks in the open-source Protenix model are redundant, so a compact model with 16 pairformer and 8 diffusion transformer blocks loses only 1-5% on benchmark LDDT. The same analysis supports replacing the MSA module with ESM2 embeddings, trading a small accuracy loss (with a larger drop on protein-protein interfaces) for much cheaper preprocessing. If all this holds, high-fidelity biomolecular structure prediction becomes feasible at roughly 20 GFLOPs instead of 93, which matters for high-throughput screening and resource-limited deployment.

What carries the argument

The mechanism is the ODE sampler defined in Algorithm 2: instead of adding noise and using a learned velocity correction with $\eta=1.5$, it directly advances $x_t \leftarrow x_t + \Delta t \cdot \mathrm{CalVelocity}(x^{\text{denoised}}_t, x_t, t)$ with $\eta=1.0$ and no noise injection, so the trajectory follows the probability-flow ODE and can be integrated in one or two steps. The second piece is block-pruning: ablations removing the first pairformer blocks show they contribute negligibly, and finetuning recovers the lost accuracy. The third is switchable pLM conditioning: at each training iteration the model randomly uses either the MSA module or ESM2 embeddings, sharing all core weights, so at inference the MSA module can be omitted entirely.

What would settle it

Evaluate the 2-step ODE sampler on RecentPDB or Posebusters entries with more than 768 tokens and compare interface LDDT against the 200-step AF3 sampler; a gap much larger than the 1-5% seen on short proteins would show the claim is limited to short chains, and a strong degradation on another AF3-style model would disprove the transfer claim.

Watch

Extended reading notes

Core claim

The central discovery is a sampler reconfiguration, not a retrained model: keeping $\gamma_0=0$ and setting $\eta=1.0$ turns the AF3 EDM sampler into a pure ODE, and the resulting two-step trajectory yields complex LDDT of 0.822 versus 0.820 for the 200-step baseline on RecentPDB proteins under 768 tokens. The claim is that AF3-style models, whether trained with EDM or flow matching, are inherently robust to drastically reduced sampling steps once the noise injection is removed and the step scale is corrected. The paper then shows the same robustness carries over to a compact architecture, Protenix-Mini, which drops redundant early pairformer blocks and uses one MSA block, producing 1-5% lower performance on benchmarks while cutting FLOPs from 93 to 20.

Load-bearing premise

All reported scores come from test proteins with fewer than 768 tokens, and the paper assumes this subset represents longer, more complex structures where the 2-step ODE and the reduced 8-block transformer may behave very differently.

Editorial extensions

If this is right

  • A pre-trained AF3-style model can be deployed with a 2-step ODE at inference, so the diffusion sampler cost shrinks by roughly two orders of magnitude without retraining.
  • The compact Protenix-Mini, at 20 GFLOPs, brings complex prediction within reach of a single GPU for proteins under 768 tokens, with 1-5% LDDT loss.
  • For applications that tolerate slightly lower accuracy, 1-step sampling remains reasonable, and the Tiny variant at 8 pairformer blocks cuts compute by around 85%.
  • Flow-matching trained models show the same few-step robustness, so the finding is not tied to one training objective.
  • ESM-only inference removes MSA search time entirely, at the cost of a roughly 10% interface LDDT drop on protein-protein interfaces; paired MSAs still matter there.

Reading between the lines

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

  • Because the evaluation is capped at 768 tokens, the paper leaves open whether long multi-chain complexes, where conditioning is harder, need more than two ODE steps; a natural next experiment is a length-stratified accuracy curve.
  • The collapse artifacts reported for some 2-step samples (e.g. the '7bnh' case) suggest the few-step ODE underfits the low-noise regime; the paper's own hybrid schedule (first steps eta=1, last steps eta=1.5) points toward a cheap fix by mixing step scales.
  • The 2-step deterministic sampler is close to a one- or two-shot regressor, so it may understate structural diversity or confidence in regions where the true posterior is multimodal; sampling time or noise injection could be reintroduced specifically for such cases.
  • A switchable MSA/ESM model could be extended to route per-protein: use the cheap ESM path when the MSA is shallow or the target is a monomer, and fall back to paired MSAs for protein-protein interfaces, where the paper measures the largest drop.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

