Artificial intelligence for toxicokinetic parameterization in environmental exposure assessment

Experimental and curated data sources

AI-based TK parameterization depends on high-quality data that integrate chemical structure, biological activity, and measured or inferred ADME/TK endpoints. The common principle of “garbage in, garbage out” is particularly relevant because prediction errors may arise not only from model limitations but also from noisy, limited, inconsistent, or poorly curated input data^[13]. Compared with pharmaceutical ADME datasets, environmental TK datasets are often more heterogeneous because of the diverse chemical structures, population-specific TK profiles, and limited experimental coverage^[14]. Inconsistent experimental designs, endpoint definitions, and study conditions can substantially reduce cross-study comparability.

TK data are typically obtained from in vivo experiments and supplemented by in vitro ADME assays^[15]. In vivo studies provide time–concentration profiles in plasma, blood, urine, feces, and tissues following oral, intravenous, dermal, inhalation, or other exposure routes. These data can be used to estimate TK parameters such as clearance, half-life, bioavailability, volume of distribution, and tissue partition coefficient^[16]. A major strength of in vivo TK studies is that they can reflect the integrated TK behavior of a given chemical in a biological system rather than considering specific TK characteristics separately^[17]. However, as noted above, these datasets are often heterogeneous and limited for generalization because reported TK values may vary by species, sex, dose, sample size, analytical method, and model-fitting assumptions. For example, the mean half-life of perfluorooctanoic acid (PFOA) was reported to be approximately 0.15-0.19 days in female rats, while 1.6-18 days in males^[18]. In practice, this variability may influence reported TK values as strongly as chemical structure itself^[19]. Moreover, due to the interspecies ADME differences, the TK parameters from in vivo animal experiments may cause systematic bias when extrapolated for human exposure assessment. Therefore, when using TK parameters from in vivo experiments for AI/ML modeling, record these sources of variability as metadata for model development and interpretation.

Compared with in vivo experimental data, in vitro ADME assays facilitate standardization of experimental settings and support high-throughput chemical testing. Common endpoints of in vitro assays may include intestinal permeability, plasma protein binding, fraction unbound, intrinsic clearance in microsomes or hepatocytes, metabolic stability, and transporter-related activity^[20]. Compared with animal-based in vivo tests for TK parameters, a key strength of in vitro assays is the adoption of human cells for testing, which can reduce interspecies TK variability and improve the reliability of results for human exposure assessment^[21]. Moreover, as in vitro assays can quantify individual TK processes, they can support targeted screening of specific TK parameters across diverse chemicals, which greatly benefits AI/ML modeling and structural generalization for data-poor chemicals and chemical families, thereby accelerating exposure characterization of emerging environmental chemicals. Nevertheless, it needs to be careful when merging in vivo and in vitro TK data into a single database due to the two systems reflecting different biological contexts and exposure conditions. In contrast to TK parameters derived from in vivo experiments, in vitro outcomes such as effective doses have to be extrapolated to in vivo equivalent doses via a reverse dosimetry approach, which introduces extra uncertainty and variability into the compiled database^[6].

Beyond individual experiments, AI/ML-based TK parameterization generally relies on curated databases that integrate measurements across multiple compounds, studies, and experimental conditions. The quantity and quality of these data can substantially impact the prediction ACC and model robustness^[22]. A representative example is the U.S. EPA’s httk R package, which includes human in vitro TK parameters (plasma protein binding and hepatic clearance), structure-derived physicochemical properties, and species-specific physiological data^[23]. Combined with a generic physiologically-based toxicokinetic (PBTK) model for in vitro- in vivo extrapolation, researchers can predict in vivo TK profiles or parameters [e.g., maximum concentration (C_max) and area under concentration curves over 24 h (AUC_24h), mean concentration (C_mean)] for diverse chemicals in humans and animals^[23,24]. Additional resources, such as PubChem and OPERA, can support chemical standardization, descriptor generation, and access to physicochemical and ADME-related properties^[25,26].

