Development of a Minimal Physiologically-Based Pharmacokinetic Modeling / Machine Learning Framework for Early Target Pharmacology Assessment

doi:10.21203/rs.3.rs-4421530/v1

Download PDF

Article

Development of a Minimal Physiologically-Based Pharmacokinetic Modeling / Machine Learning Framework for Early Target Pharmacology Assessment

https://doi.org/10.21203/rs.3.rs-4421530/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Development of antibodies often begins with the assessment and optimizing of their physicochemical properties, and their efficient engagement to 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. n the present work we propose a model-based target pharmacology assessment framework based on which optimal physicochemical properties of antibodies can be inferred from minimal physiologically based pharmacokinetic (mPBPK) modeling and machine learning (ML). Towards this goal, we aim to perform a high-throughput virtual exploration of physicochemical properties of antibody drug candidates and relate them to target occupancy (TO). 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 a known relation to PK properties. In this study, we perform an exploration of virtual antibody drug candidates with varying physicochemical properties, and virtual target candidates with varying characteristics to unravel rules for optimal antibody drug candidates and feasible drug-target interaction. We also identify that varying the antibody dose and dosing scheme, target form (soluble or membrane-bound), antibody charge, and site of action had significant effect on the optimal properties for antibody drug candidate selection. By unravelling new design rules for antibody drug properties that are dependent on model-based TO assessment, we deliver a first-in-class model-based framework towards better understanding of the biology-specific PK and ADME processes of antibody drug candidates proteins and reducing the overall time for drug development.

Target pharmacology

Antibodies

Pharmacokinetics

High-throughput ML

Decision tree classification

The discovery and development of first-in-class therapeutic antibodies often begins with the assessment of the physicochemical properties of the antibodies and their efficient engagement with a new target. Once a therapeutic target is selected, rational decisions must be made for lead optimization and candidate selection in the early drug development, which can decide the success and failure in clinical stages [1]. The failures in drug development process can often be attributed to insufficient specificity of a drug candidate towards its target. It is essential to perform a target pharmacology assessment of an antibody candidate at an early stage in this process. During the evaluation of a drug-target interaction, suitability for drug-target binding, feasibility, pharmacokinetics, pharmacodynamics, and safety are some of the factors that are important in understanding the drug’s pharmacology [2, 3]. It is relevant to understand the magnitude and duration of target engagement to enable selection of appropriate drug candidates and identification of optimal drug-target pairs. To obtain a robust target engagement, optimization of multiple properties like antibody binding affinity, targeted human efficacious dose, and dosing regimen, is required to select the best candidate with desired biological activity [4]. The binding affinity of a therapeutic antibody candidate to its target is one of the aspects that governs the target pharmacological effect [5]. The affinity and pharmacological activity are also interrelated to the physicochemical properties of the antibodies. Antibody engineering allows controlling these properties early in the drug development process. Often, the set of desired properties related to an antibody candidate required for an optimal target engagement and pharmacological response are not clearly known in the early stages of drug discovery and development. To identify desired properties, extensive in vivo pharmacology studies for each drug candidate may be required, which can be highly resource-intensive [6]. Computational model-based exploration of these engineered parameters before comprehensive experimental evaluation could provide useful early insights into desired properties and shorten lead times.

Drug-target interactions not only depend on drug and target properties but also on physiological attributes of a species [1]. It is essential to study the drug’s physicochemical (PC) and pharmacokinetic (PK) as well as the influence of non-specific interactions of the antibody within the targeted site of action to understand desired target engagement response [7, 8, 9, 10]. For instance, net surface charge and isoelectric point of antibodies leads to unintended charge-based non-specific interactions with cell components, which affects their clearance and tissue distribution, and ultimately their PK [7, 10, 9, 11]. This may have adverse effect on desired target engagement response. Many research studies have made it evident that an integrated approach to understand the relative contribution of both PC and physiology on ADME/PK properties is critical to rationally drive and inform engineering strategies to achieve desired antibody drug-target response [12, 13, 1, 4, 14, 15]. Often physiologically-based pharmacokinetic models (PBPK) and pharmacodynamic (PD) models are considered in predicting target engagement response and improve confidence in early decisions [12]. Although such models are capable of early decision making, integrating them with both in vivo pharmacology insights and high-throughput machine learning analysis can be promising and advantageous for the early drug discovery and development stages [6]. In the past, Chen et al. proposed a model-based target pharmacology assessment (mTPA) to obtain optimal combination of PK, potency, and ADME of small molecule drugs by combining PBPK/PD/ML modeling approach [6, 16]. Makowski et al. identified optimal antibody combinations of low off-target binding and low self-association using an interpretable ML approach and in vitro measurements but lacked its association to PK/TO endpoints [15]. There is still a need for an integrated modeling framework in the antibody drug research space. Such a framework can assess antibody pharmacological activity based on physiology and physiochemical properties of both drug and target to optimize antibody development as well as target engagement criteria.

In this present work, we provide a framework for model-based assessment of target engagement response of antibody drug candidates by integrating a minimal PBPK (mPBPK) model with a decision tree-based machine learning model. We generate a large number of virtual drug-target candidate pairs with varying properties using the mPBPK model. These virtual drug and target properties are categorized into optimal or non-optimal spaces according to a desired target occupancy percentage. An interpretable decision tree-based algorithm was applied for high-throughput screening of virtual drug-target properties such as binding affinity, baseline concentration, target half-life, dose, dosing scheme, and antibody charge to identify the combination of properties that are most likely to demonstrate the desired target engagement response. To the best of our knowledge, such a model-based target pharmacology assessment has not been performed before for antibodies. It can serve as a useful platform for high-throughput assessment of PK/TO and PD response of antibodies in the future.

