To defend itself against bacteria and viruses, the body depends on a group of proteins known as antibodies. Each subset of antibodies undergoes a rigorous training regimen to ensure it recognizes a single epitope well – that is, one specific region on the surface of foreign, harmful organisms.

Most antibodies stick extremely tightly to their one unique epitope, but some can also weakly bind to molecules that are vastly different from their main trained targets. This feature – known as polyreactivity – can in some cases help the immune system fight against multiple strains of viruses. On the other hand, when antibodies are designed in the laboratory to treat diseases, this characteristic can sometimes lead to the failure of pre-clinical trials. Yet it is currently unclear why some antibodies are polyreactive when others are not.

To investigate this question, Boughter et al. compared over 1,000 polyreactive and non-polyreactive antibody sequences from a large database, revealing differences in the physical properties of the region of the antibodies that attaches to epitopes. Using these defining features, Boughter et al. went on to design a new piece of freely available, automated software that could predict which antibodies would be polyreactive more than 75% of the time.

Such software could ultimately help to guide the design of antibody-based treatments, while bypassing the need for costly laboratory tests.

Antibodies are immunogenic proteins expressed by B cells that play a major role in the adaptive immune response against non-self. Upon recognition of target epitopes, these antibodies undergo multiple rounds of somatic hypermutation and affinity maturation inside a germinal center, whereby the amino acid sequence of the epitope-binding surface is selected for optimal binding to the target (Victora and Nussenzweig, 2012; Eisen and Siskind, 1964; McKean et al., 1984). The longer this affinity maturation process extends, the higher the affinity and specificity of the antibodies toward their target antigen, primarily through mutagenesis of the six complementarity determining region (CDR) loops of the antibody (Victora and Nussenzweig, 2012). Using a combination of affinity matured CDR loops, these antibodies bind strongly to the target and aid in invader neutralization. While the process of affinity maturation and somatic hypermutation of antibodies results in high-affinity and incredibly specific binders to a particular epitope, some antibodies have been shown to display signs of reactivity toward diverse off-target epitopes. This broad but low-affinity binding has been termed ‘polyreactivity’.

Antibody polyreactivity has been hypothesized to be beneficial in the early stages of antibody maturation, acting as a pool of diverse binders ready to recognize novel antigens and initiate the more stringent selection process (Dimitrov et al., 2013). To this end, a majority of B cell receptors and antibodies which have not undergone somatic hypermutation, including those on immature B cells and early ‘natural’ antibodies, have been found to be polyreactive to some extent and are suggested to have an innate-like response to pathogens (Ochsenbein et al., 1999; Wardemann et al., 2003). While these mostly unmutated polyreactive antibodies remain at low frequency in antigen-experienced individuals, a distinct population of polyreactive antibodies that have undergone selection are still expressed by mature B cells that circulate in blood (Tiller et al., 2007). In fact, some studies have found that the polyreactivity status of an antibody is mostly independent of the number of somatic hypermutations in the antibody sequence (Mouquet et al., 2010; Prigent et al., 2016). In line with this finding, only 5–10% of the repertoire of naive B cells circulating in the periphery are polyreactive, but this increases to 20–30% in the memory B cell compartment, showing a distinct capability of polyreactivity to survive selection (Tiller et al., 2007; Koelsch et al., 2007). These results suggest that polyreactivity can persist, or perhaps even be selected for during the selection process within the germinal center.

In a few notable cases, polyreactivity may in fact augment the efficacy of a given immune response. Polyreactive IgA antibodies have been shown to have an inherent reactivity to microbiota in the mouse gut, with a predicted role in host homeostasis (Bunker et al., 2017). These previously identified antibodies so far have no known primary ligands, yet play a key role in facilitating the gut immune response to the plethora of exogenous antigens encountered in the dynamic dietary and microbial environment of the gut. This implies the existence of antibodies whose primary function is to act as polyreactive sentries in the gut, yet the downstream effects of polyreactive antibodies coating commensal bacteria is so far unclear. Similar polyreactive IgA and IgG mucosal antibodies were found in the gut of human immunodeficiency virus (HIV)-infected patients, but these antibodies either had low affinity to the virus or lacked neutralization capabilities (Planchais et al., 2019). The benefit of singular antibody sequences with the ability to sample large portions of the commensal population may represent an improvement in efficiency of the homeostatic machinery of the gut.

While the precise role of these primarily polyreactive gut antibodies is still a topic of debate, polyreactivity has been suggested to augment the immune response in other immunological niches. Broadly neutralizing antibodies (bnAbs), which bind robustly to conserved epitopes on the surface glycoproteins of influenza viruses or HIV are more likely to be polyreactive (Haynes et al., 2005; Mouquet et al., 2011; Andrews et al., 2015). In one study of HIV-binding antibodies, over half of all tested bnAbs were found to be polyreactive (Prigent et al., 2018). These bnAbs have been the subject of intense study for their potential as the central components of an HIV treatment or as the byproduct of an immune response to a universal Influenza vaccine (Andrews et al., 2015; Haynes et al., 2019; Crowell et al., 2019; Li et al., 2012). One hypothesized mechanism for the capability of polyreactive antibodies to confer this broad neutralization in the face of a changing viral epitope is heteroligation, the ability of a single antibody to bind the primary target with one binding domain and use the other binding domain to bind in a polyreactive manner (Mouquet et al., 2010). This heteroligation allows the antibody to take advantage of the significant avidity increase afforded by bivalent binding, despite the low envelope protein density of HIV or a geometry which does not readily lend itself to bivalent binding on the surface of influenza viruses (Klein and Bjorkman, 2010).

Although polyreactivity may play a positive role in natural immune responses, oftentimes this same property is considered undesirable from the point of view of generating therapeutic antibodies with high specificity. Antibody-based treatments, which generally take the form of an intravenous transfusion, are sensitive to the accelerated systemic clearance of polyreactive antibodies (Hötzel et al., 2012; Kelly et al., 2015; Kelly et al., 2018; Datta-Mannan et al., 2015). In general, much work has focused on attempting to answer the question of optimizing ‘developability’ of a given antibody. These efforts have been dedicated to determining the most critical components of developability through a large array of experimental assays, in silico structural prediction-based methods, sequence-based analysis and their correlations with clearance, sequence-based SASA predictions, and sequence-based aggregation propensity predictors (Jain et al., 2017b; Raybould et al., 2019; Sharma et al., 2014; Jain et al., 2017a; Obrezanova et al., 2015). In many of these studies, polyreactivity or non-specificity in general was seen to be a negative indicator of the developability of a drug, suggesting that therapeutic antibodies should strive toward a drug-like specificity (Starr and Tessier, 2019).

In line with this goal of understanding the predominant factors involved in the specificity of therapeutic antibodies, many researchers have worked to identify the biophysical underpinnings of polyreactivity in natural immune responses. The most popular hypotheses for the primary biophysical predictors of polyreactivity have included CDR3 length (Prigent et al., 2016), CDR3 flexibility (Prigent et al., 2018), net hydrophobicity (Lecerf et al., 2019), and net charge (Rabia et al., 2018). More observational studies have found an increased prevalence of arginine and tyrosine in polyreactive antibodies (Kelly et al., 2018; Birtalan et al., 2008). While these previous studies represent substantial advances in the study of polyreactivity, they have often been limited in scope, focusing on a singular antibody source and primarily focused on CDR3H. Comparing across these individual antibody sources highlights discrepancies between the proposed predictors of polyreactivity. The aforementioned properties determined to be key to polyreactivity in previous studies were found to be statistically insignificant in studies of HIV-binding and mouse gut polyreactive antibodies (Bunker et al., 2017; Mouquet et al., 2010).

Clearly, a computational framework that would enable us to predict the polyreactivity of a given antibody a priori, whether evaluating the efficacy of a natural immune response or the potential fate of a therapeutic antibody, would be tremendously useful. Such a framework, for example, could be used to assist in the isolation of broadly neutralizing antiviral antibodies, or speed up the process of therapeutic antibody screening. To achieve this goal, a thorough understanding of the molecular features behind polyreactive binding interactions is critical. Experimental approaches utilizing next-generation sequencing and ELISA allow for the identification of hundreds of polyreactive antibody sequences. However, the systematic characterization of these antibodies is difficult. Issues immediately arise when defining the conditions by which we determine an antibody to be polyreactive. While polyreactivity may exist on some continuous spectrum, we are inclined to frame the problem as binary. This binary discretization is useful for the identification of meaningful differences, yet must be recognized as an imperfect assumption. In addition to this more philosophical challenge, experimental efforts must also overcome significant hurdles. Detailed biochemical studies of polyreactive antibodies via protein crystallography, quantitative binding experiments, and mutagenesis provide exceptional insight but are inherently low throughput. Structural modeling of these polyreactive antibodies represent a high-throughput approach, but models of flexible loops are relatively unreliable, and are unlikely to capture nuances in side-chain placement (Karami et al., 2019). A bioinformatics-based approach, centered around high-throughput analysis that minimizes structural assumptions while maintaining positional context of amino acid sequences would provide a thorough, unbiased analysis of existing data and create a powerful pipeline for future studies.