Besides these general-purpose resources, class-specific databases have also been developed for emerging contaminants whose physicochemical and toxicokinetic characteristics may not be adequately represented by conventional chemical datasets. For example, microplastic/nanoparticle TK profiles can be influenced not only by chemical composition, but also by particle-specific properties, such as size, shape, and surface chemistry^[27]. General-purpose chemical databases or molecular representations cannot adequately capture the particle-specific concentration-time curves and the determinants of ADME properties, thereby reducing the transferability and mechanistic relevance of AI-based TK parameterization for this class of microplastics. To address this limitation, a few databases tailored to microplastics/nanoparticles have been established in recent years, facilitating AI/ML development for TK estimation and improving explainability^[28-30].

In practice, researchers often need to curate and integrate datasets from multiple studies to meet specific modeling objectives. Before data merging, chemical identifiers, endpoint definitions, measurement units, species, sex, life stage, exposure route, dose, sampling time, and other biological and experimental variables should be harmonized to ensure that records represent comparable biological and chemical quantities. Following harmonization, numerical features and continuous TK endpoints may require transformation and scaling because real-world datasets often contain variables with substantially different distributions and magnitudes. Without appropriate scaling, variables with larger numerical ranges may disproportionately influence model optimization and reduce prediction reliability, particularly for scale-sensitive algorithms, such as k-nearest neighbors (KNNs), support vector machines (SVMs), and neural networks. Common approaches include min-max normalization and z-score standardization^[31]. Min–max normalization rescales values to a predefined range, whereas z-score standardization centers each feature around its mean and scales it by its standard deviation. Additional preprocessing strategies, including logarithmic transformation, robust scaling, and categorical variable encoding, may also be applied depending on the dataset characteristics and modeling objectives^[31]. A proper scaling strategy can help ML algorithms converge faster during training, avoid bias toward large magnitude variables, improve numerical stability, and produce more reliable predictions^[32].

Input representations and descriptors

The predictive performance of AI/ML models depends not only on data availability, but also on how chemical and biological information is represented for model development. In AI-based ADME/TK modeling, input representations convert chemicals into machine-readable formats, whereas molecular descriptors provide quantitative variables derived from these representations to encode structural and physicochemical properties^[33]. The choice of input representation can substantially influence prediction performance and model transferability, as different molecular representations may capture distinct structural information and perform differently across datasets and endpoints^[34]. Major input representations and descriptors are summarized in Table 1.

Traditional molecular descriptors remain the most widely used inputs for TK-related prediction. These descriptors typically include molecular weight, logP/logD, pKa, hydrogen bond donors and acceptors, topological polar surface area, rotatable bonds, formal charge, and solubility. Their main advantage is interpretability, since many of these variables are mechanistically related to the key ADME properties, such as membrane permeability, tissue binding, plasma protein binding, metabolic clearance, and excretion behavior^[35]. In addition to descriptors, molecular fingerprints, such as molecular access system (MACCS) keys and extended-connectivity fingerprints (ECFP), are frequently used to encode structural fragments and substructures into a machine-readable form^[42]. These representations are particularly useful for nonlinear machine-learning models because they capture local structural patterns efficiently, although they are generally less interpretable than descriptor-based inputs and may be more sensitive to chemical-domain mismatch.

More recently, graph-based representations have become attractive because they model molecules directly as atoms and bonds, without requiring predefined descriptors. In this framework, graph neural networks (GNNs) can learn structure-property relationships from molecular topology and local chemical environments, potentially capturing higher-order structural features relevant to ADME endpoints^[43]. For example, Ng and Lu evaluated GNNs combined with transfer learning for oral bioavailability prediction, achieving a final average ACC of 0.797, an F1 score of 0.840, and an area under the receiver operating characteristic curve (AUC-ROC) of 0.867, which outperformed previous studies using traditional molecular descriptors and fingerprints with the same test dataset^[44].

In addition, biological context variables are also important for TK prediction. Factors such as tissue composition, gastrointestinal physiology, renal function, sex, age, and life stage can strongly influence ADME behavior^[21,45]. For this reason, hybrid input strategies that combine chemical descriptors with biological and experimental metadata are often more informative than structure-only representations, particularly when the goal is to support physiologically based pharmacokinetic (PBPK) modeling and exposure assessment.

AI/ML modeling and evaluation

Following the translation of chemical structure and physicochemical features into numerical representations, AI/ML models can be developed to predict ADME- and TK-related endpoints. The current algorithmic landscape falls into two major paradigms: classical ML and DL. The representative model architectures are summarized in Figure 2.

