An in silico method to assess antibody fragment polyreactivity

Abstract

Antibodies are essential biological research tools and important therapeutic agents, but some exhibit non-specific binding to off-target proteins and other biomolecules. Such polyreactive antibodies compromise screening pipelines, lead to incorrect and irreproducible experimental results, and are generally intractable for clinical development. Here, we design a set of experiments using a diverse naïve synthetic camelid antibody fragment (nanobody) library to enable machine learning models to accurately assess polyreactivity from protein sequence (AUC > 0.8). Moreover, our models provide quantitative scoring metrics that predict the effect of amino acid substitutions on polyreactivity. We experimentally test our models' performance on three independent nanobody scaffolds, where over 90% of predicted substitutions successfully reduced polyreactivity. Importantly, the models allow us to diminish the polyreactivity of an angiotensin II type I receptor antagonist nanobody, without compromising its functional properties. We provide a companion web-server that offers a straightforward means of predicting polyreactivity and polyreactivity-reducing mutations for any given nanobody sequence.

Fig. 1. Development of a computational tool to assess and mitigate polyreactivity.

Starting from a large, naïve synthetic nanobody library, pools of nanobodies with low and high polyreactivity were isolated. Machine learning models were trained on deep sequencing data from these pools to learn sequence features of low and high polyreactive nanobodies. These algorithms were incorporated into software that quantitatively predicts polyreactivity levels and recommends substitutions that reduce it. Created with BioRender.com.

Fig. 2. Properties of purified nanobodies exhibiting varying degrees of polyreactivity.

a Spodoptera frugiperda (Sf9) insect cell PSR staining of single nanobodies isolated from FACS sorts. Data are mean +/− SEM of three independent biological experiments performed in technical triplicate. Polyreactivity levels are normalized with respect to the highest clone (Nb F02’). b CDR sequences of isolated nanobodies. c Direct ELISA assays measured the apparent EC₅₀ (EC_50APP) of five index set members and nanobody AT118i4h32 to the specified reagents. Non-specific binding, indicated by low EC_50APP values, correlates with strong binding to PSR. ELISA data are representative of two independent experiments, each performed in technical triplicates.

Fig. 3. Development of computational models to predict polyreactivity.

Supervised models were trained on pools of high and low polyreactivity sequences. a Pipeline of computational model development from raw NGS data to held-out predictions with sequence clustering for rigorous validation. b Comparison of supervised models (one-hot and k-mer logistic regression, RNN, CNN) and biochemical properties such as hydrophobicity, isoelectric point, CDR3 lengths, and number of arginine residues. c Trained parameters of a one-hot logistic regression model, showing which amino acids at specific positions are most predictive of high polyreactivity and low polyreactivity (red, negative score and blue, positive score, respectively). d Polyreactivity scores of top motifs learned from a k-mer logistic regression model that are most predictive of low and high polyreactivity (top and bottom, respectively). e Separation of high and low polyreactivity nanobodies by each of the models and biochemical properties displayed in panel b.

Fig. 4. Validation of computational model for quantitative predictions of polyreactivity and generation of rescue mutations.

a Generation of an index set of polyreactivity mutants by Spodoptera frugiperda (Sf9) insect cell membrane protein polyspecificity reagent (PSR) staining of yeast displaying 48 unique nanobodies isolated from MACS enrichment as well as non-reactive and polyreactive FACS pools. Data are mean +/− SEM of three independent biological experiments performed in technical triplicate. b New nanobody sequence(s) can be input into a webserver, which will output computational predictions of polyreactivity and biochemical properties of the sequence(s). It is also possible to input a nanobody sequence to retrieve top scoring rescue mutations predicted to decrease polyreactivity. c, e The one-hot logistic regression model and k-mer logistic regression model trained on the full NGS dataset from FACS sorts with PSR binding were used to test quantitative predictions and rankings of the index set of clones spanning a wide range of polyreactivity levels (as measured by PSR binding) (Spearman ρ of 0.77 and 0.79, respectively). A high score indicates low predicted polyreactivity, whereas a low score indicates increased polyreactivity. d, f An in silico double mutation scan (spanning substitutions, insertions, and deletions) was scored for predicted polyreactivity using both the one-hot logistic regression model and k-mer logistic regression model. From these in silico double mutation scans, a diverse set (spanning each CDR and combinations of CDRs) of high scoring mutations predicted to have low polyreactivity were selected as rescue mutations for experimental testing from two-parent clones, E10’ and D06. Created with BioRender.com.

Fig. 5. In silico designed substitutions reduce nanobody polyreactivity.

a Polyspecificity reagent (PSR) staining of yeast displaying D06 variants. For the moderately polyreactive D06 nanobody, 18 out of 21 variants that were computationally designed to decrease polyreactivity reduced levels of binding to insect cell PSR staining. Data in a comprise the mean +/− SEM of at least three independent experiments, each performed in technical triplicate. b PSR staining of yeast displaying E10’ variants. For the highly polyreactive E10’ nanobody, 15 out of 16 computationally predicted single and double substitutions reduced binding to PSR reagent. Data in b comprise the mean +/− SEM of at least three independent experiments, each performed in technical triplicate. Substitutions to CDRs 1, 2, and 3 are colored in blue, green, and orange.

Fig. 6. Development of AT118i4h32 variants with reduced polyspecificity.

a Electrostatic surface of AT118i4h32. CDR1, CDR2, and CDR3 are colored blue, green, and orange. All positions substituted to produce variants of AT118i4h32 with reduced polyreactivity are shown in sticks with atomic coloring b AT118i4h32 structure as colored in a. G26D²⁷ and T57I⁶⁵ substitutions are boxed. c PSR staining of yeast displaying AT118i4h32 variants. All amino acid substitutions decrease polyreactivity. Data in c comprise the mean +/− SEM of four independent experiments, each performed in technical triplicate. CDRs are colored as in a. d Binding of AT118i4h32 variants to HEK293 suspension cells expressing FLAG-AT1R. Cells were stained with AT118i4h32-V5-His variants, AlexaFlour-488 conjugated anti-FLAG, and AlexaFlour-647 conjugated anti-V5 antibodies, then analyzed by flow cytometry. Data in d is the average of three independent experiments performed in technical triplicate, error bars are shown as SEM. e Radioligand competition binding of AT118i4h32 variants or the small molecule antagonist losartan and [³H]-olmesartan to AT1R in cell membranes. Like WT AT118i4h32, the G26D²⁷, T57I⁶⁵, and G26D²⁷ T57I⁶⁵ variants compete with olmesartan for binding to the AT1R. Data in e is the average of three independent experiments performed in technical triplicate, error bars are shown as SEM. f Suppression of Gq-mediated inositol monophosphate production by AT118i4h32 in response to AngII stimulation. HEK293 suspension cells expressing FLAG-AT1R were treated with 5 μM AT118i4h32 or no nanobody prior to AngII stimulation. Data in d is the average of three independent experiments performed in technical triplicate, error bars are shown as SEM. K_i values are reported in Supplementary Table 6.