CoChat
Sign Up
Cited by 157SARS-CoV-2 and COVID-19 Researchvaccines and immunoinformatics approachesMonoclonal and Polyclonal Antibodies Research

Deep learning guided optimization of human antibody against SARS-CoV-2 variants with broad neutralization

Open access
Shan, Sisi, Luo, Shitong, Yang, Ziqing, Hong, Junxian, Su, Yufeng, Ding, Fan, Fu, Lili, Li, Chenyu +7 more·2022·Proceedings of the National Academy of Sciences
DOI 10.1073/pnas.2122954119PMC PMC8931377
Open this paper in CoChat — free account

Abstract

Significance: SARS-CoV-2 continues to evolve through emerging variants, more frequently observed with higher transmissibility. Despite the wide application of vaccines and antibodies, the selection pressure on the Spike protein may lead to further evolution of variants that include mutations that can evade immune response. To catch up with the virus’s evolution, we introduced a deep learning approach to redesign the complementarity-determining regions (CDRs) to target multiple virus variants and obtained an antibody that broadly neutralizes SARS-CoV-2 variants.

AI snapshot

Written from the paper’s full text.

TL;DR

A deep learning geometric attention model was used to redesign complementarity-determining regions of human antibody P36-5D2, yielding optimized antibodies with broad neutralization against SARS-CoV-2 variants. Six optimized antibodies achieved average IC50 values of 0.006–0.017 μg/mL and 10- to 600-fold improved neutralization against variants including Delta.

Key findings

  • Six optimized antibodies (HX001-013, HX001-015, HX001-020, HX001-024, HX001-033, and HX001-034) showed the most potent and broad activity, with average IC50 between 0.006 μg/mL and 0.017 μg/mL; optimized antibodies achieved 10- to 600-fold improvement in neutralizing activity against SARS-CoV-2 variants, including Delta.
  • The single mutation R103M in HCDR3 improved neutralization against all 10 pseudotyped viruses with average IC50 0.038 μg/mL; combinations improved further, e.g., HX001-013 (R103M + L104F) average IC50 0.007 μg/mL, HX001-020 (R103M + T31W + L104F) 0.006 μg/mL, HX001-024 (R103M + N57L + L104F) 0.006 μg/mL, and HX001-033 (R103M + T31W + N57L) 0.010 μg/mL. The quadruple mutant HX001-034 had IC50 0.017 μg/mL, slightly weaker than the best triple mutants.
  • SPR showed optimized antibodies bound WT RBD with KD 1.2 nM to 0.42 nM, 20- to 50-fold stronger than original P36-5D2 (KD 22.9 nM); Delta RBD KD was 6.22–26.4 nM, twofold stronger than or similar to P36-5D2 (11.4 nM). Off-rates improved from about 10⁻² to 10⁻³.
  • For Omicron-related mutations, optimized antibodies from HX001-020 were onefold to threefold higher than HX-020 against N440K and 20- to 100-fold stronger against pseudotyped virus carrying G446S.

Methods

  • An attention-based geometric neural network was trained on SKEMPI V2.0 (342 complexes, 5,318 mutations) to predict ΔΔG from protein complex structures, using split-by-complex fivefold cross-validation and benchmarking on S1131 and M1707 subsets; predictions were ensembled with GeoPPI and Rosetta.
  • In silico mutation libraries were built by enumerating CDR single and double mutations, sampling mutant complex structures with Rosetta relax and Cartesian ΔΔG, and ranking mutations by the ensemble average; top candidates were tested iteratively in pseudotyped virus neutralization assays.
  • Experimental validation used a panel of 10 pseudotyped viruses (WT 614G, Alpha, Beta, Gamma, Delta, Delta Plus, Kappa, Epsilon, Eta, and N439K) in HeLa-hACE2 cells, SPR binding kinetics to WT and Delta RBD, and antibodies produced in HEK293F cells.