Figure 2. Representative AI/ML model architectures used for toxicokinetic parameterization [Created in BioRender. Zhang, Z. (2026) https://BioRender.com/mdf4f4m]. AI/ML: Artificial intelligence and machine learning; SMILES: simplified molecular-input line-entry system; SELFIES: self-referencing embedded strings.

Among classical ML methods, ensemble tree-based algorithms are widely applied in ADME/TK prediction. Random forests (RFs) are ensembles of decorrelated decision trees trained on bootstrap samples with random feature subsets. They offer robust performance, natural resistance to overfitting in high-dimensional spaces, and built-in feature importance estimation^[46]. Gradient-boosted decision trees (e.g., XGBoost, LightGBM, and CatBoost) construct sequential ensembles where each successive tree corrects the residual errors of its predecessor^[47]. These models typically achieve superior ACC on moderate-sized datasets, though they are highly sensitive to model tuning. SVMs with radial basis function or Tanimoto kernels remain competitive for regression on continuous TK endpoints, particularly when training sets are small and high-dimensional^[48]. Gaussian process regression provides a principled Bayesian framework yielding both point predictions and predictive uncertainty estimates, though its cubic scaling with dataset size limits applicability to large training corpora^[49]. The KNN algorithm is conceptually straightforward but foundational. It underpins the US EPA’s OPERA suite, where it combines property forecasting with AD assessments based on local chemical similarity^[50]. Overall, these classical algorithms pair naturally with the engineered feature representations and remain the algorithms of choice when dataset sizes are modest (hundreds to low thousands of compounds) or when maximal interpretability is required.

The DL paradigm extends beyond classical methods by learning hierarchical feature representations directly from molecular inputs. Different neural-network structures can be paired with different molecular representations, such as fixed-length descriptors or fingerprints, molecular graphs, and sequence-based representations^[33,34]. Early deep-learning applications to ADME/TK prediction commonly used feedforward neural networks (FNNs), including multilayer artificial neural networks (ANNs) and their deeper variants, deep neural networks (DNNs), which typically operate on fixed-length molecular descriptors or fingerprints^[51]. These models retained the descriptor-based input pipeline of classical ML, but replaced linear or shallow nonlinear estimators with multilayer architectures capable of capturing higher-order interactions among descriptors. Convolutional neural networks (CNNs) use convolutional operations to learn local patterns from structured molecular encodings and can support automatic feature extraction^[52]. Using graph-based descriptors as we discussed in the last section, GNN architectures aggregate local atomic environments into expressive molecular embeddings through iterative message passing and attention-weighted readout functions^[53]. For sequence-based molecular representations, transformer-based models such as ChemBERTa and MoLFormer process sequence-based simplified molecular-input line-entry system (SMILES) inputs through multi-head self-attention layers, capturing long-range token dependencies that recurrent architectures handle less effectively^[54,55]. Additionally, these models are not always applied independently and can be integrated into hybrid models to exploit complementary molecular representations and learning strategies. For example, Limbu et al. developed a hybrid CNN-FNN model (HNN) to predict the toxicity of thousands of chemicals^[56]. The HNN model consisted of a CNN processing one-hot-encoded SMILES and an FNN processing molecular descriptors^[56]. Such integration allows structure-derived features and engineered physicochemical descriptors to contribute jointly to prediction and may improve model robustness when information from one representation is limited.

The choice of modeling strategy should ultimately depend on dataset size, endpoint characteristics, and the required balance between ACC and interpretability. As schematically illustrated in Figure 2, simpler models are often favored when datasets are small or when greater transparency is required, whereas more flexible deep-learning models may improve predictive performance when sufficiently large and diverse datasets are available^[57]. However, this increase in flexibility may also reduce interpretability and increase the risk of overfitting or poor extrapolation to novel chemical classes. Based on the task type (classification or regression), different metrics for model evaluation were used. Continuous endpoints, such as intrinsic clearance, fraction unbound, or volume of distribution, are commonly assessed using the coefficient of determination (R²), root mean squared error (RMSE), and mean absolute error (MAE). Categorical endpoints, such as CYP inhibition or substrate classification, are typically evaluated using ACC, sensitivity, specificity, the AUC-ROC, and the Matthews correlation coefficient (MCC).

