Editorial: Pharmacokinetics Modeling in the Artificial Intelligence Era

Simone Brogi^1*Patric Schyman²Vishal Babu Siramshetty³

• ¹Dept of Biotechnology, Chemistry and Pharmacy, Department of Pharmacy, University of Pisa, Pisa, Italy
• ²Eikon Therapeutics Inc, Hayward, United States
• ³Genentech Inc, South San Francisco, United States

According to the significant increase in using computational methods in drug discovery and development, the review article authored by Yao and colleagues highlighted, by a bibliometric analysis, the impact of population pharmacokinetic (PPK) model in the current scientific research focused on the drug discovery campaign. The authors, using a systematic approach, collected research and review articles to explore the usage of PKK model in a definite period of time . It is worth mentioning that this is the first comprehensive paper regarding the advancement of developing this type of model. The systematic search was conducted considering the Web of Science (WoS) database, and the collected data were analyzed utilizing different tools (i.e., CiteSpace, Bibliometrix R, Microsoft Excel). In the period considered, the authors indicated that 6,125 papers (128,856 citations) focused on PKK models were published. An in-depth analysis revealed that relevant publications experienced a typical annual growth rate of over 10%, indicating a growing interest in this particular field. Research outputs are predominantly found in North America, Western Europe, and East Asia, with the USA at the forefront with 2,340 publications and boasting the highest H-index of 93 and total citations of 54,965, whereas the subject categories for most publications are Pharmacology and Pharmacy. Interestingly, authors found that machine learning, voriconazole, tacrolimus, critically ill patients, dose optimization, external evaluation, polymyxin b, model-informed precision dosing, and extracorporeal membrane oxygenation are current research hotspots and future research trends (Yao et al., 2025).Wu and coworkers, in a research article, reported the use of AI to establish the pharmacokinetic/pharmacodynamic (PK/PD) threshold values for sitafloxacin (STX) effectiveness against targeted pathogens to inform the development of clinical breakpoints for antimicrobial susceptibility testing. This issue is pivotal in clinical settings to provide effective therapeutic interventions. In particular, a PPK model was developed, considering 342 individuals, to calculate the dosing-regimen-dependent PK factors of STX-infected patients, which were combined with in vitro PD data and PK/PD target data. The cumulative fraction of response (CFR) and probabilities of attainment (PTAs) values for different STX dosing regimens against S. pneumoniae, S. aureus, E. coli, K. pneumoniae, and P. aeruginosa were determined using the Monte Carlo method. Results indicated that STX can be used against the mentioned bacterial species/strains at different doses (50 mg/q12 h, 100 mg/q24 h, and 100 mg/q12 h) with effective outcomes against S. pneumoniae-related infections, whereas the efficacy against P. aeruginosa-related infections should be further confirmed (Wu et al., 2025).Another significant issue in the drug discovery field is related to potential drug-drug interaction. In this context, Yu and coworkers developed a multimodal feature fusion model using bi-directional gated recurrent units (BiGRU) and Transformer (BiGGT) to retrieve relevant drug-drug interactions in clinical practice. This model was conceived taking into account the drug molecule topology structure, graph information, and medical corpus (Word2vec, Mol2vec, and GCN methods were used, respectively). Notably, the BiGGT-based model is a superior approach in extracting information related to drug-drug interactions compared with the current approaches. The model showed a precision of 78.22%, indicating the feasibility of applying the multimodal deep learning (DL) approach in BiGGT model for improving the performance of extracting drug-drug interactions data (Yu et al., 2024).Morales and colleagues, by developing a ML-based model, demonstrated the possibility of effectively determining unbound drug bioavailability in the brain. The determination of brain bioavailability is crucial for developing drug candidates to treat diseases affecting the central nervous system (CNS). To this end, the authors used a lot of PK parameters to determine the brain bioavailability of drugs. These data were included in a manually curated dataset comprising 157 different molecules collected from the literature. The obtained dataset was divided into training and test subsets through the use of a clustering method. A refined dataset was used to train additional models after known substrates of Pgp and/or Breast Cancer Resistance Protein (BCRP) had been removed from the original dataset. A variety of supervised ML algorithms have been evaluated, including Gradient Boosting Machine (GBM), Support Vector Machine (SVM), Linear Discriminant Analysis (LDA), Extreme Gradient Boosting (EGB), k-nearest neighbors (kNN), DL, Random Forest (RF), and classificatory Partial Least Squares (cPLS). The development of the models adhered to proper procedures for developing predictive quantitative structure-activity relationships (QSAR). The results indicated that the model developed using the EGB method was the best performing tool in the study, with an accuracy of 85.1% for the test set. Similar results were achieved conducting a prospective validation approach in which a subset was used to compare the estimated drug brain bioavailability with experimental findings. The proposed computational approach demonstrated applicability in real-world settings, and the dataset utilized along with the molecular descriptors and code are publicly available within the article to be used by the scientific community to expand and improve the performance of the model, aiding the drug discovery pipeline (Morales et al., 2024).The last published article authored by Liu and collaborators described the application of computerbased methods to develop a tool able to identify chemical agents that could cause toxidromes (undesired features provoked by toxicological issues emerging from medical treatments), allowing a fast evaluation of potential toxicity risks of drug treatments. The authors used a DL-based approach for mapping chemicals with their possible toxidromes based on communicative message passing neural network (CMPNN) algorithms. In this way, it is possible to consider several different small datasets for model training with respect to classical deep neural networks. CMPNN-based ensemble learning demonstrated the best performance in predicting toxidrome-related chemicals with respect to a single CMPNN model. Considering toxidromes for which there is not sufficient experimental data or established molecular mechanism for the model training step, the authors used a similarity-based ensemble strategy to improve the information to generate AI-based models. Finally, a web-based interface was generated to access the data and toolset of this study (https://toxidrome.bhsai.org/) (Liu et al., 2024).We would like to thank the Frontiers editorial staff, authors, co-authors, and reviewers for their valuable contributions to this Research Topic. Their efforts made it successful. This topic is a valuable resource for researchers and students, and we anticipate that it will promote the development and application of AI-based tools and in silico models for drug discovery regarding PK issues. The Research Topic is freely consultable at the following link https://www.frontiersin.org/research-topics/56980/pharmacokinetics-modeling-in-the-artificial-intelligence-era/articles.