CoChat
Sign Up
PreprintCited by 6Protein Structure and DynamicsMachine Learning in BioinformaticsComputational Drug Discovery Methods

The DOI matches its Crossref record; 1 of the 1 references it cites were found in a registry or an index.

SPDesign: protein sequence designer based on structural sequence profile using ultrafast shape recognition

Open access
Wang, Hui, Liu, Dong, Zhao, Kai-Long, Wang, Ya-Jun, Zhang, Gui-Jun·2023·Briefings in Bioinformatics
DOI 10.1101/2023.12.14.571651PMC PMC11006797

Preprint — published as 10.1093/bib/bbae146

Open this paper in CoChat — free account

Abstract

Designing protein with specified structure and function involves a key component named sequence design, which can provide valuable insights into understanding the life systems as well for the diagnosis and therapy of diseases. Although deep learning methods have made great progress in protein sequence design, most of these studies focus on network structure optimization, while ignoring protein-specific physicochemical features. Inspired by the successful application of structure templates and pre-trained models in the field of protein structure prediction, we explored whether the representation of structural sequence profile can be used for protein sequence design. In this work, we proposed SPDesign, a method for protein sequence design based on structural sequence profile using ultrafast shape recognition. Given an input back-bone structure, SPDesign utilizes ultrafast shape recognition vectors to accelerate the search for similar protein structures (aka, structural analogs) in our in-house PAcluster80 structure database, and then extracts the sequence profile from the analogs through structure alignment. Combined with structural pre-trained knowledge and geometric features, they are further feed into an enhanced graph neural network to predict the sequence. Experimental results show that SPDesign significantly outperforms the state-of-the-art methods, such as ProteinMPNN, Pifold and LM-Design, leading to 21.89%, 15.54% and 11.4% accuracy gains in sequence recovery rate on CATH 4.2 benchmark, respectively. Encouraging results also have been achieved on the TS50 and TS500 benchmarks, with performance reaching 68.64% and 71.63%. Furthermore, detailed analysis conducted by the PDBench tool suggest that SPDesign performs well in subdivided structures such as buried residues and solenoid. More interestingly, we found that SPDesign can well reconstruct the sequences of some proteins that have similar structures but different sequences. Finally, the structural modeling verification experiment bears out that the sequences designed by our method can fold into the native structures more accurately.