After training the models with an existing database and tuning it for good performance, the selected models still demand validation at different levels: internal cross-validation, external hold-out testing, and an AD assessment^[58]. Internal validation through k-fold cross-validation (typically 5-fold or 10-fold) provides a baseline estimate of model performance but can yield optimistically biased results when applied to datasets with structural redundancy or activity^[59,60]. External hold-out testing provides the most informative measure of true predictive ability. This process evaluates the trained model on an entirely independent dataset withheld from all stages of model development. Furthermore, AD assessment identifies the specific region of chemical space where a model’s predictions are considered reliable. This assessment is a mandatory component of the Organisation for Economic Co-operation and Development (OECD) validation principles for regulatory QSAR models^[61]. It is especially critical when applying models to the highly diverse structural landscape of environmental chemicals. Common AD methods include leverage-based approaches (Williams plots), distance-to-model metrics in descriptor space, local density estimation, and conformal prediction frameworks that provide prediction-specific confidence levels^[62]. The US EPA’s OPERA suite exemplifies the integration of AD assessment into routine ADME prediction by providing per-chemical reliability indices alongside property estimates, enabling users to distinguish high-confidence predictions from extrapolations^[50].

Finally, the selected model will be examined for mechanistic interpretation, uncertainty quantification, and downstream integration strategies with the aim of translating these validated ADME predictions into actionable exposure estimates.

AI for absorption

Absorption is the process by which a chemical enters systemic circulation following external exposure. Environmental chemicals can be mainly absorbed through oral ingestion, inhalation, and dermal contact. Among these routes, oral absorption has been the most commonly studied in current AI/ML-based TK parameterization, largely because oral exposure is the most common route for conventional drugs and many tested chemicals, resulting in greater data availability and more established assay systems. Typically, oral absorption is characterized using parameters such as intestinal permeability, oral bioavailability, and the absorption rate constants. Oral bioavailability has received considerable attention, particularly from pharmaceutical research, as this parameter is a key determinant of the success or failure of candidate drugs. Multiple predictive approaches, including QSAR models, PBPK-based models, and more recent AI/ML algorithms, along with various software, have been developed for predicting oral bioavailability^[63]. Similarly, researchers have developed a few models for other oral absorption-related parameters [Table 2]. For example, Ng and Lu evaluated a GNN combined with transfer learning to predict oral bioavailability. In their framework, the model was first pretrained on a solubility prediction task using graph-based molecular representations, and then fine-tuned for oral bioavailability classification. The best transfer-learning model achieved an ACC of 0.797, an F1 score of 0.840, and an AUC-ROC of 0.867, outperforming traditional ML models using descriptor- and fingerprint-based representations^[44]. Shapley Additive exPlanations (SHAP) analysis further reported the most important factor in the quantitative estimation of drug-likeness (QED)^[44]. This study illustrates the potential advantage of learned molecular representations and knowledge transfer for improving prediction when endpoint-specific datasets are limited. However, whether this improvement generalizes to structurally diverse environmental chemicals requires further validation because the AD analysis of the transfer model suggested the absence of distinctly out-of-domain compounds in the test set. In another instance, Kamiya et al. developed a lightGBM model to predict influx and efflux apparent permeability (P_app) across Caco-2 monolayers for 218 disparate chemicals by integrating in vitro permeability coefficients with 17 and 19 descriptors, respectively, such as logD, logP, topological polar surface area, molecular weight, and basic group count^[65]. Their best lightGBM models achieved correlation coefficients of 0.83-0.84 for influx and efflux P_app (log-transformed, nm/s) prediction, indicating that AI/ML models can provide reasonably accurate estimates of intestinal permeability for structurally diverse chemicals.

