Immunological Diversity with Similarity

Abstract

A diverse immune repertoire is considered a hallmark of good health, but measuring diversity requires a framework that incorporates not only sequences’ relative frequencies but also their functional similarity to each other. Using experimentally measured dissociation constants from over 1,300 antibody-antigen and T-cell receptor (TCR)-peptide pairs, we developed a framework for functional immunological diversity based on binding and applied it to nearly 400 high-throughput antibody and TCR repertoires to reveal patterns in immunological memory, infection, vaccination, and aging. We show that functional diversity adds information that is not captured by raw diversity, revealing signatures of e.g. clonal selection, and that unlike raw diversity, functional diversity is a robust measure that does not require correction for sampling error. Finally, we show that according to functional diversity, unlike raw diversity, individuals’ repertoires overlap substantially, indicating a definable ceiling for the functional diversity of human adaptive immunity. Similarity redefines diversity in complex systems.

Repertoire-Based Diagnostics Using Statistical Biophysics

“…As proof of principle, we fit descriptor models on 31 before-and-after IgG+ repertoires (including three replicates) from 14 healthy human volunteers who were administered a seasonal trivalent influenza vaccine (Vollmers et al, 2013). We had previously shown that vaccination leads to prominent changes in both repertoires’ raw and functional diversity (Arora et al, 2018), but withheld diversity measurements from the present study in order to test the discriminatory power of the models in the absence of that additional information. Using stratified 3-fold cross-validation, we found that descriptor models distinguished between pre- and post-vaccination pairs with median AUC of 0.97 ( p =4×10 -3 ; Fig.…”

“…We note that vaccination, like many immunological perturbations, results in systems-as well as sequence-level changes; for example, changes in immunological/repertoire diversity (Jiang et al, 2013; Vollmers et al, 2013). We pre viously showed that the combination of raw and functional diversity, measured with various frequency weightings, can discriminate between pre- and post-vaccination sample pairs with high accuracy, likely in part by detecting clonal expansion with selection (Arora et al, 2018). However, changes in diversity, while potentially useful as part of a screening test, are not sufficiently specific to serve as a general diagnostic modality.…”

“…IgG (Vollmers et al, 2013), memory IgH (DeWitt et al, 2016), and TCRβ (Emerson et al, 2017) CDR3 repertoires were obtained and processed as previously described (Arora et al, 2018). For each dataset in Tests 1 and 2, 500,000 sequences were chosen at random and split 90:10 into training and test sets; for Test 3 all sequences were used.…”

“…We fit descriptor models on each of the day 0/day 7 before-and-after IgG + repertoire pairs ( n =31: n =17 from day 0, including replicates, and n =14 day 7) from the influenza vaccination dataset (Vollmers et al, 2013; Arora et al, 2018) and used a support-vector-machine (SVM) on the final models for classification (excluding length biases, which interact with the normalization constant), using the median area under the receiver-operator-characteristic curve (AUC/ROC) as the quality measure (taken over n =10,000 repeats; mean preferred over median given the observed highly skew AUC distributions expected from strong performance with outliers; Fig. 4 top inset), with stratified k -fold cross-validation (without oversampling; 17 vs. 14 was considered sufficiently balanced, but see null-model comparison below) to avoid overfitting (for k =2, 3, 5, and 10 to confirm robustness) and comparison to the AUC of randomly relabeled data as a null model (also n =10,000 repeats) to assess statistical significance.…”

Private Antibody Repertoires Are Public

“…Recently we introduced a computational method for investigating repertoire-wide antigen binding, using thousands of K d measurements for amino-acid substitutions to establish and validate a mean-behavior model for structural similarity among all pairs of antibodies in a repertoire directly from repertoire sequence. 9 That work revealed that repertoires are organized into loose overlapping classes of antibodies with similar predicted functionality, and suggested that these classes may be widely shared among repertoires from across individuals, notwithstanding low sequence overlap between repertoires.…”