ABSTRACT Quantitative systems pharmacology (QSP) provides a mechanistic framework for integrating biological and pharmacological knowledge in drug development. Simultaneously, advances in artificial intelligence (AI), particularly in multi‐modal data analysis, are transforming how biological information is derived. I argue that AI should not be viewed as a data integration tool, but as a new layer of quantitative measurement generation. This reframing positions QSP as a foundational model of biology, used here in a conceptual rather than technical sense, and AI as a source of quantitative inputs.
ABSTRACT Molecularly targeted therapeutics, such as lorlatinib, are typically initiated at fixed starting doses with subsequent tolerability‐guided modifications, although substantial interindividual variability in pharmacokinetics and drug exposure may influence both efficacy and safety outcomes. Therefore, identifying patient factors associated with individualized dose requirements is an important step toward precision dosing. The aim of this study was to apply the Systematic Covariate Observational Uncovering Technique (SCOUT) framework to explore patient characteristics that may be associated with the individualized optimal dose of lorlatinib. A previously published population pharmacokinetic model was used to estimate individual pharmacokinetic parameters and individualized optimal doses, targeting a trough concentration of 150 ng/mL 3 weeks after treatment initiation. Associations between the estimated optimal dose and patient covariates were examined through exploratory visualization and quantified using multivariate linear regression with Lindeman–Merenda–Gold relative importance analysis. Individualized optimal doses, successfully estimated in all 330 patients, exhibited a wide distribution (mean ± SD: 194 ± 134 mg/day). The exploratory analyses suggested that older age, female sex, and greater renal impairment were associated with lower estimated optimal doses, with sex explaining the largest proportion of the variance. By reframing the covariate analysis around the dose rather than pharmacokinetic parameters, this study provides a clinically interpretable perspective on interindividual variability in dosing requirements. Although exploratory and hypothesis‐generating, this study lays a quantitative foundation for future evaluations of precision dosing strategies for lorlatinib.
ABSTRACT Inter‐individual variability in drug exposure can lead to ineffective therapy or toxicity, particularly in vulnerable patients. Routine model‐informed precision dosing (MIPD) relies mainly on population pharmacokinetic (PopPK) models, Bayesian forecasting, and therapeutic drug monitoring (TDM). Physiologically based pharmacokinetic (PBPK) modeling has a complementary role in extrapolation to special populations, evaluation of drug–drug interactions, and simulation of unobserved scenarios, but it remains uncommon for bedside dose adjustment. Machine learning (ML) may provide additional support when critical illness, organ dysfunction, or polypharmacy creates complex nonlinear relationships. The increasing availability of electronic health records, real‐world data, and high‐dimensional biomedical datasets has accelerated interest in ML for precision dosing. ML can support exposure prediction, limited‐sampling AUC estimation, outcome‐risk stratification, and monitoring prioritization. Its clinical role depends on the decision target, the availability of concentration data, and whether pharmacological constraints are retained within the modeling architecture. This review covers direct prediction models, limited‐sampling AUC estimators, hybrid pharmacometric‐ML models, neural ordinary differential equation approaches, and reinforcement learning for sequential dose decisions. We examine data sources, feature engineering, validation, reporting, and governance across glycopeptides, beta‐lactams, polymyxins, triazole antifungals, and immunosuppressants. We also describe a staged clinical workflow in which concentration measurements and mechanistic models are used to calibrate data‐driven predictions, and we identify the evidence needed for clinical implementation.
ABSTRACT The target‐engagement predictor, implemented within the InSilicoTrials cloud‐based, AI‐powered platform for model‐informed drug discovery and development, enables early exploration of time‐dependent pharmacological responses. Predictions can be generated using inputs ranging from molecular structure to in vitro, animal, and clinical data. Scaling and translation across species are supported, accounting for differences in body size and plasma protein binding. This tutorial demonstrates the application of the target‐engagement predictor for benchmarking and decision‐support in the evaluation of new drug candidates. Pharmacokinetic (PK) and pharmacodynamic (PD) data for carvedilol, ivabradine, and sunitinib were used as illustrative examples. The ANDROMEDA and PLATO software were used to predict human clinical PK and target affinity parameters, respectively, based on molecular SMILES representations of the drugs. These predictions were subsequently incorporated into longitudinal PKPD simulation models. User‐friendly input wizards and interactive results dashboards were developed to facilitate clinical trial simulation for broader project teams, without requiring programming expertise. Predictions derived from molecular structure were compared with data generated from in vitro, rat, and clinical experimental settings. The target‐engagement predictor enables integration of heterogenous data sources to support drug development decisions, including early benefit–risk assessment and dose selection for Phase 1 studies. Direct prediction based on molecular structure alone represents a promising approach for early‐stage screening and evaluation of new drug candidates.
ABSTRACT Inverse problems in pharmacometrics and quantitative systems pharmacology (QSP) often involve sparse, noisy data, limited measurable states, and complex dynamical systems. Traditional parameter estimation methods can struggle with ill‐posed problems, stiffness, discontinuities, and gray‐box scenarios where only part of the system dynamics is known or observable. This tutorial provides an end‐to‐end, reproducible workflow for applying physics‐informed neural networks (PINNs) to PK/PD/QSP inverse problems and gray‐box discovery tasks. We present three worked case studies: (1) constant parameter recovery in a three‐state compartmental system, (2) gray‐box discovery of an unknown right‐hand‐side function, and (3) inference of a time‐varying chemotherapy efficacy function under partial observation. For each, we provide implementation guidance, validation checks, and practical advice on collocation design, feature expansion, optimizer selection, numerical precision, constraint enforcement, loss weighting, and residual‐based attention. We also discuss when to consider newer architectures such as Kolmogorov–Arnold networks (KANs) as drop‐in alternatives to standard multilayer perceptrons. All code and notebooks are provided through our PhINs library, short for Pharmacometrics‐Informed Networks, and are publicly available at https://github.com/NazAhmadi/PhINs/.