Besides oral absorption, inhalation and dermal absorption have received less attention in current AI/ML-based parameterization due to limited availability of curated datasets. Nevertheless, these routes are highly relevant for environmental exposure assessment, especially for volatile chemicals, aerosols, consumer product ingredients, and occupational exposures. A few studies have begun to explore AI/ML approaches for inhalation- and dermal-related TK characterization using relatively small datasets. For example, Chiu et al. developed an Extra Trees-based multitask learning model to predict the isolated perfused lung absorption rate constant (k_a,IPL, min^-1) for a dataset of 29 chemicals. To address the limited pulmonary absorption data, the model jointly learned from the larger Caco-2 and Calu-3 permeability datasets and achieved a correlation coefficient of r = 0.84 between the predicted and observed k_a,IPL values in the independent test set^[70]. Compared with single-task learning approaches, this study illustrates the advantages of multi-task learning to address limited endpoint-specific data availability by jointly learning from auxiliary tasks (i.e., Caco-2 and Calu-3 permeability in this study)^[70]. The effectiveness of such multitask learning largely depends on the biological relevance and similarity of auxiliary tasks to the target task. In another instance, Simon applied a neural ordinary differential equation (ODE) model to predict blood concentration–time profiles of two volatile organic compounds (VOCs) following inhalation exposure^[71]. In this approach, neural ODEs learn the chemical concentration changes by parameterizing the rate of change as a continuous function of time, providing a potential AI-based approach for environmental TK modeling, where chemical-specific kinetic parameters and large-scale training datasets are often limited.

For dermal exposure, Narita et al. developed a gradient boosting tree model to predict four skin permeation parameters: partition coefficients for the stratum corneum and viable epidermis/dermis (K_SC and K_VED) and diffusion coefficients in these two skin layers (D_SC and D_VED)^[68]. The predicted parameters were subsequently incorporated into a two-layer diffusion model to estimate finite-dose dermal permeation profiles. The model showed generally reasonable agreement with observed profiles for the evaluated chemicals. However, additional external validation is needed before broader application of this model in dermal pharmaceutical development and exposure assessment for topically exposed chemicals. In future applications, AI-based dermal absorption models could be used to evaluate the effectiveness of chemical neutralization and decontamination products, such as reactive skin decontamination lotion^[72]. These models could incorporate decontamination efficiency, treatment timing, chemical reactivity, and residual dermal absorption to support exposure assessment following hazardous-material incidents.

AI for distribution

Following absorption through different exposure routes, chemicals are transported from the systemic circulation into tissues, organs, and biological barriers. In exposure assessment, this process is important because toxicity is often driven more directly by target-site concentrations than by external exposure dose alone.

AI modeling for distribution mainly focuses on predicting relevant parameters, including tissue: plasma partition coefficients, apparent volume of distribution, fraction unbound in plasma, and barrier permeability [Table 3]. For example, Dawson et al. developed QSAR models to predict in vitro fraction unbound in plasma for 2,057 compounds [1,308 pharmaceuticals and 749 compounds from the U.S. EPA Toxicity Forecaster (ToxCast) program]^[73]. The best model integrated four open-source chemical descriptors, including PaDEL, OPERA, ToxPrints, and MACCS, into an RF model and achieved an R² of 0.591 with an RMSE of 0.187^[73]. This study is particularly relevant to environmental TK modeling because the predicted parameters were designed to support HTTK applications and internal dose estimation for data-poor chemicals. For highly protein-bound contaminants such as per- and polyfluoroalkyl substance (PFAS), uncertainty in predicting the unbound fraction affects the assessment of persistence and internal exposure.

Unlike general tissue partitioning, AI/ML models have also been applied to predict chemical transport across physiological barriers that regulate access to specific compartments. Blood–brain barrier (BBB) penetration is one of the most important specialized barriers, involving tight junctions, selective transport, and active efflux mechanisms. The BBB can prevent 98% of circulating molecules from entering the brain. Traditional experimental methods make it difficult to assess BBB penetration rates. AI/ML models can identify key features influencing BBB permeability, supporting TK characterization and exposure assessment. Early BBB-prediction models relied on traditional structural and physicochemical features, such as molecular weight, lipophilicity, and hydrogen-bonding properties^[85]. For example, Liu et al. developed ML and ensemble models to predict BBB permeability for 1,757 chemicals using structure-derived molecular fingerprints. They input the SMILES descriptor of chemicals into the PaDEL-Descriptor software to calculate molecular fingerprints. Combining three types of ML models, including RF, SVM, and XGBoost models, this study achieved an ACC of 0.910, an ROC-AUC of 0.957, a sensitivity (SEN) of 0.927, and a specificity of 0.867 for predicting BBB permeability^[75]. Recent advancements have applied more sophisticated DL methods, such as graph- and image-based models^[86]. For example, Nguyen et al. developed a geometric multi-color message-passing graph neural network (GMC-MPNN) for BBB permeability prediction, incorporating three-dimensional geometric information and atom-type-specific subgraphs into the graph representation^[76]. Their model achieved AUC-ROC values of 0.947 ± 0.011 and 0.9212 ± 0.0261 for BBB permeability classification on the MoleculeNet and B3DB benchmark datasets, respectively. For continuous permeability prediction using the B3DB dataset, the model achieved an RMSE of 0.5628 ± 0.0651 and a Pearson correlation coefficient of 0.6947 ± 0.0515, demonstrating the strong potential of graph-based DL for BBB-related pharmacokinetic (PK) characterization^[76]. For BBB permeability of environmental chemicals, AI/ML models can help prioritize chemicals with potential central nervous system exposure. However, current models are mainly developed using pharmaceutical compounds, and their applicability to environmental chemicals remains unclear due to differences in chemical space and exposure-relevant chemical characteristics.

