Pith. sign in

REVIEW 5 major objections 5 minor 38 references

Reshaping Biomolecular Structure Prediction through Strategic Conformational Exploration with HelixFold-S1

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read HelixFold-S1 replaces aimless sampling with contact-guided planning: it predicts an inter-chain contact probability map as a blueprint, then folds structures under the most probable contacts, reaching markedly better accuracy with an…

desk verdict A practical, well-evaluated guided-sampling method for complex structure prediction that deserves peer review, but its headline 100x efficiency claim currently rests on a thin, unreproducible benchmark. read the letter →

arxiv 2507.09251 v3 pith:TW4UM4J7 submitted 2025-07-12 q-bio.BM

Lihang Liu , Yang Liu , Xianbin Ye , Shanzhuo Zhang , Yuxin Li , Kunrui Zhu , Yang Xue , Xiaonan Zhang
show 1 more author
Xiaomin Fang
This is my paper · ORCID
classification q-bio.BM
keywords biomolecularstructurepredictionconformationalspaceexplorationcontactprobabilityguidedsamplingefficiencyprotein-antibodycomplexesconditioningmultimer
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

Aimless conformational sampling wastes compute: most generated structures are redundant and cluster near one mode of the landscape. This paper proposes replacing it with guided planning. HelixFold-S1 first runs a Contact Prediction Module that converts the model's pair representation into a matrix of inter-chain contact probabilities, treating it as a coarse blueprint of conformational space. Sampling then feeds the highest-probability, non-redundant contacts as binary constraints into a Contact Conditioning Module, generating one structure per constraint. The central quantitative claim is that on protein-antibody complexes released in 2024, 10 guided samples reach the Top-1 precision that HelixFold3 with AFSample achieves with 1,000 samples, and that guided sampling beats unguided baselines across protein-protein, protein-ligand, protein-RNA, and protein-DNA benchmarks while its probability values indicate which targets are hard and which will benefit from extra sampling.

What carries the argument

The load-bearing mechanism is the contact probability blueprint plus the contact-conditioned generator. The Contact Prediction Module (one Pairformer stack plus a sigmoid head) turns the pair representation into a matrix $p^{\rm contact}_{ij}$, the probability that tokens $i$ and $j$ from different chains have any atom pair within 5 Å. The Contact Conditioning Module adds a learned linear projection of the chosen binary contact matrix into the pair activations of the Input Embedder, so the downstream structure module folds under explicit spatial constraints. In training the two modules are activated mutually exclusively, alternating between estimating contacts and generating structures from ground-truth contacts (0-10 per complex, with 70% of samples receiving constraints), so that at inference the predicted contacts are trusted like the experimental ones. Contact selection then walks the blueprint in descending probability, and redundant contact pruning removes any contact already realized in an earlier sample, so consecutive samples explore fresh regions; once remaining probabilities fall below $0.2 \times \max p^{\rm contact}_{ij}$, sampling re-exploits previously used contact sets until the budget $S$ is reached.

What would settle it

Run HelixFold-S1's sampling with the contact ranking destroyed: feed the same number of constraints drawn uniformly at random from the predicted matrix instead of taken in descending probability order, keeping everything else fixed. If Top-1 and Top-5 precision and ensemble diversity do not drop materially, then the probability ordering, the core of the strategic targeting claim, is not what drives the gains. A complementary check is to measure accuracy only on targets whose top predicted contacts are wrong; persistence of the gains would indicate the conditioning module, not the blueprint, is responsible.

Watch

Extended reading notes

Core claim

The paper claims that strategic conformational exploration, not more sampling, is the way forward for biomolecular complex prediction. Across benchmarks of protein-antibody, protein-protein, protein-ligand, protein-RNA, and protein-DNA complexes drawn from the PDB (releases 2022-2024, clustered for diversity), HelixFold-S1 improves Top-5 precision over HelixFold3 at matched 25-sample effort on every interface type, with the largest gains on protein-antibody complexes. On a separate set of 74 antibody complexes released in 2024, the paper finds that 10 HelixFold-S1 samples match the Top-1 precision of HelixFold3 with AFSample at 1,000 samples, i.e., one percent of the sampling budget. The same contact-guided workflow transferred to the Protenix and Chai-1 folding models improves them as well, and guided ensembles show broader precision distributions, indicating greater conformational diversity.