6 major / 6 minor

Summary. This paper proposes Protenix-Mini, a compact variant of the open-source Protenix structure prediction model, with three claimed contributions: (1) replacing the default 200-step AF3 sampler with a two-step ODE sampler without retraining; (2) identifying redundant Pairformer and Diffusion Transformer blocks in Protenix and pruning them; (3) replacing the MSA module with ESM2-3B embeddings under a hybrid training scheme. The authors report benchmark results on RecentPDB and Posebusters subsets with fewer than 768 tokens, claiming only a 1-5% performance drop at roughly 20 GFLOPs versus 93 GFLOPs for the full model. The paper also compares EDM and flow-matching training and reports a Tiny variant obtained by further pruning.

Significance. If the main claims hold, the paper describes a practically useful efficiency-accuracy trade-off for AF3-style structure prediction, and the two-step ODE finding is an interesting empirical observation that could transfer to other diffusion-based structure predictors. The paper is grounded in an open-source baseline and provides detailed algorithmic pseudocode and model configurations, which are strengths. However, the significance is currently limited by internal inconsistencies in the reported evaluation and by the absence of evidence beyond the sub-768-token regime. The claimed transferability to other AF3-style models is asserted rather than demonstrated, and the quantitative claims in the abstract and conclusion are contradicted by some of the paper's own tables.

major comments (6)
  1. [Section 3.1, Figure 4] The final paragraph of Section 3.1 states that the sampler studies were conducted using a 'small-scale Protenix model,' but the 200-step baseline in Figure 4 reports complex LDDT 0.820 and Prot-Prot LDDT 0.501, identical to the full Protenix numbers in Table 2. If Figure 4 came from the full model, the small-scale statement is incorrect; if it came from a reduced model, that model's configuration is never given, so the result cannot be reproduced or attributed. This ambiguity affects the central 'two ODE steps without retraining' claim. Please specify exactly which model was used for Figures 3 and 4 and reconcile the statement with the matching numbers in Table 2.
  2. [Abstract, Section 4.2, Tables 2 and 5] The abstract and conclusion claim only a '1-5% decrease' relative to the full model, but the paper's own tables contain larger degradations. In Table 2, Protenix-Mini RNA-Prot LDDT drops from 0.363 to 0.339 (about 6.6% relative), and Protenix-Tiny Prot-Prot LDDT drops from 0.501 to 0.428 (about 14.6% relative). In Table 5, the Posebusters success rate drops from 80.0% to 72.7% for Mini and to 69.8% for Tiny. The sentence in Section 4.2 that Tiny shows 'around 2% to 3% performance drop on different interface types' is contradicted by these numbers. Please state the delta definition (relative or absolute) and correct the claims to match the reported tables.
  3. [Section 4.1, Figure 1 caption] The Figure 1 caption states that 'All the numbers in the paper are evaluated on proteins with fewer than 768 tokens in the test set.' This means the central efficiency and accuracy claims are only established for this restricted regime, yet the introduction frames the method as suitable for 'large-scale applications.' No results are reported for longer sequences, where the 2-step ODE sampler and the heavily pruned architecture may degrade more sharply. Please either add experiments on proteins with more than 768 tokens or explicitly narrow the paper's claims to this regime.
  4. [Tables 2-5] The main benchmark tables report point estimates without error bars, confidence intervals, or significance tests. Given that the paper characterizes differences as 'negligible' or 'marginal,' the absence of repeated-seed variance makes it impossible to assess whether differences such as 0.490 vs. 0.501 (Table 2) are meaningful. Please report variances or multiple-seed statistics, at least for the headline comparisons.
  5. [Section 3.3, Figure 1] Protenix-Mini-ESM relies on ESM2-3B embeddings, but the FLOPs comparison in Figure 1 appears to count only the structure prediction module. The computational cost of running ESM2-3B is substantial and is not accounted for in the efficiency comparison. If the ESM inference cost is excluded, the efficiency claim for the ESM variant is incomplete. Please state the total inference cost including the ESM model, or clarify that the reported FLOPs are for the structure module only.
  6. [Appendix D, Table 2 caption] Appendix D states that the paper reports 'the median results across 25 samples for each input sequence,' while the Table 2 caption says scores are 'average the scores of random samples without leveraging confidence scores for sample selection.' These protocols conflict. Please clarify which aggregation is actually used and ensure all table captions are consistent.
minor comments (6)
  1. [Section 3.1] The sentence 'The studies in this section was conducted using a small-scale Protenix model' has a subject-verb agreement error and should read 'were conducted.'
  2. [Section 4.1] The text contains typos: 'resmaining' should be 'remaining', 'ans' should be 'and', and 'emplyed' should be 'employed'.
  3. [Section 3.3] The phrase 'we only use the EDM module' appears to be a typo for 'the ESM module.' Also, 'thes_inputs' should be 'the s_inputs'.
  4. [Figure 1] The caption is confusing: it says 'the number of MSAs is fixed at 2048, while the right panel sets the token length to 384,' but it is unclear which panel corresponds to which setting. Please clarify the left and right panel descriptions.
  5. [Table 6] The model name 'Mini-EDM-Flow' is inconsistent with the names used elsewhere, such as 'PTX-Mini-Flow' and 'Protenix-Mini-ESM-Flow.' Please unify the nomenclature.
  6. [Section 4.2.3] The text references Figure 6 for collapsed ligands, but the figure is not included in the manuscript text; please add the figure or remove the reference.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the compact-model results are empirical measurements and ablations on an open-source codebase, not derivations from their own conclusions.