Another important biological barrier that determines chemical distribution is the placenta. AI-based predictions of placental transfer have been developed to estimate maternal-to-fetal chemical distribution for both pharmaceutical compounds and environmental chemicals. Li et al. developed a QSAR model to predict TTE, defined as the cord blood-to-maternal blood concentration ratio, using structural descriptors from 51 environmental chemicals, including organochlorine pesticides, PAHs, PCBs, PBDEs, and PFASs. The model identified molecular descriptors associated with placental transfer prediction and achieved moderate predictive performance, with an adjusted R² of 0.67 for the 10-descriptor model and 0.61 for the 5-descriptor model^[78]. More recently, Guan et al. incorporated placental pathway-related features (pathway activation scores derived from a placental gene network) together with molecular descriptors to classify 307 chemicals, both pharmaceutical compounds and environmental chemicals, with high or low placental barrier crossing potential, achieving high predictive performance in external validation (ACC_test: 0.94; F1: 0.97; AUC: 1)^[80]. The study further evaluated model applicability using Euclidean distance-based chemical space analysis, showing that the prediction test set fell within the chemical space represented by the training dataset. This model showed high potential to predict the transplacental potential of environmental chemicals. However, current AI-based placental transfer models mainly predict transfer ratios or classification outcomes rather than mechanistic placental TK parameters, and the limited availability of environmentally relevant placental kinetic datasets remains a major challenge for broader application in exposure assessment.

Beyond gestation, some toxicologists have also constructed a few AI/ML models to predict relevant parameters for the chemical transmission process during lactation^[87]. During this process, chemicals in maternal circulation can partition into breast milk and subsequently contribute to infant exposure. For example, Huang et al. developed an explainable ML model to classify chemicals with high milk-transfer risk (defined as milk-to-plasma concentration ratios ≥ 1) for 375 chemicals, with an external validation set of 198 chemicals^[82]. Using molecular descriptors and fingerprints to represent physicochemical characteristics and chemical structures, respectively, the balanced RF classifier achieved the best predictive performance with an AUC of 0.87, and an ACC of 82.67% on the internal test set and 86.36% on the external test set^[82]. The SHAP analysis further identified hydrophobicity/hydrophilicity, π-electron characteristics, pH-dependent lipophilicity, charge distribution, and conformational features as the key predictors of chemical milk-transfer risks. In particular, higher fractional hydrophobic surface area increased transfer possibility, whereas greater π-electron delocalization reduced it^[82].

AI for metabolism

The chemical hazard may be determined not only by the parent compound itself, but also by the formation, persistence, and activity of its metabolites. In TK, metabolism occurs primarily in the liver. Depending on the chemical and exposure route, it can also occur in the intestine, kidney, lung, skin, and placenta^[88]. Typically, these processes are mediated by phase I and phase II enzymes, among which cytochrome P450 enzymes, UDP-glucuronosyltransferases (UGTs), sulfotransferases (SULTs), and esterases are particularly important for the metabolic fate of xenobiotics.