In this study, we show that using just the amino acid sequences of antibodies from a database of over 1000 sequences tested for polyreactivity, unifying biophysical properties that distinguish polyreactive antibodies from non-polyreactive antibodies can be identified. We find that, while charge and hydrophobicity are in fact important determinants of polyreactivity, the characteristic feature of polyreactive antibodies appears to be a shift toward neutrality of the binding interface. In addition, loop crosstalk is more prevalent in the heavy chain of polyreactive antibodies than non-polyreactive antibodies. From these properties, a machine learning-based classification software was developed with the capability to determine the polyreactivity status of a given sequence. This software is generalizable and can be retrained on any binary classification problem and identify the key differences between two distinct populations of antibodies, T cell receptors, or MHC-like molecules at the amino acid level.

Previous research has highlighted the importance of hydrophobicity, charge, and CDR loop flexibility on antibody specificity. In this work, we expand upon these previous results with a new bioinformatic and biophysical characterization of polyreactive antibodies. The software generated for this study provides a powerful computational tool which can be utilized by researchers interested in discerning differences between populations of adaptive immune molecules in broad contexts. Building off of the efforts of our own work and that of experimental collaborators, we were able to aggregate to date one of the largest publicly available datasets of antibodies tested for polyreactivity. Differences in the germline gene frequency and amino acid frequencies show there exists some underlying differences between polyreactive and non-polyreactive antibodies. A surface level analysis of this dataset is able to discriminate certain features of polyreactive and non-polyreactive antibodies, namely that on average, polyreactive antibodies are less strongly negatively charged, less hydrophilic, and have a higher prevalence of antibodies with longer CDR loops of the heavy chain. Importantly, however, these binding surfaces do not have a net positive charge nor are they net hydrophobic.

Our results highlight an increase in $V_{H}1\text{-}69$ gene usage in polyreactive antibodies, an interesting finding given the substantial literature outlining its importance in diverse immune environments. In addition to the aforementioned role of $V_{H}1\text{-}69$ in broadly neutralizing anti-influenza and anti-HIV antibodies (Haynes et al., 2005; Mouquet et al., 2011; Andrews et al., 2015; Prigent et al., 2018), autoreactive chronic lymphocytic leukemic B cells commonly express receptors bearing $V_{H}1\text{-}69$ (Sasso et al., 1993; Forconi et al., 2010), and anti-HIV antibodies which target the membrane-proximal external region of HIV-1 envelope glycoproteins frequently utilize $V_{H}1\text{-}69$ (IAVI Protocol G Investigators et al., 2019). While previous reports suggest that the key feature permitting these auto-reactive or polyreactive interactions of $V_{H}1\text{-}69$ is an exceptionally hydrophobic CDR2H loop (Chen et al., 2019) our results suggest this does not explain the over-representation of this antibody in the polyreactive dataset, as on average the CDR2H of polyreactive antibodies is strongly hydrophilic. Instead, certain structural or dynamic features of the antibody may contribute to its out-sized role in critical biological contexts.

To dig deeper into the biophysical differences between polyreactive and non-polyreactive antibodies, we created an adaptable software for the automated analysis of large antibody datasets and the application of a new analysis pipeline for the study of polyreactive antibodies. Overall, the improvements of this software to the current state of antibody sequence analysis are sufficient to highlight key differences in the two populations with improved spatial resolution. The position sensitive sequence alignment is able to further parse through the genetic differences and show that in general, polyreactive antibodies have a tendency to have more hydrophobic residues in CDR2H, and a decreased preference for phenylalanine in CDR1H. While these observational differences provided some initial insight, a more rigorous biophysical treatment was necessary. With the addition of 62 biophysical properties analyzed using the position sensitive alignment, significant differences between the CDR3H loops in polyreactive and non-polyreactive antibodies become immediately evident, providing a more detailed depiction of the antigen binding surface of polyreactive antibodies.

These data suggest a movement toward neutrality or ‘inoffensive’ residues in the CDR loops of polyreactive antibodies: amino acids that are neither exceptionally hydrophobic nor hydrophilic and with a net charge close to 0. Previous studies have suggested that polyreactive antibodies tend to have more hydrophobic CDR loops, such that low-affinity Van der Waals interactions might be the primary means of polyreactive interactions (Prigent et al., 2018; Starr and Tessier, 2019). However, these studies counted the number of hydrophobic residues per sequence or averaged the hydrophobicity of all six CDR loops. While our results partially agree with these previous findings, our analysis extends much further into defining the biophysical basis of this phenomenon. For example, while our position sensitive representation of the sequences shows that CDR3H does become more hydrophobic in polyreactive sequences, it is still net hydrophilic on average. A highly hydrophobic binding surface would provide an avenue for non-specific interactions with other hydrophobic proteins, but it would occlude binding to highly hydrophilic ligands like DNA. A slightly hydrophilic, neutral-charged binding surface would permit weak interactions with a wide range of ligands.

Using these and other biophysical properties as input feature vectors, we were able to generate a generalizable protocol for binary comparisons between two distinct populations of Ig-domain sequences. This framework is able to successfully split all tested polyreactive and non-polyreactive antibody datasets. Care was taken to not overfit these data and a preliminary classifier built from this algorithm was able to identify the proper number of input vectors for each LDA application. While there are general features which best split the polyreactive and non-polyreactive antibodies in these datasets, including charge, hydrophobicity, and α-helix propensity, these features alone are not sufficient to discriminate between the two populations. Instead, 75 vectors taken from the position-sensitive biophysical property matrix are necessary to properly split the groups, including both simple properties like charge, hydrophobicity, flexibility, and bulkiness and more carefully curated properties like the often used Kidera factors and the hotspot detecting variables of Liu et al. (Kidera et al., 1985; Liu et al., 2018; Vihinen et al., 1994). The inability to arrive at a core few biophysical properties that could effectively distinguish polyreactive and non-polyreactive antibodies necessitated the application of further approaches, namely information theory.

The tools provided by information theory proved to be effective in the present study. The classic approach to information theory considers some input, communication of this input across a noisy channel, and then reception of a meaningful message from the resultant output. We can think of the analogous case for these antibodies, whereby the sequence and structure of the antibodies can be seen as our input, the thermal noise inherent to biological systems can complicate biochemical interactions, and the necessary output is antigen recognition, i.e. binding between the antibody and the ligand. Focusing just on the antibody side of this communication channel, we determined the underlying loop diversity through the Shannon entropy of the polyreactive and non-polyreactive datasets. This diversity was found to be nearly equivalent while the mutual information, a metric of ‘crosstalk’ across populations, between and within CDR loops was found to be increased in the heavy chain and decreased in the light chain of polyreactive antibodies. What this loop crosstalk entails physically is not immediately clear from these measurements.

The mutual information increase could come from gene usage being somehow coupled, amino acid usage coupling with the cognate ligand, or the amino acids directly interacting physically with each other. In some way, this crosstalk appears to be selected for in the polyreactive population. If this increase in mutual information manifests as an increase of charge-charge interactions, this could explain why there is a minimal change in net charge of antibodies between the two groups, yet a significant move toward neutrality in the CDR loops of polyreactive antibodies. The pairing of two charged groups would help move the binding surface of polyreactive antibodies toward a more ‘inoffensive’ binding surface. A binding surface that is neither exceptionally hydrophobic nor hydrophilic, and lacks a significant positive or negative charge, would represent a relatively appealing binding interface for a low-affinity interaction with a large array of diverse ligands. A patchwork of hydrophobic and hydrophilic non-charged residues exposed to potential ligands would provide an ideal candidate polyreactive surface. The corresponding decrease in the mutual information between the light chain CDR loops of polyreactive antibodies could be caused by a de-emphasis in the involvement of these loops due to differential binding configurations of polyreactive ligands, as has been previously hypothesized (Dimitrov et al., 2013; Sethi et al., 2006).

While further crystallographic, biochemical, and dynamic studies are necessary to identify the true source of this increase in mutual information across polyreactive antibodies, we can speculate what these results may mean in the context of the results obtained using linear discriminant analysis. In addition to standard side chain properties, many of the most important features for splitting polyreactive and non-polyreactive antibodies were structural in nature. Specifically, hotspot variables 6, 24, 25, and 41 all correspond to the structural tendencies of a given amino acid. Coupled with the increase in side chain interactions that may be implied by the increased mutual information across the loops of polyreactive antibodies, this potential for increased loop structure may suggest more rigid CDR loops in polyreactive antibodies.

In addition to the insights into polyreactivity, the computational tools developed for this study are broadly applicable to future studies of large antibody or T cell receptor repertoires. One of the strengths of this approach is a decreased emphasis on structural information when crystal structures are unavailable. Computational prediction of loop conformation is difficult, and drawing inferences from incorrect models regarding side-chain interactions and positioning could be misleading. Reliable structural information on these polyreactive antibodies will be critical to a further understanding of the mechanisms of polyreactivity, including complex structures of antibodies bound to various ligands. In the high-throughput analysis of antibody sequences, our approach strikes a careful balance of the structural assumptions that should apply consistently across antibody populations.

