Human antibody polyreactivity is governed primarily by the heavy-chain complementarity-determining regions

Abstract

Although antibody variable regions mediate antigen-specific binding, they can also mediate non-specific interactions with non-cognate antigens, impacting diverse immunological processes and the efficacy, safety, and half-life of antibody therapeutics. To understand the molecular basis of antibody non-specificity, we sorted two dissimilar human naïve antibody libraries against multiple reagents to enrich for variants with different levels of polyreactivity. Sequence analysis of >300,000 paired antibody variable regions revealed that the heavy chain primarily mediates human antibody polyreactivity, and this is due to the high positive charge, high hydrophobicity, and combinations thereof in the corresponding complementarity-determining regions, which can be predicted using a machine learning model developed in this work. Notably, a subset of the most important features governing antibody non-specific interactions, namely those that contain tyrosine, also govern specific antigen recognition. Our findings are broadly relevant for understanding fundamental aspects of antibody molecular recognition and the applied aspects of antibody-drug design.

Keywords: CP: Immunology; antibody engineering; deep sequencing; developability; machine learning; non-specific binding; nonspecificity; off-target binding; pharmacokinetic; polyspecificity; repertoire.

Figure 1.. Overview of the library sorting, deep sequencing, and model training methods used to generate machine learning models for predicting human antibody polyreactivity

(A) Two human single-chain variable fragment (scFv) libraries displayed on the surface of yeast were sorted for positive or negative binding to multiple poly-specificity reagents. The enriched libraries were deep sequenced to generate large datasets of antibody sequences and corresponding classifications for either high or low polyspecificity. One of the large human scFv datasets was used along with a smaller dataset for clinical-stage antibodies to train machine learning models to predict antibody polyspecificity. Finally, the models were tested on the second large human scFv dataset (not used for training) and additional independent datasets for preclinical and clinical-stage antibodies. (B) The human scFv library (library #1)21 was displayed on the surface of yeast and sorted successively against ovalbumin (0.13 mg/mL [2.9 μM]; fluorescence-activated cell sorting [FACS] sort #1), soluble cytosolic proteins (SCPs) from CHO cells (0.13 mg/mL; FACS sort #2), soluble membrane proteins (SMPs) from CHO cells (0.13 mg/mL; FACS sort #3), and insulin (0.13 mg/mL [22.4 μM]; FACS sort #4). FACS cytograms are shown for FACS sort #5, which was performed using the output libraries from FACS sort #4 to collect samples for deep sequencing analysis. The FACS cytograms report the antibody (scFv) expression on the x axis and non-specific binding on the y axis. The positive and negative non-specific binding populations that were selected for deep sequencing analysis are shown in red (high non-specific binding) and blue (low non-specific binding) gates. (C) Individual scFv variants were isolated after FACS sort #4, and their levels of non-specific binding were evaluated using yeast surface display and flow cytometry. The four polyspecificity reagents were used as described in (B) except at a lower concentration (0.026 mg/mL). The reported levels of non-specific binding were first normalized to their scFv expression levels and then normalized between two scFv standards (FN4 for the negative control and FP3 for positive control; Dataset S3). In (C), the data are averages of two biological replicates, and the error bars are standard errors.

Figure 2.. Germline family and sequence comparisons for antibody libraries #1 and #2 after enrichment for high and low levels of poly-reactivity

(A–D) Distribution of germline families for antibody (A and B) library #1 and (C and D) library #2. (E) Analysis of sequence differences between the two libraries. The sequences were OneHot encoded, subjected to dimensionality reduction via a truncated singular value decomposition, and embedded into two-dimensional space for visualization with t-distributed stochastic neighbor embedding (t-SNE).

Figure 3.. Molecular features that strongly differentiate between human antibodies with high and low polyreactivity

Distributions of key molecular features linked to polyreactivity for human antibodies (libraries #1 and #2) and their corresponding area under the ROC curve (AUC) values calculated using logistic regression analysis. The same features for the input antibody libraries and a human repertoire dataset were also calculated. (A and B) Fv charge (pH 7.4) distribution for (A) library #1 and (B) library #2. (C and D) Distributions of the number of tryptophan, arginine, and lysine residues in Fv for (C) library #1 and (D) library #2. (E and F) Distributions of the number of phenylalanine, isoleucine, proline, tryptophan, and tyrosine residues in Fv for (E) library #1 and (F) library #2. In (A) and (B), the net charge (pH 7.4) was calculated using charges of +1 for Arg and Lys, +0.1 for His, and −1 for Asp and Glu.

Figure 4.. Charge and hydrophobicity features based on antibody regions that include the heavy-chain CDRs strongly differentiate between antibodies with high and low levels of polyreactivity

Distributions of key molecular features for different antibody regions in Fv linked to polyreactivity for human antibodies (libraries #1 and #2) and their corresponding AUC values. The same features for the input libraries and a human repertoire dataset were also calculated. (A and B) Net charge (pH 7.4) distributions for (A) library #1 and (B) library #2. (C and D) Distributions of the number of tryptophan, arginine, and lysine residues for (C) library #1 and (D) library #2. The antibody regions are noted as V_H (variable heavy domain), V_L (variable light domain), CDRs (six CDRs for the heavy and light chains), HCDRs (three heavy-chain CDRs), and LCDRs (three light-chain CDRs).

Figure 5.. A subset of molecular features that differentiate between high- and low-poly-reactivity antibodies also differentiate between paratope and non-paratope residues

Distributions of molecular features for paratope and non-paratope residues in V_H and V_L for 468 anti-bodies and their corresponding AUC values. (A and B) Net charge (pH 7.4) distributions for (A) V_H and (B) V_L. (C and D) Distributions of the number of isoleucine, tryptophan, and tyrosine residues for (C) V_H and (D) V_L. (E and F) Distributions of the number of tryptophan, arginine, and lysine residues for (E) V_H and (F) V_L. AUC values with a negative sign next to them signify that the feature values are depleted in the antibody paratope or non-paratope residues.

Figure 6.. Evaluation of machine learning model for predicting human antibody polyreactivity

The ability of the machine learning (random forest) model developed in this work to predict the level of human antibody polyreactivity was tested using several sets of antibodies with experimentally defined levels of non-specific binding to multiple reagents. The model predictions were tested on the following antibody sets: (A) antibody set #3 (88 human scFvs), (B) antibody sets #4 and #6 (47 human scFvs), (C) antibody set #4 (20 human clinical-stage IgGs), (D) antibody set #8 (80 clinical-stage IgGs),24 and (E) 15 bococizumab variants with HCDR2 and HCDR3 mutations (antibody set #9). The reported p values were calculated using the Anderson-Darling test. In (A), (B), (D), and (E), the data are mean values, the errors are standard deviations, and the numbers of biological replicates are (A) two, (B) three, (D) three, and (E) two. In (C), the data are from a previous publication, and the number of biological replicates is unknown.