Pith. sign in

REVIEW 4 major objections 6 minor 38 references

IntFold: A Controllable Foundation Model for General and Specialized Biomolecular Structure Prediction

T0 review · 4 major / 6 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 IntFold is a biomolecular structure prediction model that aims to match AlphaFold 3's accuracy while adding adapters for allosteric states, constrained folding, and binding-affinity prediction.

desk verdict A credible AF3-scale reproduction with useful adapters, but the benchmark claims are undercut by uncontrolled ranking and self-distillation. read the letter →

arxiv 2507.02025 v2 pith:BYMY4EVQ submitted 2025-07-02 q-bio.BM

The IntFold Team: Leon Qiao , Wayne Bai , He Yan , Gary Liu , Nova Xi , Xiang Zhang , Siqi Sun This is my paper · ORCID
classification q-bio.BM
keywords biomolecularstructurepredictionAlphaFold3diffusionmodelLoRAadaptersallostericconformationbindingaffinityantibody-antigencomplexesattentionkernel
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

IntFold is a biomolecular structure prediction model that aims to match the accuracy of AlphaFold 3 across proteins, nucleic acids, and small-molecule complexes while adding a controllability layer that AlphaFold 3 lacks. The authors claim that on the FoldBench benchmark IntFold's success rates are comparable to AlphaFold 3 for protein-protein interactions, protein-ligand interactions, and antibody-antigen complexes, and that it outperforms other open models. The central innovation is modular: small trainable adapters are inserted into a frozen base model to steer predictions toward allosteric conformations, user-specified binding-site constraints, and binding-affinity estimates. If these claims hold, a single model could serve both broad structure prediction and targeted drug-design tasks without retraining the base network.

What carries the argument

The load-bearing mechanism is the adapter architecture: a frozen, AlphaFold-3-style base model with a diffusion structure module and a confidence head, into which small trainable modules are inserted. For target-type control the authors use per-layer Low-Rank Adaptation (LoRA) modules; for guided folding they add an extra embedder that encodes known binding-pocket or epitope residues; for affinity they attach a post-hoc module of four Pairformer blocks (pair-bias transformer layers) that outputs a fitness score. A second named mechanism is the FlashAttentionPairBias kernel, a Triton implementation that broadcasts the pair-bias slice on the fly to avoid materializing the full attention tensor, reducing memory and latency. A third is the model-agnostic ranking method: generate 25 structures, rank them by mean pairwise DockQ, and pick the most self-consistent one, which the authors report adds about 3 percent success on antibody-antigen benchmarks.

What would settle it

A concrete check would be to rerun the IntFold+ training recipe with the antibody-antigen distillation set replaced by experimentally determined structures or by a different model's higher-quality predictions, then compare antibody-antigen success rates on the same FoldBench subset; if the 43.2 percent success rate and the 69.0 percent guided-folding number drop toward the base model's 37.6 percent, the distillation labels are doing the work. A second check is to release the exact PLAbDab cluster split and filtering thresholds so independent groups can test whether the ranking filter selects the same targets that the benchmark rewards.

Watch

Extended reading notes

Core claim

On the paper's own terms, IntFold is a foundation model that reaches accuracy comparable to AlphaFold 3 on general biomolecular structure prediction while remaining controllable. The authors demonstrate this on FoldBench with monomer LDDT 0.88, protein-protein success 72.9 percent, antibody-antigen success 37.6 percent for the base model and 43.2 percent for the fine-tuned IntFold+, and protein-ligand success 58.5 percent, with IntFold+ reaching 61.8 percent. They also show that RNA monomer prediction exceeds AlphaFold 3 in mean LDDT, and that adapter-based specialization captures closed allosteric CDK2 conformations, doubles antibody-antigen success when epitope constraints are supplied, and yields binding-affinity predictions that beat Boltz-2 on the reported benchmarks. The conclusion is that a single frozen base model equipped with lightweight adapters can cover both general prediction and specialized drug-design tasks.