full rationale

The paper makes no derivation-style claim that could collapse into its own inputs. The few-step ODE result (Sec. 3.1) is an empirical measurement of an existing sampler under modified hyperparameters (γ0=0, η=1.0); the pruning result (Sec. 3.2) is an ablation on the open-source Protenix codebase; and the Mini/Tiny/ESM comparisons in Tables 2-5 are direct evaluations, not predictions. The baseline Protenix [20] shares authors with this paper, but it is a separately published, open-source, code-reproduced implementation, so under the cited-evidence rule it functions as real external evidence rather than a self-citation loop. The sentence 'The studies in this section was conducted using a small-scale Protenix model, but we believe the observed trends generalize well' is an explicitly hedged extrapolation; an unsupported generalization is a correctness or robustness concern, not a circular reduction. No equation is fitted to the target metric and then reported as a prediction; the 1-5% drop is a summary of measured LDDT/RMSD differences. Therefore no circular step is exhibited.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claims rest on empirical measurements rather than a formal derivation. The free parameters are the sampler hyperparameters, step count, and architecture sizes, all selected on the benchmark without a described held-out validation. The axioms are the fidelity of the Protenix reproduction, the standard ODE interpretation of the denoiser, the viability of ESM substitution, the representativeness of the short-protein benchmark subsets, and the transferability of the redundancy finding.