Further experimental assays will be necessary to more comprehensively identify the underlying mechanisms of polyreactivity, including further sequencing and biochemical analysis of polyreactive and non-polyreactive antibodies. Antibodies specific to other pathogens or those from other organisms tested for polyreactivity will help form a more complete picture and improve the generality of the results. As with any machine learning based approach, the classification algorithm is only as good as the data it is trained on. Adding further data in the training set, including more mutations and germline reversions that turn a polyreactive antibody non-polyreactive or vice-versa, will be critical for a comprehensive analysis of polyreactivity. Additionally, a more robust assay for determining polyreactivity such as a chip based screen to test for binding to many diverse targets, would greatly broaden our perspective and help understand just how broadly reactive these polyreactive antibodies are. Lastly, a more complete understanding of the germinal center and the selection processes inherent to the affinity maturation process will assist in the determination of whether polyreactivity is a byproduct or a purposeful feature of the affinity maturation process.

The software generated for this study is publicly available as a python application (see Materials and methods). The unique aspect of this software is its hybrid approach to position-sensitive amino acid sequence analysis. Structural information is implicitly encoded by the alignment strategy employed, yet these assumptions are weaker than those imposed by explicit structural prediction. Downstream analysis from this positional encoder is streamlined and can be generalized to analyze any binary or higher order classification problems. This streamlined analysis allows for the generation of each figure in this study to be applied to thousands of sequences in a matter of minutes. The classification capabilities of the software could prove particularly useful when comparing binary classes, such as T cell receptors or antibody sequences derived from healthy and diseased tissue samples. Acceptable inputs are not restricted to CDR loops of immunoglobulins, and we have shown that the software can be adapted for the analysis of MHC-like molecules. Moving forward, this MHC analysis has the potential to classify the antigen binding properties of highly-divergent MHC sequences from a broad range of species, providing insights where phylogenetic approaches prove difficult. This software represents a strong addition to the existing toolkit for repertoire analysis of diverse molecular species.

All data generated and all code used for analysis in this study has been published on GitHub at http://github.com/ctboughter/AIMS (copy archived at https://archive.softwareheritage.org/swh:1:rev:f6c855ef4a7ce63f72dba6b34e9d0e9edd9200ce/).

1. 1. Adams EJ

   (2014) Lipid presentation by human CD1 molecules and the diverse T cell populations that respond to them

   Current Opinion in Immunology 26:1–6.

   https://doi.org/10.1016/j.coi.2013.09.005

   • PubMed
   • Google Scholar
2. 1. Adams EJ
   2. Parham P

   (2001) Species-specific evolution of MHC class I genes in the higher primates

   Immunological Reviews 183:41–64.

   https://doi.org/10.1034/j.1600-065x.2001.1830104.x

   • PubMed
   • Google Scholar
3. 1. Almeida T
   2. Esteves PJ
   3. Flajnik MF
   4. Ohta Y
   5. Veríssimo A

   (2020) An ancient, MHC-Linked, nonclassical class I lineage in cartilaginous fish

   The Journal of Immunology 204:892–902.

   https://doi.org/10.4049/jimmunol.1901025

   • PubMed
   • Google Scholar
4. 1. Andrews SF
   2. Huang Y
   3. Kaur K
   4. Popova LI
   5. Ho IY
   6. Pauli NT
   7. Henry Dunand CJ
   8. Taylor WM
   9. Lim S
   10. Huang M
   11. Qu X
   12. Lee JH
   13. Salgado-Ferrer M
   14. Krammer F
   15. Palese P
   16. Wrammert J
   17. Ahmed R
   18. Wilson PC

   (2015) Immune history profoundly affects broadly protective B cell responses to influenza

   Science Translational Medicine 7:316ra192.

   https://doi.org/10.1126/scitranslmed.aad0522

   • PubMed
   • Google Scholar
5. 1. Barker M
   2. Rayens W

   (2003) Partial least squares for discrimination

   Journal of Chemometrics 17:166–173.

   https://doi.org/10.1002/cem.785

   • Google Scholar
6. 1. Birtalan S
   2. Zhang Y
   3. Fellouse FA
   4. Shao L
   5. Schaefer G
   6. Sidhu SS

   (2008) The intrinsic contributions of tyrosine, serine, glycine and arginine to the affinity and specificity of antibodies

   Journal of Molecular Biology 377:1518–1528.

   https://doi.org/10.1016/j.jmb.2008.01.093

   • PubMed
   • Google Scholar
7. 1. Borg NA
   2. Wun KS
   3. Kjer-Nielsen L
   4. Wilce MC
   5. Pellicci DG
   6. Koh R
   7. Besra GS
   8. Bharadwaj M
   9. Godfrey DI
   10. McCluskey J
   11. Rossjohn J

   (2007) CD1d-lipid-antigen recognition by the semi-invariant NKT T-cell receptor

   Nature 448:44–49.

   https://doi.org/10.1038/nature05907

   • PubMed
   • Google Scholar
8. Software

   1. Boughter CT

   (2020) AIMS - An Automated Immune Molecule Separator, version f6c855e

   GitHub.

   https://github.com/ctboughter/AIMS
9. 1. Bunker JJ
   2. Erickson SA
   3. Flynn TM
   4. Henry C
   5. Koval JC
   6. Meisel M
   7. Jabri B
   8. Antonopoulos DA
   9. Wilson PC
   10. Bendelac A

   (2017) Natural polyreactive IgA antibodies coat the intestinal Microbiota

   Science 358:eaan6619.

   https://doi.org/10.1126/science.aan6619

   • PubMed
   • Google Scholar
10. 1. Chen F
    2. Tzarum N
    3. Wilson IA
    4. Law M

    (2019) V_H1-69 antiviral broadly neutralizing antibodies: genetics, structures, and relevance to rational vaccine design

    Current Opinion in Virology 34:149–159.

    https://doi.org/10.1016/j.coviro.2019.02.004

    • PubMed
    • Google Scholar
11. 1. Cheong R
    2. Rhee A
    3. Wang CJ
    4. Nemenman I
    5. Levchenko A

    (2011) Information transduction capacity of noisy biochemical signaling networks

    Science 334:354–358.

    https://doi.org/10.1126/science.1204553

    • PubMed
    • Google Scholar
12. 1. Cordeiro MT
    2. Braga-Neto U
    3. Nogueira RM
    4. Marques ET

    (2009) Reliable classifier to differentiate primary and secondary acute dengue infection based on IgG ELISA

    PLOS ONE 4:e4945.

    https://doi.org/10.1371/journal.pone.0004945

    • PubMed
    • Google Scholar
13. 1. Crowell TA
    2. Colby DJ
    3. Pinyakorn S
    4. Sacdalan C
    5. Pagliuzza A
    6. Intasan J
    7. Benjapornpong K
    8. Tangnaree K
    9. Chomchey N
    10. Kroon E
    11. de Souza MS
    12. Tovanabutra S
    13. Rolland M
    14. Eller MA
    15. Paquin-Proulx D
    16. Bolton DL
    17. Tokarev A
    18. Thomas R
    19. Takata H
    20. Trautmann L
    21. Krebs SJ
    22. Modjarrad K
    23. McDermott AB
    24. Bailer RT
    25. Doria-Rose N
    26. Patel B
    27. Gorelick RJ
    28. Fullmer BA
    29. Schuetz A
    30. Grandin PV
    31. O'Connell RJ
    32. Ledgerwood JE
    33. Graham BS
    34. Tressler R
    35. Mascola JR
    36. Chomont N
    37. Michael NL
    38. Robb ML
    39. Phanuphak N
    40. Ananworanich J
    41. RV397 Study Group

    (2019) Safety and efficacy of VRC01 broadly neutralising antibodies in adults with acutely treated HIV (RV397): a phase 2, randomised, double-blind, placebo-controlled trial

    The Lancet HIV 6:e297–e306.

    https://doi.org/10.1016/S2352-3018(19)30053-0

    • PubMed
    • Google Scholar
14. 1. Dash P
    2. Fiore-Gartland AJ
    3. Hertz T
    4. Wang GC
    5. Sharma S
    6. Souquette A
    7. Crawford JC
    8. Clemens EB
    9. Nguyen THO
    10. Kedzierska K
    11. La Gruta NL
    12. Bradley P
    13. Thomas PG

    (2017) Quantifiable predictive features define epitope-specific T cell receptor repertoires

    Nature 547:89–93.

    https://doi.org/10.1038/nature22383

    • PubMed
    • Google Scholar
15. 1. Datta-Mannan A
    2. Lu J
    3. Witcher DR
    4. Leung D
    5. Tang Y
    6. Wroblewski VJ

    (2015) The interplay of non-specific binding, target-mediated clearance and FcRn interactions on the pharmacokinetics of humanized antibodies

    mAbs 7:1084–1093.

    https://doi.org/10.1080/19420862.2015.1075109

    • PubMed
    • Google Scholar
16. Book

    1. Davison AC
    2. Hinkley D

    (2011) Bootstrap Methods and Their Application

    Santa Barbara: Cambridge University Press.

    https://doi.org/10.1017/CBO9780511802843

    • Google Scholar
17. 1. Dimitrov JD
    2. Planchais C
    3. Roumenina LT
    4. Vassilev TL
    5. Kaveri SV
    6. Lacroix-Desmazes S

    (2013) Antibody polyreactivity in health and disease

    Journal of Immunology 191:993–999.

    https://doi.org/10.4049/jimmunol.1300880

    • Google Scholar
18. 1. Eisen HN
    2. Siskind GW

    (1964) Variations in affinities of antibodies during the immune response

    Biochemistry 3:996–1008.