Minimal PBPK Model for Monoclonal Antibodies

To study the relationship between antibody properties and ADME, PBPK models are often the models of choice since they provide quantitative descriptions of the drug disposition process in a biological system that can be scaled between species based on physiological differences [12]. The mPBPK model incorporates quantitative descriptions between physicochemical properties of antibodies such as MW, size, charge, binding affinities to FcRn, and specific targets and essential ADME processes involved in the antibody PK [11].The structure of the mPBPK model includes a plasma compartment, two lumped tissue compartments, mainly tight tissues and leaky tissues, and a lymph compartment. Briefly, the mPBPK model takes in the physiology-based parameters (eg. organ volumes, tissue lymph flows, blood flows), antibody-specific properties (eg., MW, size, charge, binding affinity), and target properties (eg. baseline expression, half-life) (Figure S1, supplementary file). More detailed information on the model kinetics and assumptions is provided in the supplementary file. The mPBPK model predicts the antibody PK profile, PK endpoints and TO%. TO% is calculated as the ratio of bound target over the total target in plasma, tight tissues, and leaky tissues (Equation S1, supplementary file). We calculated TO% at minimum and maximum concentration of drug in plasma for each scenario (supplementary file, Table S1). Besides TO%, we also calculated the PK and exposure endpoints, mainly the area under the curve at steady state (\(\text{A}\text{U}{\text{C}}_{\text{s}\text{s}}\)), minimum plasma concentration (\({ \text{C}}_{\text{m}\text{i}\text{n}}\)), maximum plasma concentration (\({\text{C}}_{\text{m}\text{a}\text{x}}\)), total soluble and membrane target suppressed. The mPBPK simulated scenarios involves three different doses, 0.1, 1, and 10 mg/kg, three different dosing regimens (once per week, once per 2 weeks, once per 4 weeks), and three different antibody charge variations (+ 5, 0, and − 5). The drug is administered intravenously at a fixed dose in the plasma compartment. We assume hypothetical variation in the net surface charge on an antibody. Our mPBPK model can relate the variable region (Fv) charge on an antibody to clearance, distribution, and non-specific cellular interaction [11]. These hypothetical charge variants are representative of charge variations in the variable domain, which is essential for antibody engineering [10]. Our decision tree-based classifier shows the influence of antibody charge on rule-based selection of drug and target properties needed for optimal target engagement response. We provide more details on scenarios used to simulate the mPBPK model and generate data in this study at the supplementary material. (Table S1, supplementary file).

Virtual Data Generation

We used a log-normalized uniform sampling method (scipy.stats.loguniform) to generate 10000 virtual candidates to perform the model-based assessment. To account for a wide range of possible target characteristics, the virtual target candidates were generated with varying target baseline values (1 pM – 1000 nM) and target half-life (1 min – 300 hour) for both soluble and membrane-bound receptors. Similarly, the virtual drug candidates were generated with varying drug-target binding affinity (1 pM – 1000 nM). Each virtual combination of drug and target candidate denotes a unique drug-target interaction. In our case, 10000 virtual candidates were considered appropriate based on comparison of decision tree-based decision rules and decision tree model performance (Table S4, supplementary file). The generated data was categorized into two classes based on TO% > 90% or < 90% at \({\text{C}}_{\text{m}\text{i}\text{n}}\) in plasma at steady state. The class distribution in the virtual data was highly imbalanced. Therefore, we applied oversampling of the minority class in our imbalanced classification dataset using Synthetic Minority Oversampling Technique (SMOTE) as part of the imbalanced-learn package in Python [17]. SMOTE works by selecting examples that are close in the feature space, drawing a line between the examples in the feature space and drawing a new sample at a point along that line [17]. This procedure was used to create as many synthetic examples for the minority class as observed in the majority class. After balancing the class for each dataset, we obtained a dataset with total size slightly over 10000 data points for each scenario. For instance, if a dataset contained 5452 data points in the majority class, SMOTE oversamples in the minority class to get 5452 data points in the minority class, which results in a total size of 10904.

Decision Tree-Based Supervised Machine Learning

A supervised classification model was applied to categorize the virtual candidate data. An interpretable decision tree-based algorithm was trained and evaluated using the virtual data for each scenario (Table S1, supplementary file). Each drug-target candidate pair was classified based on TO% calculated at \({\text{C}}_{\text{m}\text{i}\text{n}}\) into an optimal class (\(>\)90%) or non-optimal class (\(\le\)90%). We used 90% of the dataset for training the classifier and reserved 10% of the dataset for testing. A subset of training data used to train the ML model are shown as example in the supplementary file, Table S5. We used decision-tree based classifier as the algorithm for the following reasons. We compared different tree-based algorithms in the scikit-learn package in Python, namely decision trees, random forest, and gradient boosting algorithm. We compared the mean training accuracy, precision, and F1 score for 5-fold cross-validation across the three models for a fixed dose of 0.1 mg/kg (supplementary file, Table S6). We also compared these metrics on the test dataset across each model (supplementary file, Table S6). Decision trees and gradient boosting classifiers had an overall better performance compared to random forest classifiers. However, decision tree classifiers were computationally less expensive relative to other classifiers. Therefore, decision tree-based classifiers were used for different scenarios throughout this study. However, this model performance analyses may be different for other scenarios and datasets.