Load-bearing premise

The load-bearing premise is that the predicted contact probability matrix is informative enough, even on antibody-antigen targets where the contact prediction module has its lowest AUPRC, that greedily conditioning structure generation on the highest-probability contacts steers sampling toward near-native conformations rather than acting as arbitrary constraints.

Editorial extensions

If this is right

  • Protein-antibody modeling becomes cheap: the 100-fold sampling reduction means accuracy previously requiring a full GPU day of 1,000 samples is attainable with 10 samples, changing the practical economics of antibody-antigen prediction.
  • The target-level contact probability (the maximum entry of the predicted matrix) becomes a built-in difficulty signal: high-probability targets are nearly solved by a single sample, intermediate-probability targets gain most from additional sampling, and low-probability targets need large budgets, so sampling effort can be sized per target instead of uniformly.
  • Because the guidance transfers to Protenix and Chai-1, the planning strategy is separable from the base folding model; any model that accepts contact constraints could adopt it.
  • Sampling trajectories show precision declining as the algorithm works down the contact-probability list, which implies diminishing returns that could be detected on the fly and used for early stopping.

Reading between the lines

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

  • The paper's own supplementary analysis reports that the confidence score ranks conformations only weakly (C-index 0.5-0.7, barely above random); if a better ranker were available, the guided ensemble's Top-1 gains would likely exceed the reported ones, so the headline numbers are probably conservative.
  • The probability-versus-utility curve implies a resource-allocation policy the paper does not implement: sample once, read the target-level contact probability, then spend the sampling budget mainly on intermediate-probability targets while capping spend on high- and low-probability ones.
  • The greedy descending selection suggests a cheap improvement: choose the next contact not just by probability but by spatial separation from already satisfied contacts, which should further cut redundancy without the sequential cost of full redundant contact pruning.
  • If contact conditioning generalizes beyond natural complexes, the method becomes a design tool: each predicted contact is a handle that can be specified or mutated to steer a binder toward a chosen epitope, connecting prediction directly to protein design.
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

5 major / 5 minor

Summary. HelixFold-S1 extends HelixFold3 with two modules: a Contact Prediction Module (CPM) that estimates inter-chain contact probabilities from pair representations, and a Contact Conditioning Module (CCM) that injects binary contact constraints into the input embedder. At inference, contact probabilities are averaged over five runs, contacts are selected greedily in descending probability order with redundancy pruning, and each constraint is used to generate a structure that is ranked by an AF3-style confidence score. The method is evaluated on temporal splits (2022–2024) of protein–antibody, protein–protein, protein–ligand, protein–RNA, and protein–DNA complexes against HF3, HF3 with AFSample, Protenix, and Chai-1, plus S1-conditioned variants. The authors report consistent Top-K precision gains, state that 10 HelixFold-S1 samples reach the Top-1 precision of 1,000-sample HF3 with AFSample on 74 antibody complexes, and propose predicted contact probabilities as a target-difficulty and sampling-utility indicator. The paper also includes ablations of redundancy pruning, an epitope-prediction analysis, and a frank assessment of the limited discriminative power of ranking confidence.

Significance. The central idea—replacing aimless ensemble sampling with a predicted contact “blueprint” and constraint-conditioned generation—is timely and potentially impactful. If the efficiency claim holds, the method offers a practical route to reduce the GPU cost of complex prediction and a principled way to allocate sampling effort across targets. The benchmark design is careful in several respects: test complexes are temporally separated from the training cutoff, clustered with MMseqs2 to reduce redundancy, ligand filters are explicit, and multiple independent folding models (HF3, Protenix, Chai-1) are used as backbones. The cross-model transfer experiments and the RCP ablation strengthen the mechanistic story, and the contact-probability-as-difficulty analysis is a useful, falsifiable observation. I found no circularity problem: evaluation is against external PDB ground truth, and the CPM is trained on ground-truth contacts rather than on the ranking metric.