Load-bearing premise

The weakest link is the antibody-antigen distillation set: IntFold's own predictions on PLAbDab sequences, filtered by the authors' ranking method, are used as training labels for IntFold+, so if those self-generated labels are biased or inaccurate, the reported antibody-antigen gains may partly reflect self-selection rather than true generalization.

Editorial extensions

If this is right

  • IntFold is claimed to match AlphaFold 3 on protein monomers and protein-protein complexes, making it an independent alternative for general structure prediction without requiring access to the original model.
  • The adapter mechanism is claimed to turn one base model into a specialist for allosteric conformations, constrained docking, and affinity prediction, so drug-design workflows could reuse a single frozen trunk.
  • The similarity-based ranking method is claimed to be model-agnostic and training-free, so it could be bolted onto any diffusion-based structure predictor to improve success rates over random sample selection.
  • The custom attention kernel is claimed to be faster and lower-memory than existing kernels, which could make large-model training and inference cheaper.

Reading between the lines

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

  • If the self-distillation labels are as strong as the benchmarks suggest, the same loop could be extended to protein-ligand and nucleic-acid complexes, where experimental labels are sparser.
  • The consensus ranking method implicitly assumes that the model's errors are spread around the true structure; on targets where the model is confidently wrong in the same way across seeds, consensus would select the wrong pose, which could be tested on deliberately adversarial targets.
  • A testable extension is to reuse the constraint embedder for de novo design, using a desired pocket shape as the constraint and then decoding sequences from the conditioned structure.
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

4 major / 6 minor

Summary. The manuscript introduces IntFold, a diffusion-based foundation model for general biomolecular structure prediction, claiming accuracy comparable to AlphaFold 3 on the FoldBench benchmark and superior performance over other contemporary methods. It also presents a custom FlashAttentionPairBias kernel, a training-free model-agnostic ranking method based on structural similarity, and modular adapters for specialized tasks such as allosteric-state prediction, constraint-guided folding, and binding affinity estimation. The paper reports benchmark numbers, several case studies on newly released structures, and engineering insights into training stability. The central claim is that IntFold is an independent, controllable model that matches AF3-level accuracy while enabling practical task-specific adaptation.

Significance. If the central claims hold, IntFold would be a substantial contribution: a reproducible alternative to AlphaFold 3 with competitive accuracy, plus flexible adapters for drug-design-relevant tasks. The paper also provides a named consensus-ranking method that could be of independent methodological value, and the attention kernel engineering is a concrete systems contribution. These strengths are counterbalanced by the fact that the benchmark evidence is partially self-reported, the key head-to-head comparisons do not apply the same selection protocol to all models, and the antibody-antigen distillation procedure risks circular evaluation. The paper ships source code and a server, which is commendable, but the reported quantitative claims need additional experimental support to be fully convincing.