Overview

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1101/2023.12.14.571651

Cites

1 of 1 references verified0 citation errors · 0 not checkable

Statistics

Figures

Figure 1 The workflow of SPDesign. (A) The input backbone structure uses USR-V to search for structural analogs in PAcluster80, and the sequences of the matched part are extracted to create the structural sequence profile. The consensus sequence, which is converted to sequence-pattern embedding by the pre-trained model, and positional residue probability are extracted from the profile. (B) Distance map of the backbone atoms. (C) Pre-trained knowledge of structure. (D) Features are aggregated and encoded by the network and finally decoded into the predicted sequence.
Figure 1 The workflow of SPDesign. (A) The input backbone structure uses USR-V to search for structural analogs in PAcluster80, and the sequences of the matched part are extracted to create the structural sequence profile. The consensus sequence, which is converted to sequence-pattern embedding by the pre-trained model, and positional residue probability are extracted from the profile. (B) Distance map of the backbone atoms. (C) Pre-trained knowledge of structure. (D) Features are aggregated and encoded by the network and finally decoded into the predicted sequence.CC BY-NC · PubMed Central
Figure 2 (A) The left axis corresponds to the average recovery rate of designed sequences, and the right axis corresponds to the average sequence identity between the consensus sequence and native sequence. (B) Results comparison on orphan and de novo (designed) protein datasets.
Figure 2 (A) The left axis corresponds to the average recovery rate of designed sequences, and the right axis corresponds to the average sequence identity between the consensus sequence and native sequence. (B) Results comparison on orphan and de novo (designed) protein datasets.CC BY-NC · PubMed Central
Figure 3 Comparison of the similarity on the PDBench benchmark dataset. The X-axis represents various types of subdivision structures. The method has a higher similarity value on a certain type of structure, indicating a better performance on this type of structure.
Figure 3 Comparison of the similarity on the PDBench benchmark dataset. The X-axis represents various types of subdivision structures. The method has a higher similarity value on a certain type of structure, indicating a better performance on this type of structure.CC BY-NC · PubMed Central
Figure 4 (A) Comparison of the amino acid type distribution calculated by SPDesign with the native sequence amino acid type distribution. (B) Prediction bias of methods at each residue compared to the native sequence. A positive value represents that the proportion of the amino acid in the designed sequence is higher than in the native state, while a negative value indicates a proportion lower than in the native state.
Figure 4 (A) Comparison of the amino acid type distribution calculated by SPDesign with the native sequence amino acid type distribution. (B) Prediction bias of methods at each residue compared to the native sequence. A positive value represents that the proportion of the amino acid in the designed sequence is higher than in the native state, while a negative value indicates a proportion lower than in the native state.CC BY-NC · PubMed Central
Figure 5 Case study. (A) Structure of four proteins (PDB 3ney, 1kgd, 3wp0, 1ex6). The TM-score and sequence identity between pairs of these four proteins are as follows: 3ney-1kgd: 0.88, 55.2%; 3ney-3wp0: 0.77, 27.5%; 3ney-1ex6: 0.76, 31.2%; 1kgd-3wp0: 0.86, 26.5%; 1kgd-1ex6: 0.83, 26.7%; 3wp0-1ex6: 0.82, 31.5%. (B) Sequence alignment results between sequences designed by SPDesign (Designed seq) and the native sequences (marked by protein name). Conserved amino acids on the fragment sequence are underlined. Fragment-recovery represents the sequence recovery rate of SPDesign on the fragment sequence and conserved-recovery indicates the recovery rate on the conserved amino acids of the fragment. The average sequence identity between pairs of these four fragment sequences designed by SPDesign is 32.7%. (C) Alignment results between the structures predicted by AlphaFold2 using the designed sequences and the native structures.
Figure 5 Case study. (A) Structure of four proteins (PDB 3ney, 1kgd, 3wp0, 1ex6). The TM-score and sequence identity between pairs of these four proteins are as follows: 3ney-1kgd: 0.88, 55.2%; 3ney-3wp0: 0.77, 27.5%; 3ney-1ex6: 0.76, 31.2%; 1kgd-3wp0: 0.86, 26.5%; 1kgd-1ex6: 0.83, 26.7%; 3wp0-1ex6: 0.82, 31.5%. (B) Sequence alignment results between sequences designed by SPDesign (Designed seq) and the native sequences (marked by protein name). Conserved amino acids on the fragment sequence are underlined. Fragment-recovery represents the sequence recovery rate of SPDesign on the fragment sequence and conserved-recovery indicates the recovery rate on the conserved amino acids of the fragment. The average sequence identity between pairs of these four fragment sequences designed by SPDesign is 32.7%. (C) Alignment results between the structures predicted by AlphaFold2 using the designed sequences and the native structures.CC BY-NC · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Research works printed but not linked: 46