free parameters (6)
  • Sampler step scale eta = 1.0
    Selected empirically from eta in {1.5, 1.2, 1.0} by LDDT on benchmarks (Figure 3); no held-out validation described.
  • Sampler noise schedule gamma0 = 0.0
    Set to 0 to remove added noise, converting the sampler to an ODE; hand-picked modification of the AF3 sampler.
  • Diffusion inference step count = 2
    Chosen as the smallest step count with acceptable LDDT; the paper notes increased clashes at 2 steps (Table 6).
  • Architecture block counts (pairformer, diffusion transformer, MSA) = 16/8/1
    Result of an architecture search balancing LDDT and FLOPs (Section 3.2, Table 8).
  • Hybrid training probability for ESM vs MSA = 0.5
    50% probability per batch in Protenix-Mini-ESM training (Section 3.3); no sensitivity analysis reported.
  • Flow matching time-distribution Beta parameters = Beta(2.5, 2.5)
    Hand-selected training distribution for the flow matching variant (Section 4.2.2); no sensitivity analysis.
assumptions (5)
  • domain assumption Protenix is a faithful open-source reproduction of AlphaFold3
    The paper uses Protenix as the 'full-scale counterpart' and leverages its codebase; overlapping authorship with Protenix (Ref [20]) means accuracy claims relative to AF3 depend on Protenix's fidelity.
  • standard math The learned denoiser defines a probability-flow ODE that can be solved by the EDM Euler update with step scale 1.0
    Section 3.1 and Algorithm 2 reformulate the AF3 sampler as an ODE by setting gamma0=0, lambda=1, eta=1; this relies on standard EDM/score-based ODE theory (Ref [6]) without a proof in this paper.
  • domain assumption ESM2-3B embeddings can substitute for MSA features in the conditioning stack
    Section 3.3 feeds ESM2 embeddings into the input embedder and nullifies MSA; the paper's own Table 3 shows a large protein-protein interface drop, so the assumption is load-bearing and partially contradicted.
  • domain assumption The RecentPDB and Posebusters subsets with fewer than 768 tokens are representative of the full evaluation distribution
    All benchmark numbers in the paper are computed on this subset (Figure 1 caption), and no long-protein evaluation is provided.
  • ad hoc to paper Redundancy of the first pairformer blocks observed in the full Protenix model transfers to the smaller Protenix-Mini architecture
    Section 3.2/Table 8 prunes the first 8 pairformer blocks from the 16-block Mini to create Tiny; this assumes the ablation result on the 48-block model applies at the 16-block scale.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Protenix-Mini: Efficient Structure Predictor via Compact Architecture, Few-Step Diffusion and Switchable pLM." pith.science (2026). https://pith.science/paper/HQSY3I4K

@misc{pith2026250711839,
  author       = {Pith},
  title        = {Pith review of: Protenix-Mini: Efficient Structure Predictor via Compact Architecture, Few-Step Diffusion and Switchable pLM},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HQSY3I4K}},
  note         = {Machine review of arXiv:2507.11839}
}
read the original abstract

Lightweight inference is critical for biomolecular structure prediction and other downstream tasks, enabling efficient real-world deployment and inference-time scaling for large-scale applications. In this work, we address the challenge of balancing model efficiency and prediction accuracy by making several key modifications, 1) Multi-step AF3 sampler is replaced by a few-step ODE sampler, significantly reducing computational overhead for the diffusion module part during inference; 2) In the open-source Protenix framework, a subset of pairformer or diffusion transformer blocks doesn't make contributions to the final structure prediction, presenting opportunities for architectural pruning and lightweight redesign; 3) A model incorporating an ESM module is trained to substitute the conventional MSA module, reducing MSA preprocessing time. Building on these key insights, we present Protenix-Mini, a compact and optimized model designed for efficient protein structure prediction. This streamlined version incorporates a more efficient architectural design with a two-step Ordinary Differential Equation (ODE) sampling strategy. By eliminating redundant Transformer components and refining the sampling process, Protenix-Mini significantly reduces model complexity with slight accuracy drop. Evaluations on benchmark datasets demonstrate that it achieves high-fidelity predictions, with only a negligible 1 to 5 percent decrease in performance on benchmark datasets compared to its full-scale counterpart. This makes Protenix-Mini an ideal choice for applications where computational resources are limited but accurate structure prediction remains crucial.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 21 canonical work pages

  1. [1]

    J., Bambrick, J., et al

    Abramson, J., Adler, J., Dunger, J., Evans, R., Green, T., Pritzel, A., Ronneberger, O., Willmore, L., Ballard, A. J., Bambrick, J., et al. Accurate structure prediction of biomolecular interactions with alphafold 3.Nature, 630(8016):493–500, 2024

  2. [2]