major comments (4)
  1. [§5.2, Figures 2-3] The reported success rates for IntFold use the 25-sample consensus ranking (5 seeds x 5 samples) with selection by highest mean DockQ, as described in §5.2. The comparison models' metrics are either cited from FoldBench or rerun by the authors, with no indication that the same model-agnostic ranking protocol and sampling budget were applied to them. Since the ranking method is advertised as model-agnostic, fairness requires applying it to every contender; otherwise the observed gaps (e.g., 72.9% vs. 68.5% for protein-protein, 43.2% vs. 47.9% for antibody-antigen, and 58.5% vs. 55.0% for protein-ligand) could be entirely explained by the consensus selector and additional compute rather than by IntFold's learned representations. This is a load-bearing issue for the paper's headline claim of outperforming other leading approaches.
  2. [§4.1, IntFold+ and guided-folding results] The antibody-antigen distillation set is constructed by running IntFold on PLAbDab sequences and filtering predictions using the authors' own ranking method; IntFold+ is then fine-tuned on this set and evaluated on antibody-antigen benchmarks. Similarly, the guided-folding adapter is trained with structural constraints and evaluated on the same type of complexes. The reported improvements (37.6% to 43.2% for IntFold+, and 37.6% to 69.0% for guided folding) may therefore reflect a self-selection bias: the filtering step uses the same consensus criterion that later defines evaluation success, and the fine-tuning labels may contain systematic errors that the model simply learns to reproduce. The manuscript provides no independent validation set, no comparison with a model trained on experimental antibody-antigen structures only, and no leakage analysis to show that test complexes are disjoint from distillation inputs. Without such controls, these gains are not attributable to genuine generalization.
  3. [§2, Tables and Figures] The benchmark numbers are presented without confidence intervals, significance tests, or per-target breakdowns, and AlphaFold 3 metrics are cited from the original FoldBench benchmark rather than rerun in the same harness. Given the small differences on several tasks (e.g., protein monomer LDDT 0.88 for both IntFold and AF3; protein-protein success 72.9% for both), a single random or systematic variation in the evaluation protocol could change the qualitative conclusions. The paper should supply error bars, per-target lists, or a reproducibility statement describing the exact sampling and selection protocol used for each model, so that the 'comparable to AlphaFold 3' claim is supported by comparable measurement conditions.
  4. [Appendix, IntFold+ description] The appendix explicitly states that 'a detailed ablation study to quantify the individual contribution of each modification—such as the best data recipe, removal of MSA pairing, and pLDDT masking—remains an area for future work.' This admission is important because IntFold+ is presented as a core result that narrows the gap to AF3, yet the recipe combines data enrichment, loss masking, and MSA changes without any attribution. The paper should either provide the ablations or temper the claims about IntFold+'s effectiveness, since the current text leaves open the possibility that the gains come primarily from the distillation set or protocol changes rather than the model architecture.
minor comments (6)
  1. [§5.4.3] The sentence 'we required the data crops much contain at least four resolved residues' contains a typo; it should read 'must contain.'
  2. [§5.2] The 'approximately 3%' improvement from the ranking method is only quantified for the antibody-antigen benchmark; for the reader to assess the method's generality, it should be reported for all benchmark categories and ideally applied to other models.
  3. [§5.1, Figure 8] The kernel benchmark lacks details on hardware, tensor shapes, batch sizes, and precision; without this information the latency and memory comparisons cannot be reproduced or compared with alternative implementations.
  4. [§2.2, PoseBusters] The PoseBusters v2 filtered set is described as 'targets deposited after the training data cutoff of 2021-09-30,' but the paper does not explain how this cutoff relates to the training cutoffs of the comparison models (e.g., Protenix), which could affect the fairness of the comparison.
  5. [References] Several references have inconsistent formatting (e.g., [26] contains raw LaTeX artifacts), and the reference list should be cleaned up before publication.
  6. [§3.1] The CDK2 allosteric-state results are presented for a test set of 40 structures, but the paper does not state how the 5 allosteric cases were selected or whether they overlap with the training set after the described split; a clearer description of the split and evaluation protocol would strengthen this application section.

Circularity Check

2 steps flagged · score 6.0 of 10

Core FoldBench/PoseBusters accuracy claims are externally grounded, but the affinity benchmark on BindingDB is trained on BindingDB itself, and the IntFold+ antibody-antigen gains rest on self-generated distillation labels without a demonstrated disjoint split.

  1. fitted input called prediction [Section 3.3 (Predicting Protein-Ligand Binding Affinity) and Section 4.1 (Affinity Dataset)]
    "we trained a post-hoc prediction module on the curated affinity dataset from Section 4.1. The model's performance was evaluated on the standard DAVIS [11] and BindingDB [12] benchmark datasets ... We built a protein–ligand affinity set by merging activity data from ChEMBL [20], BindingDB [12], GalaxyDB [21], BioLip [22] and PubChem [23]"

    BindingDB is explicitly one of the databases merged into the training set and is also one of the two evaluation benchmarks. No split, temporal cutoff, or exclusion of BindingDB entries from the training set is described anywhere in Section 4.1 or Section 3.3. The reported AUPR on BindingDB therefore measures the post-hoc module's ability to reproduce labels it was trained on, not its ability to predict unknown affinities. This is a fitted input renamed as an evaluation prediction for the affinity application.