  1. 2019, Unified rational protein engineering with sequence-based deep representation learning, Nat. Methods, 16, 1315, 10.1038/s41592-019-0598-1 · DOI · Paper page
  2. 10.1093/nar/gkz1064 · DOI
  3. 2019, Accurate prediction of protein structures and interactions using a threetrack neural network, Science, 373, 871
  4. 2019, Ultrafast shape recognition to search compound databases for similar molecular shapes, J Comput Chem, 28, 1711
  5. 10.1006/jmbi.2000.4427 · DOI
  6. Burley, S. K. et al. (2017) Protein Data Bank (PDB): the single global macromolecular structure archive. Protein crystallography: methods and protocols, 627–641. · DOI · Paper page
  7. 2022, ProteinBERT: a universal deep-learning model of protein sequence and function, Bioinformatics, 38, 2102, 10.1093/bioinformatics/btac020 · DOI · Paper page
  8. 10.1093/bioinformatics/btad027 · DOI
  9. 10.1101/628917 · DOI
  10. 2022, Single-sequence protein structure prediction using a language model and deep learning, Nat. Biotechnol, 40, 1617, 10.1038/s41587-022-01432-w · DOI · Paper page
  11. 2021, High-level expression of palmitoylated MPP1 recombinant protein in mammalian cells, Membranes, 11, 715, 10.3390/membranes11090715 · DOI
  12. 10.1126/science.add2187 · DOI · Paper page
  13. 10.3390/ijms222111741 · DOI
  14. 10.1093/bib/bbac102 · DOI · Paper page
  15. 10.48550/arXiv.2202.01079 · DOI
  16. 10.48550/arXiv.2305.15151 · DOI
  17. 10.48550/arXiv.2209.12643 · DOI
  18. 2017, Neural message passing for quantum chemistry, ICML 2017, 3, 2053
  19. 2022, DeepUMQA: ultrafast shape recognition-based protein model quality assessment using deep learning, Bioinformatics, 38, 1895, 10.1093/bioinformatics/btac056 · DOI · Paper page
  20. Hsu, C. et al. (2022) Learning inverse folding from millions of predicted structures. ICML 2022, 8946–8970. · DOI
  21. 2023, Accurate and efficient protein sequence design through learning concise local environment of residues, Bioinformatics, 39, btad122, 10.1093/bioinformatics/btad122 · DOI · Paper page
  22. 10.48550/arXiv.2009.01411 · DOI
  23. 2020, De novo protein design for novel folds using guided conditional wasserstein generative adversarial networks, J Chem Inf Model, 60, 5667, 10.1021/acs.jcim.0c00593 · DOI
  24. 10.1074/jbc.M110792200 · DOI
  25. 10.1002/prot.24620 · DOI
  26. 10.1101/2022.07.20.500902 · DOI · Paper page
  27. 2018, Minimalist de novo design of protein catalysts, ACS Catal, 9, 9265
  28. 2018, Spin2: Predicting sequence profiles from protein structures using deep neural networks, Proteins, 86, 629, 10.1002/prot.25489 · DOI
  29. 2021, Structure-based protein design with deep learning, Curr Opin Chem Biol, 65, 136, 10.1016/j.cbpa.2021.08.004 · DOI · Paper page
  30. 10.1016/j.sbi.2021.01.007 · DOI · Paper page
  31. 10.1021/acs.jcim.0c00043 · DOI
  32. 2019, Evaluating Protein Transfer Learning with TAPE, NeurIPS 2019, 32, 9689
  33. 2020, CATH: increased structural coverage of functional space, Nucleic Acids Res, 49, D266
  34. 10.1016/j.sbi.2021.11.008 · DOI
  35. 10.48550/arXiv.2204.10673 · DOI
  36. 2021, AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space with high-accuracy models, Nucleic Acids Res, 50, D439
  37. Vaswani, A. et al. (2017) Attention is all you need. NeurIPS 2017, 30.
  38. 10.1038/s41598-018-23276-8 · DOI
  39. 2021, Representing long-range context for graph neural networks with global attention, NeurIPS 2021, 34, 13266
  40. 2023, SPIN-CGNN: Improved fixed backbone protein design with contact map-based graph construction and contact graph neural network, PLOS COMPUT BIOL, 19, e1011330, 10.1371/journal.pcbi.1011330 · DOI · Paper page
  41. 10.1093/nar/gki524 · DOI · Paper page
  42. 2020, Prodconn: Protein design using a convolutional neural network, Proteins, 88, 819, 10.1002/prot.25868 · DOI
  43. 2023, Protein structure and folding pathway prediction based on remote homologs recognition using pathreader, Commun. Biol, 6, 243, 10.1038/s42003-023-04605-8 · DOI
  44. 2021, MMpred: a distance-assisted multimodal conformation sampling for de novo protein structure prediction, Bioinformatics, 37, 4350, 10.1093/bioinformatics/btab484 · DOI
  45. 10.48550/arXiv.2302.01649 · DOI
  46. 10.1038/emboj.2011.428 · DOI

Metadata

Authors
Wang, H., Liu, D., Zhao, K.-L., Wang, Y.-J., & Zhang, G.-J.
ORCID
Venue
Briefings in Bioinformatics
Publisher
openRxiv
Published
2023-12-14
Volume
25
Issue
3
ISSN
1467-5463 (ISSN-L)
DOI
10.1101/2023.12.14.571651
PMC
PMC11006797
OpenAlex
W4394741301
Cited by
6
References
1
Citations by year
2026: 2 · 2025: 2 · 2024: 2
FWCI
0.81
Type
journal article
Language
en
Open access
Licence
cc-by-nc
Locations
  • Briefings in Bioinformatics (publishedVersion, cc-by-nc)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion)
  • PubMed Central (publishedVersion, cc-by-nc)
  • PubMed (publishedVersion)
Funders
  • National Natural Science Foundation of China (LZ20F030002, 62173304)
  • National Science and Technology Major Project (2022ZD0115103)
  • Natural Science Foundation of Zhejiang Province (LZ20F030002)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Protein Structure and Dynamics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Machine Learning in Bioinformatics
  • Life Sciences › Pharmacology, Toxicology and Pharmaceutics › Drug Discovery › Computational Drug Discovery Methods
Keywords
protein sequence design, graph neural networks, structural alignment, sequence recovery, de novo protein design, orphan proteins, protein structure prediction, geometric features
MeSH terms
Amino Acid Sequence, Proteins, Sequence Alignment, Neural Networks, Computer, Sequence Analysis, Protein

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review