(93b) Transformer-Based Hybrid PBPK Modeling: A Scalable Framework for Drug Kinetics

The success of drug development hinges not only on the therapeutic efficacy of candidate molecules but also on their pharmacokinetic (PK) profiles, which dictate how the body absorbs, distributes, metabolizes, and excretes the drug. Physiologically Based Pharmacokinetic (PBPK) modeling offers a powerful mechanistic framework for simulating drug behavior across tissues and organs, yet its application is often constrained by the need for experimentally derived PK parameters. These requirements pose a significant bottleneck in the early stages of drug screening, where large libraries of molecules must be evaluated quickly and cost-effectively [1]. In response to this challenge, we present a hybrid modeling framework that combines the predictive power of transformer-based neural networks with the mechanistic rigor of PBPK simulations, offering a scalable solution for early-stage drug kinetics analysis [2]. Our approach leverages a 12-layer encoder-only transformer model pre-trained on over a billion SMILES strings from chemical databases and fine-tuned to predict key in vitro pharmacokinetic parameters. These include intrinsic clearance, renal clearance, fraction unbound in plasma, and intestinal permeability. Unlike traditional models that rely on molecular descriptors or extensive feature engineering, our transformer directly learns structure-property relationships from raw molecular sequences using masked language modeling. The resulting molecular embeddings capture both local and global chemical contexts, allowing for accurate property prediction across a broad range of molecules [3,4].

The predicted pharmacokinetic properties are used as inputs in a whole-body PBPK model, which simulates drug disposition across major physiological compartments, including liver, kidney, brain, muscle, adipose tissue, and plasma. The model incorporates anatomical and physiological parameters such as blood flow rates, organ volumes, and tissue composition, and follows mass balance equations under perfusion-limited assumptions. Partition coefficients are calculated using mechanistic relations, embedding domain knowledge directly into the simulation and reducing the burden on data-driven components. We validated the transformer-based hybrid model on six well-characterized drugs—acetaminophen, L-arginine, riboflavin, diazepam, meloxicam, and midazolam—by comparing simulated plasma concentration-time profiles with clinical data. The model achieved area-under-the-curve (AUC) and C_max predictions within a 1.35-fold error margin for all compounds, demonstrating strong agreement with observed pharmacokinetic outcomes. These results confirm that transformer-derived inputs can enable accurate PBPK simulations even when experimental measurements are unavailable. To further demonstrate the practical utility of our framework, we applied it to a case study involving 142 HIV Integrase 1 inhibitors, previously identified through high-throughput screening for activity and ADME properties. Using only their SMILES strings, we predicted their PK parameters and simulated concentration-time profiles across all compartments. The framework successfully distinguished compounds with favorable systemic exposure, providing a more informed basis for prioritizing candidates for further testing.

Our transformer-based hybrid PBPK modeling framework significantly reduces reliance on wet-lab experimentation while preserving the interpretability and physiological insight of mechanistic models. It enables rapid, in silico evaluation of large molecular libraries and bridges the gap between deep learning and pharmacokinetic modeling. By integrating data-driven prediction with mechanistic simulation, the framework enhances early-phase drug development, minimizes late-stage failures, and accelerates the transition from discovery to preclinical testing. This work showcases how transformer-based property prediction can be effectively combined with PBPK modeling to enable scalable, mechanistically informed drug screening. The resulting transformer-based hybrid framework offers a new paradigm for pharmacokinetic analysis—one that is computationally efficient, scientifically robust, and tailored for modern drug discovery challenges.

References:

[1] G.P. Nolan, What’s wrong with drug screening today, Nat Chem Biol 3 (2007) 187–191. https://doi.org/10.1038/nchembio0407-187.

[2] S.A. Peters, GENERIC WHOLE‐BODY PHYSIOLOGICALLY BASED PHARMACOKINETIC MODELING, in: Physiologically Based Pharmacokinetic (PBPK) Modeling and Simulations, Wiley, 2021: pp. 217–223. https://doi.org/10.1002/9781119497813.ch7.

[3] A. Khambhawala, C.H. Lee, S. Pahari, J.S.-I. Kwon, Minimizing late-stage failure in drug development with transformer models: Enhancing drug screening and pharmacokinetic predictions, Chemical Engineering Journal (2025) 160423. https://doi.org/https://doi.org/10.1016/j.cej.2025.160423.

[4] A. Khambhawala, C.H. Lee, S. Pahari, P. Nancarrow, N.A. Jabbar, M.M. El-Halwagi, J.S.-I. Kwon, Advanced transformer models for structure-property relationship predictions of ionic liquid melting points, Chemical Engineering Journal 503 (2025) 158578. https://doi.org/https://doi.org/10.1016/j.cej.2024.158578.