AI-based prediction of metabolism-related processes requires not only characterization of enzyme–chemical interactions but also quantitative estimation of metabolic capacity that determines systemic exposure [Table 4]. Current models regarding metabolism-related TK parameters have mainly focused on endpoints such as enzyme affinity or interaction, intrinsic clearance, and metabolic stability^[95]. Compared with other ADME processes, metabolism-related prediction is more strongly shaped by enzyme-specific recognition and catalytic transformation. In addition to general structural and physicochemical features, metabolism is often influenced by the accessibility of reactive sites, steric hindrance, electronic properties, and the presence of functional groups susceptible to oxidation or conjugation^[96]. A representative example is the study by Sun et al., who developed QSAR models to predict CYP activity profiles of environmental chemicals using models built on drug-like compounds and then applied them to the environmental chemical space^[89]. Using the activity/inhibition data across different CYP isoforms for 17,143 drug-like compounds from PubChem, the researchers developed an SVM model to predict the CYP isozymes and applied it to Tox21 compounds. The predictions were largely accurate for CYP1A2, CYP2C9, and CYP3A4 isozymes with AUC-ROC ranging between 0.82 and 0.84^[89].

AI methods have also been applied to more mechanistic metabolism endpoints, such as site-of-metabolism and metabolite prediction. For example, Tyzack et al. developed probabilistic classifiers for CYP450 site-of-metabolism prediction using 2D topological fingerprints, showing that machine-learning approaches can identify likely metabolic sites at the atomic level^[96]. More recently, deep-learning models have been used to predict likely metabolite structures and reaction outcomes, extending metabolism modeling beyond simple clearance-related endpoints toward explicit biotransformation prediction^[97]. Yang et al. developed D-CyPre, a deep-learning framework, to predict the metabolic sites across nine human CYP450 isoforms^[91]. This model used directed message-passing neural networks to integrate atom-, bond-, and molecular-level structural information, achieving F1 scores of 0.66-0.67 on the test set and outperformed several existing metabolism predictors^[91]. These approaches provide mechanistic insights into potential metabolic pathways and may help identify metabolites with different persistence or toxicity profiles. Nevertheless, prediction of metabolite formation remains challenging because of the diversity of enzyme systems, multiple competing metabolic pathways, and limited availability of experimentally characterized metabolite datasets.

Beyond enzyme-level activity and metabolite formation, AI/ML approaches have also been applied to predict quantitative metabolism-related TK parameters, particularly hepatic intrinsic clearance (CL_int,h), which is a key determinant of hepatic metabolism and systemic exposure in TK models. This parameter is particularly important because it reflects the intrinsic metabolic capacity of the liver independent of hepatic blood flow and plasma protein binding^[98]. It is primarily determined by the activity and abundance of hepatic drug-metabolizing enzymes and therefore represents the rate at which the liver can metabolically eliminate a chemical when delivery to the liver is not limiting. Under nonsaturating conditions, CL_int,h is commonly combined with hepatic blood flow and the fraction unbound in blood or plasma to estimate hepatic clearance in mechanistic TK and PBPK modeling. Because CL_int,h is chemical-specific but often unavailable experimentally for many environmental chemicals, it has become an important target for in silico and AI-based prediction. For example, Kamiya et al. predicted the hepatic intrinsic clearance of 212 chemicals (both environmental xenobiotics and medicines), along with absorption rate constants and the systemic circulation volumes, using 17-65 structural and physicochemical descriptors selected from 1,710 calculated descriptors^[94]. The LightGBM model achieved a correlation coefficient of 0.77 and an average absolute fold error of 3.28 for hepatic intrinsic clearance. The three predicted parameters were subsequently incorporated into a simplified PBPK model. The maximum plasma concentrations and areas under the concentration–time curve generated using the in silico-estimated parameters were correlated with those generated using traditionally determined parameters, with correlation coefficients of 0.85 and 0.80, respectively^[94]. Rather than treating model outputs as standalone predictions, this study’s workflow showed that the predicted parameters can be incorporated into a mechanistic TK or PBPK framework, where their combined impact can be evaluated against observed concentration–time data of environmental chemicals.

AI for excretion

Chemical excretion is the irreversible removal of xenobiotics or their metabolites from the body and is a key determinant of overall TK behavior^[99]. Environmental chemicals are excreted through renal and biliary pathways in the general population. Other routes may include exhalation, saliva, mucosal, and sweat, depending on the compound’s physicochemical properties^[16].