    https://doi.org/10.1021/bi00895a027

    • PubMed
    • Google Scholar
19. 1. Forconi F
    2. Potter KN
    3. Wheatley I
    4. Darzentas N
    5. Sozzi E
    6. Stamatopoulos K
    7. Mockridge CI
    8. Packham G
    9. Stevenson FK

    (2010) The normal IGHV1-69-derived B-cell repertoire contains stereotypic patterns characteristic of unmutated CLL

    Blood 115:71–77.

    https://doi.org/10.1182/blood-2009-06-225813

    • PubMed
    • Google Scholar
20. 1. Haynes BF
    2. Fleming J
    3. St Clair EW
    4. Katinger H
    5. Stiegler G
    6. Kunert R
    7. Robinson J
    8. Scearce RM
    9. Plonk K
    10. Staats HF
    11. Ortel TL
    12. Liao HX
    13. Alam SM

    (2005) Cardiolipin polyspecific autoreactivity in two broadly neutralizing HIV-1 antibodies

    Science 308:1906–1908.

    https://doi.org/10.1126/science.1111781

    • PubMed
    • Google Scholar
21. 1. Haynes BF
    2. Burton DR
    3. Mascola JR

    (2019) Multiple roles for HIV broadly neutralizing antibodies

    Science Translational Medicine 11:eaaz2686.

    https://doi.org/10.1126/scitranslmed.aaz2686

    • PubMed
    • Google Scholar
22. 1. Hellberg S
    2. Sjoestroem M
    3. Skagerberg B
    4. Wold S

    (1987) Peptide quantitative structure-activity relationships, a multivariate approach

    Journal of Medicinal Chemistry 30:1126–1135.

    https://doi.org/10.1021/jm00390a003

    • Google Scholar
23. 1. Hötzel I
    2. Theil FP
    3. Bernstein LJ
    4. Prabhu S
    5. Deng R
    6. Quintana L
    7. Lutman J
    8. Sibia R
    9. Chan P
    10. Bumbaca D
    11. Fielder P
    12. Carter PJ
    13. Kelley RF

    (2012) A strategy for risk mitigation of antibodies with fast clearance

    mAbs 4:753–760.

    https://doi.org/10.4161/mabs.22189

    • PubMed
    • Google Scholar
24. 1. Hunter JD

    (2007) Matplotlib: a 2D graphics environment

    Computing in Science & Engineering 9:90–95.

    https://doi.org/10.1109/MCSE.2007.55

    • Google Scholar
25. 1. IAVI Protocol G Investigators
    2. Zhang L
    3. Irimia A
    4. He L
    5. Landais E
    6. Rantalainen K
    7. Leaman DP
    8. Vollbrecht T
    9. Stano A
    10. Sands DI
    11. Kim AS
    12. Poignard P
    13. Burton DR
    14. Murrell B
    15. Ward AB
    16. Zhu J
    17. Wilson IA
    18. Zwick MB

    (2019) An MPER antibody neutralizes HIV-1 using germline features shared among donors

    Nature Communications 10:5389.

    https://doi.org/10.1038/s41467-019-12973-1

    • PubMed
    • Google Scholar
26. 1. Jain T
    2. Boland T
    3. Lilov A
    4. Burnina I
    5. Brown M
    6. Xu Y
    7. Vásquez M

    (2017a) Prediction of delayed retention of antibodies in hydrophobic interaction chromatography from sequence using machine learning

    Bioinformatics 33:3758–3766.

    https://doi.org/10.1093/bioinformatics/btx519

    • PubMed
    • Google Scholar
27. 1. Jain T
    2. Sun T
    3. Durand S
    4. Hall A
    5. Houston NR
    6. Nett JH
    7. Sharkey B
    8. Bobrowicz B
    9. Caffry I
    10. Yu Y
    11. Cao Y
    12. Lynaugh H
    13. Brown M
    14. Baruah H
    15. Gray LT
    16. Krauland EM
    17. Xu Y
    18. Vásquez M
    19. Wittrup KD

    (2017b) Biophysical properties of the clinical-stage antibody landscape

    PNAS 114:944–949.

    https://doi.org/10.1073/pnas.1616408114

    • PubMed
    • Google Scholar
28. 1. Karami Y
    2. Rey J
    3. Postic G
    4. Murail S
    5. Tufféry P
    6. de Vries SJ

    (2019) DaReUS-Loop: a web server to model multiple loops in homology models

    Nucleic Acids Research 47:W423–W428.

    https://doi.org/10.1093/nar/gkz403

    • PubMed
    • Google Scholar
29. 1. Kelley LA
    2. Mezulis S
    3. Yates CM
    4. Wass MN
    5. Sternberg MJE

    (2015) The Phyre2 web portal for protein modeling, prediction and analysis

    Nature Protocols 10:845–858.

    https://doi.org/10.1038/nprot.2015.053

    • Google Scholar
30. 1. Kelly RL
    2. Sun T
    3. Jain T
    4. Caffry I
    5. Yu Y
    6. Cao Y
    7. Lynaugh H
    8. Brown M
    9. Vásquez M
    10. Wittrup KD
    11. Xu Y

    (2015) High throughput cross-interaction measures for human IgG1 antibodies correlate with clearance rates in mice

    mAbs 7:770–777.

    https://doi.org/10.1080/19420862.2015.1043503

    • PubMed
    • Google Scholar
31. 1. Kelly RL
    2. Le D
    3. Zhao J
    4. Wittrup KD

    (2018) Reduction of nonspecificity motifs in synthetic antibody libraries

    Journal of Molecular Biology 430:119–130.

    https://doi.org/10.1016/j.jmb.2017.11.008

    • PubMed
    • Google Scholar
32. 1. Kidera A
    2. Konishi Y
    3. Oka M
    4. Ooi T
    5. Scheraga HA

    (1985) Statistical analysis of the physical properties of the 20 naturally occurring amino acids

    Journal of Protein Chemistry 4:23–55.

    https://doi.org/10.1007/BF01025492

    • Google Scholar
33. 1. Klein JS
    2. Bjorkman PJ

    (2010) Few and far between: how HIV may be evading antibody avidity

    PLOS Pathogens 6:e1000908.

    https://doi.org/10.1371/journal.ppat.1000908

    • PubMed
    • Google Scholar
34. Conference

    1. Kluyver T
    2. Ragan-kelley B
    3. Fernando P
    4. Granger B
    5. Bussonnier M
    6. Frederic J
    7. Kelley K
    8. Hamrick J
    9. Grout J
    10. Corlay S
    11. Ivanov P
    12. Avila D
    13. Abdalla S
    14. Willing C

    (2016)

    Jupyter Notebooks—a publishing format for reproducible computational workflows

    Positioning and Power in Academic Publishing: Players, Agents and Agendas.

    • Google Scholar
35. 1. Koelsch K
    2. Zheng NY
    3. Zhang Q
    4. Duty A
    5. Helms C
    6. Mathias MD
    7. Jared M
    8. Smith K
    9. Capra JD
    10. Wilson PC

    (2007) Mature B cells class switched to IgD are autoreactive in healthy individuals

    Journal of Clinical Investigation 117:1558–1565.

    https://doi.org/10.1172/JCI27628

    • PubMed
    • Google Scholar
36. 1. Lecerf M
    2. Kanyavuz A
    3. Lacroix-Desmazes S
    4. Dimitrov JD

    (2019) Sequence features of variable region determining physicochemical properties and polyreactivity of therapeutic antibodies

    Molecular Immunology 112:338–346.

    https://doi.org/10.1016/j.molimm.2019.06.012

    • PubMed
    • Google Scholar
37. 1. Li GM
    2. Chiu C
    3. Wrammert J
    4. McCausland M
    5. Andrews SF
    6. Zheng NY
    7. Lee JH
    8. Huang M
    9. Qu X
    10. Edupuganti S
    11. Mulligan M
    12. Das SR
    13. Yewdell JW
    14. Mehta AK
    15. Wilson PC
    16. Ahmed R

    (2012) Pandemic H1N1 influenza vaccine induces a recall response in humans that favors broadly cross-reactive memory B cells

    PNAS 109:9047–9052.

    https://doi.org/10.1073/pnas.1118979109

    • PubMed
    • Google Scholar
38. 1. Liu Q
    2. Chen P
    3. Wang B
    4. Zhang J
    5. Li J

    (2018) Hot spot prediction in protein-protein interactions by an ensemble system

    BMC Systems Biology 12:8.

    https://doi.org/10.1186/s12918-018-0665-8

    • Google Scholar
39. 1. Luoma AM
    2. Castro CD
    3. Mayassi T
    4. Bembinster LA
    5. Bai L
    6. Picard D
    7. Anderson B
    8. Scharf L
    9. Kung JE
    10. Sibener LV
    11. Savage PB
    12. Jabri B
    13. Bendelac A
    14. Adams EJ

    (2013) Crystal structure of vδ1 T Cell Receptor in Complex with CD1d-Sulfatide Shows MHC-like Recognition of a Self-Lipid by Human γδ T Cells

    Immunity 39:1032–1042.

    https://doi.org/10.1016/j.immuni.2013.11.001

    • PubMed
    • Google Scholar
40. 1. Ma Y
    2. Vilanova D
    3. Atalar K
    4. Delfour O
    5. Edgeworth J
    6. Ostermann M
    7. Hernandez-Fuentes M
    8. Razafimahatratra S
    9. Michot B
    10. Persing DH
    11. Ziegler I
    12. Törös B
    13. Mölling P
    14. Olcén P
    15. Beale R
    16. Lord GM

