Leveraging machine learning models in evaluating ADMET properties for drug discovery and development doi: https://doi.org/10.5599/admet.2772 1 ADMET & DMPK 13(3) (2025) 2772; doi: https://doi.org/10.5599/admet.2772 Open Access : ISSN : 1848-7718 http://www.pub.iapchem.org/ojs/index.php/admet/index Review Leveraging machine learning models in evaluating ADMET properties for drug discovery and development Magesh Venkataraman, Gopi Chand Rao, Jeevan Karthik Madavareddi and Srinivas Rao Maddi* Department of Pharmacology, Acubiosys Private Limited, Hyderabad, Telangana, India Corresponding Author: E-mail: *srinivasm@acubiosys.com Received: April 4, 2025; Revised: June 6, 2025; Published: June 7, 2025 Abstract Background and purpose: The evaluation of ADMET properties remains a critical bottleneck in drug discovery and development, contributing significantly to the high attrition rate of drug candidates. Traditional experimental approaches are often time-consuming, cost-intensive, and limited in scalability. This review aims to investigate how recent advances in machine learning (ML) models are revolutionizing ADMET prediction by enhancing accuracy, reducing experimental burden, and accelerating decision-making during early-stage drug development. Experimental approach: This article systematically examines the current landscape of ML applications in ADMET prediction, including the types of algorithms employed, common molecular descriptors and datasets used, and model development workflows. It also explores public databases, model evaluation metrics, and regulatory considerations relevant to computational toxicology. Emphasis is placed on supervised and deep learning techniques, model validation strategies, and the challenges of data imbalance and model interpretability. Key results: ML-based models have demon- strated significant promise in predicting key ADMET endpoints, outperforming some traditional quantitative structure - activity relationship (QSAR) models. These approaches provide rapid, cost-effective, and repro- ducible alternatives that integrate seamlessly with existing drug discovery pipelines. Case studies discussed in this review illustrate the successful deployment of ML models for solubility, permeability, metabolism, and toxicity predictions. Conclusion: Machine learning has emerged as a transformative tool in ADMET prediction, offering new opportunities for early risk assessment and compound prioritization. While challenges such as data quality, algorithm transparency, and regulatory acceptance persist, continued integration of ML with experimental pharmacology holds the potential to substantially improve drug development efficiency and reduce late-stage failures. ©2025 by the authors. This article is an open-access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/4.0/). Keywords ADMET prediction, AI/ML, pharmacokinetics, computational toxicology, molecular descriptors Introduction The typical timeframe for drug discovery and development of a new drug spans from 10 to 15 years of rigorous research and testing [1]. The sheer volume of potential drug candidates renders traditional wet lab experiments impractical. However, advancements in data science over the past decade, coupled with enhanced computational capabilities, have paved the way for in silico methodologies for screening extensive drug libraries. This preliminary step, preceding preclinical studies, significantly reduces costs and expands the scope of drug discovery efforts [2]. Machine learning (ML) is a method of data analysis involving the of new algorithms and models capable of interpreting a multitude of data. Within this framework, ML techniques https://doi.org/10.5599/admet.2772 https://doi.org/10.5599/admet.2772 http://www.pub.iapchem.org/ojs/index.php/admet/index mailto:srinivasm@acubiosys.com http://creativecommons.org/licenses/by/4.0/ M. Venkataraman et al. ADMET & DMPK 13(3) (2025) 2772 2 have emerged as pivotal tools in the pharmaceutical drug discovery and development field [3]. Recent progress in ML algorithms, coupled with the accessibility of extensive proprietary and public absorption, distribution, metabolism, excretion and toxicity (ADMET) datasets, has sparked enthusiasm among academic and pharmaceutical science circles in predicting pharmacokinetic and physicochemical endpoints in early drug discovery [4]. It has been widely recognized that ADMET should be evaluated as early as possible [5]. Figure 1 illustrates the schematic representation of the ADME process for orally administered drugs, illustrating key pharmacokinetic phases including absorption in the gastrointestinal tract, metabolism in the liver, distribution via systemic circulation, and excretion through renal clearance. In silico ADME-Toxicity (ADME- Tox) evaluation models have been developed as an additional tool to assist medicinal chemists in the design and optimization of leads [6]. The majority of the problems arising during the process of drug discovery include unfavourable ADMET properties, which have been known to be a major cause of failure of the potential molecules in the drug development pipeline, contributing to large consumption of time, capital and human resources [7]. Figure 1. This figure depicts the absorption, distribution, metabolism, and excretion (ADME) process for drugs administered orally. When a tablet or capsule is ingested, it disintegrates in the gastrointestinal (GI) tract, releasing drug molecules. These molecules either dissolve for absorption or remain in a precipitated state, eventually being excreted. Absorbed drug molecules must cross the gut wall, where they may be transported back into the intestinal lumen or metabolized by enzymes. Those that successfully traverse the gut barrier enter the liver via the portal circulation. The liver plays a crucial role in drug metabolism, utilizing phase I (modification) and phase II (conjugation) enzymatic reactions to increase the hydrophilicity of xenobiotics, facilitating their elimination through the kidneys. Drugs that escape metabolism enter systemic circulation, though a portion binds to plasma proteins, limiting their bioavailability. Only the free, unbound drug and its metabolites can reach target cells and interact with biomolecules to exert therapeutic effects. Meanwhile, some drug molecules are rapidly cleared by the kidneys. The drug’s efficacy is determined by its ability to reach and maintain an optimal concentration at the site of action while navigating these physiological processes (Courtesy:NIH Bioart; Bioicons). This has increased the interest in the early-stage prediction of ADMET properties of drug candidates so that the success rate of a compound reaching the later stages of drug development can be enhanced. ML has ADMET & DMPK 13(3) (2025) 2772 Machine learning models for ADMET prediction in drug development doi: https://doi.org/10.5599/admet.2772 3 been effectively utilized to develop models and prediction tools for ADMET properties. Apart from property predictions, ML has also contributed to early phases of drug discovery, like de novo designing of chemical compounds and peptides [8]. Moreover, companies involved in clinical research have ascertained that revising the research strategies by introducing ML-based techniques has resulted in greater success rates in both preclinical and clinical trials [9]. After high-throughput screening, the chosen compounds, often referred to as hit compounds, are analysed for their biological activity via in vitro studies. The most potent compounds obtained from in vitro activity data are developed into lead compounds through a lead optimization process [10]. During the lead optimization phase, the compounds are modified to improve their bioavailability, solubility, partition coefficient, and stability as these factors can have a direct impact on the drug's therapeutic efficacy and potency [11]. The molecules with optimized ADMET properties are then further evaluated for their effectiveness using suitable animal models [12]. The optimized compounds are tested in human subjects in order to validate and confirm the potency, thera- peutic efficacy, ADMET and possible adverse drug reactions through a four-step process called clinical trials, in which each step is carried out in a varying number of human subjects in a randomized control manner [13]. This review mainly focuses on the ML-based tools used in ADMET properties. We have given a general introduction to the drug discovery process, Machine learning and DL techniques, followed by specific examples and discussion on various ML and artificial intelligence (AI) based tools for drug development. We also pointed out some notable success stories in the use of AI and ML in ADMET. 2. Fundamentals of machine learning in drug discovery 2.1. Basics of machine learning models ML starts with obtaining a suitable dataset, often from publicly available sources, and has become increasingly prominent in ADMET prediction, offering valuable tools for drug development and toxicity [14]. ML methods are generally divided into supervised and unsupervised approaches. In supervised learning, models are trained using labelled data to make predictions, such as predicting properties like pharmacokinetic (PK) properties, based on input attributes like chemical descriptors of new compounds. On the other hand, unsupervised learning aims to find patterns, structures, or relationships within a dataset without using labelled or predefined outputs. Its goal is to uncover inherent structures and insights, which are sometimes missed in supervised learning approaches because they rely on pre-defined answers [3] Common ML algorithms used in this field include supervised methods such as support vector machines, random forests, decision trees, and neural networks, as well as unsupervised approaches like Kohonen's self- organizing maps [15] (Figure 2). These methods can be applied to various types of input data, ranging from chemical structural descriptors to transcriptome analysis, enhancing prediction accuracy [16]. The selection of appropriate ML techniques depends on the characteristics of available data and the specific ADMET property being predicted [15]. The development of a robust machine learning model for ADMET predictions begins with raw data collection, which includes both labelled and unlabelled datasets. This data undergoes preprocessing, ensuring quality and consistency before being split into training and testing datasets. Various ML algorithms, such as supervised, unsupervised, and deep learning approaches, are then applied to the training data to develop predictive models. To enhance model accuracy and generalizability, feature selection and hyperparameter optimization are performed, followed by cross-validation techniques like k-fold validation. Finally, the optimized model is tested using an independent dataset to evaluate its performance based on classification and regression metrics. Figure 3 illustrates the stepwise workflow for generating an ML model, detailing each phase from data preprocessing to model evaluation. https://doi.org/10.5599/admet.2772 M. Venkataraman et al. ADMET & DMPK 13(3) (2025) 2772 4 Figure 2. Commonly used AI/ML algorithms for developing ADMET prediction models Figure 3. The diagram illustrates the stepwise workflow for generating a machine learning model, starting from raw data collection to model evaluation. Initially, raw data, which can include both labeled and unlabeled datasets, undergoes preprocessing and structuring. This iterative step ensures data quality before proceeding further. Once the data is refined, it is split into training and test sets, commonly in an 80:20 ratio of the total volume of the data. The 80 % training data is then subjected to various machine learning algorithms, including supervised, unsupervised, reinforcement learning, and deep learning approaches. Through an iterative process, the best-performing candidate model is selected. The chosen candidate model is further refined through hyperparameter optimization and feature selection, leading to the final optimized model. This model is then validated using cross-validation techniques such as 5-fold cross-validation to ensure robustness. Following validation, the model is tested using the 20 % test dataset. The predicted outcomes (y-values) are analysed, and the model's performance is evaluated using classification and regression metrics, ensuring its reliability and accuracy for real-world applications (Courtesy: Bioicons; Flaticon). 