Overcoming PK/PD Modeling Challenges

Pharmacokinetic (PK)/pharmacodynamic (PD) modeling of small-molecule drug candidates provides information essential for selecting and progressing optimal drug candidates. Effective PK/PD modeling, however, requires specialized expertise, time, and resources that are not always readily available.

Use of advanced artificial intelligence (AI) and machine learning (ML) algorithms can ameliorate resource limitations while also enabling the identification of patterns not readily apparent to human analysts, thus streamlining development while providing more robust predictions of in vivo drug performance.

PK/PD modeling plays important role in drug development

Understanding the absorption, distribution, metabolism, elimination, and toxicity (ADMET) properties of small-molecule drug candidates can inform the evaluation of small-molecule drug candidates through discovery and development efforts. Detailed information about these attributes can help weed out undesirable molecules and move forward candidates with a higher likelihood of success. In vitro and in vivo PK/PD screening studies are time-consuming and costly, however, with ADMET assessments often becoming a bottleneck in drug development programs (1).

In silico approaches for determination of ADMET properties have, therefore, become crucial to accelerating drug discovery and development efforts, allowing rapid evaluation of large numbers of compounds. “PK/PD modeling provides a quantitative framework that is essential for effective, data-driven drug development,” states Pauline Traynard, product manager for MonolixSuite at SimulationsPlus. “By integrating diverse data from laboratory experiments and clinical studies, these models allow us to understand the complex interplay between a drug’s concentration in the body and its therapeutic effect. This mechanistic insight is critical for optimizing dosing regimens, designing more efficient and informative clinical trials, and making crucial go/no-go decisions,” she explains.

Ultimately, Traynard observes, PK/PD modeling helps reduce late-stage attrition rates and supports a more streamlined, cost-effective, and successful path to regulatory approval and clinical use.

Increasing complexity creating challenges

Greater understanding of disease mechanisms and advances in rapid synthesis of compound libraries based on increasingly complex molecules, many which have limited solubility (and therefore bioavailability) under physiological conditions, has created several challenges to achieving effective PK/PD modeling.

“Emerging trends such as highly potent molecules with narrow therapeutic indices and complex drug delivery technologies (e.g., long-acting injectables or lipid nanoparticles) pose significant modeling challenges,” says Traynard. She also notes that many advanced drug candidates exhibit nonlinear kinetics, require specialized absorption models, or interact with biological barriers in non-intuitive ways.

An additional significant challenge, Traynard adds, arises during clinical development. “Understanding and predicting the high degree of variability in how different patients respond to a drug is difficult to decipher from the limited samples per patient typically collected during clinical trials,” she comments.

AI/ML algorithms overcome many modeling challenges

AI/ML algorithms are instrumental in overcoming many of these modern PK/PD challenges, according to Traynard, due to their superior ability to identify complex patterns in high-dimensional data where mechanistic understanding is still incomplete.

The most impactful applications for AI/ML-based PK/PD modeling are, Traynard observes, in early-stage discovery and the enhancement of mechanistic models. “In discovery, AI/ML-driven prediction of ADME and toxicity is invaluable for rapidly screening vast compound libraries to de-risk projects before significant investment is made,” she explains.

For instance, Traynard notes that AI/ML models can rapidly and accurately predict a full suite of ADME properties directly from a chemical structure, allowing scientists to screen and optimize thousands of candidates virtually and focus resources on the most promising ones. In clinical development, meanwhile, ML-based models can in early development be employed to analyze sparse patient data for efficient identification of factors contributing to variability in drug response, leading to more robust population PK models and better-informed dosing strategies, according to Traynard.