We used a Decision tree classifier (decisiontreeclassifier) from sklearn package in Python 3 using default hyperparameters, except max_depth = 5, criterion = ‘gini’, min_samples_split = 3, splitter = ‘best’, class_weight = ‘balanced’, which were obtained from a hyperparameter search using gridsearchCV. We trained the decision tree classifier using a stratified 5-fold cross-validation. Additionally, we also performed a three-label classification using decision trees on virtual dataset, which was generated for a fixed dose of 1 mg/kg Q2W (IV). We classified the decision tree-based rules for drug-target properties for a low (\(\le\)50%), medium (50% − 90%), and high TO% (\(>\)90%).

mPBPK / ML Model Framework

We built a model-based target pharmacology assessment framework that combines a minimal physiologically based pharmacokinetic (mPBPK) model, a target-mediated drug disposition (TMDD) model, and a machine learning (ML) model to infer the optimal drug and target characteristics responsible for desired target engagement and occupancy. Our model-based framework was implemented to study the relationship between physicochemical properties of the antibodies like charge and drug-target binding (\({\text{K}}_{\text{D}}\)), ADME characteristics, and target properties like baseline (\({\text{T}}_{0}\)), half-life (\({\text{t}}_{1/2}\)), and soluble or membrane-bound forms of the target. Figure 1 shows the steps involved in the model development process. Firstly, we generated many virtual candidate pairs of antibody drugs and their specific targets. Each virtual candidate signifies a unique drug-target interaction. A virtual drug candidate was generated by varying dose level, dose regimen, surface charge, and target binding affinity. A virtual target candidate was generated by varying target baseline expression and target turnover or half-life. Secondly, we used the mPBPK model (Figure S2, supplementary file) to simulate the PK profile and PK/exposure endpoints (\(\text{A}\text{U}{\text{C}}_{\text{s}\text{s}}, {\text{C}}_{\text{m}\text{i}\text{n}}, {\text{C}}_{\text{m}\text{a}\text{x}} \text{e}\text{t}\text{c}.\)) for each virtual candidate [11]. We assess the magnitude of target engagement response for each virtual candidate by calculating the target occupancy percentage (TO%). A decision tree-based machine learning algorithm is used to classify each virtual candidate (or drug-target interaction) into an optimal or non-optimal interaction based on the calculated TO%. The ML classifier helps to infer the optimal combination of drug and target properties that is more likely to provide the desired TO endpoint and eventually desired pharmacological effect.

We used the virtual data to train the ML classifier for different dose variations, dosing frequencies, antibody charge variations, and across different forms of target (soluble or membrane-bound). A complete list of scenarios considered to generate data for ML model development is provided in Table S1 the supplementary file. In Fig. 2, we provide the results from the decision tree-based model using 10000 virtual drug-target candidates, where the antibody was administered via intravenous (IV) route at 1 mg/kg once every 2 weeks (Q2W) and bound to only soluble forms of target. The decision tree model uses appropriate cut-off value for each input feature (\({\text{K}}_{\text{D}}, {\text{T}}_{\text{s}0}, {\text{t}}_{1/2\text{s}}\)) to classify each virtual antibody drug-target candidate into an optimal (blue) or non-optimal (orange) class based on the TO% observed at minimum concentration in plasma at steady state. The decision tree-based model suggests that for a drug administered at 1 mg/kg once every two weeks should have greater than 90% target engagement when the binding constant (\({\text{K}}_{\text{D}}\)) is below 9 nM for a class of soluble targets that have baseline values below 4 nM in plasma (Fig. 2). The decision tree-based rules for optimal class prediction were chosen based on Gini impurity close to 0 and accuracy close to 100%. It is important to choose the best rule with the highest accuracy as it increases the likelihood for true prediction for other unseen datasets. \({\text{K}}_{\text{D}}\) and \({\text{T}}_{\text{s}0}\) are the most important properties to predict more than 90% target occupancy as a representative target criteria. \({\text{t}}_{1/2\text{s}}\) parameter seemed to have the least influence in decision making in this scenario. However, class of soluble receptors with low half-life and high baseline may be the least ideal combination of target properties to achieve greater than 90% target occupancy in plasma (Fig. 2). The binary classification decision tree for the above scenario is provided in Figure S2 in the supplementary file. Each decision tree-based model was evaluated on a separate validation data set based on the following evaluation metrics such as classification metric, confusion matrix, and area under the receiver-operating curve (ROC) (Figure S2, supplementary file). The set of optimal and non-optimal rules used in decision tree-based binary classification for this scenario are also provided in the supplementary file.

Optimal Drug-Target Properties for Target Engagement for Different Dose and Regimen

Figure 3 shows the difference in optimal rules and values of drug-target properties for different doses and dosing schemes. The decision tree-based ML model provided an optimal region for drug-target binding constant (\({\text{K}}_{\text{D}}\)), target baseline, and target half-life values for bolus and repeated IV dose of 0.1, 1, and 10 mg/kg (Fig. 3). The optimal rule for low dose (0.1 mg/kg) scenario suggests a \({\text{K}}_{\text{D}}\) value below 1 nM. A drug administered at such low dose may achieve 90% target occupancy with class of soluble receptors with baseline values below 2 nM (Fig. 3). As dose is increased from 0.1 mg/kg to 1 mg/kg, the cut-off for optimal \({\text{K}}_{\text{D}}\) value increases to 9 nM as well as target baseline values increases to 4 nM (Fig. 3). Similarly, at high dose of 10 mg/kg, the optimal cut-off for \({\text{K}}_{\text{D}}\) increased to 82 nM. At lower administered dose the drug clears faster and undergoes target-mediated disposition. Therefore, to achieve greater than 90% target occupancy at such a low dose, drug-target affinity should be higher (\({\text{K}}_{\text{D}}\)values should be lower). The optimal rules suggested that \({\text{K}}_{\text{D}}\) and target baseline were relatively more important to classify a given drug-target pair based on target engagement response. It is known that target baseline concentration and target half-life are relevant target properties that govern drug-target interaction and desired target pharmacology [1, 13]. However, to achieve more than 90% target occupancy at different doses, target half-life was the least influential in the decision-making process. On the other hand, we also observed the effect of dosing scheme on optimal TO% when drug is administered at a repeated dose of 1 mg/kg once every 1 week (Q1W), once every 2 weeks (Q2W), and once every 4 weeks (Q4W) (Fig. 3). The decision tree-based rules deemed that as dosing frequency increases, the optimal range for \({\text{K}}_{\text{D}}\), \({\text{T}}_{\text{s}0},\)and \({\text{t}}_{1/2\text{s}}\) increases. The effect of dosing frequency on optimal drug and target properties can be observed in the density plot, Fig. 3. For Q4W, Q2W, and Q1W dosing scheme, \({\text{K}}_{\text{D}}\) values below 4.25, 9, and 15. 4 nM will provide greater than 90% target occupancy for a given class of soluble receptors with baseline values below 2, 4, and 8.3 nM, respectively. Target half-life has the least dependence on dosing scheme and was insufficient in solely determining the optimal class.