    (2013) Genome-wide sequencing of cellular microRNAs identifies a combinatorial expression signature diagnostic of Sepsis

    PLOS ONE 8:e75918.

    https://doi.org/10.1371/journal.pone.0075918

    • PubMed
    • Google Scholar
41. 1. McKean D
    2. Huppi K
    3. Bell M
    4. Staudt L
    5. Gerhard W
    6. Weigert M

    (1984) Generation of antibody diversity in the immune response of BALB/c mice to influenza virus hemagglutinin

    PNAS 81:3180–3184.

    https://doi.org/10.1073/pnas.81.10.3180

    • PubMed
    • Google Scholar
42. Software

    1. McKinney W

    (2015)

    Pandas - Powerful Python Data Analysis Toolkit

    Pandas - Powerful Python Data Analysis Toolkit.
43. 1. Mora T
    2. Walczak AM
    3. Bialek W
    4. Callan CG

    (2010) Maximum entropy models for antibody diversity

    PNAS 107:5405–5410.

    https://doi.org/10.1073/pnas.1001705107

    • PubMed
    • Google Scholar
44. 1. Mouquet H
    2. Scheid JF
    3. Zoller MJ
    4. Krogsgaard M
    5. Ott RG
    6. Shukair S
    7. Artyomov MN
    8. Pietzsch J
    9. Connors M
    10. Pereyra F
    11. Walker BD
    12. Ho DD
    13. Wilson PC
    14. Seaman MS
    15. Eisen HN
    16. Chakraborty AK
    17. Hope TJ
    18. Ravetch JV
    19. Wardemann H
    20. Nussenzweig MC

    (2010) Polyreactivity increases the apparent affinity of anti-HIV antibodies by heteroligation

    Nature 467:591–595.

    https://doi.org/10.1038/nature09385

    • PubMed
    • Google Scholar
45. 1. Mouquet H
    2. Klein F
    3. Scheid JF
    4. Warncke M
    5. Pietzsch J
    6. Oliveira TY
    7. Velinzon K
    8. Seaman MS
    9. Nussenzweig MC

    (2011) Memory B cell antibodies to HIV-1 gp140 cloned from individuals infected with clade A and B viruses

    PLOS ONE 6:e24078.

    https://doi.org/10.1371/journal.pone.0024078

    • PubMed
    • Google Scholar
46. 1. Murugan A
    2. Mora T
    3. Walczak AM
    4. Callan CG

    (2012) Statistical inference of the generation probability of T-cell receptors from sequence repertoires

    PNAS 109:16161–16166.

    https://doi.org/10.1073/pnas.1212755109

    • PubMed
    • Google Scholar
47. 1. Neu KE
    2. Guthmiller JJ
    3. Huang M
    4. La J
    5. Vieira MC
    6. Kim K
    7. Zheng NY
    8. Cortese M
    9. Tepora ME
    10. Hamel NJ
    11. Rojas KT
    12. Henry C
    13. Shaw D
    14. Dulberger CL
    15. Pulendran B
    16. Cobey S
    17. Khan AA
    18. Wilson PC

    (2019) Spec-seq unveils transcriptional subpopulations of antibody-secreting cells following influenza vaccination

    The Journal of Clinical Investigation 129:JCI121341.

    https://doi.org/10.1172/JCI121341

    • Google Scholar
48. 1. Obrezanova O
    2. Arnell A
    3. de la Cuesta RG
    4. Berthelot ME
    5. Gallagher TR
    6. Zurdo J
    7. Stallwood Y

    (2015) Aggregation risk prediction for antibodies and its application to biotherapeutic development

    mAbs 7:352–363.

    https://doi.org/10.1080/19420862.2015.1007828

    • PubMed
    • Google Scholar
49. 1. Ochsenbein AF
    2. Fehr T
    3. Lutz C
    4. Suter M
    5. Brombacher F
    6. Hengartner H
    7. Zinkernagel RM

    (1999) Control of early viral and bacterial distribution and disease by natural antibodies

    Science 286:2156–2159.

    https://doi.org/10.1126/science.286.5447.2156

    • PubMed
    • Google Scholar
50. Preprint

    1. Pedregosa F
    2. Varoquaux G
    3. Gramfort A
    4. Michel V
    5. Thirion B

    (2011) Scikit-learn: machine learning in Python

    arXiv.

    https://arxiv.org/abs/1201.0490

    • Google Scholar
51. 1. Piertney SB
    2. Oliver MK

    (2006) The evolutionary ecology of the major histocompatibility complex

    Heredity 96:7–21.

    https://doi.org/10.1038/sj.hdy.6800724

    • PubMed
    • Google Scholar
52. 1. Planchais C
    2. Kök A
    3. Kanyavuz A
    4. Lorin V
    5. Bruel T
    6. Guivel-Benhassine F
    7. Rollenske T
    8. Prigent J
    9. Hieu T
    10. Prazuck T
    11. Lefrou L
    12. Wardemann H
    13. Schwartz O
    14. Dimitrov JD
    15. Hocqueloux L
    16. Mouquet H

    (2019) HIV-1 envelope recognition by polyreactive and Cross-Reactive intestinal B cells

    Cell Reports 27:572–585.

    https://doi.org/10.1016/j.celrep.2019.03.032

    • PubMed
    • Google Scholar
53. 1. Prigent J
    2. Lorin V
    3. Kök A
    4. Hieu T
    5. Bourgeau S
    6. Mouquet H

    (2016) Scarcity of autoreactive human blood IgA⁺ memory B cells

    European Journal of Immunology 46:2340–2351.

    https://doi.org/10.1002/eji.201646446

    • PubMed
    • Google Scholar
54. 1. Prigent J
    2. Jarossay A
    3. Planchais C
    4. Eden C
    5. Dufloo J
    6. Kök A
    7. Lorin V
    8. Vratskikh O
    9. Couderc T
    10. Bruel T
    11. Schwartz O
    12. Seaman MS
    13. Ohlenschläger O
    14. Dimitrov JD
    15. Mouquet H

    (2018) Conformational plasticity in broadly neutralizing HIV-1 antibodies triggers polyreactivity

    Cell Reports 23:2568–2581.

    https://doi.org/10.1016/j.celrep.2018.04.101

    • PubMed
    • Google Scholar
55. 1. Qiao Z
    2. Zhou L
    3. Huang JZ

    (2009)

    Sparse linear discriminant analysis with applications to high dimensional low sample size data

    IAENG International Journal of Applied Mathematics 39:2.

    • Google Scholar
56. 1. Rabia LA
    2. Zhang Y
    3. Ludwig SD
    4. Julian MC
    5. Tessier PM

    (2018) Net charge of antibody complementarity-determining regions is a key predictor of specificity

    Protein Engineering, Design and Selection 31:409–418.

    https://doi.org/10.1093/protein/gzz002

    • Google Scholar
57. 1. Raybould MIJ
    2. Marks C
    3. Krawczyk K
    4. Taddese B
    5. Nowak J
    6. Lewis AP
    7. Bujotzek A
    8. Shi J
    9. Deane CM

    (2019) Five computational developability guidelines for therapeutic antibody profiling

    PNAS 116:4025–4030.

    https://doi.org/10.1073/pnas.1810576116

    • PubMed
    • Google Scholar
58. 1. Román-Roldán R
    2. Bernaola-Galván P
    3. Oliver J

    (1996) Application of information theory to DNA sequence analysis: a review

    Pattern Recognition 29:1187–1194.

    https://doi.org/10.1016/0031-3203(95)00145-X

    • Google Scholar
59. 1. Sasso EH
    2. Willems van Dijk K
    3. Bull AP
    4. Milner EC

    (1993) A fetally expressed immunoglobulin VH1 gene belongs to a complex set of alleles

    Journal of Clinical Investigation 91:2358–2367.

    https://doi.org/10.1172/JCI116468

    • PubMed
    • Google Scholar
60. 1. Sayers EW
    2. Beck J
    3. Brister JR
    4. Bolton EE
    5. Canese K
    6. Comeau DC
    7. Funk K
    8. Ketter A
    9. Kim S
    10. Kimchi A
    11. Kitts PA
    12. Kuznetsov A
    13. Lathrop S
    14. Lu Z
    15. McGarvey K
    16. Madden TL
    17. Murphy TD
    18. O'Leary N
    19. Phan L
    20. Schneider VA
    21. Thibaud-Nissen F
    22. Trawick BW
    23. Pruitt KD
    24. Ostell J

    (2020) Database resources of the national center for biotechnology information

    Nucleic Acids Research 48:D9–D16.

    https://doi.org/10.1093/nar/gkz899

    • PubMed
    • Google Scholar
61. 1. Sethi DK
    2. Agarwal A
    3. Manivel V
    4. Rao KV
    5. Salunke DM

    (2006) Differential epitope positioning within the germline antibody paratope enhances promiscuity in the primary immune response

    Immunity 24:429–438.

    https://doi.org/10.1016/j.immuni.2006.02.010

    • PubMed
    • Google Scholar
62. 1. Shannon CE

    (1948) A mathematical theory of communication

    Bell System Technical Journal 27:379–423.