Limitations

  • The paper does not include a formal limitations section.
  • Evident from the text: optimization and validation relied primarily on pseudotyped viruses and binding assays, with no reported authentic SARS-CoV-2 challenge or in vivo protection data.
  • Evident from the text: Omicron work is described as “initial promising studies” and tested only N440K and G446S mutations, not the full Omicron variant; one designed antibody, HX001-041, was not produced.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1073/pnas.2122954119

Cites

60 of 60 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 34 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    Each of six double mutations achieved over an order of magnitude improvement in neutralizing activity over the best single mutation R103M in the first round.

    “Each of those six double mutations achieved over an order of magnitude improvement in neutralizing activity over the best single mutation R103M in the first round, especially for HX001-013 (R103M and L104F) and HX001-015 (R103M and T31W), for which average IC50 values increased to 0.007μg/mL and 0.013μg/mL, respectively (Fig. 2).”
    Quote in the paperSubstantial claim
  • Findingown data

    In first-round neutralization, the original antibody P36-5D2 neutralizes SARS-CoV-2 WT, Alpha, Beta, and Gamma, but is affected by Delta and other viruses carrying L452R.

    “In the first-round neutralization shown in Fig. 2, the original antibody P36-5D2 neutralizes SARS-CoV-2 WT, Alpha, Beta, and Gamma, but is affected by Delta and other viruses carrying L452R, consistent with our structural analysis and predictions.”
    Quote in the paperSubstantial claim
  • Findingown data

    Through iterative modeling and experimental validation, optimized antibodies, including six antibodies, achieved 10- to 600-fold improvement in neutralizing activity against multiple SARS-CoV-2 variants, including Delta.

    “Through an iterative process of modeling and experimental validation, we were able to obtain six optimized antibodies with substantially improved potency of about 10- to 600-fold against multiple variants, including Delta.”
    • Also stated in: § Summary. · plain
      “Optimized antibodies with single mutations or their combinations can achieve 10- to 600-fold improvement in neutralizing activity against SARS-CoV-2 variants, including Delta.”
    • Also stated in: § Discussion · plain
      “Through an iterative optimization procedure, we found that optimized antibodies exhibited broader and much more potent neutralizing activity compared to the original antibody.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: SARS-CoV-2 continues to evolve through emerging variants, more frequently observed with higher transmissibility.
    • Rests on an assumption: Selection pressure on the Spike protein may lead to further evolution of variants that include mutations that can evade immune response.
    • Rests on an assumption: Efficient optimization of antibodies that confer broad and potent neutralizing activity against diverse variants is urgently needed.
  • Findingown data

    Antibodies HX001-020, HX001-024, HX001-033, and HX001-034 with three or four mutations were stronger than HX001-013 with only two mutations.

    “Antibodies HX001-020, HX001-024, HX001-033, and HX001-034 with three or four mutations were also stronger than HX001-013 with only two mutations.”
    Quote in the paperSubstantial claim
  • Findingown data

    R103M in HCDR3 significantly improved neutralizing activity of P36-5D2 against all 10 pseudotyped viruses with average IC50 reaching 0.038 μg/mL.

    “R103M in HCDR3, which was the most surprising mutation, significantly improved neutralizing activity of P36-5D2 against all 10 pseudotyped viruses with average half-maximal inhibitory concentration (IC50) reaching 0.038 μg/mL.”
    Quote in the paperSubstantial claim
  • May combine two claims. 29
  • Classified as probably not a substantial claim. 7

  • May combine two claims.

    Findingown data

    The model makes predictions with moderate to high correlation with experimental binding data and outperforms GeoPPI and other recently proposed methods for predicting single-mutation effects.

    “We found that our model is able to make predictions with moderate to high correlation with experimental binding data and also outperforms GeoPPI (41), the current state-of-the-art method, as well as a few other recently proposed methods for predicting single-mutation effects.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [41] X. Liu, Y. Luo, P. Li, S. Song, J. Peng, Deep geometric representations for modeling effects of mutations on protein-protein binding affinity. PLoS Comput. Biol. 17, e1009284 (2021). 34347784 PMC8366979 source

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    HX001-020 and HX001-024 displayed IC50 values similar to REGN10987 in general and were better than REGN10987 in neutralizing the N439K mutant.