Effect of Antibody Charge on Optimal Drug-Target Engagement

The net surface charge, charge distribution, and isoelectric point (pI) of an antibody affect nonspecific cellular uptake and degradation [10]. The surface charge of a therapeutic protein is a property of the amino acid sequence of the protein and the pH of its surroundings [18]. The net surface charge leads to non-specific interactions with the charged extracellular matrix components and membrane proteins in the cells, resulting in enhanced pinocytotic uptake and degradation, suggesting that charge can be a relevant descriptor of antibody ADME processes [19]. Hence, charge or isoelectric point is a relevant and inherent property of an antibody which can affect its specificity and target engagement response. We incorporated the effect of charge in the decision tree classifier and observed a significant contribution on optimal (> 90%) target occupancy (Fig. 4). Figure 4 quantifies the best ML-derived rules. \({\text{K}}_{\text{D}}\) values should be lower than 2.1 nM for when the targeted class of soluble receptors have a baseline below 4 nM approximately for relatively positively charged antibody. Similarly, cut-off for \({\text{K}}_{\text{D}}\) values increased with more negative charge on the antibody. This agrees with our previous findings from the mPBPK model, where positive charge on an antibody had significant effect on clearance and distribution [11]. A positively charged antibody with faster clearance has relatively lower exposure in plasma and restricts the amount and duration of engagement with its specific target. Also, at lower baseline values of a given soluble target, half-life did not affect rule-based classification (Figure S3, supplementary file). However, if the potential soluble targets have a higher baseline expression, it is critical to account for the target half-life (Figure S3). A high baseline and longer half-life of a target is an optimal combination of target properties to achieve high TO%.

Assessment of Target Form on Optimal Target Engagement in Tissues

We simulated the mPBPK model for two different forms of target (soluble or membrane-bound). Virtual drug-target candidates were same for the above simulations. The predicted TO% at minimum concentration in leaky tissues was used as a criterion for binary classification. The optimal TO% (> 90%) for antibody engagement with soluble target and membrane-bound target are shown in Fig. 5. The scatter plot (Fig. 5) shows the optimal values of target baseline and half-life to achieve TO% > 90% in plasma with soluble and membrane receptors. Rules indicated that for given drug candidate with optimal \({\text{K}}_{\text{D}}\), a class of membrane-bound receptors with baseline below 4 nM and half-life greater than 20 min is suitable. However, for given drug candidate with optimal \({\text{K}}_{\text{D}}\), a class of soluble receptors with baseline below 7 nM is more suitable, whereas half-life doesn’t affect the target candidate selection. The membrane-bound complex internalizes at the same rate as the target degradation, which is dependent on the target half-life. Therefore, a lower half-life means a higher target degradation and higher internalization rate of the drug-target complex. Both target degradation and complex internalization affect desired TO%, as shown in Equation S1 in the supplementary file. The soluble drug-target complex internalizes at the same rate as drug clearance, which is a likely reason observed differences in distribution of optimal properties between soluble and membrane-bound receptors. Moreover, in each scenario, when target baseline is high and half-life is low, neither soluble or nor membrane-bound target engagement is feasible (Fig. 5). The decision tree-based rules suggested slight differences in drug-target binding constant (\({\text{K}}_{\text{D}}\)) values between soluble and membrane-bound receptors. \({\text{K}}_{\text{D}}\) must be lower than 11 nM and 9 nM to achieve more than 90% TO with soluble and membrane receptors, respectively.

Optimal Drug-Target Properties for Target Engagement at Different Sites of Action

