1. Evaluation Summary:

    CR6261 and CR9114 are two antibodies that bind to the conserved stem of influenza hemagglutinin (HA) through their VH regions and differ by 14-18 mutations from their inferred germline sequences. The authors constructed large combinatorial libraries containing all combinations of 11 and 16 binding-surface mutations for CR6261 and CR9114. These were used in yeast surface display titrations to infer individual and epistatic contributions to binding diverse HAs and to infer possible evolutionary trajectories going from germline to the mature antibodies. The study provides a wealth of knowledge on amino acid contributions to binding affinity. The study informs our understanding of biochemical epistasis, and could potentially serve as a starting point for a more detailed understanding of antibody affinity maturation more generally.

    (This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #1 and Reviewer #3 agreed to share their names with the authors.)

    Read the original source
    Was this evaluation helpful?
  2. Reviewer #1 (Public Review):

    The experiments are exceptionally well done and the analysis is detailed and comprehensive. The figures are mostly good although a few are a bit complex, and I found the force-directed graphs to be not very helpful in the static figures, although extremely useful in the interactive browser (maybe the interactive version could be called out more clearly in the figure legends).

    Here are some suggestions for revisions that might further improve the manuscript:

    • In both the abstract and beginning of results, it would be helpful to describe the libraries a bit more clearly: all combinations of mutations separating the germline and mature antibody in the VH domain among sites contacting the epitope (at least that's my understanding).

    • It is also never clearly stated in results whether it is a single chain (scFv) antibody, and if the HA is trimeric. If so, is there a potential for avidity so that the measurements are Kd, apparent for the multivalent interaction rather than monomeric Kd?

    • The calculating of the probabilities of different paths is cool! Is anything known about the actual maturation of these antibodies? I know there are longitudinal data on FI6v3 from the patient from whom the antibody was isolated, but wasn't sure for these two antibodies.

    • I personally was very curious how the specific epistasis models the authors used compare to global epistasis models. This is well explained in the appendix, but might be helpful to mention in another sentence or two in the main text as well.

    • The figures are really informative but are also numerous and dense. I really appreciated panels like Figure 5A,B, which are easy to interpret and pretty simple!

    • In the non-global-epistasis model, is anything done to handle the censoring of the data at the high and low end of the affinity scale? I was confused by this because line 115 says values outside the range are pinned to the boundaries, but then in Appendix (line 1282) it seems to suggest censoring isn't an issue.

    Read the original source
    Was this evaluation helpful?
  3. Reviewer #2 (Public Review):

    The binding landscapes of two broadly neutralizing antibodies (bnAbs), CR9114 and CR62621, which bind to the conserved stem region of influenza hemagglutinin through their VH regions and display varying levels of breadth, were studied. CR9114 and CR6261 contain 14 and 18 mutations in the VH chain between somatic sequences and constructed germline sequences. To decipher the contribution of individual mutations as well as pairwise and higher-order interactions between mutations, the authors made combinatorially complete mutational libraries for both the antibodies. Libraries excluded mutations distant from antigen contacts in the crystal structure (CR9114: 2, CR6261:3). Single-chain variable fragment libraries were constructed and displayed on the yeast surface. Their equilibrium binding affinities were determined using the Tite-seq method against selected antigens (CR9114: H1, H3, and Influenza B HA; CR6261: H1 and H9 HA). Both CR9114 and CR6261 libraries showed expectedly different patterns of breadth. Breadth conferring mutations in CR9114 exhibit nested structure, giving rise to the hierarchical structure of the binding landscape. Examination of the extensive pairwise and higher-order epistasis between mutations revealed key sites with strong synergistic interactions that are highly similar across antigens for CR-6261 and different for CR-9114. It was suggested that features of the binding affinity landscapes strongly favor the sequential acquisition of affinity to diverse antigens for CR-9114 and are highly constrained, while the acquisition of breadth to more similar antigens for CR-6261 is less constrained. The study indicates that the evolutionary pathways to bnAbs are highly dependent on epistatic and pleiotropic effects of mutations. The experimental and computational procedures used to generate and analyze the data are clearly described and there is wealth of binding affinity data that will be of interest to those studying antibody:antigen and more generally protein:protein interactions.