    “HX001-020 and HX001-024 displayed IC50 values similar to the reference antibody REGN10987 in general and were better than REGN10987 in neutralizing the N439K mutant.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    N57 on HCDR2 is not on the Delta RBD interaction interface and has no direct contact with Y449 or R452.

    “As shown in Fig. 3B, N57 on HCDR2 is not on the Delta RBD interaction interface and has no direct contact with Y449 or R452.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    HX001-034 had potent neutralizing activity, with an IC50 of 0.017 μg/mL, yet was slightly weaker than three mutation combinations: HX001-020, HX001-024, and HX001-033.

    “HX001-034 had potent neutralizing activity, with an IC50 of 0.017μg/mL, yet was slightly weaker than three mutation combinations: HX001-020, HX001-024, and HX001-033 (Fig. 2).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Three mutations, N52D and N55L in HCDR2, and G100P in HCDR3, obtained enhanced neutralization against Delta but reduced neutralization against WT.

    “Three mutations, N52D, N55L in HCDR2, and G100P in HCDR3, obtained enhanced neutralization against Delta but reduced neutralization against WT.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The dissociation constant (KD) of the optimized antibodies to the SARS-CoV-2 WT RBD was 1.2 nM to 0.42 nM, which is 20- to 50-fold stronger than the original P36-5D2 antibody.

    “The dissociation constant (KD) of the optimized antibodies to the SARS-CoV-2 WT RBD was 1.2 nM to 0.42 nM, which is 20- to 50-fold stronger than the original P36-5D2 antibody.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The trained geometric neural network model effectively extracts interresidue interaction features and predicts changes in binding affinity due to single or multiple amino acid substitutions to the antigen.

    “Based on a large collection of antibody–antigen complex structures and binding affinity data, we trained a geometric neural network model, recently developed in computer vision, that effectively extracts interresidue interaction features and makes predictions of changes in binding affinity due to single or multiple amino acid substitutions to the antigen.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The off-rate (kd) values of the six optimized antibodies against WT RBD all reached 10−3, compared to around 10−2 for the original antibody, signifying longer half-life binding period and higher binding stability.

    “While the off-rate (kd) value of the original antibody against WT RBD was around 10−2, the off-rate (kd) values of the six optimized antibodies all reached 10−3, signifying a longer half-life binding period and higher binding stability.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Out of 12 single-mutation sites, only 4 (T30, T31, R103, and Q105) were positioned on the paratope of P36-5D2 bound to the RBD, while the remaining 8 were outside and did not directly interact with the RBD.

    “Out of the 12 single-mutation sites, only 4 (T30, T31, R103, and Q105) were positioned on the paratope of P36-5D2 bound to the RBD, while the remaining 8 were outside and did not directly interact with the RBD.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The dissociation constant of optimized antibodies with the SARS-CoV-2 Delta RBD was 6.22 nM to 26.4 nM, which is twofold stronger than or similar to P36-5D2, while the KD with Delta RBD is still slightly weaker than with WT RBD.

    “Notably, the dissociation constant of optimized antibodies with the SARS-CoV-2 Delta RBD was 6.22 nM to 26.4 nM, which is twofold stronger than or similar to P36-5D2, while the KD of optimized antibodies with the SARS-CoV-2 Delta RBD is still slightly weaker than with WT RBD.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    HX001-013 has an IC50 similar to reference antibody REGN10987 and is more potent than REGN10987 in neutralizing the N439K variant.

    “HX001-013 has an IC50 similar to the reference antibody REGN10987 and is more potent than REGN10987 in neutralizing the N439K variant.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    A deep learning approach was introduced to redesign complementarity-determining regions (CDRs) to target multiple virus variants and obtained an antibody that broadly neutralizes SARS-CoV-2 variants.