The ML-derived rules for different scenarios were tested using the mPBPK model and Monolix/Simulux model (Table S2). A combination of optimal and non-optimal drug and target properties were chosen based on decision tree-based rules for different scenarios such as dose, dosing scheme etc. The mPBPK model predicted TO% accurately for all combinations used for validation as shown in Table S2 in the supplementary file. The Monolix/Simulux model also predicted correctly for most combinations (Table S2). The validation of ML-derived rules against mPBPK model is expected as virtual data was generated using the mPBPK model. However, the validation of these rules against Monolix/Simulux predictions is also relevant as model structure and assumptions of the Monolix/Simulux model may be different from the mPBPK model. These validations increase confidence in the robustness and performance of the ML model. We also validated the ML-derived rules using real data for clinically approved monoclonal antibodies. Table S3 in supplementary file provides a list of literature-derived properties of monoclonal antibodies and their respective target, as well as recommended clinical dose and regimen [20, 21, 22, 23, 24, 25, 26, 27, 28, 29]. In our study, the trained ML model can make predictions for a fixed dose and regimen administered intravenously. Considering this limitation of the ML model, we made some necessary but reasonable assumptions to validate our model against real data. We assumed a dose of for a dose of 0.1, 1, or 10 mg/kg that was nearest to the recommended clinical dose for each antibody. These assumed dose values for each antibody are provided in Table S3 in supplementary file. We used the trained ML model to predict the optimal \({\text{K}}_{\text{D}}, {\text{T}}_{0}, {\text{a}\text{n}\text{d} \text{t}}_{1/2}\) values for each antibody at the assumed dose and recommended regimen administered intravenously. The optimal drug-target properties were validated for the following scenarios, 0.1 mg/kg bolus, 1 mg/kg Q4W, 1 mg/kg Q2W, and 10 mg/kg Q2W (Fig. 7). The experimental properties of most antibodies lie within the predicted optimal boundaries of drug-target properties when administered at their respective dose and regimen. Except for Abciximab and its target, \({\text{K}}_{\text{D}} \text{a}\text{n}\text{d} {\text{T}}_{0}\)values lie outside the predicted optimal bounds (Fig. 7A). We believe this is because the assumed dose is lower than the clinical dose (0.25 mg/kg bolus). It is likely that the optimal values of \({\text{K}}_{\text{D}} \text{a}\text{n}\text{d} {\text{T}}_{0}\) will be higher for a dose greater than 0.1 mg/kg, as seen from the dose-dependence in Fig. 3. Moreover, the optimal \({t}_{1/2}\) values for each scenario (Fig. 7B) were validated very well against the experimental observation for each antibody. While these preliminary validations are helpful to validate the ML-derived rules, more antibody and target data can help increase confidence in ML predictions.

The optimality criterion for target occupancy may vary across diseases and treatments. The threshold for target occupancy % can be different than 90%. The ML model can be modified to accommodate such differences. As an example, we performed a multi-label classification of drug-target properties based on a low (< 50%), medium (50–90%), and high (> 90%) target occupancy. We used virtual data for drug administration at 1 mg/kg once every two weeks (Q2W) and categorized the data into three classes based on TO% at minimum concentration in plasma at steady state. The performance metrics for multi-label classifier are provided in the supplementary file (Figure S4). Figure 8 shows the optimal region for selection of drug-target binding constant for a given target with baseline and half-life that lead to low, medium, and high target occupancy. We observed clear boundaries for separation of virtual drug-target properties for respective classes. Our results agree that higher binding affinity (lower \({\text{K}}_{\text{D}}\) values) is suitable to achieve higher TO%. Optimal \({\text{K}}_{\text{D}}\) values should be below 9 nM to achieve > 90% TO and must be between 9 nM and 82 nM to achieve 50–90% TO. A high drug-target engagement can be achieved with respective optimal \({\text{K}}_{\text{D}}\) values when soluble target baseline is below 24 nM. A moderate target engagement response can be expected when soluble target baseline level is below 24 nM and half-life is below 19 hours. An insufficient and low target engagement response can be expected when target baseline expression is more than 24 nM and half-life is between 32–300 hours. The multi-label decision tree classifier had an accuracy of 99%, and class precision of 99%, 98%, and 99% for low, medium, and high classes, respectively (Figure S4, supplementary file).

Antibodies represent a diverse array of large molecules that have revolutionized the treatment of a wide range of diseases and have several advantages over small-molecule drugs such as high specificity, high tolerability, and longer half-lives [30, 31]. Despite recent progress, only a small percentage of drug candidates in development enter the clinic and ultimately reach the market, often due to lack of desired pharmacological activity [32]. To assess the desired pharmacology of an antibody candidate at an early stage in this process, drug-target interaction, suitability for drug-target binding, feasibility, PK/PD, and safety are taken into consideration [2, 3]. The physicochemical properties of antibodies such as size, charge, drug-target binding affinity influence their ADME, PK, PD and ultimately their desired pharmacological response [19]. Antibody engineering allows optimization of these properties early in the drug development process. In the past few decades, computational modeling techniques such as PBPK and PKPD modeling have supported the drug design and development significantly [6]. In addition, machine learning has found its place in supporting the drug development pipeline in numerous ways [6, 16, 15, 33, 34]. In this study, we present a model-based pharmacology assessment framework that using both minimal PBPK and ML modeling to determine physicochemical properties of antibodies and specific target characteristics relevant to achieve a desired target engagement response. It is critical to understand the magnitude and duration of target engagement to enable selection of appropriate drug candidates. We successfully achieved such a relationship by first simulating the mPBPK model with many virtual drug and target candidates with varying properties and then using a decision tree-based ML classifier to unravel rules for these properties based on a model-based TO% as a criterion. We used a decision tree-based classification model due to its interpretable nature; it draws a relationship between input (drug-target properties) and output (TO%) using recursive partitioning in a tree-like manner (Figure S2). This interpretable ML model delineates the process of obtaining the predicted class from the given input, which makes it preferable over other black-box machine learning algorithms such as artificial neural networks [6, 35].

The results from our proposed model-based framework showed a quantitative relationship between antibodies’ physicochemical properties on ADME, PK, and TO for different intravenous doses (Fig. 3a, 3b), dosing schemes (Fig. 3c, 3d), and at different sites of action (Fig. 6). We found that decreasing dose and/or frequency of dosing necessitates the selection of antibody candidates with stronger binding affinity with class of specific receptors. As increasing dose and frequency of administration results in higher concentration and accumulation of the drug in both plasma and tissues. To achieve greater than 90% target occupancy, the model also presented rules for antibody candidate properties and target candidates that were different depending on the site of action, such as plasma, leaky tissues, or tight tissues. To sum up, our high-throughput screening technique allows to narrow down the drug candidates with the highest potential based on a combination of their structure-based properties, ADME, PK, and TO. For instance, the ML model reduced the initially available search space for antibodies from a total of 11620 candidates to 590 optimal candidates based on TO%, which is approximately 5% of the total virtual candidates. The decision tree-based ML is robust and reasonably accurate when validated against model-based and observed data.

