Another string to your bow: machine learning prediction of the pharmacokinetic properties of small molecules

Figures & data

Figure 1. Three archetypes of methods for PK parameter prediction, and visual presentations of selected PK parameters. Panel A: graphical representation of the three archetypal methods for PK property prediction: machine learning (ML), in vitro-in vivo extrapolation (IVIVE), and physiologically based PK (PBPK) modeling. Panels B-E: graphical representations of selected PK parameters that are of common interest. See Table 1 for definitions and explanations. Panel a was created with BioRender.com. Panels B-E were made by JD Zhang and released under GNU public license. The source code and the figure can be found at https://github.com/Accio/2024-02-PK-parameters-visualized. JD Zhang explicitly authorizes the use of his work for this review.

Table 1. Selected PK parameters obtainable through a single-dose PK experiment.

Property	Symbol	Common units	Definition and comments
Area Under the Curve	AUC	h*ng/mL	The actual body exposure to a drug after administration of a dose of it, represented as concentration integrated over time.
Plasma clearance	CLp	L/h(/kg) mL/min(/kg)	The rate of the body eliminating a drug, represented as the volume of plasma from which a drug is completely removed per unit time.
Volume of distribution at steady-state	Vss	L	Ratio of the dose of the drug to the concentration that is measured in blood plasma, when the steady state is reached, i.e. no net flux of drug between the central and peripheral compartments.
Bioavailability	F, or F%	–	Fraction of an administered drug that reaches the systemic circulation.
Plasma half-life	t1/2	h	The time required for a drug’s plasma concentration to diminish by 50% due to elimination.
Maximal serum concentration	Cmax	ng/mL	It can be measured empirically.
Time to peak drug concentration	tmax	h	It can be measured empirically.

Figure 2. Relationship between ADME processes in vivo (a), compound’s ADME properties (b), and selected in vitro assays available to assess ADME properties (c). The in vivo outcomes are typically governed by multiple ADME compound properties. Many in vitro experiments are available to measure these properties, including physicochemical, biochemical, and cellular assays. in vitro assays can be used to deconvolute observed ADME processes or to predict ADME properties prospectively.

Table 2. Advantages and potential limitations of ADME assays as predictors of drug disposition in vivo and tools for rational drug design, as well as possible mitigation measures.

Advantages of ADME assays	Potential limitations	Possible mitigation measures
They provide mechanistic insights into individual ADME properties, and enable predictions for and interpretation of in vivo outcomes.	Some assays are devoid of the physiological and therapeutic context. Some assays have limited predictivity of in vivo outcomes.	Applying IVIVE approaches or PBPK models.
They support rational drug design. Optimization cycles integrating in vitro assays help reduce preclinical animal studies.	While ADME properties can be measured individually, they can hardly be tuned individually, as chemical modifications often change more than one ADME properties at once. Intuition, experience, and sometimes serendipity are essential to improve molecule design based on assay readouts.	Multiparametric Optimization (MPO) approaches, including in silico property prediction with machine learning, can be used.
The assay readouts allow comparing, ranking, and prioritizing compounds based on standardized readouts.	It can be misleading to compare and rank compounds with a single property.	Readouts from multiple assays as well as in silico predictions can be used by either machine-learning or PBPK models to make holistic decisions.
Assay operation and data analysis can be standardized and automated. Historical datasets can be leveraged for computational predictions, for instance with ML methods.	As chemical space evolves, new chemotypes can be outside of the applicability domain of both experimental and computational models.	Operation of the assays, data quality, and the applicability domain need regular scrutiny. The assay can be updated and the associated machine learning model can be retrained if necessary.

Figure 3. Evolution of machine-learning models for PK parameter prediction. The panels illustrate key parameters of ML models, including the year of publication (x-axis) and performance metrics (y-axis), species of predictions (colors), number of compounds used (size of the bubbles), and root-mean-square errors (RMSE) indicated as numbers near the bubbles. In cases RMSE values are not available, NA (not available) is indicated. Panel a, c, and e visualize models reviewed that predict Vss, CLp, and bioavailability, respectively, using AAFE (absolute average fold error), which is also known as Geometric mean Fold error (GMFE), as the metric of performance. Panel b, d, and f show models that predict the same properties as panel a, c, and e, however using coefficients of determination (R²) as the metric. Some studies only report either AAFE or R² but not the other, therefore the studies included in pairs of panels for the same property (for instance panel a and c for Vss prediction) overlap only partially. Red dashed lines indicate boundary conditions at which an ideal model that makes perfect predictions would perform: an AAFE of 1.0 or an R² of 1.0.

Table 3. Overview of the three approaches for PK predictions.

Method	Requirements	Limitations	Advantages
Machine learning (ML)	Abundant high-quality data; Infrastructure to collect and analyze data: database, software, resources, experience, etc.; User training on using the model and interpreting the results; With new projects covering broader chemical spaces, models need regular retraining, predictions need to be compared with measurements, and feedback is to be collected to inform the evolution of models. Assessment of applicability domain, quantification of uncertainty, and ideally model interpretability;	Model quality is largely determined by the target of prediction and the quality and quantity of data. Community-driven development is hampered by the lack of openly available, high-quality benchmark datasets that cover new chemical spaces. Off-the-shelf models are usually not enough: fine-tuning and adaptations are often required. Predictions often lack causal interpretation.	Predictions can be made even before the molecule is synthesized. Compared with alternative methods, ML predictions are fast and scalable since they involve little cost. When the quality and quantity of training data are high, the model can confidently and reliably predict inside its applicability domain. Given enough data, models can effectively simulate any linear or non-linear relationship between input features and the property to predict. Models can be updated and retrained with new data.
In vitro-in vivo extrapolation (IVIVE)	Relevant biological processes must be captured. Extrapolations are validated with compounds of known properties before use. Devices, resources, and expertise are required.	Identification of fit-for-purpose models consumes time and resources. Both technical robustness and biological variability need to be assessed. Trade-off between model complexity, throughput, and fidelity. Not all relevant biological processes and organs can be captured yet. Modeling population-level variance (donor-to-donor variability) can be challenging.	When a ‘function-alike’ (not only ‘look-alike’) in vitro system is established, it becomes a useful representation of the biological environment analyzed. One model may serve more than one purpose, for instance safety evaluation, MoA understanding, and biomarker discovery. An established model can be manipulated for the modeling and evaluation of both homeostatic (healthy) and pathological states. The evaluation of casualty in the biological system is more straightforward than in silico approaches.
Physiologically-based pharmacokinetic modeling (PBPK)	Models require experimental or predicted data as input. Models must be validated with in vivo or human data. Database, software, resources, and expertise are required.	Not all required data may be available. Skills and expertise required to scale up the analysis to serve many projects with diverse questions.	Tunable on the single molecule or experiment examined. Relatively scalable, depending on the level of detail with which each molecule has to be evaluated. Relatively explainable results: they come from a combination of multiple, but known, differential equations. Historically, PBPK has demonstrated its value and served as a gold standard during both preclinical and clinical phases.