    “To catch up with the virus’s evolution, we introduced a deep learning approach to redesign the complementarity-determining regions (CDRs) to target multiple virus variants and obtained an antibody that broadly neutralizes SARS-CoV-2 variants.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: SARS-CoV-2 continues to evolve through emerging variants, more frequently observed with higher transmissibility.
    • Rests on an assumption: Selection pressure on the Spike protein may lead to further evolution of variants that include mutations that can evade immune response.
    • Rests on an assumption: Efficient optimization of antibodies that confer broad and potent neutralizing activity against diverse variants is urgently needed.

  • May combine two claims.

    Findingown data

    Four mutations, T31W in HCDR1, A53F and N57L in HCDR2, and L104F in HCDR3, improved overall neutralizing activity, with average IC50 of 0.479, 0.547, 0.025, and 0.104 μg/mL, respectively.

    “Four mutations, T31W in HCDR1, A53F, and N57L in HCDR2, and L104F in HCDR3, also improved the overall neutralizing activity, with average IC50 of 0.479 μg/mL, 0.547 μg/mL, 0.025 μg/mL, and 0.104 μg/mL, respectively.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    Based on a deep learning method, promising CDR mutation candidates can be selected and effectively combined to achieve efficient antibody optimization.

    “Based on a deep learning method, we are able to select promising CDR mutation candidates and effectively combine them to achieve efficient antibody optimization.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Through structure analysis, the G446S mutation was located on the epitope of antibody P36-5D2, and N440K was near (within 5 Å) the epitope.

    “Through structure analysis, we found that the G446S mutation was located on the epitope of antibody P36-5D2, and N440K was near (within 5 Å) the epitope.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Optimized antibodies were onefold to threefold higher than HX-020 in neutralizing against N440K, and 20- to 100-fold stronger in the pseudotyped virus carrying the G446S mutation.

    “It can be seen from Fig. 2C that the optimized antibodies were onefold to threefold higher than HX-020 in neutralizing against N440K, and 20- to 100-fold stronger in the pseudotyped virus carrying the G446S mutation.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Since both R103 and R346 have very long sidechains and carry positive charges, proximity between the two may introduce strong repulsion which may greatly reduce the binding affinity between the antibody and the antigen.

    “Since both R103 and R346 have very long sidechains and carry positive charges, proximity between the two may introduce strong repulsion which may greatly reduce the binding affinity between the antibody and the antigen.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Interpretationown data

    These results highlight the power of deep learning approaches for antibody optimization and their potential application to a wide range of other protein molecules.

    “These results highlight the power of deep learning approaches for antibody optimization and their potential application to a wide range of other protein molecules.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Attention-based networks are extremely powerful for modeling complicated interactions between entities, making them suitable for modeling proteins whose functions are largely determined by interactions between amino acids.

    “Among many successful neural network architectures, attention-based networks are extremely powerful for modeling complicated interactions between entities, making them suitable for modeling proteins whose functions are largely determined by interactions between amino acids (34–37).”
    Quote in the paperSubstantial claim0 of 4 sources checked
    • cites [34] J. Jumper, Highly accurate protein structure prediction with AlphaFold. Nature 596, 583– 589 (2021). 34265844 PMC8371605 source

      Not checked in the free run.

    • cites [35] A. Vaswani, “Attention is all you need” in Proceedings of the 31st International Conference on Neural Information Processing Systems, U. von Luxburg, I. Guyon, Eds. (Curran Associates Inc., Long Beach, CA, 2017), pp 6000– 6010.

      Not checked in the free run.

    • cites [36] R. Rao, MSA Transformer. bioRxiv [Preprint] (2021). 10.1101/2021.02.12.430858 (Accessed 15 February 2022). source

      Not checked in the free run.

    • cites [37] R. Rao, J. Meier, T. Sercu, S. Ovchinnikov, A. Rives, Transformer protein language models are unsupervised structure learners. bioRxiv [Preprint] (2020). 10.1101/2020.12.15.422761 (Accessed 15 February 2022). source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    In vitro affinity maturation methods, such as random mutagenesis with display technologies, improve antibody binding against target proteins but are time consuming and labor intensive.

