(S-093) Machine Learning-Integrated Physiology-Based Pharmacokinetic Modeling for Optimizing Nanoparticle Design

Joseph Cave – Mathematics in Medicine Program – Weill Cornell Medicine; Houston Methodist Research Institute; Carmine Schiavone – Houston Methodist Research Institute; University of Naples Federico II; Zhihui Wang – Houston Methodist Research Institute; Weill Cornell Medicine; Vittorio Cristini – Houston Methodist Research Institute; Weill Cornell Medicine; Prashant Dogra – Weill Cornell Medicine; Houston Methodist Research Institute

Objectives: Since the advent of liposomal Doxorubicin (DOXIL®) as a clinical cancer treatment[1], nanomedicine has advanced significantly, allowing for precise control over nanoparticle (NP) properties such as size, morphology, zeta potential (ζ), surface chemistry, and composition[2]. Yet, despite the potential of NP-based therapies, their clinical implementation remains modest[3], primarily attributed to the rapid clearance by the liver[4] and spleen[5], leading to suboptimal tumor accumulation and safety concerns from unintended exposure. To this end, we present a machine learning (ML)–driven framework that predicts the mechanistic parameters of a minimal physiologically-based pharmacokinetic (PBPK) model based on NP design and experimental conditions, thereby guiding the rational design of NPs for enhanced bioaccumulation.

Methods: We compiled a literature-derived dataset of 346 whole-body NP biodistribution profiles in tumor-bearing mice, encompassing a range of NP physicochemical properties and experimental conditions. An eight-compartment minimal PBPK model (plasma, liver, spleen, lungs, kidneys, heart, tumor, clearance pathways)[6] was fitted to each study using MATLAB’s ode15s solver to estimate compartmental unbound fractions (f_i), excretion rates (k_(e,i)), and initial plasma concentration (c_(p,0)).
Nineteen multi-target regression models—including linear, clustering, decision tree, boosting ensemble, SVM, and neural network approaches—were then trained on these derived PBPK parameters via 10-fold nested cross validation (80/20 train/test split) to ensure generalizability. Subsequently, the top five models underwent SHAP and Partial Dependence Plot (PDP) analyses to quantify the influence of individual NP features on each PBPK parameter and extract explicit mathematical relationships for NP design.

Results: Literature survey: One of the largest harmonized preclinical NP biodistribution datasets, spanning seven physicochemical properties (size, ζ, shape, composition, surface coating, targeting ligand, drug loading) and three experimental conditions (tumor cell line, murine strain, biodistribution analysis method).

Predictive performance: Tree and boosting based regressors outperformed other architectures. The top gradient boosting model (CatBoost) achieved average RMSEs of 0.277 for f_i, 0.342 for k_(e,i), and 0.286 for c_(p,0).

Feature importance: SHAP analysis identified NP size, ζ, shape, and tumor cell line as the most influential predictors across all PBPK outputs. NP size, surface charge, and targeting ligand were especially critical for tumor accumulation parameters.

Design rules: PDPs quantified empirical mathematical relationships between NP size and PBPK parameters, providing quantitative guidelines to rationalize NP design for optimized in vivo behavior.

Conclusions: We introduce a robust, interpretable ML–PBPK pipeline that predicts key PK parameters from NP design features, enabling the rational engineering of NPs with tailored biodistribution. By elucidating how physicochemical properties govern tissue-specific accumulation and clearance, we provide actionable design rules to enhance the therapeutic efficacy and safety of nanomedicines.

Citations: [1] a)A. Gabizon, R. Isacson, E. Libson, B. Kaufman, B. Uziely, R. Catane, C. G. Ben-Dor, E. Rabello, Y. Cass, T. Peretz, et al., Acta Oncol 1994, 33, 779; b)J. Patel, Journal of Oncology Pharmacy Practice 1996, 2, 201.
[2] A. A. Yetisgin, S. Cetinel, M. Zuvin, A. Kosar, O. Kutlu, Molecules 2020, 25, 2193.
[3] a)M. J. Mitchell, M. M. Billingsley, R. M. Haley, M. E. Wechsler, N. A. Peppas, R. Langer, Nature Reviews Drug Discovery 2021, 20, 101; b)D. Bobo, K. J. Robinson, J. Islam, K. J. Thurecht, S. R. Corrie, Pharm Res 2016, 33, 2373.
[4] K. M. Tsoi, S. A. MacParland, X. Z. Ma, V. N. Spetzler, J. Echeverri, B. Ouyang, S. M. Fadel, E. A. Sykes, N. Goldaracena, J. M. Kaths, J. B. Conneely, B. A. Alman, M. Selzner, M. A. Ostrowski, O. A. Adeyi, A. Zilman, I. D. McGilvray, W. C. Chan, Nat Mater 2016, 15, 1212.
[5] M. Cataldi, C. Vigliotti, T. Mosca, M. Cammarota, D. Capone, Int J Mol Sci 2017, 18.
[6] M. M. Parrot, J. Cave, M. Pelaez, H. Ghandehari, P. Dogra, V. K. Yellepeddi, medRxiv 2024, 2024.09.18.24313941.

Keywords: Machine learning, Nanoparticles, Physiology-Based Pharmacokinetic (PBPK) Modeling