Applications to Early Stages of Antibody Discovery and Development

The implementation of the proposed mPBPK/ML modeling framework would enable researchers to screen and optimize thousands of antibody candidates based on optimal ADME, PK, and drug engagement to specific targets, and reduce attrition rate and overall discovery time for lead candidates. To speed up hit-to-lead, this work can identify which candidate hits to move forward. Such a high-throughput screening and evaluation of antibody properties has not been done previously and will allow candidate evaluation against target occupancy and engagement very early in the drug discovery and development pipeline. The outcome from the mPBPK/ML modeling framework would provide scientists with optimal starting values and range of properties for antibody engineering/development.

The modeling framework can be used for the following potential applications and future improvements. For target selection, the model-based prediction of the optimal bounds could assist scientists in selecting the desired target with optimal properties. For lead candidate generation, this framework can be used to determine optimal candidate properties for desired target engagement that are needed to progress to hit-to-lead stage. These advanced high-throughput screening techniques enables scientists to narrow down the candidates with highest potential based on a combination of their structure-based properties, ADME, PK, and TO. The mPBPK/ML framework provides a tabulated list of combination of drug candidate properties and specific target properties to scientists for all the candidates that achieve greater than 90% target engagement, which can be used by scientists as an initial reference sheet to focus on the candidates with most potential. Chemists can use this framework for a priori knowledge of dose-driven design of antibodies, to prioritize the candidates with the highest potential to achieve target engagement and enter lead stage at the lowest dose. The mPBPK/ML framework can also be used by teams for in vitro ADME screening efforts during hit-to-lead or early lead optimization, visualization of the optimal spaces of candidate properties to define the most critical combination of properties for screening [6, 30]. The mPBPK/ML framework can identify optimal drug-target pairs based on drug-target engagement and interaction, identification of these drug-target pairs have many benefits such as drug repositioning [36, 37]. ML models for prediction of drug-target interaction pairs has been done previously based on chemical structure or sequence and not necessarily based on molecular properties and TO% [36, 37].

Further, we aim to explore more advanced interpretable ML algorithms. The current framework serves as a foundation that can incorporate, as additions, other in vitro or in silico derived properties of the antibodies, other pharmacokinetic endpoints such as clearance, pharmacodynamic effect, and safety to train the ML model.

Challenges and Limitations

The proposed modeling framework have several challenges and some limitations that can be overcome with necessary data availability and improvements. Currently, the validity and robustness of the ML outcome depends on the accuracy of mPBPK model, which provides the necessary PK/TO output for respective virtual candidate properties. At early stages of drug discovery, a detailed experimental validation of target interaction, target occupancy %, and drug’s mechanism of action may not be available and must rely on model-based prediction. Moreover, our framework is also limited to assessment of drug properties against a PK or TO endpoint and does not incorporate the PD effect and efficacy of the drug. The desired target occupancy is very much dependent on the target of interest and can be variable among different diseases and disease states. In this work, we used a typical value of 90% TO that is occasionally used as a typical value indicating efficacy when no other information is available. Despite this limitation, our results were able to capture optimal ranges from various compounds targeting different targets for various diseases (Fig. 7). Despite this limitation, the mPBPK/ML model identifies candidate properties that achieve the desired target engagement response and indicates bounds for different properties to achieve same desired effect at reduced dose, dosing frequency etc. The preliminary analysis for early discovery of antibodies using virtual data and model-based PK/TO endpoints should be reasonable to evaluate potential candidates. The mPBPK/ML model may further be improved to use other PK endpoints such as clearance, potency etc. as criteria to identify optimal properties based on the research question at hand. The quality and accuracy of the mPBPK/ML model should also improve with more scientific investment in early discovery efforts. With advancement in the early discovery pathway, the modeling framework can be updated and re-trained as new data becomes available.

In conclusion, we deliver a first-in-class model-based framework that integrates mPBPK and ML model to better understand the biology-specific PK and ADME processes and desired target pharmacology of monoclonal antibodies. By unravelling new ML-based design rules for selection of drug and target properties based on target engagement response, we can reduce the overall time for drug candidate selection in the early discovery stages.

Data Availability

The virtual datasets generated during and/or analyzed during the current study are available in public GitHub repository (https://github.com/Sanofi-GitHub/DMPK-US-ML-mTPA).

Code Availability

The computer codes used for this study have been deposited in GitHub (https://github.com/Sanofi-GitHub/DMPK-US-ML-mTPA).

Acknowledgments

The study was funded by Sanofi.

Author Contributions

All authors contributed to the conception and design of the study. Data generation, collection, and analysis were performed by Krutika Patidar. ML model training and testing in Python were performed by Krutika Patidar. Model validation using the mPBPK model and the Monolix/Simulx model was performed by Krutika Patidar and Saroj Dhakal, respectively. All authors interpreted the results from the study. The first draft of the manuscript was written by Krutika Patidar, and all authors reviewed and edited the manuscript. All authors read and approved the final manuscript.

Competing interests

All authors were employed by Sanofi while the manuscript was written.