    “In vitro affinity maturation methods, such as random mutagenesis with display technologies, has been shown to improve antibody binding against target proteins, but such approaches are time consuming and labor intensive (27–32).”
    Quote in the paperSubstantial claim0 of 6 sources checked
    • cites [27] H. Gram, In vitro selection and affinity maturation of antibodies from a naive combinatorial immunoglobulin library. Proc. Natl. Acad. Sci. U.S.A. 89, 3576– 3580 (1992). 1565653 PMC48911 source

      Not checked in the free run.

    • cites [28] M. J. Hammerling, In vitro ribosome synthesis and evolution through ribosome display. Nat. Commun. 11, 1108 (2020). 32111839 PMC7048773 source

      Not checked in the free run.

    • cites [29] J. R. Jackson, G. Sathe, M. Rosenberg, R. Sweet, In vitro antibody maturation. Improvement of a high affinity, neutralizing antibody against IL-1 beta. J. Immunol. 154, 3310– 3319 (1995). 7897213 source

      Not checked in the free run.

    • cites [30] M. Ho, I. Pastan, In vitro antibody affinity maturation targeting germline hotspots. Methods Mol. Biol. 525, 293– 308, (2009). 19252855 PMC3473074 source

      Not checked in the free run.

    • cites [31] B. Li, In vitro affinity maturation of a natural human antibody overcomes a barrier to in vivo affinity maturation. MAbs 6, 437– 445 (2014). 24492299 PMC3984332 source

      Not checked in the free run.

    • cites [32] J. Wang, In vitro affinity maturation of antibody against membrane-bound GPCR molecules. Appl. Microbiol. Biotechnol. 103, 7703– 7717 (2019). 31359103 PMC6719327 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    REGN-10987 was included as a positive control during experimental evaluation.

    “During experimental evaluation, we also included REGN-10987 (44), an ultrapotent antibody approved for emergency use authorization, as the positive control.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [44] J. Hansen, Studies in humanized mice and convalescent humans yield a SARS-CoV-2 antibody cocktail. Science 369, 1010– 1014 (2020). 32540901 PMC7299284 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductcitation

    The model was evaluated by split-by-complex fivefold cross-validation over the SKEMPI V2.0 dataset.

    “This model was evaluated by split-by-complex fivefold cross-validation over the Structural Kinetic and Energetic database of Mutant Protein Interactions (SKEMPI) V2.0 dataset (38).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [38] J. Jankauskaite, B. Jiménez-García, J. Dapkunas, J. Fernández-Recio, I. H. Moal, SKEMPI 2.0: An updated benchmark of changes in protein-protein binding energy, kinetics and thermodynamics upon mutation. Bioinformatics 35, 462– 469 (2019). 30020414 PMC6361233 source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Study design and conductcitation

    A subset consisting of 1,131 single-point mutations (S1131) was used to benchmark the model and other baselines.

    “A subset consisting of 1,131 single-point mutations (S1131) (39) was used to benchmark the model and other baselines.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [39] P. Xiong, C. Zhang, W. Zheng, Y. Zhang, BindProfX: Assessing mutation-induced binding affinity change by protein interface profiles with pseudo-counts. J. Mol. Biol. 429, 426– 434 (2017). 27899282 PMC5963940 source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    Compared to traditional approaches, the deep learning search space is theoretically much larger and is easily applicable to targeting multiple variants simultaneously via multiobjective optimization.

    “Compared to traditional approaches, the deep learning search space is theoretically much larger and is also easily applicable in targeting multiple variants simultaneously via multiobjective optimization.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Targeted optimization toward one particular variant may result in loss of neutralizing activity against other variants.

  • May combine two claims.

    Interpretationown data

    After substituting N57 with L57, a new interaction between L57's sidechain and Y449 is observed, which potentially explains the enhanced binding affinity of this mutation.

    “After substituting N57 with L57, we observed a new interaction between L57’s sidechain and Y449, which potentially explains the enhanced binding affinity of this mutation.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductown data

    An established panel of 10 pseudotyped viruses was used to evaluate neutralizing breadth and potency.

