Overcoming challenges in PK/PD modeling

Enlarger with pills on blue background | Image credit: ©pixel -shot -stock.adobe.com

Pharmacokinetics (PK)/pharmacodynamics (PD) modeling of small molecule drug candidates provides essential information for the selection and progression of the optimal drug candidate. However, effective PK/PD modeling requires specialized expertise, time and resources that are not always readily available.

The use of advanced artificial intelligence (AI) and machine learning (ML) algorithms improve resource limitations, while also enabling the identification of patterns that are not readily apparent to human analysts, providing more robust predictions while streamlining development, in vivo Drug performance.

PK/PD modeling plays an important role in drug development

Understanding the absorption, distribution, metabolism, excretion 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 remove unwanted molecules and advance candidates with a high probability of success. In vitro and in vivo, screening studies for PK/PD are time-consuming and expensive, but the evaluation of ADMET is often a bottleneck in drug development programs (1).

Therefore, the silico approach for determining ADMET properties will become important to accelerate biological discovery and development efforts, allowing for the rapid evaluation of numerous compounds. “PK/PD modeling provides a quantitative framework essential for effective, data-driven drug development,” said Pauline Traynard, product manager for MonolixSuite at SimulationPlus. “By integrating diverse data from laboratory experiments and clinical research, these models allow us to understand the complex interactions between drug concentrations and their therapeutic effects. This mechanical insight is important for optimizing dosing regimens, designing more efficient and beneficial clinical trials, and making Crucial Go/No-Go decisions.”

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

Increased complexity creates challenges

A deeper understanding of disease mechanisms and advances in the rapid synthesis of increasingly complex molecule-based compounds libraries with limited solubility under physiological conditions (and therefore bioavailability) has created several challenges to achieve effective PK/PD modeling.

“New trends such as narrow therapeutic indexes and extremely powerful molecules with complex drug delivery techniques (such as long-acting injectables and lipid nanoparticles) pose important modeling challenges,” says Traynard. She also points out that many advanced drug candidates exhibit nonlinear dynamics and require specialized absorption models or interact with biological barriers in a non-intuitive way.

More important challenges arise during clinical development, Traynard adds. “It is usually difficult to understand patients' responses to different drugs during clinical trials to decipher them from a limited sample per patient,” she commented.

AI/ML algorithms overcome many modeling challenges

According to Traynard, AI/ML algorithms can help you overcome many of these latest PK/PD challenges. This is due to its excellent ability to identify complex patterns in high-dimensional data where mechanical understanding is still incomplete.

The most influential applications of AI/ML-based PK/PD modeling are observed in early stage discovery and enhancement of mechanical models. “In discovery, AI/ML-driven ADME and toxicity predictions are invaluable for rapid screening of risky projects before significant investments are made,” she explains.

For example, Traynard points out that AI/ML models can quickly and accurately predict a complete suite of ADME properties from chemical structures, allowing scientists to focus thousands of candidates on virtually the most promising candidates. Meanwhile, clinical development employs ML-based models in early development to analyze sparse patient data for efficient identification of factors contributing to drug response fluctuations, leading to more robust population PK models and more capable dosing strategies.

AI/ML techniques are especially valuable for classes of APIs that face limitations due to the biological or chemical complexity that traditional modeling methods are based on. Traynard highlights compounds with low solubility and permeability (Biopharmaceutics Classification System Class II/IV). Here, AI/ML can effectively model the complex relationship between drug formulation and in vivo absorption to predict bioavailability. “AI/ML is also used in APIs with complex safety profiles or nonlinear pharmacokinetics because they are skilled at integrating diverse biological data to identify safety signals that are difficult to predict with simpler models,” she says.

The integration of these approaches also offers great advantages when using ML models in conjunction with in vitro-in vivo extrapolation and physiologically based pharmacokinetics (PBPK) modeling (2). According to Traynard, the second important application, the enhancement of established PBPK and population PK/PD models with model structure learning and parameter estimation. As an example, such a modified model can help identify influential covariates or nonlinear relationships within clinical data.

The third important application of advanced algorithms in PK/PD modeling is automation of model development workflows. Traynard believes that implementing ML guide model selection, fit optimization and diagnostics can reduce time-intensive manual steps.

Higher quality data is required

The biggest challenge to enable the widespread use of AI/ML algorithms to improve PK/PD modeling is the same challenge facing the development of AI/ML models for all other applications in the biopharmaceutical industry. There is a need for large amounts of high quality, reliable data. “The limited access to large, high-quality datasets required to train trustworthy AI models is a major challenge for wider AI/ML adoption in PK/PD modeling, primarily due to limitations and experimental variation,” says Traynard.

Another important concern pointed out in Traynard is the inherent “black box” nature of some algorithms.

Therefore, hybrid approaches combining established mechanical models such as collective PK/PD and PBPK with interpretable AI components have gained traction. “This strategy eradicates the powerful pattern recognition capabilities of AI in the context of known biology, making the results more interpretable, scientifically plausible and reliable for both scientists and regulators,” commented Traynard.

Explainable algorithms are the most impactful

Of course, regulatory acceptance is extremely important, and there is growing interest from regulators for population simulation, dose selection and the possibility of decision support for AI/ML, particularly when integrated into the PBPK or exposure response models used in regulatory applications.

“The core regulatory concerns are model transparency, validation, and management of bias, and bias to ensure that AI/ML tools are suitable for their purposes.” As a result, she says, she focuses on establishing “good machine learning practices” and developing explanatory AI that allows regulators to understand the fundamentals of model prediction.

Going forward, Traynard believes that the evolution of AI/ML in PK/PD modeling will likely focus on deeper integration with mechanical science, creating powerful hybrid systems that are more predictive and explainable. “As these predictive models become more robust, we anticipate advances in automated clinical trial simulations and design. Here, AI can be used to predict trial results in various scenarios and optimize the protocol parameters of silico before one patient is enrolled,” she adds.

More specifically, Traynard points out that leveraging AI to provide truly personalized healthcare represents an important frontier. “Ambition” is to “use sophisticated algorithms to integrate vast patient-specific datasets to allow for more accurate, personalized predictions of individuals' unique responses to drugs, paving the way for individual treatment optimization.”

reference

Myung, Y. ; desá. AGC;Ascher, DB Deep-PK: Deep learning open access for small molecule pharmacokinetics and toxicity prediction. Nucl. Acid rsch. 2024 52 (W1), pp. W469– W475. doi: 10.1093/nar/gkae254
Bassani, D. ;Parrot, New Jersey; Manevski, N. Zhang, JD Another String to Your Bow: Machine Learning Prediction of the Pharmacokinetic Properties of Small Molecules. Experts' opinions regarding drug discovery 2024 19(6), 683–698. doi:10.1080/17460441.2024.2348157