    https://doi.org/10.1002/j.1538-7305.1948.tb01338.x

    • Google Scholar
63. 1. Sharma VK
    2. Patapoff TW
    3. Kabakoff B
    4. Pai S
    5. Hilario E
    6. Zhang B
    7. Li C
    8. Borisov O
    9. Kelley RF
    10. Chorny I
    11. Zhou JZ
    12. Dill KA
    13. Swartz TE

    (2014) In silico selection of therapeutic antibodies for development: viscosity, clearance, and chemical stability

    PNAS 111:18601–18606.

    https://doi.org/10.1073/pnas.1421779112

    • Google Scholar
64. 1. Sneath PHA

    (1966) Relations between chemical structure and biological activity in peptides

    Journal of Theoretical Biology 12:157–195.

    https://doi.org/10.1016/0022-5193(66)90112-3

    • Google Scholar
65. 1. Starr CG
    2. Tessier PM

    (2019) Selecting and engineering monoclonal antibodies with drug-like specificity

    Current Opinion in Biotechnology 60:119–127.

    https://doi.org/10.1016/j.copbio.2019.01.008

    • PubMed
    • Google Scholar
66. 1. Tiller T
    2. Tsuiji M
    3. Yurasov S
    4. Velinzon K
    5. Nussenzweig MC
    6. Wardemann H

    (2007) Autoreactivity in human IgG+ memory B cells

    Immunity 26:205–213.

    https://doi.org/10.1016/j.immuni.2007.01.009

    • PubMed
    • Google Scholar
67. 1. Victora GD
    2. Nussenzweig MC

    (2012) Germinal centers

    Annual Review of Immunology 30:429–457.

    https://doi.org/10.1146/annurev-immunol-020711-075032

    • PubMed
    • Google Scholar
68. 1. Vihinen M
    2. Torkkila E
    3. Riikonen P

    (1994) Accuracy of protein flexibility predictions

    Proteins: Structure, Function, and Genetics 19:141–149.

    https://doi.org/10.1002/prot.340190207

    • PubMed
    • Google Scholar
69. 1. Vinga S

    (2014) Information theory applications for biological sequence analysis

    Briefings in Bioinformatics 15:376–389.

    https://doi.org/10.1093/bib/bbt068

    • PubMed
    • Google Scholar
70. 1. Virtanen P
    2. Gommers R
    3. Oliphant TE
    4. Haberland M
    5. Reddy T
    6. Cournapeau D
    7. Burovski E
    8. Peterson P
    9. Weckesser W
    10. Bright J
    11. van der Walt SJ
    12. Brett M
    13. Wilson J
    14. Millman KJ
    15. Mayorov N
    16. Nelson ARJ
    17. Jones E
    18. Kern R
    19. Larson E
    20. Carey CJ
    21. Polat İ
    22. Feng Y
    23. Moore EW
    24. VanderPlas J
    25. Laxalde D
    26. Perktold J
    27. Cimrman R
    28. Henriksen I
    29. Quintero EA
    30. Harris CR
    31. Archibald AM
    32. Ribeiro AH
    33. Pedregosa F
    34. van Mulbregt P
    35. SciPy 1.0 Contributors

    (2020) SciPy 1.0: fundamental algorithms for scientific computing in Python

    Nature Methods 17:261–272.

    https://doi.org/10.1038/s41592-019-0686-2

    • PubMed
    • Google Scholar
71. 1. Wardemann H
    2. Yurasov S
    3. Schaefer A
    4. Young JW
    5. Meffre E
    6. Nussenzweig MC

    (2003) Predominant autoantibody production by early human B cell precursors

    Science 301:1374–1377.

    https://doi.org/10.1126/science.1086907

    • PubMed
    • Google Scholar
72. 1. Whittle JRR
    2. Zhang R
    3. Khurana S
    4. King LR
    5. Manischewitz J
    6. Golding H

    (2011) Broadly neutralizing human antibody that recognizes the receptor-binding pocket of influenza virus hemagglutinin

    PNAS 108 :14216–14221.

    https://doi.org/10.1073/pnas.1111497108

    • Google Scholar
73. 1. Wrammert J
    2. Koutsonanos D
    3. Li G-M
    4. Edupuganti S
    5. Sui J
    6. Morrissey M
    7. McCausland M
    8. Skountzou I
    9. Hornig M
    10. Lipkin WI
    11. Mehta A
    12. Razavi B
    13. Del Rio C
    14. Zheng N-Y
    15. Lee J-H
    16. Huang M
    17. Ali Z
    18. Kaur K
    19. Andrews S
    20. Amara RR
    21. Wang Y
    22. Das SR
    23. O'Donnell CD
    24. Yewdell JW
    25. Subbarao K
    26. Marasco WA
    27. Mulligan MJ
    28. Compans R
    29. Ahmed R
    30. Wilson PC

    (2011) Broadly cross-reactive antibodies dominate the human B cell response against 2009 pandemic H1N1 influenza virus infection

    Journal of Experimental Medicine 208:181–193.

    https://doi.org/10.1084/jem.20101352

    • Google Scholar
74. 1. Ye J
    2. Ma N
    3. Madden TL
    4. Ostell JM

    (2013) IgBLAST: an immunoglobulin variable domain sequence analysis tool

    Nucleic Acids Research 41:W34–W40.

    https://doi.org/10.1093/nar/gkt382

    • PubMed
    • Google Scholar
75. Software

    1. Ziegler E
    2. Zaytsev YV
    3. Waskom MT
    4. Botvinnik O
    5. Hobson P
    6. Cole JB
    7. Halchenko Y
    8. Hoyer S
    9. Miles A
    10. Augspurger T
    11. Yarkoni T
    12. Megies T
    13. Coelho LP
    14. Wehner D
    15. Waskom M

    (2014)

    Seaborn: v0.5.0

    Seaborn: v0.5.0.

Thank you for submitting your article "Biochemical Patterns of Antibody Polyreactivity Revealed Through a Bioinformatics-Based Analysis of CDR Loops" for consideration by eLife. Your article has been reviewed by three peer reviewers, and the evaluation has been overseen by a Reviewing Editor and Satyajit Rath as the Senior Editor. The following individuals involved in review of your submission have agreed to reveal their identity: Brian Baker (Reviewer #1); Bart Haynes (Reviewer #2).

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

As the editors have judged that your manuscript is of interest, but as described below, we suggest substantial revisions. We would like to draw your attention to changes in our revision policy that we have made in response to COVID-19 (https://elifesciences.org/articles/57162). First, because many researchers have temporarily lost access to the labs, we will give authors as much time as they need to submit revised manuscripts. We are also offering, if you choose, to post the manuscript to bioRxiv (if it is not already there) along with this decision letter and a formal designation that the manuscript is "in revision at eLife". Please let us know if you would like to pursue this option. (If your work is more suitable for medRxiv, you will need to post the preprint yourself, as the mechanisms for us to do so are still in development.)

Summary:

The authors present a bioinformatics pipeline to analyze a large aggregate data set of nearly 1,500 antibody sequences, seeking to uncover the biophysical underpinning of antibody polyreactivity – i.e., broad low-affinity binding to diverse epitopes. This is an important topic, relevant not only to Ab therapeutics, but to other assessments of immune repertoires and considerations of protein-protein interaction hubs in general. The unbiased informatics approach that is taken, largely separated from structural considerations, is a distinctive feature of the work, and an attractive aspect of the study. The authors identify two key determinants of polyreactivity: First, the binding interface tends to be neutral, i.e., being neither strongly hydrophobic nor hydrophilic, and lacking significant positive or negative charge. Second, crosstalk (measured by mutual information) between and within CDR loops increases in the heavy chain of polyreactive antibodies. The method of sequence alignment leaves out most structural information (except aligning by the center of each CDR loop) yet retaining the positional context.

The research goal of identifying biophysical properties that allow for promiscuity of a protein selected for strong binding to a particular target is of general interest and biomedical importance, and the approach of combining statistical analyses and information metrics is appealing. The lack of a clear smoking gun is an important finding, as it clearly illustrates the subtleties at play. A classifier is described at the end of the paper, showing promise for sequence-based predictions and applications to therapeutic Abs.

We believe the paper may be suitable for eLife with revisions to address the following concerns.

Essential revisions:

1) The reviewers have concerns about the binary distinction between polyreactive vs. non-polyreactive. The authors essentially sum up this concern in the third paragraph of the subsection “Systematic Determination of the Key Contributions to Polyreactivity”, as well as in the ninth paragraph of the Discussion. The fact that pharma has nonetheless articulated a test for polyreactivity does not make the problem binary, which is clearly recognized by the authors. One way around this would be to repeat the analyses in Figures 1-3 on the more stringently separated (albeit now smaller) dataset used for Figure 4. How do things change? Are there any stronger signals?

Also, this limitation, while articulated well in the Discussion, could be brought up in the Introduction. Forcing a non-binary problem to be binary is by no means unique to this study, but it is a limitation best addressed upfront. As mentioned above, a strength of the paper is the demonstration of the scope of the challenge.