2.2. Data requirements and preprocessing The standard ML methodology starts with obtaining a suitable dataset, often from publicly available repo- sitories tailored for drug discovery. The quality of data is crucial for successful ML tasks, as it directly impacts model performance. Various databases provide pharmacokinetic and physicochemical properties, enabling robust model training and validation [4]. A comprehensive list of such databases is presented in Table 1. Data preprocessing, including cleaning, normalization, and feature selection, is essential for improving data quality ADMET & DMPK 13(3) (2025) 2772 Machine learning models for ADMET prediction in drug development doi: https://doi.org/10.5599/admet.2772 5 and reducing irrelevant or redundant information [17]. Feature quality, such as relevant, informative, and predictive a specific feature is within a dataset has been shown to be more important than feature quantity, with models trained on non-redundant data achieving higher accuracy (>80 %) compared to those trained on all features [18]. When dealing with imbalanced datasets, combining feature selection and data sampling techniques can significantly improve software defect prediction performance. Empirical results suggest that feature selection based on sampled data outperforms feature selection based on original data [19]. These findings highlight the importance of carefully considering data quality, feature selection, and handling of imbalanced datasets in ML tasks to achieve optimal model performance. Table 1. A list of databases that contain pharmacokinetic data for machine learning analyses Database Source Ref. PK-DB- pharmacokinetics database https://pk-db.com/ [20] e-Drug3D https://chemoinfo.ipmc.cnrs.fr/MOLDB/index.php [21] Drugbank https://go.drugbank.com [22] ChEMBL https://www.ebi.ac.uk/chembl/ [23] Therapeutic target database https://db.idrblab.net/ttd/ [24] CvTdb https://github.com/USEPA/CompTox-PK-CvTdb [25] PubChem https://pubchem.ncbi.nlm.nih.gov/ [26] Zinc https://zinc15.docking.org/ [27] SuperCYP https://insilico-cyp.charite.de/SuperCYPsPred/ [28] The ADMET prediction database http://modem.ucsd.edu/adme/databases/databases.htm [29] 2.3. Feature engineering in ADME-tox prediction Feature engineering plays a crucial role in improving ADMET prediction accuracy. Traditional approaches rely on fixed fingerprint, i.e. an efficient and quick means of portraying fixed-length data, ignoring the internal substructures within representations of molecules [30]. However, recent advancements involve learning task-specific features by representing molecules as graphs, where atoms are nodes and bonds are edges. Graph convolutions applied to these explicit molecular representations have achieved unprecedented accuracy in ADMET property prediction [31]. Feature selection methods can help determine relevant properties for specific classification or regression tasks, alleviating the need for time-consuming experimental assessments [32]. Filter methods are employed during the pre-processing stage to select features from the dataset without relying on any specific machine learning algorithm. These methods swiftly identify and eliminate duplicated, correlated, and redundant features, making them highly efficient in computational terms [33]. They excel at isolating individual features for evaluation, which proves beneficial when features operate independently. However, they fall short in addressing multicollinearity, as they do not mitigate the interdependencies between features. Despite their speed and cost-effectiveness, filter methods may not capture the potential performance enhancements achievable through feature combinations [34]. In a study by Ahmed and Ramakrishnan [35], correlation-based feature selection (CFS), a type of filter method to identify fundamental molecular descriptors for predicting oral bioavailability. Out of 247 physicochemical descriptors from 2279 molecules, 47 were found to be major contributors to oral bioavailability, as confirmed by the logistic algorithm with a predictive accuracy exceeding 71 %. Wrapper methods, also known as greedy algorithms, iteratively train the algorithm using subsets of features. These methods dynamically add and remove features based on insights gained during previous model training iterations [36]. Unlike filter methods, wrapper methods offer an optimal feature set for model training, leading to superior accuracy. However, their computational demands are higher compared to filter methods due to the iterative nature of the process [33]. In embedded methods, the feature selection algorithm is integrated into the learning algorithm, possessing inherent feature selection capabilities. The models combine filtering and wrapping techniques to https://doi.org/10.5599/admet.2772 https://pk-db.com/ https://chemoinfo.ipmc.cnrs.fr/MOLDB/index.php https://go.drugbank.com/ https://www.ebi.ac.uk/chembl/ https://db.idrblab.net/ttd/ https://github.com/USEPA/CompTox-PK-CvTdb https://pubchem.ncbi.nlm.nih.gov/ https://zinc15.docking.org/ https://insilico-cyp.charite.de/SuperCYPsPred/ http://modem.ucsd.edu/adme/databases/databases.htm M. Venkataraman et al. ADMET & DMPK 13(3) (2025) 2772 6 optimize feature selection. Initially, these models utilize a filter-based approach to reduce the feature space dimensionality. Subsequently, the best subset of features identified through the filter-based step is incorporated using a wrapper technique [37]. Embedded methods combine the strengths of filter and wrapper techniques while mitigating their respective drawbacks. They inherit the speed of filter methods while surpassing them in accuracy [38]. 2.4. Molecular descriptors Molecular descriptors (MD) are crucial components in in-silico research, serving as numerical representations that accurately convey the structural and physicochemical attributes of compounds based on their 1D, 2D, or 3D structures [39]. Various software tools are available for the calculation of molecular descriptors, facilitating the extraction of relevant features for predictive modelling. Table 2 provides a comprehensive list of software packages commonly used in cheminformatics and computational drug discovery for descriptor generation. These programs offer a wide array of over 5000 descriptors, encom- passing constitutional descriptors as well as more intricate 2D and 3D descriptors that capture various geometric, connectivity, and physicochemical properties [40]. The choice of molecular representation determines whether a molecule is described using experimental descriptors or theoretical descriptors. Experimental descriptors encompass all measurements obtained through experiments, such as the octanol- water partition coefficient, molar refractivity, polarizability, and various other physicochemical properties obtained through specific experimental procedures [41]. Conversely, theoretical molecular descriptors span 0D, 1D, 2D, 3D and 4D molecular descriptors, which are derived from defined chemoinformatic algorithms applied to a clear molecular representation [42] Table 2. List of software packages for the calculation of Molecular descriptors Name Organization/institution Availability RDKit GitHub https://github.com/rdkit PaDELPy University of Massachusetts Lowell https://github.com/ecrl/padelpy ADMET Predictor Simulations Plus, Inc https://www.simulations-plus.com/ CODESSA™ Semichem http://www.semichem.com/codessa/default.php DRAGON Talete SRL https://www.talete.mi.it/products/dragon_description.htm EPISUITE™ United States Environmental Protection Agency https://www.epa.gov/tsca-screening-tools/epi-suitetm- estimation-program-interface MOE Chemical Computing Group https://www.chemcomp.com/Products.htm Molconn-Z™ EduSoft http://www.edusoft-lc.com/molconn/ MOLD2 National Center for Toxicological Research https://www.fda.gov/science-research/bioinformatics- tools/mold2 MOLGEN University of Bayreuth https://www.molgen.de/ PowerMV National Institute of Statistical Sciences https://www.niss.org/research/software/powermv Alvadesc Alvascience https://www.alvascience.com/alvadesc/ CORAL Mario Negri Institute for Pharmacological Research http://www.insilico.eu/coral/SOFTWARECORAL.html The 0D molecular descriptors are straightforward and convenient to compute and interpret because they don't require structural information or connectivity between atoms. They are independent of molecular con- formation and optimization [43]. While they offer limited information content, they still play a vital role in modelling various physicochemical properties or contributing to more complex models [44]. Descriptors that compute information from fractions of a molecule fall into the 1D molecular descriptors category, often repre- sented as fingerprints - binary vectors where 1 indicates a substructure's presence and 0 its absence [45]. Similar to 0D descriptors, they're straightforward to calculate, interpretable, and conformation-independent [46]. Two- dimensional descriptors describe properties computable from 2D molecular representations, relying on graph https://github.com/rdkit https://github.com/ecrl/padelpy https://www.simulations-plus.com/ http://www.semichem.com/codessa/default.php https://www.talete.mi.it/products/dragon_description.htm https://www.epa.gov/tsca-screening-tools/epi-suitetm-estimation-program-interface https://www.epa.gov/tsca-screening-tools/epi-suitetm-estimation-program-interface https://www.chemcomp.com/Products.htm http://www.edusoft-lc.com/molconn/ https://www.fda.gov/science-research/bioinformatics-tools/mold2 https://www.fda.gov/science-research/bioinformatics-tools/mold2 https://www.molgen.de/ https://www.niss.org/research/software/powermv https://www.alvascience.com/alvadesc/ http://www.insilico.eu/coral/SOFTWARECORAL.html ADMET & DMPK 13(3) (2025) 2772 Machine learning models for ADMET prediction in drug development doi: https://doi.org/10.5599/admet.2772 7 theory and maintaining theoretical properties through isomorphism. They're sensitive to molecular charac- teristics like size, shape, and chemical information [46]. They're divided into structural-topological indices, encoding adjacency and distance, and topochemical indices, quantifying topology and atomic properties [47]. Three-dimensional descriptors relate to the 3D representation of molecules, incorporating molecular conformations, bond distances, angles, and dihedral angles to describe stereochemical properties [39]. Popular types include pharmacophore representations, characterizing steric and electronic features crucial for interactions with biological targets [48]. Grid-based descriptors, also called 4D, introduce a fourth dimension to capture interactions between molecules, their conformations, and biological receptor active sites. By considering ligand conformational variation and interactions within binding pockets, they aim to enhance quantitative structure-activity relationship (QSAR) model reliability [49]. 3. Machine learning applications in predicting ADME properties 3.1. Absorption prediction Machine learning models have shown promise in predicting intestinal absorption and permeability of compounds. Various approaches have been employed, including artificial neural networks for oligopep- tides [50], support vector machines for general compounds [51], and ensemble methods combining support vector machines (SVM), random forest (RF), and gradient boosting for natural products [52] (A summary of machine learning-based models for absorption prediction is presented in Table 3.). These models have demonstrated high predictive accuracy, with support vector machines achieving up to 91.54 % accuracy [51]. Computational models based on molecular descriptors, such as polar surface area, have also been developed to predict intestinal permeability [53]. These in silico methods offer quick and cost-effective alternatives to experimental techniques like Caco-2 cell assays. Table 3. Summary of machine learning-based models for absorption prediction No. of comp. Target Descriptors Modelling method Performance Ref. 1242 HIA 1D and 2D molecular descriptors SVM Accuracy: Training set = 90.38 %; Test set = 91.54 %; MCC = 0.80, AUC = 0.885 [51] 67 HIA 1D - 3D theoretical descriptors plus one of Abraham’s solvation param. MARS Whole data set: RMSE = 7.2 % Whole data set: R2 = 0.93 [54] 31 Ka MOE descriptors XGBoost Training set: RMSE = 0.0023 h−1 Prediction set: RMSE = 0.0021 h−1 [52] 160 HIA 0D - 3D Dragon theoretical descriptors Multilayer perceptron-artificial neural network, SVM Training set: R2= 0.8; RMSE = 0.18 Test set: R2= 0.66; RMSE = 0.21 [55] 552 HIA Adriana Code and Cerius2 0D - 2D theoretical descriptors Genetic algorithm, partial least squares regression, SVM Training set: R2= 0.66; RMSE = 12.5 Test set: R2= 0.77; RMSE = 16 [56] 1593 HIA 1D - 2D theoretical descriptors SVM Accuracy: Training set = 98.5 %, Test set = 99 % [57] 970 HIA 2D - 3D descriptors, molecular fingerprints and structural fragments Random forest Training set: SE = 0.89; SP = 0.85; Q = 0.89 Test set: SE = 0.88; SP = 0.81; Q = 0.87 [58] HIA - human