  2. fitted input called prediction [Section 4.1 (Antibody-Antigen Distillation), Appendix (IntFold+), Section 2.1 (Ab-Ag results)]
    "On the resulting pairs, we performed large-scale structural predictions using IntFold. The distillation set was formed by filtering predictions to examples selected by our ranking method. ... we augmented the training data by including our curated antibody-antigen distillation set 4.1. ... IntFold+ improves the success rate to 43.2%"

    IntFold+ is fine-tuned on antibody-antigen labels produced by IntFold itself and filtered by the paper's own ranking method, then the resulting Ab-Ag success rate is reported on the same task family. The paper does not establish that the FoldBench Ab-Ag evaluation targets are disjoint from the PLAbDab-derived distillation set, and the distillation labels are not independent experimental ground truth. Without such a disjointness guarantee, the 37.6% to 43.2% gain can be explained by the model being trained to reproduce its own ranking-filtered outputs rather than by new external evidence.

full rationale

The general structure-prediction claims (monomer LDDT, protein-protein, protein-ligand, nucleic-acid success rates on FoldBench, and PoseBusters v2 after a 2021-09-30 cutoff) are evaluated against external experimental structures, so those central accuracy comparisons are not circular. The circularity is confined to specialized sub-claims. First, the affinity module is trained on a merged dataset that includes BindingDB and then evaluated on BindingDB; the quoted passages contain no train/test split, making the BindingDB AUPR a fit. Second, IntFold+ is trained on antibody-antigen predictions generated by IntFold and filtered by the same ranking method, and its improved Ab-Ag numbers are presented without a demonstrated disjoint split from that distillation set, so the specialized Ab-Ag gain is not independently grounded. The ranking method's inconsistent application to baselines is a comparison-fairness concern rather than a circularity of derivation; it is noted but not scored. Overall, the main benchmark results stand on external data, but the affinity and Ab-Ag specialization claims contain at least one prediction that reduces to its own training input, giving a partial circularity score of 6.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central accuracy claims are empirical, not derived, so the free-parameter ledger is dominated by the trained state of the model and by hand-chosen data-curation thresholds rather than by physical constants. No new physical entities are introduced; 'IntFold+' and the adapters are software artifacts, not conjectured entities. The main structural risk is the self-distillation loop in Section 4.1, where the model generates its own antibody-antigen training labels and the same ranking method filters them.

free parameters (2)
  • Trained base model, adapter, and affinity module weights = not disclosed
    The reported benchmark and application results are produced by learned weights. No parameter count, training loss curves, or checkpoint hashes are provided, so the fitted state of the model cannot be independently reproduced.
  • Data curation thresholds = multiple thresholds, e.g., pLDDT > 85, 30% clustering, 60% Ab-Ag clustering, ligand similarity 0.5
    Hand-chosen thresholds in data curation (Section 4.1) affect which training examples are used and hence model behavior; their values are reported but not justified by ablations.
assumptions (3)
  • domain assumption The AlphaFold 3 design (pairformer, diffusion, confidence head) is a sound basis for an all-atom structure prediction model.
    Section 5 states the architecture and training procedures are based on AlphaFold 3 supplementary materials; the base design is inherited from prior work, not proven here.
  • domain assumption FoldBench and PoseBusters are fair, complete external benchmarks for the claims made.
    All headline accuracy comparisons rest on these benchmarks; their composition and scoring rules are taken as given.
  • ad hoc to paper The antibody-antigen distillation predictions generated by IntFold and filtered by the ranking method are accurate enough to serve as training labels.
    Section 4.1 uses large-scale structural predictions using IntFold filtered by the authors' ranking method to construct the Ab-Ag distillation set; this self-training loop is assumed to help, and no independent validation is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of IntFold: A Controllable Foundation Model for General and Specialized Biomolecular Structure Prediction." pith.science (2026). https://pith.science/paper/BYMY4EVQ

