Development of an mPBPK machine learning framework for early target pharmacology assessment of biotherapeutics

Abstract

Development of antibodies often begins with the assessment and optimization of their physicochemical properties, and their efficient engagement with the target of interest. Decisions at the early optimization stage are critical for the success of the drug candidate but are constrained due to the limited knowledge of the antibody and target pharmacology. In the present work, we propose a machine learning-based target pharmacology assessment framework that utilizes minimal physiologically based pharmacokinetic (mPBPK) modeling and machine learning (ML) to infer optimal physicochemical properties of antibodies and their targets. We use a mPBPK model previously developed by our group that incorporates a multivariate quantitative relationship between antibodies' physicochemical properties such as molecular weight (MW), size, charge, and in silico + in vitro derived descriptors with their PK properties. In this study, we perform a high-throughput exploration of virtual antibody drug candidates with varying physicochemical properties (binding affinity, charge, etc.), and virtual target candidates with varying characteristics (baseline expression, half-life, etc.) to unravel rules for antibody drug candidate selection that achieve favorable drug-target interaction, which is defined by target occupancy (TO) percentage. We identified that variations in the antibody dose and dosing scheme, target form (soluble or membrane-bound), antibody charge, and site of action had a significant effect on the TO and selection criteria for antibody drug candidates. By unraveling new design rules for antibody drug properties that are dependent on ML-based TO assessment, we deliver a first-in-class ML-based target pharmacology assessment framework toward better understanding of the biology-specific PK and ADME processes of antibody drug candidate proteins and reduce the overall time for drug development.

Keywords: Target pharmacology, Antibodies, Pharmacokinetics, High-throughput ML, Decision tree classification..

Fig. 1

Representation of steps involved in the development of the minimal physiologically based pharmacokinetic / Machine learning modeling framework. mAb: monoclonal antibody, Ag: Antigen or target, MW: Molecular weight, Q: antibody charge, KD: Binding constant, t_1/2: target half-life, T₀: target baseline, TO: target occupancy, C_min: minimum plasma concentration, AUC_ss: area under the curve at steady state, C_max: maximum plasma concentration, and C_ss: plasma concentration at steady state.

Fig. 2

Representation of results from a specific scenario. Antibody is administered at 1 mg/kg through intravenous (IV) route once every two weeks. Target Occupancy percentage (TO %) is measured in plasma for drug and soluble target engagement. The pairwise scatter plot (top panel) and pairwise Kernel density plot (bottom panel) shows the pairwise distribution of properties (baseline, binding constant, and half-life). Each property is classified into optimal (blue) and non-optimal class (orange) based on TO% calculated at minimum concentration at steady state.

Fig. 3

The effect of different doses (A, B) and dosing schemes (C, D) on drug-target properties for optimal target engagement. (A, B): The tree-based ML model was trained and validated for synthetic data based on different doses of 0.1, 1, and 10 mg/kg administered once every 2 week (Q2W). (C, D): The tree-based ML model was trained / validated for synthetic data based on a fixed dose of 1 mg/kg administered once every week (Q1W), once every two weeks (Q2W), and once every 4 weeks (Q4W). Only optimal properties dependent on target occupancy > 90% at minimum plasma concentration at steady state are shown.

Fig. 4

The effect of antibody charge on optimal drug and target properties needed for optimal target engagement response. The bar plot represents the maximum cut-off values of binding constant (nM) and target baseline (nM) as predicted by the tree-based classifier needed to achieve TO% > 90% in plasma. Each antibody has a net surface charge of + 5, 0, and -5, and is administered at 1 mg/kg once every 2 weeks.

Fig. 5

The effect of different forms of target, soluble (green) or membrane-bound (orange). The decision tree-based boundaries for target properties (half-life and baseline) needed for optimal target occupancy (TO% > 90%) for each target form are shown. The fitted dashed black line (y = 0.324x) separates the optimal properties needed for soluble and membrane-bound receptors that achieve 90% target occupancy.

Fig. 6

The ML-derived rules for optimal properties for different sites of action (plasma, leaky tissues, tight tissues). The distribution of drug properties (binding constant, (nM)) and target properties (half-life, , baseline, (nM)) is provided as a pairwise kernel density plot (A, B). The ML-derived rules are quantified as optimal cut-off values for binding constant and target baseline for each site of action (C).

Fig. 7

The ML-derived optimal (TO% > 90%) properties for dose of 0.1 mg/kg bolus (pink region), 1 mg/kg Q4W (IV) (blue region), 1 mg/kg Q2W (orange region), and 10 mg/kg Q2W (green region) are shown. Experimentally derived properties of clinically approved monoclonal antibodies for the respective dose regimen are shown as data points ^–. The antibody data is color-coded based on dosing regimen.

Fig. 8

Multi-label classification of drug-target properties. Pairwise scatterplot of drug and target properties classified based on target occupancy % (TO%) <  = 50% (orange), 50–90% (green), and > 90% (purple).