    S., Jude, K

    Cao, L., Coventry, B., Goreshnik, I., Huang, B., Sheffler, W., Park, J. S., Jude, K. M., Marković, I., Kadam, R. U., Verschueren, K. H., et al. Design of protein-binding proteins from the target structure alone.Nature, 605 (7910):551–560, 2022

  3. [3]

    Cheng, H., Zhang, M., and Shi, J. Q. A survey on deep neural network pruning: Taxonomy, comparison, analysis, and recommendations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

  4. [4]

    H., Zielinski, M., Sargeant, T., et al

    Cheng, J., Novati, G., Pan, J., Bycroft, C., Žemgulyt˙ e, A., Applebaum, T., Pritzel, A., Wong, L. H., Zielinski, M., Sargeant, T., et al. Accurate proteome-wide missense variant effect prediction with alphamissense.Science, 381 (6664):eadg7492, 2023

  5. [5]

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

    Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., Tunyasuvunakool, K., Bates, R., Žídek, A., Potapenko, A., et al. Highly accurate protein structure prediction with alphafold.Nature, 596(7873): 583–589, 2021

  6. [6]

    Elucidating the design space of diffusion-based generative models

    Karras, T., Aittala, M., Aila, T., and Laine, S. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565–26577, 2022

  7. [7]

    Guiding a diffusion model with a bad version of itself.Advances in Neural Information Processing Systems, 37:52996–53021, 2024

    Karras, T., Aittala, M., Kynkäänniemi, T., Lehtinen, J., Aila, T., and Laine, S. Guiding a diffusion model with a bad version of itself.Advances in Neural Information Processing Systems, 37:52996–53021, 2024

  8. [8]

    Crafting papers on machine learning

    Langley, P. Crafting papers on machine learning. In Langley, P. (ed.),Proceedings of the 17th International Conference on Machine Learning (ICML 2000), pp. 1207–1216, Stanford, CA, 2000. Morgan Kaufmann

Show all 27 references
  1. [9]

    Wide neural networks of any depth evolve as linear models under gradient descent.Advancesin neural information processing systems, 32, 2019

    Lee, J., Xiao, L., Schoenholz, S., Bahri, Y., Novak, R., Sohl-Dickstein, J., and Pennington, J. Wide neural networks of any depth evolve as linear models under gradient descent.Advancesin neural information processing systems, 32, 2019

  2. [10]

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

    Lin, Z., Akin, H., Rao, R., Hie, B., Zhu, Z., Lu, W., Smetanin, N., Verkuil, R., Kabeli, O., Shmueli, Y., et al. Evolutionary-scale prediction of atomic-level protein structure with a language model.Science, 379(6637): 1123–1130, 2023

  3. [11]

    T., Ben-Hamu, H., Nickel, M., and Le, M

    Lipman, Y., Chen, R. T., Ben-Hamu, H., Nickel, M., and Le, M. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022

  4. [12]

    Rectified flow: A marginal preserving approach to optimal transport.arXiv preprint arXiv:2209.14577, 2022

    Liu, Q. Rectified flow: A marginal preserving approach to optimal transport.arXiv preprint arXiv:2209.14577, 2022

  5. [13]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Liu, X., Gong, C., et al. Flow straight and fast: Learning to generate and transfer data with rectified flow. In The Eleventh International Conference on Learning Representations, 2023

  6. [14]

    Flowgrad: Controlling the output of generative odes with gradients

    Liu, X., Wu, L., Zhang, S., Gong, C., Ping, W., and Liu, Q. Flowgrad: Controlling the output of generative odes with gradients. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 24335–24344, 2023

  7. [15]

    Instaflow: One step is enough for high-quality diffusion-based text-to-image generation

    Liu, X., Zhang, X., Ma, J., Peng, J., et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. InThe TwelfthInternational Conference on Learning Representations, 2023

  8. [16]

    Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps.Advances in Neural Information Processing Systems, 35:5775–5787, 2022

    Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., and Zhu, J. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps.Advances in Neural Information Processing Systems, 35:5775–5787, 2022

  9. [17]

    Inference-time scaling for diffusion models beyond scaling denoising steps.arXiv preprint arXiv:2501.09732, 2025