intestinal absorption; SVM - support vector machines; MCC - Matthews correlation coefficient; AUC - area under the curve; MARS - multi- variate adaptive regression splines; RMSE- root mean square error; Ka - absorption rate constant; SE - sensitivity; SP - specificity; Q - accuracy The ability to predict intestinal absorption and permeability can significantly aid in drug development by facilitating the selection of promising candidates and potentially reducing attrition rates in clinical trials [52]. Bei et al. [59] developed an XGBoost model to predict the subcutaneous absorption rate constant of monoclonal antibodies using only their primary sequence. Kamiya et al. [60] used machine learning to estimate key physiologically based pharmacokinetic model parameters, including absorption rate constants, for 212 diverse chemicals. Karalis [61] applied machine learning techniques to identify the maximum plasma concentration (Cmax) / time to reach Cmax (Tmax) ratio as a potentially superior metric for absorption rate in https://doi.org/10.5599/admet.2772 M. Venkataraman et al. ADMET & DMPK 13(3) (2025) 2772 8 bioequivalence studies. Kumar et al. [62] implemented a graph convolutional neural network model to predict 18 early ADME properties across an enterprise-wide drug discovery pipeline. An integrated strategy combining multiple machine learning models and physiologically-based pharmacokinetic models achieved promising results in predicting human oral bioavailability directly from chemical structure [63]. Artificial Neural Networks outperformed support vector machines in predicting food effects on bioavailability, with key factors including octanol water partition coefficient, hydrogen bond donors, topological polar surface area, and dose [64]. Graph Neural Networks with transfer learning have also shown potential in predicting oral bioavailability, outperforming previous studies by automatically extracting important features from molecular structures [65]. These studies demonstrate the potential of machine learning to enhance the prediction of absorption kinetics and other pharmacokinetic parameters, potentially accelerating drug development processes. 3.2. Distribution prediction This summary examines models for predicting drug distribution in the body, focusing on blood-brain barrier (BBB) penetration and volume of distribution. Physiologically-based pharmacokinetic (PBPK) models are commonly used to predict tissue: plasma partition coefficients and volume of distribution [66]. Key factors affecting BBB penetration include permeability-surface area product (PS), unbound fraction in plasma (fu, plasma), and brain tissue (fu, brain) [67] (see Table 4 for an overview of machine learning-based models for distribution prediction). In vivo methods, such as brain microdialysis and in situ brain perfusion, are valuable for assessing brain drug distribution [68]. Table 4. Summary of machine learning-based models for distribution prediction No. of comp. Target Descriptors Modelling method Performance Ref. 529 Kp Fragmental descriptors Feed-forward back propagation neural network Q = 0.815, RMSEcv = 0.318 [69] 6741 PPB 2D, 3D, and fingerprints Consensus of K-nearest neighbours, support vector regression, RF, boosted trees and gradi- ent boosting regressor MAE = 0.089, RMSE = 0.153; R2= 0.738 [70] 21 Kp Not explained BIOiSIM Test set: AFE = 0.96 (Cmax), 0.89 (AUC), 0.69 (Vdss); AAFE = 1.20 (Cmax), 1.30 (AUC), 1.71 (Vdss); R2= 0.99 (Cmax), 0.98 (AUC), 0.99 (Vdss) [71] 227 PPB Constitutional descriptors, topological descriptors, geometric descriptors, molecular properties and RDF descriptors QSAR, convolutional neural network, feed- forward neural network Training set: MAE= 0.066, R2= 0.905, MSE= 0.011 Test set: MAE= 0.068, R2= 0.945, MSE= 0.007 [72] 1970 Kp Daylight fingerprints, atomic and ring multiplicities, simple molecular parameters and chemical descriptors RF, SVM Overall accuracy of 95 %, mean square contingency coefficient () of 0.74 [73] 310 BBB Topological descriptors, geometrical descriptors, electrostatic and quantum chemical descriptors SVM, genetic algorithm- partial least squares Training set: R2 = 0.98, RMSE = 0.117 Test set: R2 = 0.98, RMSE = 0.118 [74] 208 Kp Constitutional descriptors, topological descriptors, geometric descriptors, electrostatic descriptors and quantum chemical descriptors Least squares SVM Training set: R2 = 0.97, RMSE = 0.0226 Test set: R2 = 0.97, RMSE = 0.0289 [75] Kp - tissue-to-plasma partition coefficient; PPB - plasma protein binding; RF -random forest, MAE -mean absolute error; MSE - mean square error; AUC - area under the curve; AFE -average fold error; AAFE - absolute average fold error; Vdss - volume of distribution at steady-state. A simple two-descriptor model using molecular volume and polar surface area has been proposed for predicting BBB penetration [67]. Overall, current methods predict drug volume of distribution with an average 2-fold error [66]. Rapid brain equilibration requires high BBB permeability and low brain tissue binding [67], highlighting the importance of considering multiple drug design and selection factors. Iwata et al. [76] developed ML models for total body clearance and steady-state volume of distribution (Vd using imputed ADMET & DMPK 13(3) (2025) 2772 Machine learning models for ADMET prediction in drug development doi: https://doi.org/10.5599/admet.2772 9 nonclinical data, achieving accuracies comparable to animal scale-up models. Parrott et al. [77] integrated ML- predicted properties into physiologically-based pharmacokinetic (PBPK) models for highly lipophilic compounds, showing promising results for Vd predictions. Antontsev et al. [71] demonstrated a hybrid approach combining ML optimization with mechanistic modelling to predict tissue-plasma partition coefficients accurately. Mulpuru and Mishra [78] developed an AutoML model for predicting plasma unbound fraction, achieving a coefficient of determination of 0.85. Cao et al. [79] created a model to classify Polyfluorinated alkyl substances (PFAS) binding fractions in plasma, with 92 % accuracy. Riedl et al. [80] introduced a descriptor-free deep learning model using bidirectional encoder representations from transformers (BERT) for predicting plasma unbound fraction, offering flexibility and minimal domain expertise requirements. These studies demonstrate the potential of machine learning in predicting drug binding and distribution properties. 3.3. Metabolism prediction Machine learning techniques have shown promising results in predicting drug metabolism and interactions related to cytochrome P450 (CYP) enzymes. Various approaches, including k-nearest neighbours, decision trees, random forests, artificial neural networks, and support vector machines, have been employed to classify CYP activities and predict interactions with high accuracy [81] (For a structured overview of ML models in drug metabolism, see Table 5). Machine learning has also been applied to model relationships between chemical structure and metabolic fate, addressing complex endpoints such as metabolic stability and in vivo clearance [82]. Table 5. Summary of machine learning-based models for metabolism prediction No. of comp. Target Descriptors Modelling method Performance Ref. 4545 Metabolic pathway Molecular finger- prints, physicochemi- cal properties, structural descriptors Graph convolutional network and RF Single-class classification 95.16 % Multi-class classification 97.61 % [83] 1917 Metabolic pathway Physicochemical properties and others RF On external test: ACC 0.74, MCC 0.48, Sensitivity 0.70, Specificity 0.86, PPV 0.94, NPV 0.46 [84] 26138 Metabolic Stability 2D descriptors Principal com- ponent analysis, XGBoost Test set: ACC 93.6 % [85] 16613 Cytochrome inhibition 2D descriptors RMSE, XGBoos Test set: ACC 97.6 % [86] (Substrate, inhibitor) data: CYP1A2-(396, 13459) CYP2C9-(518, 12677) CYP2C19-(628, 13162) CYP2D6-(714, 13732) CYP3A4-(1584, 12990) Metabolic DDIs 2D descriptors, CATS, ECFP4 and MACCS RF, XGBoos Internal validation: ACC 0.8, AUC 0.9 External validation: ACC 0.795 Multi-level validation: ACC 0.793 PPV - positive predictive value; NPV - negative predictive value; ACC - accuracy Recent advancements have led to the development of consensus models for predicting metabolic drug-drug interactions (DDIs) related to five important CYP450 isozymes (CYP1A2, 2C9, 2C19, 2D6, 3A4), achieving high accuracy and robustness in both internal and external validations [86]. These in silico methods have become valuable tools in drug discovery and development, offering efficient alternatives to time-consuming and costly experimental assessments. Mamada et al. [87] developed a novel combination model using DeepSnap-Deep Learning and conventional ML, achieving high accuracy in predicting rat clearance. Keefer et al. [88] compared ML and mechanistic in vitro-in vivo extrapolation (IVIVE) models, finding that ML IVIVE models performed comparably or better than mechanistic counterparts for human intrinsic clearance prediction. Rodríguez-Pérez et al. [89] introduced a multitask graph neural network architecture for multispecies intrinsic clearance prediction, approaching experimental variability in performance. Andrews-Morger et al. [90] explored ML strategies to improve rat clearance predictions for physiologically based pharmacokinetic modelling, https://doi.org/10.5599/admet.2772 M. Venkataraman et al. ADMET & DMPK 13(3) (2025) 2772 10 demonstrating enhanced accuracy compared to standard in vitro bottom-up approaches. These studies highlight the potential of ML models in improving clearance predictions across species, contributing to more efficient drug discovery and development processes. 3.4. Excretion prediction Machine learning models have emerged as powerful tools for predicting drug metabolism and excretion in early drug discovery and development [91]. Researchers have developed in-silico prediction systems for renal excretion and clearance, incorporating factors such as the fraction unbound in plasma to improve accuracy [92]. ML models for pharmacokinetic (PK) prediction complement established approaches like IVIVE and PBPK models [93]. Ongoing research focuses on improving model accuracy, addressing limitations, and integrating ML approaches into drug discovery workflows to enhance efficiency and clinical success rates [93]. Recent studies have explored machine learning approaches for predicting pharmacokinetic parameters, including clearance (CL), elimination rate constant (ke), and half-life (t½) (Table 6 outlines key machine learning models developed for predicting drug excretion parameters). Seal et al. [94] created PKSmart, an open-source model for predicting human PK parameters, including volume of distribution at steady-state (VDss), CL, and t½, using molecular fingerprints and animal PK data. Fan et al. [95] focused on predicting drug half-life using ensemble and consensus machine learning methods, with XGBoost outperforming other individual models and a consensus model further enhancing prediction performance. These studies demonstrate the potential of machine learning approaches in improving the accuracy and efficiency of PK parameter prediction in drug discovery and development. Table 6. Summary of machine learning-based models for excretion prediction No. of comp. Target Descriptors Modelling method Performance Ref. 244 Intrinsic clearance Molecular fingerprints, physico- chemical properties, and 3D quantum chemical descriptors Partial least squares, RF, multi- label classification, principal component analysis R2 = 0.96, Q = 48 [96] 748 Total clearance The chemical structure was represented as graph Deep learning Test data set: Geometric mean fold error = 2.68 [97] 1114 Total clearance 2D SMARTS-based descriptors Model building using - StarDrop RF, radial basis function Whole data set: R2 = 0.55, RMSE=0.332 [98] 112 Intrinsic clearance 233 molecular descriptors Artificial neural network Training set: R2 = 0.953, RMSE = 0.236 Test set: R2 = 0.804, RMSE = 0.544 [99] 349 Renal clearance 195 descriptors Partial least squares, RF Training data: R2 =0.93, RMSE = 0.32 Test data: R2 = 0.63, RMSE = 0.63 [100] 1352 Renal clearance 2D and 3D descriptors and 49 fingerprints SVM, gradient boosting machine, XGBoost, RF Training set: R2 = 0.882, RMSE = 0.239 Test set: R2 = 0.875, RMSE = 0.103 [101] 4. Machine learning approaches in toxicity prediction 4.1. In silico toxicity models Machine learning models have become increasingly popular for predicting various toxicity endpoints, including hepatotoxicity, nephrotoxicity, cardiotoxicity, and genotoxicity [102]. Various ML models for predicting drug toxicity are summarized in Table 7. These models utilize physicochemical properties and in vitro assays to predict drug-induced liver injuries (DILI), which are major causes of drug attrition [103]. Support vector machines and random forests are among the most commonly used algorithms for toxicity prediction [104]. Khan et al. [104] developed an ensemble model integrating ML and deep learning algorithms, achieving 80.26% accuracy in predicting hepatotoxicity. Ancuceanu et al. [105] used the DILIrank dataset to create 78 models, which were then stacked for improved performance. Lu et al. [106] focused on predicting hepatotoxicity of drug metabolites using an ensemble approach based on support vector machines, achieving 78.47 % balanced accuracy. For specific drugs like colistin, machine learning models using electronic health records have been developed to predict nephrotoxicity, identifying key risk factors and dose thresholds [107]. ADMET & DMPK 13(3) (2025) 2772 Machine learning models for ADMET prediction in drug development doi: https://doi.org/10.5599/admet.2772 11 Table 7. Summary of machine learning-based models for toxicity prediction No. of comp. Target Descriptors Modelling method Performance Ref. Mice = 6,226 Rat = 6,238 Acute oral toxicity Molecular fingerprints, molecular descriptors Graph neural network, RF, SVM, artificial neural network Accuracy: Mice = 0.9586; Rat = 0.9335 MCC: Mice = 0.5514; Rat = 0.4929 AUROC: Mice = 0.7778; Rat = 0.7442 [108] 575 Hepatotoxicity 1D and 2D molecular descriptors RF Accuracy= 0.631 [109] 7889 Cardiotoxicity MOE and Mol2vec descriptors Multitask A AUC: training set = 0.944; validation set= 0967 [110] 641 Genotoxicity Molecular fingerprints, molecular descriptors SVM, RF Accuracy = 0.937 [111] 6512 Mutagenicity Molecular fingerprints, molecular descriptors SVM AUC= 0.93 [112] 863 Carcinogenicity Mol2vec, Mold2, MACCS Deep learning MCC= 0.432 [113] AUROC - area under the receiver operating characteristic curve A radiomics-based ML model for predicting radiotherapy-induced cardiotoxicity in breast cancer patients demonstrated high performance, with AUC up to 97 % when combining dosimetric, demographic, clinical, and imaging features [114]. Another study introduced cardioToxCSM, a web-based tool that predicts six types of cardiac toxicity outcomes for small molecules, achieving AUC values up to 0.898 [115]. The MultiFlow® DNA damage assay (MFA), which measures four mechanistic markers at two time points, has been combined with ML to enhance genotoxicity assessment and predict the mode of action of DNA-damaging agents [116]. Recent studies have achieved high accuracies in predicting genotoxicity, with models reaching up to 95 % accuracy on training data and 92 % on external test sets [116]. These studies demonstrate the potential of ML in toxicity prediction. The availability of large toxicology databases has facilitated the development of more accurate models [117]. However, challenges remain, such as the need for benchmarking datasets due to inconsistencies in toxicity assignments across different sources [102]. Despite these challenges, computational toxicology has made significant progress over the past decade, with machine learning models showing promise in predicting various toxicity endpoints and potentially reducing the need for costly and time-consuming in vivo studies [102]. 4.2. Adverse drug reactions Machine learning approaches have shown promising potential in predicting toxicological properties and adverse drug reactions (ADRs) of pharmaceutical agents [118]. Recent advancements include the development of AI models for precise prediction of compound off-target interactions, which can be used to differentiate drugs and classify compound toxicity [102]. The MAESTER framework integrates diverse features to predict tissue-specific adverse events with high accuracy, sensitivity, and specificity [119]. Similarly, the Off-targetP ML framework uses deep learning and automated machine learning to predict off-target panel activities directly from compound structures, aiding in drug design and discovery [120]. These computational methods offer efficient, low-cost tools for early assessment of compound safety and toxicity, potentially reducing costly failures in drug development and identifying toxic [102]. 4.3. Machine learning for predicting dose-dependent toxicity Machine learning is revolutionizing toxicology by enhancing predictive capabilities across various dosing ran- ges and improving safety assessments. ML models can analyse large datasets to predict drug toxicity, environ- mental hazards, and off-target effects, offering more efficient and accurate risk evaluations [121-123-87] . These models can be applied early in drug discovery to identify potential safety liabilities and filter out problematic compounds [117]. The integration of ML with DNA-encoded libraries (DELs) shows promise for modelling binding to off-targets and improving predictive toxicology [123]. Various toxic endpoints, including acute oral toxicity, hepatotoxicity, and cardiotoxicity, can be predicted using ML methods, although performance varies depending on the dataset and chemical space covered [117]. Despite challenges, ML in https://doi.org/10.5599/admet.2772 M. Venkataraman et al. ADMET & DMPK 13(3) (2025) 2772 12 toxicology offers significant potential for enhancing risk assessment, determining clinical toxicities, and detecting harmful side effects of medications [121]. 5. Overview of common machine learning techniques in ADME-Tox prediction ML is a method of data analysis involving the development of new algorithms and models capable of interpreting a multitude of data [14]. The algorithms used in recent years have successively improved their performance with the increase in both the quantitative and qualitative aspects of data available for learning [124]. Figure 2 illustrates commonly used AI/ML algorithms for developing ADMET prediction models. ML is considered one of the best options available when applied to solve problems for which a big amount of data and various variables are available to the individual, but a model or formula relating these various variables amongst themselves, along with the expected result, is not known [3,4]. However, when drug discovery moved into an era of a large amount of data, ML approaches evolved into DL approaches, which are more powerful and efficient in dealing with the massive amounts of data generated from modern drug discovery approaches [125]. DL is a subset of ML based on artificial neural networks that use multiple layers to progressively extract higher-level features from raw input. Due to its ability to learn from data and the environment, DL and neural network (NN), also known as artificial neural networks (ANN) named after its artificial representation of the working of a human nervous system, have become one of the most successful techniques in various AI research areas [126]. Different types of machine learning models with varying degrees of complexity can predict molecular properties, such as similarity-based models, linear models, kernel-based models, Bayesian models, tree- based models, and neural networks. 5.1. Traditional machine learning approaches 5.1.1. Random forest Random forest (RF) is widely used for ADMET due to its ability to handle high-dimensional data and complex, non-linear relationships. Random forest is an ensemble learning method that utilizes multiple decision trees to make predictions. It operates by constructing a multitude of decision trees during training and outputting the mode of the classes (classification) or the mean prediction (regression) of the individual trees [127]. RF introduces randomness both in the selection of data points used to build each tree and in the selection of features used at each split point. This randomness helps to decorrelate the trees, making the ensemble more robust and less prone to overfitting [128]. It has been used to predict toxicity endpoints [109], metabolic stability [84], and solubility [129]. RF has emerged as a powerful machine learning technique for predicting absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties of compounds. RF models have been successfully applied to classify toxicity datasets [130] and predict maximum recommended daily pharmaceutical doses [131]. These models utilize substructure fingerprints as descriptors, which encode the presence or absence of specific molecular substructures. RF's ability to identify important substructure features provides insights into structure-toxicity relationships [130]. The predictive performance of RF models can be further improved through rigorous model selection processes, as demonstrated in the Tox21 Challenge [132]. RF has also been employed in computational studies to evaluate ADMET properties of natural products, such as Papua red fruit flavonoids, aiding in the assessment of their potential as bioactive compounds in functional foods [133]. These applications highlight RF's versatility and effectiveness in ADMET prediction and toxicity evaluation. ADMET & DMPK 13(3) (2025) 2772 Machine learning models for ADMET prediction in drug development doi: https://doi.org/10.5599/admet.2772 13 5.1.2. Support vector machines Support vector machines (SVMs) are effective for binary classification tasks in ADMET modelling, especially for toxicology studies (like classifying compounds as toxic or non-toxic). SVMs have emerged as a powerful tool in predicting ADMET properties in drug discovery. It works by finding the optimal hyperplane that best separates different classes or predicts the continuous target variable by maximizing the margin between the classes [121]. SVMs can handle both linear and nonlinear data by using appropriate kernel functions. SVM can be extended to handle nonlinear data by mapping the input features into a higher- dimensional space using a kernel function [134]. The kernel function computes the dot product between the feature vectors in the higher-dimensional space without explicitly transforming the data. Common kernel functions include linear, polynomial, radial basis function (RBF), and sigmoid kernels, which allow SVM to capture complex nonlinear relationships in the data. SVMs, along with other machine learning techniques like Random Forests and Decision Trees, have become dominant methods in predictive toxicology due to their ability to handle complex datasets [121]. Studies have shown that SVMs can accurately classify compounds based on human intestinal absorption, using molecular descriptors such as topological polar surface area and predicted octanol-water distribution coefficient. SVMs have demonstrated competitive performance compared to other state-of-the-art techniques in pharmaceutical classification tasks [135]. However, challenges remain in fully integrating predictive ADMET modelling into drug discovery processes, and there is a need for larger, high-quality datasets and improved molecular descriptors to fully realize the potential of machine learning techniques in this field [136]. 5.1.3. K-nearest neighbours The k-nearest neighbours (k-NN) algorithm has shown promise in predicting various aspects of absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties of chemicals. K-nearest neighbour is a simple and intuitive supervised learning method used for classification and regression tasks. It operates on the principle that objects (e.g. data points) with similar characteristics are often found near each other in the feature space. The k-NN algorithm classifies or predicts the label of a new data point by considering the labels of its k-nearest neighbours, where k is a user-defined parameter [137]. Studies have demonstrated its effectiveness in predicting sub-chronic oral toxicity in rats [138], acute contact toxicity of pesticides in honey- bees [139], and chronic toxicity based on acute toxicity data [140]. These k-NN models have achieved reason- nable accuracy, with external validation results ranging from 65 to 77 % for different endpoints. The approach has been valuable in prioritizing chemical safety assessments and potentially reducing animal testing [139]. Incorporating ADMET screening earlier in the drug discovery process has become crucial for identifying poorly behaved compounds and improving the success rate of new chemical entities reaching the market. These computational models can play a significant role in predicting toxicity and supporting future risk assessments. 