major comments (5)
  1. [Results, Fig. 2c; Methods, Inference Regime] The headline efficiency claim—that 10 HF-S1 samples match the Top-1 precision of 1,000 HF3 w/ AFSample samples—is supported only by two independent runs, as stated in Methods (“shaded areas indicate the variability between the two runs”). No per-target paired comparison, confidence interval, or statistical test is reported for this specific comparison, even though Fig. 2d reports significance asterisks for a related comparison. Because the curves in Fig. 2c are averaged over n=74 targets, a two-run average can hide substantial per-target variability. Please provide per-target paired differences (e.g., bootstrap or Wilcoxon signed-rank test on 10-sample HF-S1 versus 1,000-sample baseline Top-1 values), report the number of seeds or runs, and show the distribution rather than only the mean curve.
  2. [Methods, Table 1; Results, Fig. 2c] The comparison against “HF3 w/ AFSample” uses a self-implemented mixture of three hyperparameter settings (Table 1) rather than the original AFSample code or a widely validated AFSample configuration. The text states that this follows “the AlphaFold settings used in AFsample,” but no evidence is given that this mixture reproduces the original method's behavior. If the baseline is weaker than the field-standard massive-sampling configuration, the reported 100x advantage over unguided sampling is inflated. Please either run the original AFSample implementation on these targets or report the sensitivity of the conclusion to the baseline settings, for example the three settings individually and the published AFSample recipe.
  3. [Results, Fig. 2c; Methods, Inference Regime] The statement that 10 samples represent “just 1% of the computational cost” of 1,000 samples equates sample count with cost, but HF-S1's inference pipeline is not cost-equivalent to one AFSample pass: the Contact Prediction stage is executed five times per target, contact selection with RCP is sequential, and each conditioned generation may involve additional overhead. No wall-clock time, GPU-hours, or FLOP comparison is reported. Please measure actual runtime on the same hardware and target set, and report the overhead of CPM and sequential sampling; the efficiency claim should be stated in terms of measured compute, not sample count alone.
  4. [Results, Fig. 2e and Fig. 2c] The contact prediction module has its lowest AUPRC on protein–antibody complexes (Fig. 2e), yet the strongest gains—including the 10-sample versus 1,000-sample result—are on antibody targets. This creates a tension for the proposed mechanism: the blueprint is least reliable exactly where the claimed efficiency gain is largest. Please stratify the antibody targets by CPM AUPRC or by target-level contact probability and show the 10-sample versus 1,000-sample comparison within strata, or report the per-target correlation between contact-prediction accuracy and the sampling-efficiency gain. This would directly test whether the guided-sampling benefit is robust when the predicted contacts are noisy.
  5. [Code Availability] The source code, trained weights, and inference scripts are withheld until acceptance, and the web service may not expose the exact sequential contact-selection and pruning protocol used in Figs. 2–4. Given that the central claim rests on a specific inference algorithm, the review process and readers need at least the inference code and weights to verify the efficiency comparison. Please provide these artifacts to reviewers or deposit them in a public repository at revision time.
minor comments (5)
  1. [Abstract; Fig. 2c caption] The abstract says “reducing sampling requirements by an order of magnitude,” but the concrete claim in Fig. 2c is 10 samples versus 1,000 samples, i.e., a 100x reduction (two orders of magnitude). Please make the claimed reduction factor consistent and state explicitly whether “order of magnitude” is meant loosely.
  2. [Supplementary Fig. S4 caption] Fig. S4 reports n=103 for the protein–antibody 2024 dataset, while the main text and Fig. 2c/Fig. S6 use n=74 with the <800-residue filter. Please reconcile these numbers or clarify that Fig. S4 uses the unfiltered 2024 set.
  3. [Methods, Inference Regime] The inference description says the contact prediction task is “executed five times” and averaged, but it is unclear whether these five repeats are included in the sampling budget S or in the reported sample count. Please clarify the accounting of these five passes in the 25-sample and 10-sample experiments.
  4. [Supplementary Fig. S6 caption] The caption contains a typo: “orcale ensemble DockQs” should be “oracle ensemble DockQs.” Please also define the oracle metric in the caption for completeness.
  5. [References, [32]] Reference [32] is missing volume and page details; please complete the bibliographic entry.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the guided-sampling gains are measured against external PDB ground truth, and the contact prediction and conditioning modules are trained on ground-truth contacts rather than on the benchmark ranking metric.