    Major points:

    1. Unlike other studies where antibodies were isolated from single cell sorting of memory B-cells, the present bNAbs were isolated from phage display libraries. These libraries (Throsby et al, 2008; Dreyfus et al, 2012) were constructed from pooled IgM+ from 10 (not 3 as incorrectly stated in manuscript) (CR6261) or 3 (CR9114) healthy donors and scFv fragments were cloned and screened using phage display against H5HA for CR6261 (Throsby et al, 2008) and against sequential panning against Has from H1, H3 and both B lineages for CR9114 (Dreyfus et al, 2012) respectively. Use of phage display methodology which involves multiple rounds of PCR as well as panning, means that mutations can be introduced during library construction. Hence the resulting sequences isolated need not accurately reflect antibody gene sequences present in the donors. Further the multi-round panning process with diverse HAs (especially for CR9114) biases the resulting sequences selected, altering the order of the panning steps might result in a different selected sequence. Additionally it is unclear what the natural selection pressures are. It is likely that they would be for increased neutralization breadth and potency which is not straightforwardly related to improved breadth of binding to soluble HA where the stem accessibility is much higher than it is in the viral context. The known polyreactivity of many bNAbs including stem directed bNAbs (PMID: 33049994) is another confounding factor.

    2. CR6261 was selected by panning against H5 HA whereas the vaccinated individuals presumably primarily had experienced H1. Since the antibody binds to the conserved stem of HA what is relevant to interpret the mutational landscapes is not the overall antigenic diversity of HA (Figure 1B) but the diversity of the stem epitopic region. If these regions are similar in H1 and H9 HA that would explain the observation that multiple mutations in the antibody are tolerated in both cases. In the case of CR9114, the mature antibody binds best to H1, less well to H3 and weakest to B HA. Unsurprisingly the same pattern is reflected in the mutational data. Given the greater diversity in HA stem epitope sequence between H1, H3 and B relative to that between H1 and H9 it is also unsurprising that there should be sequential acquisition of breadth in the latter CR9114 case (see also 1 above).

    Minor points:

    1. It would be useful to have list the amount of surface area each residue in the paratope for both antibodies contributes to binding where such information is available.

    2. Is it possible to convert the effect scores into a free energy of binding contribution? Are there error estimates for the effect scores that would allow one to assess whether apparent differences in effect scores are statistically significant? It is surprising to see significant second order effects at relatively large distances (Figure 2F). What is the suggested explanation?

    3. Larger epistatic effect scores occur between residues with the largest contributions to binding. Is this expected?

    Read the original source
    Was this evaluation helpful?
  4. Reviewer #3 (Public Review):

    Phillips et al. aimed to characterize the binding affinity landscapes of two influenza bnAbs, CR9114 and CR6261. Mutant libraries of CR9114 and CR6261 that contained all possible evolutionary intermediates back to their corresponding unmutated germline sequences were constructed. Tite-seq was then used to measure the binding affinity of all variants in the mutant library of CR9114 against HAs from H1 (group 1), H3 (group 2) and influenza B, and the mutant library of CR6261 against HAs from H1 and H9 (both are group 1). The binding landscape of CR9114 showed a hierarchical structure, where most variants bind H1, but only a subset of high affinity variants to H1 bound H3 and a subset of high affinity variants to H3 bound influenza B. In contrast, the binding affinities of CR6261 variants to H1 and H9 correlated well. Mutations with large first order effect are at the HA-binding interface, whereas mutations with strong pairwise interactions tend to be close in space. Additionally, a subset of mutations that dominated the higher-order interactions largely determined the overall structure of the binding affinity landscape. At the end, the authors demonstrated that sequential antigen selection is important for the affinity maturation of CR9114 to cross-react with H3 and influenza B HAs, but not so much for that of CR6261 to the two group 1 HAs (H1 and H9).

    Overall, this data in this study is of exceptionally high quality. The authors' claims are mostly well supported. But some additional explanations need to be made.

    Read the original source
    Was this evaluation helpful?