T. Dunlap and Y. Cao, "Physiological Considerations for Modeling in vivo Antibody-Target Interactions.," Frontiers in pharmacology, vol. 13, 2022.
C. H. Emmerich, L. M. Gamboa, M. C. J. Hofmann, M. Bonin-Andresen, O. Arbach, P. Schendel, B. Gerlach, K. Hempel, A. Bespalov, U. Dirnagl and M. J. Parnham, "Improving target assessment in biomedical research: the GOT-IT recommendations.," Nat Rev Drug Discov, vol. 20, p. 64–81, 2021.
M. H. Linaraju, "Target Assessment in Drug Discovery and Development," June 2020. [Online]. Available: https://www.linkedin.com/pulse/target-assessment-drug-discovery-development-lingaraju-lings-m-h-.
A. P. Chimalakonda, R. Yadav and P. Marathe, "Factors influencing magnitude and duration of target inhibition following antibody therapy: implications in drug discovery and development.," The AAPS journal, pp. 717-727, 2013.
T. P. Kenakin, "Pharmacology in Drug Discovery and Development (Second Edition)," in Chapter 1 - Pharmacology: The Chemical Control of Physiology, Academic Press, 2017.
E. P. Chen, R. W. Bondi and P. J. Michalski, "Model-based Target Pharmacology Assessment (mTPA): An Approach Using PBPK/PD Modeling and Machine Learning to Design Medicinal Chemistry and DMPK Strategies in Early Drug Discovery," Journal of Medicinal Chemistry, vol. 64, pp. 3185-3196, 2021.
A. Datta-Mannan, J. Lu, D. R. Witcher, D. Leung, Y. Tang and V. J. Wroblewski, "The interplay of non-specific binding, target-mediated clearance and FcRn interactions on the pharmacokinetics of humanized antibodies.," mAbs, vol. 1084–1093, p. 7, 2015.
H. M. Jones, Z. Zhang, P. Jasper, H. Luo, L. B. Avery, L. E. King, H. Neubert, H. A. Barton, A. M. Betts and R. Webster, ", A Physiologically-Based Pharmacokinetic Model for the Prediction of Monoclonal Antibody Pharmacokinetics From In Vitro Data," CPT Pharmacometrics Syst. Pharmacol., 2019.
S. Liu and D. K. Shah, "Physiologically Based Pharmacokinetic Modeling to Characterize the Effect of Molecular Charge on Whole-Body Disposition of Monoclonal Antibodies.," AAPS Journal, 2023.
S. Liu, A. Verma, H. Kettenberger, W. F. Richter and D. K. Shah, "Effect of variable domain charge on in vitro and in vivo disposition of monoclonal antibodies," mAbs, 2021.
K. Patidar, N. Pillai, S. Dhakal, L. B. Avery and P. D. Mavroudis, "A minimal physiologically based pharmacokinetic model to study the combined effect of antibody size, charge, and binding affinity to FcRn/antigen on antibody pharmacokinetics," Journal of Pharmacokinetics and Pharmacodynamics , 2024.
D. K. Shah and A. M. Betts, "Towards a platform PBPK model to characterize the plasma and tissue disposition of monoclonal antibodies in preclinical species and human.," Journal of pharmacokinetics and pharmacodynamics, p. 67–86., 2012.
M. Ovacik and K. Lin, "Tutorial on Monoclonal Antibody Pharmacokinetics and Its Considerations in Early Development," Clin Transl Sci, vol. 11, 2018.
D. Yuan, F. Rode and Y. Cao, "A Minimal Physiologically Based Pharmacokinetic Model with a Nested Endosome Compartment for Novel Engineered Antibodies.," The AAPS journal, 2018.
E. Makowski, T. Wang, J. Zupancic, J. Huang, L. Wu, J. Schardt, A. De Groot, S. Elkins, W. Martin and P. Tessier, "Optimization of therapeutic antibodies for reduced self-association and non-specific binding via interpretable machine learning," Nat. Biomed. Eng., vol. 8, p. 45–56, 2024.
E. P. Chen, R. W. Bondi, C. Zhang, D. J. Price, M.-H. Ho, K. A. KiraArmacost and M. P. DeMartino, "Applications of Model-Based Target Pharmacology Assessment in Defining Drug Design and DMPK Strategies: GSK Experiences," Journal of Medicinal Chemistry, vol. 65, pp. 6926-6939, 2022.
G. Lemaitre, F. Nogueira and C. K. Aridas, "Imbalanced-learn: A Python Toolbox to Tackle the Curse of Imbalanced Datasets in Machine Learning," Journal of Machine Learning Research, vol. 18, pp. 1-5, 2017.
J. Tibbitts, D. Canter, R. Graff, A. Smith and L. A. Khawli, "Key factors influencing ADME properties of therapeutic proteins: A need for ADME characterization in drug discovery and development.," mAbs, pp. 229 - 245., 2015.
X. Jing, Y. Hou, W. Hallett, C. G. Sahajwalla and P. Ji, "Key Physicochemical Characterisitcs Infuencing ADME Properties of Therapeutic Proteins," Adv Exp Med Biol, 2019.
L. Peng, V. Oganesyan, H. Wu, W. F. Dall'Acqua and M. M. Damschroder, "Molecular basis for antagonistic activity of anifrolumab, an anti-interferon-α receptor 1 antibody," MAbs, vol. 7, pp. 428-39, 2015.
O. Le Tilly, P. Gatault, C. Baron, T. Bejan-Angoulvant, M. Buchler, G. Paintaud and D. Ternant, "Is CD25 blockade optimal in kidney transplant patients treated with basiliximab? A target-mediated drug disposition model," Br J Clin Pharmacol, vol. 88, pp. 3500-3505, 2022.
P. M. Glassman and J. P. Balthasar, "Physiologically-based modeling to predict the clinical behavior of monoclonal antibodies directed against lymphocyte antigens," MAbs, vol. 9, pp. 297-306, 2017.
M. D. Pescovitz, "Rituximab, an anti-cd20 monoclonal antibody: history and mechanism of action," Am J Transplant, vol. 6, pp. 859-66, 2006.
S. Basu, Y. T. K. Lien, V. Vozmediano, J. F. Schlender, T. Eissing, S. Schmidt and C. Niederalt, "Physiologically Based Pharmacokinetic Modeling of Monoclonal Antibodies in Pediatric Populations Using PK-Sim," Front Pharmacol, vol. 11, p. 868, 2020.
D. E. Mager, M. A. Mascelli, N. S. Kleiman, D. J. Fitzgerald and D. R. Abernethy, "Simultaneous modeling of abciximab plasma concentrations and ex vivo pharmacodynamics in patients undergoing coronary angioplasty," J Pharmacol Exp Ther, vol. 307, pp. 969-76, 2003.
T. D. Liu, S. H. Ren, X. Ding, Z. L. Xie and Y. Kong, "A Short Half-Life alpha(IIb)beta(3) Antagonist ANTP266 Reduces Thrombus Formation," Int J Mol Sci, vol. 19, 2018.
P. Bloomingdale, D. Bumbaca-Yadav, J. Sugam, S. Grauer, B. Smith, S. Antonenko, M. Judo, G. Azadi and K. L. Yee, "PBPK-PD modeling for the preclinical development and clinical translation of tau antibodies for Alzheimer's disease," Front Pharmacol, vol. 13, p. 867457, 2022.
Z. Tsilogianni, G. Hillas, P. Bakakos, L. Aggelakis, E. Konstantellou, A. L. Papaioannou, A. Papaporfyriou, S. Papiris, N. Koulouris, S. Loukides and K. Kostikas, "Sputum interleukin-13 as a biomarker for the evaluation of asthma control," Clin Exp Allergy, vol. 46, pp. 923-31, 2016.
A. J. Okragly, A. Ryuzoji, I. Wulur, M. Daniels, R. D. Van Horn, C. N. Patel and R. J. Benschop, "Binding, Neutralization and Internalization of the Interleukin-13 Antibody, Lebrikizumab," Dermatology and Therapy, vol. 13, pp. 1535-1547, 2023.
L. B. Avery, J. Wade, M. Wang, A. Tam, N. Piche-Nicholas, M. S. Kavosi, S. Penn, D. Cirelli, J. C. Kurz, M. Zhang, O. Cunningham, R. Jones, B. J. Fennel, B. McDonnell, P. Sakorafas, W. J. Finlay, L. Lin, L. Bloom and D. M. O'Hara, "Establishing in vitro in vivo correlations to screen monoclonal antibodies for physicochemical properties related to favorable human pharmacokinetics," MABS, vol. 10, 2018.
J. Bolleddula, K. Brady, G. Bruin, A. Lee, J. A. Martin, M. Walles, K. Xu, T. Y. Yang, X. Zhu and H. Yu, "Absorption, Distribution, Metabolism, and Excretion (ADME) of Therapeutic Proteins," Drug Metab Dispos, 2022.
S. Akhondzadeh, "The Importance of Clinical Trials in Drug Development. .," Avicenna journal of medical biotechnology, 2016.
P. D. Mavroudis, D. Teutonico, A. Abos and N. Pillai, "Application of machine learning in combination with mechanistic modeling to predict plasma exposure of small molecules," Frontiers in Systems Biology, vol. 3, 2023.
A. Gruber, F. Fuhrer, S. Marz, H. Diedam, A. H. Goller and S. Schneckener, "Prediction of Human Pharmacokinetics From Chemical Structure: Combining Mechanistic Modeling with Machine Learning," Journal of Pharmaceutical Sciences, 2023.
V. Rodriguez-Galiano, M. Sanchez-Castillo, M. Chica-Olmo and M. Chica-Rivas, "Machine learning predictive models for mineral prospectivity: An evaluation of neural networks, random forest, regression trees and support vector machines," Ore Geology Reviews, vol. 71, pp. 804-818, 2015.
P. N. Hameed, K. Verspoor, S. Kusljic and S. Halgamuge, "A two-tiered unsupervised clustering approach for drug repositioning through heterogeneous data integration," BMC Bioinformatics, vol. 19, 2018.
T. Pahikkala , A. Airola, S. Pietilä, S. Shakyawar, A. Szwajda, J. Tang and T. Aittokallio, "Toward more realistic drug-target interaction predictions.," Brief Bioinform., vol. 16, pp. 325-37, 2015.