    “To evaluate the neutralizing breadth and potency of the optimized antibodies, we used an established panel of 10 pseudotyped viruses, including SARS-CoV-2 WT 614G, four VOCs—Alpha, Beta, Gamma, Delta—as well as Delta Plus, two VOIs—Kappa, Epsilon which carried L452R, and Eta which carried E484K—and the variant N439K which effected reference antibody REGN10987.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    The L452R mutation is located at the periphery of the receptor binding motif (RBM) and reduces neutralizing activity by antibodies.

    “The L452R mutation is located at the periphery of the receptor binding motif (RBM) and is found to reduce neutralizing activity by antibodies.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [22] B. Hie, E. D. Zhong, B. Berger, B. Bryson, Learning the language of viral evolution and escape. Science 371, 284– 288 (2021). 33446556 source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    P36-5D2 was initially isolated from a convalescent patient and demonstrated reasonably strong potency and breadth against Alpha, Beta, and Gamma but not Delta.

    “To demonstrate the utility of our approach, we sought to optimize a human neutralizing antibody P36-5D2, which was initially isolated from a convalescent patient, and demonstrated reasonably strong potency and breadth against Alpha, Beta, and Gamma (33) but not Delta, due to Delta’s L452R but not T478K mutation through computational structure analysis.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [33] S. Shan, A potent and protective human neutralizing antibody against SARS-CoV-2 variants. Front. Immunol. 12, 766821 (2021). 34966387 PMC8710476 source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Omicron carries 36 total mutations in the spike protein and 15 mutations in the RBD, which greatly changed the immunogenicity of the RBD and caused many neutralizing antibodies to lose their neutralization.

    “Omicron carries 36 total mutations in the spike protein and 15 mutations in the RBD, which greatly changed the immunogenicity of the RBD and caused many neutralizing antibodies to lose their neutralization (45–47).”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [45] L. Liu, Striking antibody evasion manifested by the Omicron Variant of SARS-CoV-2. Nature 602, 676–681 (2022). 35016198

      Not checked in the free run.

    • cites [46] D. Planas, Considerable escape of SARS-CoV-2 Omicron to antibody neutralization. Nature 602, 671–675 (2022). 35016199

      Not checked in the free run.

    • cites [47] E. Cameroni, Broadly neutralizing antibodies overcome SARS-CoV-2 Omicron antigenic shift. Nature 602, 664–670 (2022). PMC9531318 35016195

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    After substituting M103, no direct interaction with R346 on the Delta RBD is observed, which might explain the greatly improved neutralization against the Delta variant.

    “After substituting M103, which has a much smaller sidechain, we no longer observe a direct interaction with R346 on the Delta RBD. This factor might explain the greatly improved neutralization against the Delta variant.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    The SKEMPI V2.0 dataset is the largest antibody-antigen binding affinity dataset, containing 342 complexes and 5,318 mutations.

    “The training and validation datasets were constructed from the SKEMPI V2.0 dataset (38), the largest antibody–antigen binding affinity dataset, which contains 342 complexes and 5,318 mutations in total.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [38] J. Jankauskaite, B. Jiménez-García, J. Dapkunas, J. Fernández-Recio, I. H. Moal, SKEMPI 2.0: An updated benchmark of changes in protein-protein binding energy, kinetics and thermodynamics upon mutation. Bioinformatics 35, 462– 469 (2019). 30020414 PMC6361233 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductown data

    The model was trained and evaluated using split-by-complex fivefold cross-validation over the SKEMPI V2.0 dataset, with five subsets without overlapping complexes.

    “The dataset was split evenly into five subsets without overlapping complexes. Accordingly, five models were trained, using each subset for validation.”
    • Also stated in: § Evaluating Model Performance. · plain
      “The model was evaluated by split-by-complex fivefold cross-validation over the entire SKEMPI V2.0 dataset.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Many SARS-CoV-2 variants, including Delta and Omicron, have adopted different mutation strategies to evade immune recognition, leading to the failure of neutralizing antibodies and vaccines.