The primary excretion pathway depends on the chemical molecular structure and physicochemical features. For polar and ionizable environmental chemicals, renal excretion is typically the dominant pathway, which involves glomerular filtration, active tubular secretion, and reabsorption processes^[100]. Besides the physiological features common to the other three processes, several specific molecular determinants may influence renal clearance, such as interaction efficiencies with organic cation transporter 2 (OCT2), organic anion transporters (OAT1/3), and multidrug and toxin extrusion proteins (MATEs)^[100-103]. For example, the chemicals within the PFAS family exhibit substantial variability in renal elimination rates, driven by differences in carbon chain length, protein-binding fractions, and interactions with transporters^[104].

In contrast, some chemicals are primarily eliminated through the biliary pathway, particularly those with relatively large molecular weight (typically > 300-500 Da) and higher lipophilicity (e.g., logP > 2-3). A statistically derived threshold suggests that compounds with molecular weight above approximately 350 Da are more likely to undergo significant biliary elimination^[105]. This process requires active secretion of chemicals and/or their metabolites from hepatocytes into bile, followed by elimination in feces, which is mediated by hepatic transporters and influenced by molecular weight, polarity, and conjugation status^[106-108].

These mechanistic determinants form the basis for AI-driven prediction of excretion-related parameters. Renal excretion has received comparatively more attention, with models developed for quantitative renal clearance (CL_R) and the fraction of a chemical excreted unchanged in urine (fe). For example, Paine et al. developed 3 statistical models [partial least squares, RF, as well as classification and regression trees (CARTs)] to predict the human renal clearance rate of 349 compounds [acids (N = 59), bases (N = 124), zwitterions (N = 112), and neutral (N = 54)]^[109]. Using the molecular structures of these compounds, a total of 195 molecular descriptors were generated to characterize the key molecular features, including lipophilicity, hydrogen-bonding, size, charge/polarity, and topology. Among the tested models, the RF model showed moderate to high predictive performance across compound classes, with particularly strong performance for acids and zwitterions (r_obf² = 0.79). Top influencing descriptors included lipophilicity-based and positive charge-related descriptors^[109]. A later study developed a two-stage in silico framework for human renal excretion using datasets of 411 compounds for fe and 401 compounds for CL_R^[110]. The first model classified whether a compound was predominantly renally excreted, achieving a balanced ACC of 0.74. The subsequent models further separated compounds by renal excretion type and incorporated the fraction unbound in plasma (f_u,p) as an additional descriptor. Inclusion of measured or predicted f_u,p improved quantitative clearance prediction, with 78.6% of compounds in the high-CL_R group predicted within twofold error when predicted f_u,p was used^[110]. This example illustrates the advantage of incorporating a physiologically relevant TK parameter into structure-based renal clearance prediction rather than relying on molecular structure alone.

Compared with renal clearance, AI-based prediction of biliary excretion remains less developed, partly because quantitative human biliary-excretion data are relatively limited and transporter-mediated processes are difficult to characterize experimentally^[111]. One representative QSAR study predicted the percentage of biliary excretion of 217 compounds using molecular descriptors. A simple regression tree model generated using the CART algorithm provided the best predictive performance (MAE_test = 0.373), and model interpretation indicated that larger molecular size, ionic character, and moderate lipophilicity favored biliary excretion. The study also identified a statistically supported molecular-weight threshold of approximately 348 Da for significant biliary excretion and found poorer model performance for compounds with extreme lipophilicity (logP > 5.35) or molecular weight below approximately 280 Da.

Overall, current AI applications for excretion appear more mature for renal endpoints than for biliary elimination. Renal models benefit from relatively well-defined quantitative endpoints such as renal clearance and fraction excreted unchanged in urine, whereas biliary excretion involves sequential hepatic uptake, intracellular handling, transporter-mediated canalicular secretion, and possible enterohepatic recirculation, making a single structure–endpoint relationship more difficult to establish. In addition, most existing models were developed primarily from pharmaceutical datasets, and their generalizability to environmental chemical space remains insufficiently evaluated. Future models may therefore benefit from integrating molecular descriptors with transporter-specific information and other mechanistically relevant TK parameters rather than relying on chemical structure alone.