No competing interests reported.

PM051424mTPAprojectsupplementary.docx

Download PDF

Editorial decision: Revision requested
09 Jun, 2024
Reviews received at journal
07 Jun, 2024
Reviews received at journal
30 May, 2024
Reviewers agreed at journal
21 May, 2024
Reviewers agreed at journal
19 May, 2024
Reviewers invited by journal
16 May, 2024
Editor assigned by journal
16 May, 2024
Editor invited by journal
16 May, 2024
Submission checks completed at journal
16 May, 2024
First submitted to journal
14 May, 2024

You are reading this latest preprint version

Development of a Minimal Physiologically-Based Pharmacokinetic Modeling / Machine Learning Framework for Early Target Pharmacology Assessment

Status:

Version 1

Abstract

Figures

Introduction

Methods

Minimal PBPK Model for Monoclonal Antibodies

Virtual Data Generation

Decision Tree-Based Supervised Machine Learning

mPBPK / ML Model Framework

Results

Optimal Drug-Target Properties for Target Engagement for Different Dose and Regimen

Assessment of Target Form on Optimal Target Engagement in Tissues

Optimal Drug-Target Properties for Target Engagement at Different Sites of Action

Discussion

Applications to Early Stages of Antibody Discovery and Development

Challenges and Limitations

Declarations

Data Availability

Code Availability

Acknowledgments

Author Contributions

Competing interests

References

Additional Declarations

Supplementary Files

Status:

Version 1