5.1.4. Gradient boosting machines and extreme gradient boosting Recent studies have demonstrated the effectiveness of gradient boosting algorithms in predicting Absorption, Distribution, Metabolism, Excretion, and Toxicity (ADMET) properties of drug compounds. Tian et al. [141] developed ADMET boost, a web server utilizing extreme gradient boosting (XGBoost) for accurate ADMET prediction, achieving top rankings in multiple benchmark tasks. An et al. [142] compared various machine learning models, including XGBoost and light gradient boosting machines (LGBM), for predicting estrogen receptor alpha (ERα) bioactivity and ADMET properties, finding high accuracy and robustness in their approach. Li et al. [143] applied LGBM to predict ADMET properties of anti-breast cancer compounds, reporting superior performance compared to other algorithms. These studies highlight the potential of gradient boosting techniques in drug discovery and development, offering accurate predictions of crucial https://doi.org/10.5599/admet.2772 M. Venkataraman et al. ADMET & DMPK 13(3) (2025) 2772 14 pharmacokinetic and toxicological properties. The integration of such models into web-based tools further enhances their accessibility and utility in the biopharmaceutical field. 5.2. Deep learning approaches 5.2.1. Fully Connected neural networks Fully connected neural networks (FCNNs) and other artificial neural network architectures have shown significant potential in predicting ADMET properties of chemical compounds. These models can achieve high accuracy in predicting various toxicological parameters, with some studies reporting accuracies between 74 and 98% [144]. FCNNs have demonstrated superior performance in multitask learning approaches, improving predictive quality for properties like human metabolic stability [145]. Neural networks outperform traditional methods in predicting several ADMET properties, including acute toxicity, carcinogenicity, and hepatic clearance [144]. The application of neural networks in ADMET prediction allows for early consideration of these properties in drug development, potentially reducing late-stage failures and improving pharmaceutical industry efficiency [146]. As artificial intelligence continues to advance, neural networks are expected to play an increasingly important role in toxicology research and drug discovery [147]. 5.2.2. Convolutional neural networks Convolutional neural networks (CNNs) and other deep learning models have shown promising results in predicting absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties of drug candidates. These models can analyse large datasets to identify relevant features and evaluate hidden trends among multiple ADMET parameters [145]. Deep neural networks have demonstrated superior performance compared to traditional methods in predicting properties such as microsomal stability, passive permeability, and log D [148]. Various ADMET prediction models have been developed and made publicly available, offering rapid assessments of important drug properties like cytotoxicity, mutagenicity, and drug-drug interactions [149]. The integration of ADMET screening earlier in the drug discovery process helps eliminate poorly behaved compounds, reducing costly failures in later stages[5]. As artificial intelligence continues to advance, neural networks are expected to play an increasingly important role in toxicology research and the development of accurate biosensors for toxic substance detection [147]. 5.2.3. Recurrent neural networks and long short-term memory Recent research has explored the application of recurrent neural networks (RNNs) and long short-term memory (LSTM) models in predicting and analysing ADMET properties of drugs and peptides. Wang et al. [15] demonstrated that LSTM networks can accurately model complex pharmacokinetic-pharmacodynamics relationships . Wenzel et al. [145] developed multitask deep neural networks for predicting multiple ADME-Tox properties simultaneously, showing improved performance over single-task models. González-Díaz [150] reviewed multi-output QSPR models for predicting ADMET processes, including drug-target interactions and nanoparticle toxicity. In the realm of peptide design, Müller et al. [151] utilized LSTM RNNs to generate novel antimicrobial peptide sequences with a higher predicted activity rate compared to randomly sampled sequences. These studies highlight the potential of RNN and LSTM models in drug discovery, ADMET prediction, and peptide design, offering promising tools for pharmaceutical research and development. 5.2.4. Graph neural networks Recent advancements in Graph neural networks (GNNs) have significantly improved the prediction of ADMET properties in drug discovery. De Carlo et al. [152] developed an attention-based GNN that processes molecular information from substructures to whole molecules, effectively predicting ADMET properties without relying on molecular descriptors. Aburidi and Marcia [153] introduced an optimal transport-based ADMET & DMPK 13(3) (2025) 2772 Machine learning models for ADMET prediction in drug development doi: https://doi.org/10.5599/admet.2772 15 graph kernel approach that outperformed state-of-the-art GNNs on multiple ADMET datasets. Feinberg et al. [31] demonstrated that graph convolutions applied to explicit molecular representations achieve unprecedented accuracy in ADMET prediction, enabling both interpolation and extrapolation to new chemical spaces. Wenzel et al. [145] explored multitask deep neural networks for ADME-Tox modelling, showing improved performance compared to single-task models and introducing a "response map" visualization technique for interpreting model predictions. These studies collectively highlight the potential of GNNs and deep learning approaches to enhance ADMET property prediction in drug development. 5.3. Generative models for ADMET-optimized molecule design 5.3.1. Variational autoencoders and generative adversarial networks Variational autoencoders (VAEs) and generative adversarial networks (GANs) are powerful machine learning models with applications in various fields. Adversarial Variational Bayes (AVB) unifies VAEs and GANs, allowing for more expressive inference models [154]. The connection between VAEs, GANs, and minimum Kantorovitch estimators has been explored from an optimal transport perspective [155]. In the field of drug development and toxicology, multi-output QSPR models have been used to predict absorption, distribution, metabolism, excretion and toxicity (ADMET) properties for drugs, pollutants, and nano- particles [150]. ADME profiling has significantly reduced pharmacokinetic drug failures in clinical trials and has become crucial for safety and toxicity prediction across industries [156]. The integration of ADME information with in vitro results and computer modelling is essential for developing quantitative in vitro to in vivo extrapolations and integrated testing strategies in toxicology [156]. 5.3.2. Reinforcement learning (RL) Generative AI models combined with reinforcement learning can efficiently design drug candidates with suitable ADMET properties [157]. Quantum-informed molecular representation learning has improved ADMET property prediction, achieving state-of-the-art results in multiple tasks [158]. The REMEDI framework uses reinforcement learning to model bile acid metabolism adaptations in primary sclerosing cholangitis, demonstrating potential for exploring treatments [159]. RL is emerging as a powerful tool in drug discovery and development, offering potential to accelerate and optimize the process. RL algorithms can improve sample efficiency and policy optimization for de novo drug design [160]. These approaches enable simultaneous optimization of molecules for multiple goals through structure-based drug design and high- throughput screening; one such example is where RF has been applied to optimize failed anticancer drugs by considering multiple properties simultaneously, including binding affinity and toxicity profiles [161]. 6. Recent developments in ADMET modeling 6.1. Integration of machine learning methods with physiologically-based pharmacokinetic and quantitative structure-activity relationship models Physiologically-based pharmacokinetic (PBPK) models are mathematical representations of how chemicals enter the body through various routes such as inhalation, ingestion, or dermal exposure. These models describe how much of the chemical enters the bloodstream, its distribution among different tissues, and how the body metabolizes and eliminates it [162]. They incorporate information about the body's anatomy, physiology, and biochemical processes. PBPK models can range from simple versions with few features to complex ones that capture intricate details about chemical movement and fate in the body [163]. However, creating PBPK models for new chemicals is challenging due to their complexity and the numerous parameters involved [163]. To address this, some researchers have proposed an integrated approach that combines a simplified PBPK model with machine learning based quantitative structure-activity relationship (QSAR) models [164]. This integrated https://doi.org/10.5599/admet.2772 M. Venkataraman et al. ADMET & DMPK 13(3) (2025) 2772 16 approach aims to estimate plasma and tissue concentrations, as well as various pharmacokinetic (PK) parameters, by leveraging databases containing in vivo and in vitro data along with the structural and physicochemical properties of selected compounds. ML models are trained using these databases to predict ADME parameters, which are then incorporated into the PBPK model to simulate time-concentration profiles and calculate PK parameters such as Area under the Curve (AUC) and maximum concentration (Cmax). The performance of the integrated ML-based PBPK model is assessed against in vivo PK data, and if satisfactory, the model can be used to generate simulation data for further refinement and validation [165]. This approach aims to enhance efficiency in drug development and reduce animal testing [166]. A novel computational platform combining ML and PBPK models has demonstrated improved accuracy in predicting pharmacokinetic profiles without experimental data, potentially accelerating early drug discovery [167]. Additionally, adapting PBPK models for ML applications has shown promise in recapitulating summary pharmacokinetic parameters, although limitations in the underlying PBPK models may affect prediction accuracy [169]. Despite challenges such as the need for diverse training data and improved interpretability of ML models, the integration of ML/AI approaches with PBPK modelling is expected to facilitate more efficient and robust ADMET predictions for a wide range of chemicals [165]. 6.2. Generalization 6.2.1. Multi-task learning While each ADMET parameter can be studied independently, they are interconnected and influence each other throughout the drug development process. Therefore, ADMET is not a sequential process but rather a holistic approach that considers the interplay between absorption, distribution, metabolism, excretion, and toxicity to predict drug behaviour accurately [14]. Considering this, it would be practical to develop a reliable multitask model that could simultaneously predict multiple ADME endpoints. Despite the advancements made by classical single-task learning in predicting individual ADMET endpoints using abundant labelled data, multi-task learning (MTL) emerges as a promising paradigm, offering a solution that reduces reliance on endpoint labels while simultaneously predicting multiple ADMET endpoints [169]. By leveraging shared information and underlying correlations between different ADMET properties, MTL provides a more comprehensive and efficient framework for predictive modelling in ADMET studies. For example, Wenzel et al. [145] introduced an industrialized approach for optimizing deep neural network (DNN) models to predict ADME-Tox properties, utilizing up to 50,000 compounds from diverse databases. The study highlights the significance of DNN hyperparameters and molecular descriptors in model success, demonstrating the superiority of multitask DNNs in predictive performance across various datasets. For instance, multitask DNNs showed improved predictive quality compared to single-task models, with an increase in R2 from 0.6 to 0.7 for human metabolic stability data in external validation sets. In another study, S. Zhang et al. [170] employ machine learning techniques, including random forest (RF) and artificial neural network (ANN), to develop multi-task (MT) models for predicting tissue-to-blood partition coefficient (Ptb) values across various mammalian tissues. Compared to single-task (ST) models, the MT approach consistently outperforms, with ANN-based MT models exhibiting the highest prediction accuracy, showcasing determination coefficients ranging from 0.704 to 0.886 and low root mean square errors and mean absolute errors across multiple endpoints. The study by Walter et al. [171] investigates multi-task machine learning models for predicting ADME and animal PK endpoints using in-house data of 28 endpoints. It reveals the superior performance of multi-task graph-based neural networks, attributing this