    “Many such variants, including Delta and the most recent Omicron, have adopted different mutation strategies to evade immune recognition, leading to the failure of neutralizing antibodies and vaccines (18, 23, 45–48).”
    Quote in the paperSubstantial claim0 of 6 sources checked
    • cites [18] D. Corti, L. A. Purcell, G. Snell, D. Veesler, Tackling COVID-19 with neutralizing monoclonal antibodies. Cell 184, 4593– 4595 (2021). 34416148 PMC8375150 source

      Not checked in the free run.

    • cites [23] D. Planas, Reduced sensitivity of SARS-CoV-2 variant Delta to antibody neutralization. Nature 596, 276– 280 (2021). 34237773 source

      Not checked in the free run.

    • cites [45] L. Liu, Striking antibody evasion manifested by the Omicron Variant of SARS-CoV-2. Nature 602, 676–681 (2022). 35016198

      Not checked in the free run.

    • cites [46] D. Planas, Considerable escape of SARS-CoV-2 Omicron to antibody neutralization. Nature 602, 671–675 (2022). 35016199

      Not checked in the free run.

    • cites [47] E. Cameroni, Broadly neutralizing antibodies overcome SARS-CoV-2 Omicron antigenic shift. Nature 602, 664–670 (2022). PMC9531318 35016195

      Not checked in the free run.

    • cites [48] S. Cele; NGS-SA; COMMIT-KZN Team, Omicron extensively but incompletely escapes Pfizer BNT162b2 neutralization. Nature, 10.1038/s41586-021-04387-1 (2021). PMC8866126 35016196 source

      Not checked in the free run.

33 of 33

  • The paper relies on this without support.

    Targeted optimization toward one particular variant may result in loss of neutralizing activity against other variants.

    “Targeted optimization toward one particular variant may also result in loss of neutralizing activity against other variants.”

  • The paper relies on this without support.

    SARS-CoV-2 continues to evolve through emerging variants, more frequently observed with higher transmissibility.

    “SARS-CoV-2 continues to evolve through emerging variants, more frequently observed with higher transmissibility.”

  • The paper relies on this without support.

    Selection pressure on the Spike protein may lead to further evolution of variants that include mutations that can evade immune response.

    “Despite the wide application of vaccines and antibodies, the selection pressure on the Spike protein may lead to further evolution of variants that include mutations that can evade immune response.”

  • The paper relies on this without support.

    Efficient optimization of antibodies that confer broad and potent neutralizing activity against diverse variants is urgently needed.

    “Efficient optimization of antibodies that confer broad and potent neutralizing activity against diverse variants is therefore urgently needed.”

  • The paper relies on this without support.

    Directed maturation toward a new variant may cause the antibody to lose neutralization against previous variants.

    “Directed maturation toward a new variant may also cause the antibody to lose neutralization against previous variants.”

  • The paper relies on this without support.

    More diverse SARS-CoV-2 variants with greater immune pressure will continue to emerge in the infected population.

    “As the virus continues to evolve, more and more diverse variants will emerge in the infected population with greater immune pressure.”
6 of 6

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Shan et al. developed SARS-CoV-2 antibodies by redesigning CDR regions utilizing a model involving learning graph embeddings of protein structures to predict the change in ddG upon introducing specific mutations in the loops.

    “Shan et al. [55] developed severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) antibodies by redesigning CDR regions utilizing a model involving learning graph embeddings of protein structures to predict the change in ddG upon introducing specific mutations in the loops.”

    From Computational and artificial intelligence-based methods for antibody development (2023) · cites this paper as [55]

  • There exist approaches that address subsets of the components of the problem of joint generation of antibody sequence and structure conditioned on the target epitope.

    “There has been progress towards this goal and there exist approaches that address subsets of the components of this problem [53][54][55][56][63][64][65][68][69][70].”

