Optimization of therapeutic antibodies for reduced self-association and non-specific binding via interpretable machine learning

Abstract

Antibody development, delivery, and efficacy are influenced by antibody-antigen affinity interactions, off-target interactions that reduce antibody bioavailability and pharmacokinetics, and repulsive self-interactions that increase the stability of concentrated antibody formulations and reduce their corresponding viscosity. Yet identifying antibody variants with optimal combinations of these three types of interactions is challenging. Here we show that interpretable machine-learning classifiers, leveraging antibody structural features descriptive of their variable regions and trained on experimental data for a panel of 80 clinical-stage monoclonal antibodies, can identify antibodies with optimal combinations of low off-target binding in a common physiological-solution condition and low self-association in a common antibody-formulation condition. For three clinical-stage antibodies with suboptimal combinations of off-target binding and self-association, the classifiers predicted variable-region mutations that optimized non-affinity interactions while maintaining high-affinity antibody-antigen interactions. Interpretable machine-learning models may facilitate the optimization of antibody candidates for therapeutic applications.

Figure 1.. Approach for predicting non-affinity interactions for clinical-stage antibodies.

Experimental measurements of non-affinity interactions (self-association and non-specific binding) were acquired for 80 clinical-stage antibodies. Structural modeling and Gaussian naïve Bayes classification were used to develop accurate and interpretable models to predict the level of each type of non-affinity antibody interaction. The interpretable models were used to identify mutations that reduce non-affinity interactions for suboptimal therapeutic antibodies while maintaining high affinity.

Figure 2.. Experimental classification of clinical-stage antibodies with different levels of self-association and non-specific binding.

(A) Experimental measurements of self-association in a typical antibody formulation condition (pH 6, 10 mM histidine) using charge-stabilized self-interaction nanoparticle spectroscopy (CS-SINS) and non-specific binding in a physiological condition (pH 7.4, PBS) using the polyspecificity particle (PSP) assay. High levels of each type of non-affinity interaction are indicated by cutoffs established previously.^, (B) The number of approved antibody drugs in each class of behavior. Of the 80 clinical-stage antibodies in this our study, 26 were approved drugs as of July 2022. (C) The percentage of approved drugs formulated without ionic additives, such as sodium chloride and arginine. (D-G) The (D) linear clearance rates in humans, (E) sequence-based isoelectric points of the Fv regions, (F) discovery platform and (G) most common germlines of the clinical-stage antibodies. In (A), the measurements are averages of three replicates and the error bars are standard deviations. In (D), independent two-sided t-tests were performed to determine significance and the p-values are < 0.05 (*), < 0.01 (**) and < 0.001 (***). In (F-G), the marker size corresponds to the number of samples in each category and the color corresponds to the fraction of each class (Class I-IV).

Figure 3.. Interpretable models for predicting the levels of self-association and non-specific binding for clinical-stage antibodies.

(A-B) Gaussian naïve Bayes classification models were trained using a set of 80 measurements of antibody self-association (pH 6, 10 mM histidine) and non-specific binding (pH 7.4, PBS) using standard cross-validation methods. The models for predicting (A) self-association and (B) non-specific binding each contain three features based on homology models of the Fv regions. In (A-B), the data points are colored according to the experimental classification based on previously reported cutoff values.^, Model prediction probabilities are depicted via the background color of the graph, with the black decision boundary plotted where there is an equal likelihood of classification as either class. Model features are normalized between 0 and 1 for optimized model fitting. The range of x-axis values for each feature are given.

Figure 4.. Prediction of mutations that co-optimize affinity and non-affinity interactions for three clinical-stage antibodies.

(A) The selected sub-optimal antibodies (Class II: panitumumab; Class III: gantenerumab; and Class IV: cinpanemab) were selected for co-optimization. (B) Crystal structures were used to identify all residues in the paratope (< 5 Å, green) in each Fv region. Sites in the Fv region that were identified for mutation based on several criteria are shown in red. (C, D) Predicted impacts of multi-site mutations on the levels of (C) self-association and (B) non-specific binding of each clinical-stage antibody. Parental antibodies are depicted as diamonds and mutants are shown as circles.

Figure 5.. Predicted mutations in clinical-stage antibodies display co-optimized levels of affinity and non-affinity interactions.

(A) Antibody variants display reduced self-association and/or non-specific binding. Most (12 out of 17) of the predicted mutants display optimal combinations of non-affinity interactions (Class I mutants, indicated by gold stars). (B) Most antibody variants also maintain high affinity despite that the mutations are in the CDRs and closely neighboring framework regions. In (A) and (B), Pwt is wild-type panitumumab and P1-P7 are multi-site mutants of panitumumab, GWT is wild-type gantenerumab and G1-G4 are multi-site mutants of gantenerumab, and CWT is wild-type cinpanemab and C1-C7 are multi-site mutants of cinpanemab. In (A), the measurements are averages of three replicates. In (B), the measurements are averages of two replicates, the reported EC50 values are averages, and the errors are standard deviations.