2) A major concern is that, with the biased removal of antibodies that are harder to classify, the true statistics are distorted. This occurs in two places. First, to avoid overfitting of LDA, only input vectors with the largest average differences between the two populations are kept. But this necessarily introduces bias which simplifies the classification problem and distorts the statistical structure in the original data set. Second, upon observing that an intermediate exists between the two classes when applying LDA to a subsampled data set, the authors removed the antibodies that bind a moderate number of select ligands from the analysis. But this continuum of polyreactivity is a real, and likely important, data feature. Since all the subsequent analyses act on the input data, it is necessary to address, with a clear description and supporting evidence, that the main findings remain valid when using the full data set. After this is addressed, direct comparisons – LDA vs. PCA, and with other methods – should be provided to justify the choice of LDA over PCA, and the performance of this newly developed pipeline relative to existing methods.

3) There seems to be a jump from the LDA to mutual information, then back to LDA in the application to therapeutic antibodies. While the physical hints the mutual information gives as mentioned in the Discussion is appreciated, and this can be built on, is this incorporated at all into the analysis of the therapeutic antibodies?

4) Related to point 3, it is not clear why mutual information is a good choice for detecting correlated changes between residues, and there is no comparison to other methods and previous studies. Moreover, mutual information won't be able to distinguish between coevolution and crosstalk between residues. For instance, Direct Coupling Analysis (DCA) indicates that coevolving residues are often in physical proximity. This points to a weakness of lacking the structural information in this work. Thus, it is unclear whether increased mutual information really indicates crosstalk, or instead signals coevolution between residues. This ambiguity undermines the second determinant of polyreactivity.

5) The analysis is finally applied to therapeutic antibodies, which is considered a logical next step. But this is not well motivated – why, in the first place, shall one expect correlations between polyreactivity status of naturally-derived antibodies and the acceptance or discontinuation of a therapeutic antibody? Hence, this section didn't lend validation support to the approach. Perhaps the manuscript could be shortened by deletion of the therapeutic dataset. These antibodies are not natural and it is not clear what approved versus discontinued antibody means.

6) There are other major omissions. First, no performance comparison between this new software and other available methods of antibody sequence analysis is presented (all is said in words, with no supporting figures). Second, it is stated that the approach has been adapted and applied to analyze MHC-like molecules, but no data or references are provided.

7) While in general the manuscript is quite clear, some sections are less so. Is there nothing to be discerned from the identity of the top 10 weights shown in Figure 4? We are only given the identity of three (subsection “Systematic Determination of the Key Contributions to Polyreactivity”, last paragraph), and even then these are only mentioned in a cursory manner. How do these and the other heavily weighted terms relate to the biophysical clues gleaned from the earlier analyses?

8) In the Abstract, the word "offensive" is not defined and would best be "neutral charge" for reader understanding.

9) Introduction, last paragraph, it is not clear in the first sentence if there are 1500 polyreactive abs or a total of 1500 abs studied. It is clarified later but not clear here.

10) The authors use a relatively small panel of antigens to define polyreactivity, whereas a chip type of assay with more proteins may be more precise. This might be mentioned as a way to make the discrimination between PR and non-PR abs more precise.

11) Figure 1 is terrific and very interesting. The fact that V_H1-69 in over-represented in the polyreactive group of antibodies is important, both because of the propensity of polyreactive/rheumatoid factor B cells in fetal liver to be V_H1-69 and because the polar region (PR)-distal HIV gp41 MPER Abs frequently use V_H1-69. Perhaps these associations could be referenced and discussed.

12) In the fifth paragraph of the subsection “A Surface-Level Analysis of Polyreactive Antibody Sequences”, can statistics be applied to determine if any difference between 27% and 17%?

13) Figure 3, were these p values corrected for multiple tests?

Essential revisions:

1) The reviewers have concerns about the binary distinction between polyreactive vs. non-polyreactive. The authors essentially sum up this concern in the third paragraph of the subsection “Systematic Determination of the Key Contributions to Polyreactivity”, as well as in the ninth paragraph of the Discussion. The fact that pharma has nonetheless articulated a test for polyreactivity does not make the problem binary, which is clearly recognized by the authors. One way around this would be to repeat the analyses in Figure 1-3 on the more stringently separated (albeit now smaller) dataset used for Figure 4. How do things change? Are there any stronger signals?

We appreciate the reviewers’ cognizance of the difficult nuances inherent to the current definitions of polyreactivity available to us and the broader field. We agree that including the analysis of Figures 1-3 applied to the parsed dataset would help contextualize our findings, and these data have been added to the supplement (Figure 1—figure supplement 1, Figure 2—figure supplement 1, and Figure 3—figure supplement 1). In these new figures, we see stronger signals in the same direction as the general trends found in the analysis of the full dataset. These stronger signals are discussed across multiple sections (subsections “A Surface-Level Analysis of Polyreactive Antibody Sequences” and “A Position Sensitive Matrix Representation of Sequences Provides Further Insights into Polyreactivity”). An interesting difference can be seen in the mutual information analysis, where loop crosstalk is increased in all loops, not just those of the heavy chain. This is highlighted in newly added text in the subsection “An Information Theoretic Approach” and Figure 5—figure supplement 1. Now that the parsed dataset is discussed earlier in the manuscript, a description of the parsed dataset has been relocated to the subsection “Database”.

Also, this limitation, while articulated well in the Discussion, could be brought up in the Introduction. Forcing a non-binary problem to be binary is by no means unique to this study, but it is a limitation best addressed upfront. As mentioned above, a strength of the paper is the demonstration of the scope of the challenge.

This difficulty of reframing a non-binary problem as binary is now discussed in the Introduction, in addition to the pre-existing statements found in the Results section and Discussion section. This updated section of the Introduction can be found in the seventh paragraph, when discussing the many difficulties of the systemic characterization of polyreactive antibodies

2) A major concern is that, with the biased removal of antibodies that are harder to classify, the true statistics are distorted. This occurs in two places. First, to avoid overfitting of LDA, only input vectors with the largest average differences between the two populations are kept. But this necessarily introduces bias which simplifies the classification problem and distorts the statistical structure in the original data set.

First, to hopefully alleviate some concerns for the reviewers, we have included Figure 4—figure supplement 2, using LDA to classify the full dataset, removing the identified bias. While performance clearly suffers, we are still able to successfully identify antibody polyreactivity with 71% accuracy, or a 65% accuracy when using LDA as a canonical classifier.

Second, when discussing LDA in this manuscript, it is important to differentiate between the two distinct modes used. In mode 1, our goal is explicitly to find the vectors which best discriminate between polyreactive and non-polyreactive antibodies. So here the “bias” is simply a search method. However, we can show that using PCA to reduce dimensionality rather than the maximum difference algorithm results in only a slight decrease in the reported accuracies (shown in Author response image 1).

In mode 2, the more canonical classification mode, this concern of simplifying the classification problem would be valid if the selection of the vectors happened before the splitting of data into “test” and “training” sets. However, this step occurs once the data has already been split. In other words, for each replica of the classification testing, the vectors selected with the largest average differences are changing, albeit slightly. Conceptually, this is no different from using other approaches to reduce the dimensionality of the system. Rather than using PCA, which identifies the axes of maximal variance using a linear combination of the input vectors, we are identifying axes of maximal difference.

The strength of this approach is that the resultant classifier is a linear combination of initial input vectors, rather than a linear combination of vectors that are linear combinations of the input vectors. This increased clarity makes the classifier more interpretable.

Author response image 1

Download asset Open asset

Second, upon observing that an intermediate exists between the two classes when applying LDA to a subsampled data set, the authors removed the antibodies that bind a moderate number of select ligands from the analysis. But this continuum of polyreactivity is a real, and likely important, data feature. Since all the subsequent analyses act on the input data, it is necessary to address, with a clear description and supporting evidence, that the main findings remain valid when using the full data set. After this is addressed, direct comparisons – LDA vs. PCA, and with other methods – should be provided to justify the choice of LDA over PCA, and the performance of this newly developed pipeline relative to existing methods.

It should be pointed out, the full dataset is used for the majority of the analysis in the paper (Figures 1-3, Figure 5). Please see the subsection “Database”, for this clarification. As was addressed in reviewer comment 1, these analyses have now been applied to the parsed datasets and show stronger effects.

Likewise, the use of PCA was shown to be ineffective in separating the parsed data in Figure 4—figure supplement 1. For completeness, the PCA analysis of the full dataset is now also included in Figure 4—figure supplement 2. PCA, whether applied to the parsed or full dataset, is incapable of separating the two classes.

Lastly, to our knowledge, no methods to discriminate polyreactive and non-polyreactive antibodies in a systematic way currently exist.

3) There seems to be a jump from the LDA to mutual information, then back to LDA in the application to therapeutic antibodies. While the physical hints the mutual information gives as mentioned in the Discussion is appreciated, and this can be built on, is this incorporated at all into the analysis of the therapeutic antibodies?

As addressed in comment 5, we have removed the therapeutic antibody section, helping improve the flow of the manuscript.

4) Related to point 3, it is not clear why mutual information is a good choice for detecting correlated changes between residues, and there is no comparison to other methods and previous studies. Moreover, mutual information won't be able to distinguish between coevolution and crosstalk between residues. For instance, Direct Coupling Analysis (DCA) indicates that coevolving residues are often in physical proximity. This points to a weakness of lacking the structural information in this work. Thus, it is unclear whether increased mutual information really indicates crosstalk, or instead signals coevolution between residues. This ambiguity undermines the second determinant of polyreactivity.