    From Computational and artificial intelligence-based methods for antibody development (2023) · cites this paper as [55]

Self-check

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

Each contradiction between two passages listed here was confirmed by an independent check.

  • Third-round mutation list changes N55L to A55L

    Minor

    The third-round list names A55L, although the single mutation selected and used in the preceding rounds is N55L.

  • Some optimized antibodies bind Delta RBD more weakly than the original

    Minor

    The text describes the optimized antibodies' Delta RBD binding as stronger than or similar to the original's, but HX001-015 has a higher KD (26.4 versus 11.4 nM), indicating weaker binding.

  • Reported Delta binding-affinity range excludes HX001-034

    Minor

    The stated minimum Delta RBD KD is 6.22 nM, but Table 1 reports 5.16 nM for HX001-034.

  • Accounts disagree on whether T478K contributes to reduced Delta activity

    Minor

    The introduction attributes the reduced activity to L452R but explicitly not T478K; the Antibody of Choice section says the reduction was likely attributable to both mutations.

  • Methods and results give different mutations at position 53

    Minor

    The first-round results identify A53F, whereas the Methods list of initial mutations identifies N53F.

  • Methods and results give different mutations at position 32

    Minor

    The first-round results identify Y32P, whereas the Methods list of initial mutations identifies Y32O.

6 of 6

Cites

Flagged first, then by citations

Research works printed but not linked: 1

  1. A. Vaswani “Attention is all you need” in Proceedings of the 31st International Conference on Neural Information Processing Systems U. von Luxburg I. Guyon Eds. (Curran Associates Inc. Long Beach CA 2017) pp 6000–6010.

Cited by

Sorted by citations

Related

Authors
Shan, S., Luo, S., Yang, Z., Hong, J., Su, Y., Ding, F., Fu, L., Li, C., Chen, P., Ma, J., Shi, X., Zhang, Q., Berger, B., Zhang, L., & Peng, J.
Affiliations
  • Tsinghua University (CN)
  • HeliXon Limited, Beijing 100084, China
  • University of Illinois Urbana-Champaign (US)
  • MIT Computer Science and Artificial Intelligence Laboratory (US)
Corresponding author
  • Berger, Bonnie
  • Zhang, Linqi
  • Peng, Jian
ORCID
Venue
Proceedings of the National Academy of Sciences
Publisher
National Academy of Sciences
Published
2022-03
Volume
119
Issue
11
Pages
e2122954119-e2122954119
ISSN
  • 0027-8424 (ISSN-L)
  • 1091-6490
DOI
10.1073/pnas.2122954119
PMC
PMC8931377
OpenAlex
W4214886481
Cited by
157
References
60
Citations by year
2026: 26 · 2025: 43 · 2024: 40 · 2023: 34 · 2022: 13 · 2012: 1
FWCI
15.86
Type
journal article
Language
en
Open access
Licence
Locations
  • Proceedings of the National Academy of Sciences (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DSpace@MIT (Massachusetts Institute of Technology) (submittedVersion, cc-by-nc)
Funders
  • Tsinghua University (20201080053)
  • Beijing Municipal Science and Technology Commission (Z201100005420019)
  • Beijing Advanced Innovation Center for Structural Biology, Tsinghua University
Topics
  • Health Sciences › Medicine › Infectious Diseases › SARS-CoV-2 and COVID-19 Research
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › vaccines and immunoinformatics approaches
  • Health Sciences › Medicine › Radiology, Nuclear Medicine and Imaging › Monoclonal and Polyclonal Antibodies Research
Keywords
SARS-CoV-2, human monoclonal antibodies, antibody design, deep learning, hypervariable region, viral variants, spike protein, immune evasion, vaccine escape
MeSH terms
Protein Domains, Deep Learning, Broadly Neutralizing Antibodies, COVID-19, SARS-CoV-2, COVID-19 Vaccines, Antibodies, Viral, Epitopes, Humans, Immunotherapy, Neutralization Tests, Complementarity Determining Regions, Antibodies, Neutralizing, Spike Glycoprotein, Coronavirus

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.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-10.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-10.

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