    Ma, N., Tong, S., Jia, H., Hu, H., Su, Y.-C., Zhang, M., Yang, X., Li, Y., Jaakkola, T., Jia, X., et al. Inference-time scaling for diffusion models beyond scaling denoising steps.arXiv preprint arXiv:2501.09732, 2025. 9

  10. [18]

    Bindcraft: one-shot design of functional protein binders.bioRxiv, pp

    Pacesa, M., Nickel, L., Schellhaas, C., Schmidt, J., Pyatova, E., Kissling, L., Barendse, P., Choudhury, J., Kapoor, S., Alcaraz-Serna, A., et al. Bindcraft: one-shot design of functional protein binders.bioRxiv, pp. 2024–09, 2024

  11. [19]

    Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    Song, J., Meng, C., and Ermon, S. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

  12. [20]

    Team, B. A. A., Chen, X., Zhang, Y., Lu, C., Ma, W., Guan, J., Gong, C., Yang, J., Zhang, H., Zhang, K., Wu, S., Zhou, K., Yang, Y., Liu, Z., Wang, L., Shi, B., Shi, S., and Xiao, W. Protenix - advancing structure prediction through a comprehensive alphafold3 reproduction.bioR...

  13. [21]

    D., Boitreaud, J., Dent, J., McPartlon, M., Meier, J., Reis, V., Rogozhonikov, A., and Wu, K

    team, C. D., Boitreaud, J., Dent, J., McPartlon, M., Meier, J., Reis, V., Rogozhonikov, A., and Wu, K. Chai-1: Decoding the molecular interactions of life.BioRxiv, pp. 2024–10, 2024

  14. [22]

    Scientific discovery in the age of artificial intelligence.Nature, 620(7972):47–60, 2023

    Wang, H., Fu, T., Du, Y., Gao, W., Huang, K., Liu, Z., Chandak, P., Liu, S., Van Katwyk, P., Deac, A., et al. Scientific discovery in the age of artificial intelligence.Nature, 620(7972):47–60, 2023

  15. [23]

    L., Juergens, D., Bennett, N

    Watson, J. L., Juergens, D., Bennett, N. R., Trippe, B. L., Yim, J., Eisenach, H. E., Ahern, W., Borst, A. J., Ragotte, R. J., Milles, L. F., et al. De novo design of protein structure and function with rfdiffusion.Nature, 620 (7976):1089–1100, 2023

  16. [24]

    Boltz-1: Democratizing biomolecular interaction modeling.bioRxiv, pp

    Wohlwend, J., Corso, G., Passaro, S., Reveiz, M., Leidal, K., Swiderski, W., Portnoi, T., Chinn, I., Silterra, J., Jaakkola, T., et al. Boltz-1: Democratizing biomolecular interaction modeling.bioRxiv, pp. 2024–11, 2024

  17. [25]

    Diffusion models: A comprehensive survey of methods and applications.ACM Computing Surveys, 56(4):1–39, 2023

    Yang, L., Zhang, Z., Song, Y., Hong, S., Xu, R., Zhao, Y., Zhang, W., Cui, B., and Yang, M.-H. Diffusion models: A comprehensive survey of methods and applications.ACM Computing Surveys, 56(4):1–39, 2023

  18. [26]

    Good subnetworks provably exist: Pruning via greedy forward selection

    Ye, M., Gong, C., Nie, L., Zhou, D., Klivans, A., and Liu, Q. Good subnetworks provably exist: Pruning via greedy forward selection. InInternational Conference on Machine Learning, pp. 10820–10830. PMLR, 2020

  19. [27]

    Are all layers created equal?Journal of Machine Learning Research, 23 (67):1–28, 2022

    Zhang, C., Bengio, S., and Singer, Y. Are all layers created equal?Journal of Machine Learning Research, 23 (67):1–28, 2022. 10 A Training Details EDM Training Objectives Denote x0 as random noise, andx1 as the target structure, the EDM training objectives contain several term...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.