success to the influence of endpoints with larger data sets, such as physicochemical endpoints and microsomal clearance. ADMET & DMPK 13(3) (2025) 2772 Machine learning models for ADMET prediction in drug development doi: https://doi.org/10.5599/admet.2772 17 6.2.2. Transfer learning Developing highly accurate models is essential for swiftly evaluating pharmacokinetic (PK) properties during the drug development process. Traditional models, which rely on a substantial amount of existing domain knowledge and constructing these models is rather time-consuming; thus, with transfer learning, which aims to generalize the knowledge gained from one task to another task to enhance its applicability and decision- making. Leveraging the common features learned from a similar source domain, transfer learning has been demonstrated to be able to develop the models without learning from scratch [172]. Ye et al. [173] introduced an integrated transfer learning and multitask learning approach utilizing three deep neural networks to predict pharmacokinetic parameters. The consensus model, DeepPharm utilized transfer learning from three pre- trained deep neural networks: DeepPharm-BA, DeepPharm-PPBR, and DeepPharm-VDss&HL and achieved the highest accuracies of 27.78, 44.22, 63.33 and 68.39 % in predicting oral bioavailability (BA), plasma protein binding rate (PPBR), apparent VDss and elimination half-life (HL), respectively, outperforming conventional machine learning methods. In a study proposed by Abbasi et al. [174] explores the transfer of knowledge across different physiological and biophysical domains, evaluating its effectiveness in predicting compound activity with limited labelled data. By leveraging source datasets such as Tox21, ToxCast, SIDER, HIV, and BACE, the proposed approach transfers knowledge between related or semi-related tasks, demonstrating improved performance in target tasks. Additionally, the study highlighted the importance of selecting appropriate source datasets and revealed that knowledge transfer between tasks within the same category yields better results compared to tasks from different categories. In another study S. Wang et al. [175] introduced a semi-supervised model named SMILES-BERT for molecular property prediction, utilizing deep learning techniques and a large- scale of unlabelled data. The model utilized an attention mechanism-based Transformer Layer and underwent pre-training via a Masked SMILES Recovery task on the ZINC dataset. This pre-training process significantly improved the model's generalization capability, as evidenced by achieving an exact recovery rate of 82.85 % on the validation dataset. During fine-tuning, the model was trained using various learning rates and optimization strategies, resulting in high prediction performance across three datasets: log P, PM2, and PCBA-686978. SMILES-BERT surpassed state-of-the-art methods, highlighting its ability to effectively leverage unlabelled data and its potential for molecular property prediction tasks with varying dataset sizes and properties. Similarly, X. Li and Fourches [176] introduced MolPMoFiT, an inductive transfer learning method for molecular activity prediction in Quantitative Structure-Activity Relationship (QSAR) modelling. The approach utilized a pre-trained Molecular Structure Prediction Model (MSPM) using one million unlabelled molecules from ChEMBL, fine- tuning it for specific QSAR tasks. This method achieved strong performance across four benchmark datasets (lipophilicity, FreeSolv, HIV, and blood-brain barrier penetration), when compared to state-of-the-art techniques reported in the literature. The approach showcased its potential for improving next-generation QSAR models, particularly for smaller datasets with challenging endpoints. 6.2.3. Pretrained models Recent advancements in machine learning have significantly improved the prediction of absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties in drug discovery. Pretrained models and self-supervised learning approaches have shown promising results in this field. Zhang et al. [177] developed HelixADMET, a system incorporating self-supervised learning that achieved a 4 % improvement over existing ADMET systems. Jung et al. [178] utilized the pretrained ChemBERTa model for ADMET prediction, exploring various architectures. Wenzel et al. [179] demonstrated the effectiveness of multitask deep neural networks in predicting ADME-Tox properties, showing improved performance compared to single-task models. Kumar et al. [62] implemented an enterprise-wide predictive model, gTPP, which outperformed commercial ADME models and automatic model builders. These studies highlight the potential https://doi.org/10.5599/admet.2772 M. Venkataraman et al. ADMET & DMPK 13(3) (2025) 2772 18 of advanced machine learning techniques, particularly pretrained models and self-supervised learning, in enhancing ADMET property prediction and facilitating early-stage drug development. 6.3. Interpretable and explainable ADMET models As artificial intelligence and machine learning models grow in complexity, a significant challenge has surfaced within the field: the absence of transparency and interpretability [180]. While these models demonstrate remarkable predictive capabilities, elucidating the rationale behind their predictions remains difficult. This absence of interpretability can pose challenges for researchers and regulatory authorities in relying on AI and ML-driven predictions, particularly in drug discovery [157]. Moreover, assessing and prioritizing discovered targets or compounds becomes cumbersome without understanding the decision- making process of AI algorithms. In order to foster trust in AI/ML systems, it is imperative that models are transparent and understandable to users thus efforts are being made to enhance the interpretability by embracing explainable artificial intelligence (XAI), which tries to offer clear and intelligible justifications for the predictions made by AI and ML models of machine learning models in ADME studies [181]. Interpretability in AI exposes the inner workings of these systems, allowing for the detection of issues like information leakage, model bias, robustness, and causality [182]. 6.3.1. Local interpretable model-agnostic explanations and Shapley additive explanations Recent research has explored the application of explainable artificial intelligence (XAI) techniques, particularly LIME (Local Interpretable Model-agnostic Explanations) and SHAP (SHapley Additive exPlana- tions), in interpreting complex machine learning models for medical applications such as Alzheimer's disease detection [183]. The paper introduced LIME, a method designed to provide interpretable explanations for complex machine learning models by approximating their behaviour locally. LIME achieves this by generating local surrogate models around specific instances, enabling users to understand model predictions on individual data points. A novel extension, KG-LIME, has been developed to predict individualized risk of adverse drug events in multiple sclerosis therapy, leveraging knowledge graphs for more interpretable explanations [184]. Another study by Gabbay et al. [185] presents a LIME-based explainable machine learning model for predicting the severity level of COVID-19 diagnosed patients. By employing LIME, the model provides interpretable insights into the factors influencing severity prediction, aiding in understanding and decision-making in COVID-19 management. Another explainable technique called SHAP (SHapley Additive exPlanations) methodology has emerged as a powerful tool for interpreting machine learning models in ADMET prediction and drug design. SHAP enables the identification and prioritization of molecular features that influence compound activity and potency predictions, regardless of model complexity [186]. This approach has been applied to various ADMET proper- ties, including metabolic stability [187] and general ADME profiles [188]. SHAP analysis can be used to interpret predictions from diverse machine learning algorithms, such as random forests, support vector machines, and deep neural networks [189]. By providing insights into the contribution of specific structural features to model outcomes, SHAP aids in compound optimization and supports experts in drug candidate selection [188]. This interpretability enhances confidence in machine learning applications within pharmaceutical research. These advancements demonstrate the growing importance of interpretable machine learning models in drug discovery and optimization, offering researchers valuable tools for understanding and improving ADMET predictions. ADMET & DMPK 13(3) (2025) 2772 Machine learning models for ADMET prediction in drug development doi: https://doi.org/10.5599/admet.2772 19 7. Machine learning techniques in clinical trial designs In silico prediction of absorption, distribution, metabolism, and excretion (ADME) properties, as well as model-informed drug discovery and development (MID3)1 strategies. MID3 includes providing quantitative predictions for aspects such as pharmacokinetics, pharmacodynamics, efficacy and safety end points, and disease progression [190]. By leveraging these models, researchers can optimize dosing strategies, inform clinical trial designs, and obtain robust quantitative assessments regarding drug efficacy and safety. The mainstay of modelling activities for drug development includes an empirical compartmental model built from sparsely sampled PK/PD datasets [191]. In this respect, AI/ML provides new ways for pharmacometricians to think about their models. There have been a number of approaches proposed in using feed-forward NNs [192] for modelling of PK(/PD) data. However, these did not tackle the more complex problem of extrapolating outside the range of observed data. In fact, the main limitation of such models is that they do not explicitly encode causality relationships among dose, PKs, and PDs and, hence, cannot enable robust predictions of new dosing regimens. In the 1990s, the availability of biological reagents and liquid chromatography mass spectrometry dramatically reduced the attrition of small-molecule drugs due to PK considerations. Currently, attrition due to poor clinical exposure is rare, with preclinical toxicology, clinical intolerability, or insufficient efficacy being the major sources of attrition [7]. Reagents, such as microsomes, cryopreserved hepatocytes, recombinant drug metabolizing enzymes, and cells overexpressing specific transporters, have enabled drug metabolism and PK departments to generate large quantities of in vitro ADME data over the last 15 to 20 years. These data serve two specific functions: first, in vitro data related to metabolic stability, plasma protein binding, permeability, efflux, and CYP inhibition can be used for the design (i.e. prior to synthesis) of small molecules with superior ADME properties, along with other parameters, such as biochemical and cellular potency and selectivity data; second, archived data can be used to build ML models to predict these properties (In silico optimization). 8. Data sources and challenges in machine learning based ADME-Tox prediction 8.1. Key databases for ADME-Tox data The development of predictive models for ADME-Tox properties is crucial in drug discovery, but it relies heavily on the availability and quality of data. Several databases and resources have emerged to address this need. Canault et al. [193] proposed an interactive network of databases to facilitate finding relevant ADME-Tox data sources. Ekins and Williams [194] advocated for making preclinical ADME-Tox data freely available on the web, suggesting the expansion of databases like ChemSpider. Pawar et al. [195] conducted a comprehensive review of over 900 databases relevant to in silico toxicology, categorizing them based on various criteria. To assist in compound filtering, Miteva et al. [196] developed FAF-Drugs, an online service that allows users to process their compound collections using simple ADME-Tox filtering rules. These resources collectively aim to improve drug development processes by enhancing access to and utilization of ADME-Tox data. 8.2. Data quality and availability issues Data quality issues have become increasingly critical in the era of big data, affecting various domains and applications [197]. These issues encompass multiple dimensions, including accuracy, completeness, consistency, and currency [198]. Poor data quality can significantly impact organizational efficiency and decision-making, leading to financial losses and credibility issues [198]. Data cleaning, a crucial process in addressing these problems, is particularly important when integrating heterogeneous data sources and in data warehouse environments [200]. Researchers have proposed various methodologies and techniques to tackle data quality challenges, drawing from fields such as data mining, probability theory, and machine https://doi.org/10.5599/admet.2772 M. Venkataraman et al. ADMET & DMPK 13(3) (2025) 2772 20 learning [201]. These approaches often involve the use of data quality rules and algorithms for detecting and correcting errors, as well as managing issues like data deduplication and information completeness [198]. 