@misc{pith2026250702025,
  author       = {Pith},
  title        = {Pith review of: IntFold: A Controllable Foundation Model for General and Specialized Biomolecular Structure Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BYMY4EVQ}},
  note         = {Machine review of arXiv:2507.02025}
}
read the original abstract

We introduce IntFold, a controllable foundation model for general and specialized biomolecular structure prediction. Utilizing a high-performance custom attention kernel, IntFold achieves accuracy comparable to the state-of-the-art AlphaFold 3 on a comprehensive benchmark of diverse biomolecular structures, while also significantly outperforming other leading all-atom prediction approaches. The model's key innovation is its controllability, enabling downstream applications critical for drug screening and design. Through specialized adapters, it can be precisely guided to predict complex allosteric states, apply user-defined structural constraints, and estimate binding affinity. Furthermore, we present a training-free, similarity-based method for ranking predictions that improves success rates in a model-agnostic manner. This report details these advancements and shares insights from the training and development of this large-scale model.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 33 canonical work pages

  1. [1]

    Highly accurate protein structure prediction with AlphaFold

    John Jumper et al. “Highly accurate protein structure prediction with AlphaFold”. In:nature 596.7873 (2021), pp. 583–589

    work page 2021

  2. [2]

    AccuratestructurepredictionofbiomolecularinteractionswithAlphaFold 3

    JoshAbramsonetal.“AccuratestructurepredictionofbiomolecularinteractionswithAlphaFold 3”. In:Nature 630.8016 (2024), pp. 493–500. 14 IntFold: A Controllable Foundation Model for General and Specialized Biomolecular Structure Prediction

    work page 2024

  3. [3]

    Challenge for Deep Learning: Protein Structure Prediction of Ligand- Induced Conformational Changes at Allosteric and Orthosteric Sites

    Gustav Olanders et al. “Challenge for Deep Learning: Protein Structure Prediction of Ligand- Induced Conformational Changes at Allosteric and Orthosteric Sites”. In:Journal of Chemical Information and Modeling 64.22 (2024), pp. 8481–8494

    work page 2024

  4. [4]

    FoldBench: An All-atom Benchmark for Biomolecular Structure Prediction

    Sheng Xu et al. “FoldBench: An All-atom Benchmark for Biomolecular Structure Prediction”. In: bioRxiv (2025), pp. 2025–05

    work page 2025

  5. [5]

    Boltz-1: Democratizing Biomolecular Interaction Modeling

    Jeremy Wohlwend et al. “Boltz-1: Democratizing Biomolecular Interaction Modeling”. In: bioRxiv (2024), pp. 2024–11

    work page 2024

  6. [6]

    Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction

    Saro Passaro et al. “Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction”. In: bioRxiv (2025), pp. 2025–06

    work page 2025

  7. [7]

    Chai-1: Decoding the molecular interactions of life

    Chai Discovery team et al. “Chai-1: Decoding the molecular interactions of life”. In:BioRxiv (2024), pp. 2024–10

    work page 2024

  8. [8]

    Technical report of HelixFold3 for biomolecular structure prediction

    Lihang Liu et al. “Technical report of HelixFold3 for biomolecular structure prediction”. In: arXiv preprint arXiv:2408.16975 (2024)

    arXiv 2024

Show all 38 references
  1. [9]

    Protenix-advancing structure prediction through a comprehensive AlphaFold3 reproduction

    ByteDance AML AI4Science Team et al. “Protenix-advancing structure prediction through a comprehensive AlphaFold3 reproduction”. In:bioRxiv (2025), pp. 2025–01