AI/ML techniques are particularly valuable for classes of APIs where traditional modeling methods face limitations due to underlying biological or chemical complexity. Traynard highlights compounds with poor solubility and permeability (Biopharmaceutics Classification System Class II/IV), where AI/ML can effectively model the intricate relationship between a drug’s formulation and its in vivo absorption to predict bioavailability. “It is also used for APIs with complex safety profiles or non-linear pharmacokinetics, as AI/ML is adept at integrating diverse biological data to identify safety signals, such as the potential for organ injury, that are difficult to predict with simpler models,” she says.

Significant benefits can also be obtained when using ML models in conjunction with in vitro-in vivo extrapolation and physiologically based pharmacokinetic (PBPK) modeling because integration of these approaches provides a more comprehensive picture of a drug’s ADMET profile (2). Augmentation of established PBPK and population PK/PD models with model structure learning and parameter estimation is, says Traynard, a second key application. As an example, such modified models are useful for identifying influential covariates or nonlinear relationships within clinical data.

A third important application of advance algorithms in PK/PD modeling is the automation of model development workflows, which Traynard believes can reduce time-intensive manual steps by implementing ML-guided model selection, fit optimization, and diagnostics.

More high-quality data needed

The biggest challenge to enabling the widespread use of AI/ML algorithms to improve PK/PD modeling is the same issue facing the development of AI/ML models for all other applications in the biopharmaceutical industry: the need for large quantities of high-quality, reliable data. “Limited access to large, high-quality datasets needed to train reliable AI models is a major challenge to broader AI/ML adoption in PK/PD modeling, largely due to restrictions and experimental variability,” observes Traynard.

Another important concern noted by Traynard is the inherent “black box” nature of some algorithms, which may lack the transparency needed for regulatory or internal scientific confidence.

Hybrid approaches combining established mechanistic models like population PK/PD or PBPK with interpretable AI components are therefore gaining traction. “This strategy grounds the AI’s powerful pattern-recognition abilities in the context of known biology, making the results more interpretable, scientifically plausible, and trustworthy for both scientists and regulators,” Traynard comments.

Explainable algorithms will be most impactful

Regulatory acceptance is crucial, of course, and there is growing interest from regulators in AI/ML’s potential for population simulation, dose selection, and decision support, according to Traynard, particularly when integrated into PBPK or exposure-response models used in regulatory filings.

“The core regulatory concerns,” observes Traynard, “are model transparency, validation, and managing bias to ensure that any AI/ML tool is fit-for-purpose.” As a result, she says, there is a strong emphasis on establishing “good machine learning practice” and developing explainable AI that allows regulators to understand the basis of a model’s predictions.

Going forward, Traynard believes the evolution of AI/ML in PK/PD modeling is likely to center on a deeper integration with mechanistic science, creating powerful hybrid systems that are both more predictive and explainable. “As these predictive models become more robust, we anticipate advances in automated clinical trial simulation and design, where AI could be used to forecast trial outcomes under different scenarios and optimize protocol parameters in silico before a single patient is enrolled,” she adds.

Looking further out, Traynard notes that leveraging AI to deliver truly personalized medicine represents an important frontier. “The ambition,” she explains, “is to use sophisticated algorithms to integrate vast, patient-specific datasets to enable more accurate personalized predictions of an individual’s unique response to a drug, paving the way for individual treatment optimization.”

References

1. Myung, Y.; de Sá. A.G.C.; and Ascher, D.B. Deep-PK: Deep Learning for Small Molecule Pharmacokinetic and Toxicity Prediction Open Access. Nucl. Acids Rsch. 2024 52 (W1), pp. W469–W475. DOI: 10.1093/nar/gkae254
2. Bassani, D.; Parrott, N. J.; Manevski, N.; and Zhang, J. D. Another String to your Bow: Machine Learning Prediction of the Pharmacokinetic Properties of Small Molecules. Expert Opinion on Drug Discovery 2024 19(6), 683–698. DOI:10.1080/17460441.2024.2348157

About the author

Cynthia A. Challener, PhD, has been a freelance technical writer for more than 25 years and is a contributing editor to Pharmaceutical Technology®.