8.3. Integration of multi-omics data Multi-omics data integration is crucial for understanding complex biological systems and improving clinical outcomes. Various strategies have been developed, including early, mixed, intermediate, late, and hierarchical integration approaches [202]. Machine learning algorithms have been applied to multi-omics data to produce diagnostic and classification biomarkers [203]. Tools like OmicsNet use multilayer networks to integrate hetero- geneous omics data, facilitating functional analysis, biomarker discovery, and drug response prediction [204]. The integration of multi-omics data has applications in disease subtyping, biomarker prediction, and deriving biological insights [205]. Despite progress in the field, challenges remain in developing computational methods for the proper integration of multi-omics datasets [204]. Researchers have developed numerous software tools and methods to address these challenges and improve clinical outcome predictions [205]. 8.4. Privacy and ethical concerns Privacy and ethical concerns in research and technology adoption have gained significant attention. Key issues include ensuring meaningful user notice, access control, data anonymization, and algorithm validation to prevent harm [206]. The increasing use of technology in learning processes raises concerns about learner tracking, necessitating principles for trust, accountability, and transparency in learning analytics [207]. RFID technology, while promising for data collection, presents challenges for privacy due to its ability to track individual products [208]. Researchers are exploring situations where privacy may not always be optimal, con- sidering the balance between privacy and other competing values in research and design [209]. These studies emphasize the importance of addressing ethical and privacy concerns in various contexts, from social-behavi- oural research to technological implementations, to ensure responsible data use and protect individuals' rights. 9. Case studies of machine learning in ADME-Tox prediction 9.1. Successful machine learning applications in drug development Machine learning techniques are increasingly applied across various stages of drug discovery and development to accelerate the process and reduce failure rates [3]. ML approaches have shown promise in target validation, biomarker identification, and digital pathology analysis [3]. Specific applications include SNP discoveries, drug repurposing, virtual screening, lead identification, QSAR modelling, and ADMET analysis [210]. Algorithms such as support vector machines, random forests, and artificial neural networks have demonstrated success in predicting human intestinal absorption and identifying novel compounds for cancer treatment [210]. The Janssen gTPP model, employing graph convolutional neural networks, has shown superior performance in predicting early ADME properties compared to commercial models [62]. However, challenges remain in the interpretability and repeatability of ML-generated results, necessitating systematic data generation and validation of ML approaches [3,4]. 9.2. Real-world use cases Machine learning techniques have been increasingly applied to predict ADME-Tox properties of drug candidates, helping to streamline the drug discovery process and reduce costs [211]. These computational methods rely heavily on high-quality experimental data to generate accurate models. Flexible approaches combining multiple technologies, such as Bio-Rad's KnowItAll ADME/Tox system with support vector machine platforms, have shown promise in improving prediction performance and overcoming limitations of individual methods [212]. Despite the potential of in silico approaches, regulatory requirements for ADME and ADMET & DMPK 13(3) (2025) 2772 Machine learning models for ADMET prediction in drug development doi: https://doi.org/10.5599/admet.2772 21 toxicokinetic data vary widely across different chemical frameworks, with some areas having minimal or no requirements [213]. Incorporating ADME/TK information early in toxicity testing can enhance study design, support 4R goals, and ultimately improve risk assessment and characterization of chemical safety [213]. 9.3. Drug repurposing applications Drug repurposing, the process of using existing drugs for new indications, has gained attention due to its potential to reduce costs and development time [214]. This approach leverages existing ADMET data to expedite drug development [215]. Chemical structure modifications and nanotechnology applications can improve ADME-Tox properties of drug candidates, enhancing absorption, permeability, distribution, and stability while reducing toxicity [216]. ABC transporters play a crucial role in ADMET, influencing drug resistance and passage through cellular barriers [217]. Various models and assay systems, including in vitro, in vivo, and in silico approaches, can be used to analyze drug interactions with ABC transporters and predict ADMET profiles [217]. These strategies collectively contribute to more efficient drug discovery and development processes, potentially leading to improved therapeutic outcomes. 10. Challenges and limitations in machine learning based ADME-Tox prediction 10.1. Data scarcity and imbalance Machine learning techniques have been increasingly applied to ADME-Tox prediction, offering promising tools for toxicity screening and compound profiling [211]. However, data scarcity and class imbalance pose significant challenges in developing accurate models. Studies have shown that class imbalance can significantly impact model performance, particularly affecting recall and F1 scores [218]. To address these issues, various strategies have been explored, including resampling methods and transfer learning. Resampling techniques have demonstrated improvements in sensitivity and specificity for nuclear receptor profiling [219]. Additionally, transfer learning approaches have shown success in predicting drug activity and toxicity for targets with insufficient data by leveraging information from data-rich targets [220]. These methods, along with appropriate evaluation metrics and hyperparameter tuning, can enhance the performance of toxicity classification models and improve predictions for understudied targets [221]. 10.2. Model interpretability and trustworthiness Data scarcity and imbalance pose significant challenges for deep learning models, particularly in high-stakes domains [222]. These issues can lead to reduced model performance and trustworthiness. To address data scarcity, various techniques have been proposed, including transfer learning, self-supervised learning, and generative adversarial networks [222]. For imbalanced datasets, interpretable machine learning approaches can help identify class prototypes, sub-concepts, and outlier instances [223]. However, class imbalance can adversely affect the stability of interpretation methods like LIME and SHAP, particularly in credit scoring applications [224]. When evaluating model interpretability and trustworthiness, it's crucial to consider the inductive bias of different algorithms. For instance, in generalized additive models (GAMs), tree-based appro- aches offer a good balance of sparsity, fidelity, and accuracy, making them potentially more trustworthy [225]. 10.3. Generalization across chemical space Machine learning models in chemistry often face challenges due to data scarcity and imbalance, which can lead to overfitting and poor generalization. Several strategies have been proposed to address these issues. Farthest point sampling in chemical feature spaces can generate well-distributed training datasets, enhancing model performance across various algorithms [226]. Latent space enrichment, combining disparate data sources in joint prediction tasks, improves prediction in data-scarce applications [227]. https://doi.org/10.5599/admet.2772 M. Venkataraman et al. ADMET & DMPK 13(3) (2025) 2772 22 Similarity-based machine learning enables on-the-fly data selection and model training for specific queries, requiring only a fraction of data to achieve competitive performance [228]. When dealing with class imbalance and data scarcity in toxicity classification models, appropriate resampling algorithms, evaluation metrics, and hyperparameter tuning are crucial for optimal performance [218]. These approaches collectively offer promising solutions to enhance machine learning model performance in chemistry, particularly when faced with limited or imbalanced datasets. 10.4. Model validation and regulatory acceptance Data scarcity and class imbalance pose significant challenges in developing and validating machine learning models, particularly for safety-critical applications. These issues can affect model performance metrics and potentially invalidate underlying assumptions [229]. Studies have shown that class imbalance significantly impacts recall and F1 scores, while hyperparameter tuning can improve performance on imbalanced datasets [218]. To address data scarcity, various techniques have been proposed, including transfer learning, self-supervised learning, and generative adversarial networks [222]. For regulatory acceptance of models, it is crucial to consider model domain, uncertainty, validity, and predictability [230]. Researchers emphasize the importance of using appropriate evaluation metrics, tuning hyperparameters, and ensuring the trustworthiness of training datasets [222]. These considerations are essential for developing reliable and effective models in fields such as toxicity prediction, aviation, and medical imaging. 11. Future directions and innovations 11.1. Integrating AI with experimental approaches Recent research highlights the integration of AI with experimental approaches in various fields. AI and ML models can serve as fast surrogates for time-consuming experiments or computational models, enhancing predictive capabilities while reducing data requirements [231]. In molecular design, AI techniques are being combined with experimental validation and chemistry automation, although these efforts are still in early stages [232]. The development of cell therapies is benefiting from AI and ML methods, which can generate predictive models and design rules based on high-throughput screening data [233]. The integration of AI with geography, termed GeoAI, is providing novel approaches for addressing environmental and societal problems [234]. Future directions include the automatic discovery of physical laws, active learning for optimal experiment design, and the integration of multi-fidelity data from various computational models and experimental instruments [231]. 11.2. Advances in explainable AI Explainable AI (XAI) is an emerging field aimed at increasing the interpretability and transparency of machine learning models. Recent research highlights the integration of diverse approaches to advance XAI. Experimental psychology methods can contribute to XAI by applying cognitive modelling techniques to artificial black boxes [235]. Formal methods, such as algebraic decision diagrams, can enhance explainability by providing precise characterizations of AI outcomes [236]. The arts offer valuable contributions to address limitations in explainable AI, fostering collaborations between scientists and artists to investigate human- machine entanglements [237]. Current XAI approaches in deep learning encompass various applications, evaluation metrics, and challenges, with ongoing research focusing on improving trust, accountability, and interoperability of complex neural networks [238]. These multidisciplinary efforts aim to create more transparent and understandable AI systems across diverse domains. ADMET & DMPK 13(3) (2025) 2772 Machine learning models for ADMET prediction in drug development doi: https://doi.org/10.5599/admet.2772 23 11.3. Potential of quantum computing and machine learning Recent research explores the synergies between quantum computing, AI and ML. Quantum computing shows significant potential in revolutionizing drug discovery and development. It offers faster and more accurate molecular characterization through quantum simulation, outperforming classical quantum chemistry methods [239]. Quantum machine learning (QML) algorithms are emerging as strong competitors to classical approaches, particularly in the early stages of drug discovery for identifying novel drug-like molecules [240]. QC's ability to perform complex calculations efficiently could accelerate the drug discovery process, making it more cost-effective and accurate [241]. Recent applications of QC in drug development include protein structure prediction, molecular docking, quantum simulation, and quantitative structure- activity relationship models [242]. While current quantum devices are still susceptible to noise and errors, hybrid quantum-classical approaches and quantum-inspired devices like quantum annealers have demonstrated quantum advantage [242]. Further research and development are needed to fully leverage QC's potential in drug discovery [241]. 