    2025

  2. [10]

    PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences

    Martin Buttenschoen, Garrett M Morris, and Charlotte M Deane. “PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences”. In: Chemical Science 15.9 (2024), pp. 3130–3139

    2024

  3. [11]

    Comprehensive analysis of kinase inhibitor selectivity

    Mindy I Davis et al. “Comprehensive analysis of kinase inhibitor selectivity”. In:Nature biotech- nology 29.11 (2011), pp. 1046–1051

    2011

  4. [12]

    BindingDB: a web-accessible database of experimentally determined protein– ligand binding affinities

    Tiqing Liu et al. “BindingDB: a web-accessible database of experimentally determined protein– ligand binding affinities”. In:Nucleic acids research 35.suppl_1 (2007), pp. D198–D201

    2007

  5. [13]

    Assessmentofpharmaceuticalprotein-ligandposeandaffinitypredictions in CASP16

    MichaelGilsonetal.“Assessmentofpharmaceuticalprotein-ligandposeandaffinitypredictions in CASP16”. In:Assessment 15 (2025), p. 4

    2025

  6. [14]

    The protein data bank

    Helen M Berman et al. “The protein data bank”. In:Biological Crystallography 58.6 (2002), pp. 899–907

    2002

  7. [15]

    AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space with high-accuracy models

    Mihaly Varadi et al. “AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space with high-accuracy models”. In:Nucleic acids research 50.D1 (2022), pp. D439–D444

    2022

  8. [16]

    Simulating 500 million years of evolution with a language model

    Thomas Hayes et al. “Simulating 500 million years of evolution with a language model”. In: Science (2025), eads0018

    2025

  9. [17]

    MGnify: the microbiome analysis resource in 2020

    Alex L Mitchell et al. “MGnify: the microbiome analysis resource in 2020”. In:Nucleic acids research 48.D1 (2020), pp. D570–D578

    2020

  10. [18]

    Protein complex prediction with AlphaFold-Multimer

    Richard Evans et al. “Protein complex prediction with AlphaFold-Multimer”. In:biorxiv (2021), pp. 2021–10

    2021

  11. [19]

    The Patent and Literature Antibody Database (PLAbDab): an evolving reference set of functionally diverse, literature-annotated antibody sequences and structures

    Brennan Abanades et al. “The Patent and Literature Antibody Database (PLAbDab): an evolving reference set of functionally diverse, literature-annotated antibody sequences and structures”. In: Nucleic Acids Research 52.D1 (2024), pp. D545–D551

    2024

  12. [20]

    The ChEMBL Database in 2023: a drug discovery platform spanning multiple bioactivity data types and time periods

    Barbara Zdrazil et al. “The ChEMBL Database in 2023: a drug discovery platform spanning multiple bioactivity data types and time periods”. In:Nucleic acids research 52.D1 (2024), pp. D1180–D1192

    2024

  13. [21]

    Structure-aware multimodal deep learning for drug–protein interaction prediction

    Penglei Wang et al. “Structure-aware multimodal deep learning for drug–protein interaction prediction”. In:Journal of chemical information and modeling 62.5 (2022), pp. 1308–1317. 15 IntFold: A Controllable Foundation Model for General and Specialized Biomolecular Structure Prediction

    2022

  14. [22]

    BioLiP: a semi-manually curated database for biologically relevant ligand–protein interactions

    Jianyi Yang, Ambrish Roy, and Yang Zhang. “BioLiP: a semi-manually curated database for biologically relevant ligand–protein interactions”. In:Nucleic acids research 41.D1 (2012), pp. D1096–D1103

    2012

  15. [23]

    PubChem 2023 update

    Sunghwan Kim et al. “PubChem 2023 update”. In: Nucleic acids research 51.D1 (2023), pp. D1373–D1380

    2023

  16. [24]