full rationale

The paper's central claim is that guided contact-based sampling achieves higher structural accuracy than unguided sampling while using far fewer samples. That claim is evaluated with DockQ, iLDDT, and ligand RMSD against experimentally determined PDB structures, which are external to the method's fitted parameters. The contact prediction module is trained with a supervised cross-entropy loss against ground-truth inter-chain contacts extracted from experimental structures (Methods, L_contact), and the contact conditioning module is trained with 0-10 ground-truth contacts per complex. Neither module is fitted to the Top-K precision or sampling-efficiency metrics that constitute the paper's headline result. The 10-samples-versus-1000-samples comparison in Fig. 2c is an empirical measurement on an external 74-complex antibody benchmark, not a quantity that appears in the training objective. The difficulty-indicator and sampling-utility analyses in Fig. 3 are post hoc stratifications of observed predictions and accuracies; they are descriptive, not used to construct the predictions, so they are not self-fulfilling. The paper cites the authors' own HelixFold3 as the base architecture, but HF3 is used as a tool and is compared against other external models (Protenix, Chai-1); the self-citation does not supply the evidence for the new guided-sampling claim. The paper honestly reports limitations, including weak ranking-confidence discrimination (C-index 0.5-0.7 in Fig. S3), diminishing returns of greedy contact selection, and two-run variability in the extensive-sampling figure. These are robustness and verifiability concerns, not circularity. The evaluation is self-contained against external benchmarks, so a low circularity score is appropriate.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical entities or fitted physical parameters. It relies on inherited HF3 weights, a standard contact definition, and a small number of hand-set inference hyperparameters. The main unaccounted cost is the pretrained model itself, which is the authors' own HelixFold3.

free parameters (2)
  • Enrichment threshold factor = 0.2 times the maximum predicted contact probability
    Hand-chosen cutoff that switches sampling from new contacts to re-sampling previously used contact sets (Methods, Inference Regime). The central qualitative results are unlikely to depend on its exact value, but it shapes the sampling trajectory.
  • Contact prediction repeats = 5
    Inference averages five contact probability matrices to reduce variance (Methods, Inference Regime). Simple design choice, not tuned against the test set.
assumptions (4)
  • domain assumption HelixFold3 faithfully reproduces AlphaFold3 behavior as the base model.
    The paper builds on HF3 [9], authored by the same group; no independent verification of HF3 is provided in this manuscript.
  • domain assumption The temporal-plus-clustering split prevents train/test leakage.
    Methods describe sequence identity filtering (40% identity clustering and ligand CCD exclusion), and Supp. F shows weak correlation between sequence identity and contact probability, but no formal leakage audit is given.
  • domain assumption A contact defined by any atom pair within 5 Å is an adequate blueprint for guiding sampling.
    This coarse definition is standard in the field and is used for both training supervision and inference; if it were too coarse, conditioning would not reliably steer the generator.
  • domain assumption The diffusion module retains enough stochasticity to produce diverse structures under a fixed contact constraint.
    The method generates multiple conformations per constraint; if the generator collapsed to one output per constraint, diversity would be lost. The reported diversity in Fig. 4 supports this assumption empirically, but it is not proven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reshaping Biomolecular Structure Prediction through Strategic Conformational Exploration with HelixFold-S1." pith.science (2026). https://pith.science/paper/TW4UM4J7

@misc{pith2026250709251,
  author       = {Pith},
  title        = {Pith review of: Reshaping Biomolecular Structure Prediction through Strategic Conformational Exploration with HelixFold-S1},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TW4UM4J7}},
  note         = {Machine review of arXiv:2507.09251}
}
read the original abstract