One of the key novel approaches in our manuscript is the quantification of antibody sequences using position sensitive Shannon Entropy and Mutual Information. While Shannon Entropy has been used previously in repertoire analysis, notably in the generation of “sequence logos” [Schneider and Stephens, Nucleic Acids Research 1990], analysis using Mutual Information has been less common. However, this lack of previous use does not preclude mutual information from being a useful tool. The novelty here is part of the strength of our manuscript.

The reviewers point regarding mutual information’s inability to determine whether the signal is originating from coevolved residues and direct crosstalk is correct, and in fact this is mentioned in the sixth paragraph of the Discussion. While direct coupling analysis may find that coevolving residues are often in physical proximity, that does not explain the relative increase in the mutual information for polyreactive antibodies. There is no reason to believe, a priori, that mutual information should be increased across and within the loops of polyreactive antibodies compared to those of non-polyreactive antibodies. Regardless of whether this increase in signal is due to coevolution throughout the process of VDJ recombination and somatic hypermutation or direct contact between side chains, the signal is stronger in polyreactive antibodies. Importantly, this signal becomes even more pronounced when analyzing the parsed dataset (Figure 5—figure supplement 1). Hopefully, future studies can address this using structural and biochemical approaches, but these are outside of the scope of the present manuscript.

5) The analysis is finally applied to therapeutic antibodies, which is considered a logical next step. But this is not well motivated – why, in the first place, shall one expect correlations between polyreactivity status of naturally-derived antibodies and the acceptance or discontinuation of a therapeutic antibody? Hence, this section didn't lend validation support to the approach. Perhaps the manuscript could be shortened by deletion of the therapeutic dataset. These antibodies are not natural and it is not clear what approved versus discontinued antibody means.

As pointed out by the reviewers, while polyreactivity can play an important role in the success of an antibody therapeutic, the precise link between natural polyreactivity outlined in this manuscript and polyreactivity in engineered antibodies is not necessarily clear. Indeed, the data from the original Supplementary Figure 8 (now removed) seems to suggest there is little to no correlation between the two.

The original goal of this section was to give an example of the flexibility of our approach and its utility as a general tool. However, tackling a problem as complex as the success of or failure of an antibody therapeutic in clinical trials was likely overly ambitious. As suggested by the reviewers, we have removed this section. We now exhibit the flexibility of our approach with the new MHC and MHC-like analysis section.

6) There are other major omissions. First, no performance comparison between this new software and other available methods of antibody sequence analysis is presented (all is said in words, with no supporting figures). Second, it is stated that the approach has been adapted and applied to analyze MHC-like molecules, but no data or references are provided.

To our knowledge, no other software or available methods for antibody sequence analysis exists that was either designed specifically for, or is flexible enough to include, the classification of polyreactive and non-polyreactive antibodies. While previous researchers have attempted to find patterns in polyreactivity (Introduction, sixth paragraph) none of these studies claim to be able to predict polyreactivity in antibodies from their studies. Other potential approaches (Introduction, seventh paragraph) have been attempted at some level, but likewise have no claims of prediction.

Those algorithms that have been generated for prediction primarily focus on structural considerations of antibody-antigen interactions; i.e. epitope prediction, binding interface prediction, and ligand identification [Graves et al. Antibodies 2020]. While in theory these approaches could be used to classify polyreactive antibodies, they were not designed with this application in mind.

Last, a manuscript that utilizes the application of the software to analyze MHC and MHC-like molecules is currently in preparation. However, to demonstrate the potential application, a simple test case has been outlined in a new section of the manuscript (subsection “Extension of the Analysis to MHC and MHC-Like Molecules”, Figure 6). Here we are tasked with the very simple analysis of separating peptide-binding and lipid-binding MHC/MHC-like molecules, and then classifying molecules identified in a recent publication. In addition, the GitHub page hosting this software has an example test case for the easy application of this analysis mode.

7) While in general the manuscript is quite clear, some sections are less so. Is there nothing to be discerned from the identity of the top 10 weights shown in Figure 4? We are only given the identity of three (subsection “Systematic Determination of the Key Contributions to Polyreactivity”, last paragraph), and even then these are only mentioned in a cursory manner. How do these and the other heavily weighted terms relate to the biophysical clues gleaned from the earlier analyses?

We appreciate the reviewer’s suggestion to more closely consider the conclusions from our analysis. These weights were not so strongly discussed in the original manuscript due to the speculative nature of some of our conclusions. Without a crystal structure of a polyreactive antibody, our conclusions can only be strung together using the bioinformatic results we have. Our revised discussion of these data can now be found in the eighth paragraph of the Discussion section.

8) In the Abstract, the word "offensive" is not defined and would best be "neutral charge" for reader understanding.

We thank the reviewer for this suggestion. This wording has been changed in the Abstract.

9) Introduction, last paragraph, it is not clear in the first sentence if there are 1500 polyreactive abs or a total of 1500 abs studied. It is clarified later but not clear here.

This point has now been clarified. Additionally, since the therapeutic antibodies are no longer considered in this manuscript, references to the number of antibodies in the study have been changed to “over 1000”.

10) The authors use a relatively small panel of antigens to define polyreactivity, whereas a chip type of assay with more proteins may be more precise. This might be mentioned as a way to make the discrimination between PR and non-PR abs more precise.

One major drawback to the currently available experimental data is the relatively limited scope of the antigens tested to determine polyreactivity. While this panel of 7 antigens is the standard in the field, a protein microarray of some sort, as is often used when testing for polyreactivity in therapeutic antibodies, would provide a more detailed picture of polyreactivity. A reference to this proposed improvement in the analysis of polyreactivity has been added to the tenth paragraph of the Discussion.

11) Figure 1 is terrific and very interesting. The fact that V_H1-69 in over-represented in the polyreactive group of antibodies is important, both because of the propensity of polyreactive/rheumatoid factor B cells in fetal liver to be V_H1-69 and because the polar region (PR)-distal HIV gp41 MPER Abs frequently use V_H1-69. Perhaps these associations could be referenced and discussed.

We thank the reviewer for their appreciation of Figure 1, and for their additional important context on V_H1-69, which we have now included. These and other conclusions regarding the results of Figure 1 can now be found in the second paragraph of the Discussion.

12) In the fifth paragraph of the subsection “A Surface-Level Analysis of Polyreactive Antibody Sequences”, can statistics be applied to determine if any difference between 27% and 17%?

While we agree with the reviewer that the application of statistics to these percentages would strengthen the findings, it is unclear how statistics could be applied here. Since Figure 1 and the data associated with it are direct reports of the “count” of some property, here the count of a specific gene, it is difficult to acquire a standard error for this metric. Ideally, future experiments may quantify the gene usage of polyreactive antibodies across individual mice or humans, allowing for a more precise understanding of the deviation in these values.

13) Figure 3, were these p values corrected for multiple tests?

Originally, the p-values of Figure 3 were not corrected for multiple tests. Figure 3A and B now contain p-values at a level determined by the Bonferroni correction. It should be noted, when applying the Bonferroni correction to the data of Figure 3C and D, all differences fail to satisfy the corrected p-value level. However, a known shortcoming of this correction is a tendency towards an overly conservative analysis of the data. In this case where the number of tested points is 104, we believe that the multiple test correction is not a proper correction to our statistics. For increased data transparency, we have included the data profiles of the null distributions used for statistical comparisons in Figure 3—figure supplement 2.

Author details

1. Christopher T Boughter

   Graduate Program in Biophysical Sciences, University of Chicago, Chicago, United States

   Contribution

   Conceptualization, Resources, Data curation, Software, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing - original draft, Writing - review and editing

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-7106-4699

2. Marta T Borowska

   Department of Biochemistry and Molecular Biology, University of Chicago, Chicago, United States

   Contribution

   Conceptualization, Data curation, Writing - review and editing

   Competing interests

   No competing interests declared

3. Jenna J Guthmiller

   Department of Medicine, Section of Rheumatology, University of Chicago, Chicago, United States

   Contribution

   Resources, Data curation, Writing - review and editing

   Competing interests

   No competing interests declared

4. Albert Bendelac

   1. Committee on Immunology, University of Chicago, Chicago, United States
   2. Department of Pathology, University of Chicago, Chicago, United States

   Contribution

   Conceptualization, Resources, Data curation, Writing - review and editing

   Competing interests

   No competing interests declared

5. Patrick C Wilson

   1. Department of Medicine, Section of Rheumatology, University of Chicago, Chicago, United States
   2. Committee on Immunology, University of Chicago, Chicago, United States

   Contribution

   Resources, Data curation, Writing - review and editing

   Competing interests

   No competing interests declared

6. Benoit Roux

   Department of Biochemistry and Molecular Biology, University of Chicago, Chicago, United States

   Contribution

   Conceptualization, Resources, Data curation, Supervision, Funding acquisition, Project administration, Writing - review and editing

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-5254-2712

7. Erin J Adams

   1. Department of Biochemistry and Molecular Biology, University of Chicago, Chicago, United States
   2. Committee on Immunology, University of Chicago, Chicago, United States

   Contribution

   Conceptualization, Resources, Data curation, Supervision, Funding acquisition, Project administration, Writing - review and editing

   For correspondence

   ejadams@uchicago.edu

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-6271-8574

• 3,429

  Page views

• 492

  Downloads

• 11

  Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.