    ExCAPE-DB: an integrated large scale dataset facilitating Big Data analysis in chemogenomics

    Jiangming Sun et al. “ExCAPE-DB: an integrated large scale dataset facilitating Big Data analysis in chemogenomics”. In:Journal of cheminformatics 9 (2017), pp. 1–9

    2017

  17. [25]

    Hidden Markov model speed heuristic and iterative HMM search procedure

    L Steven Johnson, Sean R Eddy, and Elon Portugaly. “Hidden Markov model speed heuristic and iterative HMM search procedure”. In:BMC bioinformatics 11 (2010), pp. 1–8

    2010

  18. [26]

    UniProt: the Universal Protein Knowledgebase in 2023

    The UniProt Consortium. “UniProt: the Universal Protein Knowledgebase in 2023”. In:Nucleic Acids Research 51.D1 (2022), pp. D523–D531. i s sn: 0305-1048. d oi: 10 . 1093 / nar / gkac1052. eprint: https://academic.oup.com/nar/article- pdf/51/D1/D523/ 48441158/gkac1052.pdf. u r ...

    2022 doi

  19. [27]

    ColabFold: making protein folding accessible to all

    Milot Mirdita et al. “ColabFold: making protein folding accessible to all”. In:Nature methods 19.6 (2022), pp. 679–682

    2022

  20. [28]

    MMseqs2 enables sensitive protein sequence search- ing for the analysis of massive data sets

    Martin Steinegger and Johannes Söding. “MMseqs2 enables sensitive protein sequence search- ing for the analysis of massive data sets”. In:Nature biotechnology 35.11 (2017), pp. 1026– 1028

    2017

  21. [29]

    FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning

    Tri Dao. “FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning”. In: The Twelfth International Conference on Learning Representations . 2024

    2024

  22. [30]

    Lora: Low-rank adaptation of large language models

    Edward J Hu et al. “Lora: Low-rank adaptation of large language models.” In:ICLR 1.2 (2022), p. 3

    2022

  23. [31]

    On layer normalization in the transformer architecture

    Ruibin Xiong et al. “On layer normalization in the transformer architecture”. In:International conference on machine learning. PMLR. 2020, pp. 10524–10533

    2020

  24. [32]

    Glu variants improve transformer

    Noam Shazeer. “Glu variants improve transformer”. In:arXiv preprint arXiv:2002.05202 (2020)

    2020 arXiv

  25. [33]

    Cogview: Mastering text-to-image generation via transformers

    Ming Ding et al. “Cogview: Mastering text-to-image generation via transformers”. In:Advances in neural information processing systems 34 (2021), pp. 19822–19835

    2021

  26. [34]

    Query-Key Normalization for Transformers

    Alex Henry et al. “Query-Key Normalization for Transformers”. In:Findings of the Association for Computational Linguistics: EMNLP 2020 . 2020, pp. 4246–4253

    2020

  27. [35]

    Accurate, large minibatch sgd: Training imagenet in 1 hour

    Priya Goyal et al. “Accurate, large minibatch sgd: Training imagenet in 1 hour”. In:arXiv preprint arXiv:1706.02677 (2017)

    2017 arXiv

  28. [36]

    Delving deep into rectifiers: Surpassing human-level performance on imagenet classification

    Kaiming He et al. “Delving deep into rectifiers: Surpassing human-level performance on imagenet classification”. In:Proceedings of the IEEE international conference on computer vision . 2015, pp. 1026–1034

    2015

  29. [37]

    Efficient backprop

    Yann LeCun et al. “Efficient backprop”. In:Neural networks: Tricks of the trade . Springer, 2002, pp. 9–50

    2002

  30. [38]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. “Scalable diffusion models with transformers”. In:Proceedings of the IEEE/CVF international conference on computer vision . 2023, pp. 4195–4205. 16 IntFold: A Controllable Foundation Model for General and Specialized Biomolecular Structure Pred...

    2023

Pith tools

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