Generating large ensembles of candidate conformations is standard for improving biomolecular structure prediction. Yet aimless sampling is inefficient and costly, producing many redundant conformations with limited diversity, particularly for complex multimeric assemblies. Here, we present HelixFold-S1, a guided planning approach specifically designed to enhance the structural prediction of biomolecular complexes by strategically targeting the most informative regions of conformational space to produce accurate conformations. For each complex, predicted inter-chain contact probabilities serve as a blueprint of the conformational space, guiding computational effort toward higher-probability, low-redundancy contacts that constrain structure generation. Across diverse biomolecular complex benchmarks, HelixFold-S1 achieves markedly higher structural accuracy than traditional unguided methods while reducing sampling requirements by an order of magnitude. Predicted contact probabilities also provide a rough indicator of prediction difficulty and sampling utility. These results demonstrate that guided planning reshapes conformational exploration and enables more efficient and accurate structural inference.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 37 canonical work pages

  1. [1]

    nature596(7873), 583–589 (2021)

    Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., Tunyasuvunakool, K., Bates, R., ˇZ ´ ıdek, A., Potapenko, A.,et al.: Highly accurate protein structure prediction with alphafold. nature596(7873), 583–589 (2021)

    work page 2021

  2. [2]

    biorxiv, 2021–10 (2021)

    Evans, R., O’Neill, M., Pritzel, A., Antropova, N., Senior, A., Green, T., ˇZ ´ ıdek, A., Bates, R., Blackwell, S., Yim, J., et al.: Protein complex prediction with alphafold-multimer. biorxiv, 2021–10 (2021)

    work page 2021

  3. [3]

    Nature, 1–3 (2024)

    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, 1–3 (2024)

    work page 2024

  4. [4]

    Science373(6557), 871–876 (2021)

    Baek, M., DiMaio, F., Anishchenko, I., Dauparas, J., Ovchinnikov, S., Lee, G.R., Wang, J., Cong, Q., Kinch, L.N., Schaeffer, R.D.,et al.: Accurate prediction of protein structures and interactions using a three-track neural network. Science373(6557), 871–876 (2021)

    work page 2021

  5. [5]

    Science384(6693), 2528 (2024)

    Krishna, R., Wang, J., Ahern, W., Sturmfels, P., Venkatesh, P., Kalvet, I., Lee, G.R., Morey-Burrows, F.S., Anishchenko, I., Humphreys, I.R.,et al.: Generalized biomolecular modeling and design with rosettafold all-atom. Science384(6693), 2528 (2024)

    work page 2024

  6. [6]

    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)

    work page 2023

  7. [7]

    Nature Machine Intelligence5(10), 1087–1096 (2023)

    Fang, X., Wang, F., Liu, L., He, J., Lin, D., Xiang, Y., Zhu, K., Zhang, X., Wu, H., Li, H.,et al.: A method for multiple-sequence-alignment-free protein structure prediction using a protein language model. Nature Machine Intelligence5(10), 1087–1096 (2023)

    work page 2023

  8. [8]

    HelixFold-Multimer: Elevating Protein Complex Structure Prediction to New Heights

    Fang, X., Gao, J., Hu, J., Liu, L., Xue, Y., Zhang, X., Zhu, K.: Helixfold-multimer: Elevating protein complex structure prediction to new heights. arXiv preprint arXiv:2404.10260 (2024)

    work page Pith review arXiv 2024

Show all 38 references
  1. [9]

    arXiv preprint arXiv:2408.16975 (2024)

    Liu, L., Zhang, S., Xue, Y., Ye, X., Zhu, K., Li, Y., Liu, Y., Gao, J., Zhao, W., Yu, H., et al.: Technical report of helixfold3 for biomolecular structure prediction. arXiv preprint arXiv:2408.16975 (2024)

    2024 arXiv

  2. [10]

    Nature methods19(6), 679–682 (2022)

    Mirdita, M., Sch¨ utze, K., Moriwaki, Y., Heo, L., Ovchinnikov, S., Steinegger, M.: Colabfold: making protein folding accessible to all. Nature methods19(6), 679–682 (2022)

    2022

  3. [11]

    Science, 0018 (2025)

    Hayes, T., Rao, R., Akin, H., Sofroniew, N.J., Oktay, D., Lin, Z., Verkuil, R., Tran, V.Q., Deaton, J., Wiggert, M., et al.: Simulating 500 million years of evolution with a language model. Science, 0018 (2025)

    2025

  4. [12]

    Bioinformatics39(9), 573 (2023)

    Wallner, B.: Afsample: improving multimer prediction with alphafold using massive sampling. Bioinformatics39(9), 573 (2023)

    2023

  5. [13]

    Communications Biology8(1), 373 (2025)

    Kalakoti, Y., Wallner, B.: Afsample2 predicts multiple conformations and ensembles with alphafold2. Communications Biology8(1), 373 (2025)

    2025

  6. [14]

    PLOS Computational Biology18(8), 1010483 (2022)

    Stein, R.A., Mchaourab, H.S.: Speach af: Sampling protein ensembles and conformational heterogeneity with alphafold2. PLOS Computational Biology18(8), 1010483 (2022)

    2022

  7. [15]

    Protein Science33(1), 4865 (2024)

    Yin, R., Pierce, B.G.: Evaluation of alphafold antibody–antigen modeling with implications for improving predictive accuracy. Protein Science33(1), 4865 (2024)

    2024

  8. [16]

    Research Square, 3 (2025)

    Xing, E., Zhang, J., Wang, S., Cheng, X.: Leveraging sequence purification for accurate prediction of multiple conformational states with alphafold2. Research Square, 3 (2025)

    2025

  9. [17]

    nature communications15(1), 2464 (2024) 27

    Silva, G., Cui, J.Y., Dalgarno, D.C., Lisi, G.P., Rubenstein, B.M.: High-throughput prediction of protein conformational distributions with subsampled alphafold2. nature communications15(1), 2464 (2024) 27

    2024

  10. [18]

    Nature625(7996), 832–839 (2024)

    Wayment-Steele, H.K., Ojoawo, A., Otten, R., Apitz, J.M., Pitsawong, W., H¨ omberger, M., Ovchin- nikov, S., Colwell, L., Kern, D.: Predicting multiple conformations via sequence clustering and alphafold2. Nature625(7996), 832–839 (2024)

    2024

  11. [19]

    Nature Communications 15(1), 7328 (2024)

    Bryant, P., No´ e, F.: Structure prediction of alternative protein conformations. Nature Communications 15(1), 7328 (2024)

    2024

  12. [20]

    Protein crystallography: methods and protocols, 627–641 (2017)

    Burley, S.K., Berman, H.M., Kleywegt, G.J., Markley, J.L., Nakamura, H., Velankar, S.: Protein data bank (pdb): the single global macromolecular structure archive. Protein crystallography: methods and protocols, 627–641 (2017)

    2017

  13. [21]

    Nature biotechnology35(11), 1026–1028 (2017)

    Steinegger, M., S¨ oding, J.: Mmseqs2 enables sensitive protein sequence searching for the analysis of massive data sets. Nature biotechnology35(11), 1026–1028 (2017)

    2017

  14. [22]

    BioRxiv, 2025–06 (2025)

    Passaro, S., Corso, G., Wohlwend, J., Reveiz, M., Thaler, S., Ram Somnath, V., Getz, N., Portnoi, T., Roy, J., Stark, H., et al.: Boltz-2: Towards accurate and efficient binding affinity prediction. BioRxiv, 2025–06 (2025)

    2025

  15. [23]

    BioRxiv, 2025–01 (2025)

    Team, B.A.A., Chen, X., Zhang, Y., Lu, C., Ma, W., Guan, J., Gong, C., Yang, J., Zhang, H., Zhang, K., et al.: Protenix-advancing structure prediction through a comprehensive alphafold3 reproduction. BioRxiv, 2025–01 (2025)

    2025

  16. [24]

    BioRxiv, 2024–10 (2024)

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

    2024

  17. [25]

    The Protein Journal27(1), 59–70 (2008)

    Yan, C., Wu, F., Jernigan, R.L., Dobbs, D., Honavar, V.: Characterization of protein–protein interfaces. The Protein Journal27(1), 59–70 (2008)

    2008

  18. [26]

    Journal of The Royal Society Interface10(79), 20120835 (2013)

    Kastritis, P.L., Bonvin, A.M.J.J.: On the binding affinity of macromolecular interactions: daring to ask why proteins interact. Journal of The Royal Society Interface10(79), 20120835 (2013)

    2013

  19. [27]

    Chemical Reviews118(4), 1691–1741 (2018)

    Zhou, H.-X., Pang, X.: Electrostatic interactions in protein structure, folding, binding, and condensa- tion. Chemical Reviews118(4), 1691–1741 (2018). PMID: 29319301

    2018

  20. [28]

    Beilstein Journal of Organic Chemistry17, 1–10 (2021)

    Seychell, B.C., Beck, T.: Molecular basis for protein–protein interactions. Beilstein Journal of Organic Chemistry17, 1–10 (2021)

    2021

  21. [29]

    Angewandte Chemie136(34), 202405823 (2024)

    Redrado-Hern´ andez, S., Mac ´ ıas-Le´ on, J., Castro-L´ opez, J., Bel´ en Sanz, A., Dolader, E., Arias, M., Gonz´ alez-Ram ´ ırez, A.M., S´ anchez-Navarro, D., Petryk, Y., Farkaˇ s, V.,et al.: Broad protection against invasive fungal disease from a nanobody targeting the acti...

    2024

  22. [30]

    Nature structural & molecular biology30(8), 1172–1182 (2023)

    Omura, S.N., Nakagawa, R., S¨ udfeld, C., Villegas Warren, R., Wu, W.Y., Hirano, H., Laffeber, C., Kusakizako, T., Kise, Y., Lebbink, J.H.,et al.: Mechanistic and evolutionary insights into a type vm crispr–cas effector enzyme. Nature structural & molecular biology30(8), 1172–...

    2023

  23. [31]

    arXiv preprint arXiv:1412.6980 (2014)

    Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

    2014 arXiv

  24. [32]

    Nucleic Acids Research50, 1368–1372

    Schneider, C., Raybould, M.I.J., Deane, C.M.: SAbDab in the age of biotherapeutics: updates including SAbDab-nano, the nanobody structure tracker. Nucleic Acids Research50, 1368–1372

  25. [33]

    Nucleic Acids Research28(1), 235–242 (2000)

    Berman, H.M., Westbrook, J., Feng, Z., Gilliland, G., Bhat, T.N., Weissig, H., Shindyalov, I.N., Bourne, P.E.: The Protein Data Bank. Nucleic Acids Research28(1), 235–242 (2000)

    2000

  26. [34]

    PLOS ONE11(8), 1–9 (2016)

    Basu, S., Wallner, B.: Dockq: A quality measure for protein-protein docking models. PLOS ONE11(8), 1–9 (2016)

    2016

  27. [35]

    Bioinformatics29(21), 2722–2728 (2013) 28

    Mariani, V., Biasini, M., Barbato, A., Schwede, T.: lddt: a local superposition-free score for comparing protein structures and models using distance difference tests. Bioinformatics29(21), 2722–2728 (2013) 28

    2013

  28. [36]

    http://www.rdkit.org

    Landrum, G.: RDKit: Open-source Cheminformatics. http://www.rdkit.org

  29. [37]

    Bioinformatics 32(2), 298–300 (2015)

    Dunbar, J., Deane, C.M.: Anarci: antigen receptor numbering and receptor classification. Bioinformatics 32(2), 298–300 (2015)

    2015

  30. [38]

    BMC Bioinformatics17(1), 81 (2016) 29

    Daily, J.: Parasail: Simd c library for global, semi-global, and local pairwise sequence alignments. BMC Bioinformatics17(1), 81 (2016) 29

    2016

Pith tools

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