11.4. Regulatory and industry trends Artificial intelligence is revolutionizing drug discovery and preclinical research by integrating with experimental approaches. AI techniques, such as machine learning and neural networks, are improving the efficiency and effectiveness of drug candidate identification and optimization [243]. The integration of virtual and experimental screening methods, including high-throughput screening and DNA-encoded libraries, is enhancing early-phase drug discovery [243]. AI combined with new experimental technologies is expected to make drug discovery faster, cheaper, and more effective [244]. Additionally, integrated approaches to testing and assessment (IATA) are being developed to replace animal testing in toxicology, incorporating in vitro, in- silico, and in vivo methods [245]. While AI-driven approaches offer significant benefits, challenges remain, such as the need for high-quality databases and addressing regulatory requirements [245]. 12. Conclusion The application of artificial intelligence and machine learning in absorption, distribution, metabolism, excretion and toxicity (ADMET) predictions has revolutionized drug discovery and development. Traditional in vitro and in vivo approaches, though essential, are often time-consuming, costly, and sometimes unreliable in predicting human responses. AI and ML-driven computational methods provide a faster, more accurate, and cost-effective alternative by leveraging large-scale datasets and predictive modelling techniques such as deep learning, support vector machines, and ensemble learning. Molecular descriptors, ranging from 0D to 4D, alongside feature selection methods like filter, wrapper, and embedded techniques, play a crucial role in refining model accuracy and interpretability. These advancements facilitate high-throughput virtual screening, enabling the early identification of drug candidates with favourable pharmacokinetic and toxicity profiles, thereby reducing the likelihood of late- stage failures. Additionally, deep learning architectures have shown significant promise in predicting complex metabolic and toxicity pathways. However, challenges persist, including data quality, model generalizability, and regulatory acceptance. Standardized datasets, improved interpretability of AI models, and rigorous validation protocols are necessary for broader adoption. Regulatory agencies must establish clear guidelines for integrating AI-driven ADMET predictions into drug development workflows. Moving forward, integrating AI-driven predictions with experimental validation will be key to optimizing drug development pipelines. By overcoming current limitations and fostering collaboration between https://doi.org/10.5599/admet.2772 M. Venkataraman et al. ADMET & DMPK 13(3) (2025) 2772 24 computational and experimental research, AI and ML can drive more efficient, precise, and cost-effective drug discovery, ultimately leading to safer and more effective therapeutics. ABBREVIATIONS ADMET: Absorption, distribution, metabolism, excretion and toxicity ADRs: Adverse drug reactions ANN: Artificial neural network AVB: Adversarial variational bayes BA: Bioavailability BBB: Blood brain barrier BERT: Bidirectional encoder representations from transformers CFS: Correlation-based feature selection CL: Clearance CNNs: Convolutional neural networks CYP: Cytochrome DDIs: Drug-drug interactions DELs: DNA-encoded libraries DILI: Drug-induced liver injuries DNN: Deep neural network ERα: Estrogen receptor alpha FCNNs: Fully connected neural networks Fu: Unbound fraction GANs: Generative Adversarial networks GNNs: Graph neural networks IATA: Integrated approaches to testing and assessment IVIVE: in vitro-in vivo extrapolation Ke: Elimination rate constant k-NN: k-nearest neighbours LSTM: long short-term memory MD: Molecular descriptors ML: Machine Learning MTL: Multi-task learning NN: Neural network PBPK: Physiologically-based pharmacokinetic PFAS: Polyfluorinated alkyl substances PK: Pharmacokinetics PS: Permeability-surface QML: Quantum machine learning QSAR: Quantitative structure-activity relationship RBF: Radial basis function RF: Random forest RNNs: Recurrent neural networks SVM: Support vector machines VAEs: Variational autoencoders Vd: Volume of distribution XAI: Explainable artificial intelligence Conflict of interest: The authors declare no conflicts of interest related to this review article. Author contributions: All authors contributed to the study conception and design. Material preparation, data collection, and analysis were performed by Gopi Chand Rao, Jeevan Karthik Madavareddi, and Magesh Venkataraman. Magesh Venkataraman also contributed to methodology development and project administration. Srinivas Rao Maddi and Magesh Venkataraman supervised the study and contributed to the review and editing of the manuscript. The original draft was written by Magesh Venkataraman, Gopi Chand Rao, and Jeevan Karthik Madavareddi. All authors read and approved the final manuscript. Funding sources: This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. References [1] N. Berdigaliyev, M. Aljofan. An overview of drug discovery and development. Future Medicinal Chemistry 12 (2020) 939-947. https://doi.org/10.4155/fmc-2019-0307 [2] Z. Pei. Computer-aided drug discovery: From traditional simulation methods to language models and quantum computing. Cell Reports Physical Science 5 (2024) 11-21. https://doi.org/10.1016/j.xcrp.2024.102334 [3] J. Vamathevan, D. Clark, P. Czodrowski, I. Dunham, E. Ferran, G. Lee, B. Li, A. Madabhushi, P. Shah, M. Spitzer, S. Zhao. Applications of machine learning in drug discovery and development. Nature Reviews Drug Discovery 18 (2019) 463-477. https://doi.org/10.1038/s41573-019-0024-5 [4] S. Dara, S. Dhamercherla, S.S. Jadav, C.M. Babu, M.J. Ahsan. Machine Learning in Drug Discovery: A Review. Artificial Intelligence Review 55 (2022) 1947-1999. https://doi.org/10.1007/s10462-021- 10058-4 [5] G. Xiong, Z. Wu, J. Yi, L. Fu, Z. Yang, C. Hsieh, M. Yin, X. Zeng, C. Wu, A. Lu, X. Chen, T. Hou, D. Cao. ADMETlab 2.0: An integrated online platform for accurate and comprehensive predictions of ADMET properties. Nucleic Acids Research 49 (2021) W5-W14. https://doi.org/10.1093/nar/gkab255 https://doi.org/10.4155/fmc-2019-0307 https://doi.org/10.1016/j.xcrp.2024.102334 https://doi.org/10.1038/s41573-019-0024-5 https://doi.org/10.1007/s10462-021-10058-4 https://doi.org/10.1007/s10462-021-10058-4 https://doi.org/10.1093/nar/gkab255 ADMET & DMPK 13(3) (2025) 2772 Machine learning models for ADMET prediction in drug development doi: https://doi.org/10.5599/admet.2772 25 [6] J.A. Pradeepkiran, S.B. Sainath. Brucella Melitensis: Identification and Characterization of Potential Drug Targets, Elsevier, Texas, US, (2021) 133-176. https://doi.org/10.1016/C2020-0-03079-3 [7] D. Sun, W. Gao, H. Hu, S. Zhou. Why 90% of clinical drug development fails and how to improve it? Acta Pharmaceutica Sinica B 12 (2022) 3049-3062. https://doi.org/10.1016/j.apsb.2022.02.002 [8] J. Jiménez-Luna, F. Grisoni, N. Weskamp, G. Schneider. Artificial intelligence in drug discovery: recent advances and future perspectives. Expert Opinion on Drug Discovery 16 (2021) 949-959. https://doi.org/10.1080/17460441.2021.1909567 [9] S. Harrer, P. Shah, B. Antony, J. Hu. Artificial Intelligence for Clinical Trial Design. Trends in Pharmacological Sciences 40 (2019) 577-591. https://doi.org/10.1016/j.tips.2019.05.005 [10] M.J. Wildey, A. Haunso, M. Tudor, M. Webb, J.H. Connick. High-Throughput Screening. Annual Reports in Medicinal Chemistry 50 (2017) 149-195. https://doi.org/10.1016/bs.armc.2017.08.004 [11] S. Nag, A.T.K. Baidya, A. Mandal, A.T. Mathew, B. Das, B. Devi, R. Kumar. Deep learning tools for advancing drug discovery and development. 3 Biotech 12 (2022) 110. https://doi.org/10.1007/s13205-022-03165-8 [12] C.R. Hooijmans, R.B.M. De Vries, M. Ritskes-Hoitinga, M.M. Rovers, M.M. Leeflang, J. IntHout, K.E. Wever, L. Hooft, H. de Beer, T. Kuijpers, M.R. Macleod, E.S. Sena, G. ter Riet, R.L. Morgan, K.A. Thayer, A.A. Rooney, G.H. Guyatt, H.J. Schünemann, M.W. Langendam. Facilitating healthcare decisions by assessing the certainty in the evidence from preclinical animal studies. PLoS ONE 13(1) (2018) e0187271. https://doi.org/10.1371/journal.pone.0187271 [13] N. Singh, P. Vayer, S. Tanwar, J.-L. Poyet, K. Tsaioun, B.O. Villoutreix. Drug discovery and development: introduction to the general public and patient groups. Frontiers in Drug Discovery 3 (2023) 1201419. https://doi.org/10.3389/fddsv.2023.1201419 [14] F. Wu, Y. Zhou, L. Li, X. Shen, G. Chen, X. Wang, X. Liang, M. Tan, Z. Huang. Computational Approaches in Preclinical Studies on Drug Discovery and Development. Frontiers in Chemistry 8 (2020) 726. https://doi.org/10.3389/fchem.2020.00726 [15] D. Wang, W. Liu, Z. Shen, L. Jiang, J. Wang, S. Li, H. Li. Deep learning based drug metabolites prediction. Frontiers in Pharmacology 10 (2020) 1586. https://doi.org/10.3389/fphar.2019.01586 [16] Y. Wu, G. Wang. Machine learning based toxicity prediction: From chemical structural description to transcriptome analysis. International Journal of Molecular Sciences 19(8) (2018) 2358. https://doi.org/10.3390/ijms19082358 [17] K. Maharana, S. Mondal, B. Nemade. Data pre-processing and data augmentation techniques. Global Transitions Proceedings 3 (2022) 91-99. https://doi.org/10.1016/j.gltp.2022.04.020 [18] K. Bhayani, D. Tanna, V. Maan, Dhiraj, S. Kumar. An exploration of the impact of Feature quality versus Feature quantity on the performance of a machine learning model. in: IEEE Int. Conf. Contemp. Comput. Commun., IEEE, Bangalore, India, (2023): pp.1-5 https://doi.org/10.1109/InC457730.2023.10262824 [19] S.C. Rathi, S. Misra, R. Colomo-Palacios, R. Adarsh, L.B.M. Neti, L. Kumar. Empirical evaluation of the performance of data sampling and feature selection techniques for software fault prediction. Expert Systems with Applications 223 (2023) 119806. https://doi.org/10.1016/j.eswa.2023.119806 [20] F. Caloni, I. De Angelis, T. Hartung. Replacement of animal testing by integrated approaches to testing and assessment (IATA): a call for in vivitrosi. Archives of Toxicology 96 (2022) 1935-1950. https://doi.org/10.1007/s00204-022-03299-x [21] J. Grzegorzewski, J. Brandhorst, K. Green, D. Eleftheriadou, Y. Duport, F. Barthorscht, A. Köller, D.Y.J. Ke, S. De Angelis, M. König. PK-DB: Pharmacokinetics database for individualized and stratified computational modeling. Nucleic Acids Research 49 (2021) D1358-D1364. https://doi.org/10.1093/nar/gkaa990 [22] E. Pihan, L. Colliandre, J.F. Guichou, D. Douguet. E-Drug3D: 3D structure collections dedicated to drug repurposing and fragment-based drug design. Bioinformatics 28 (2012) 1540-1541. https://doi.org/10.1093/bioinformatics/bts186 https://doi.org/10.5599/admet.2772 https://doi.org/10.1016/C2020-0-03079-3 https://doi.org/10.1016/j.apsb.2022.02.002 https://doi.org/10.1080/17460441.2021.1909567 https://doi.org/10.1016/j.tips.2019.05.005 https://doi.org/10.1016/bs.armc.2017.08.004 https://doi.org/10.1007/s13205-022-03165-8 https://doi.org/10.1371/journal.pone.0187271 https://doi.org/10.3389/fddsv.2023.1201419 https://doi.org/10.3389/fchem.2020.00726 https://doi.org/10.3389/fphar.2019.01586 https://doi.org/10.3390/ijms19082358 https://doi.org/10.1016/j.gltp.2022.04.020 https://doi.org/10.1109/InC457730.2023.10262824 https://doi.org/10.1016/j.eswa.2023.119806 https://doi.org/10.1007/s00204-022-03299-x https://doi.org/10.1093/nar/gkaa990 https://doi.org/10.1093/bioinformatics/bts186 M. Venkataraman et al. ADMET & DMPK 13(3) (2025) 2772 26 [23] D.S. Wishart, Y.D. Feunang, A.C. Guo, E.J. Lo, A. Marcu, J.R. Grant, T. Sajed, D. Johnson, C. Li, Z. Sayeeda, N. Assempour, I. Iynkkaran, Y. Liu, A. MacIejewski, N. Gale, A. Wilson, L. Chin, R. Cummings, Di. Le, A. Pon, C. Knox, M. Wilson. DrugBank 5.0: A major update to the DrugBank database for 2018. Nucleic Acids Research 46 (2018) D1074-D1082. https://doi.org/10.1093/nar/gkx1037 [24] A. Gaulton, L.J. Bellis, A.P. Bento, J. Chambers, M. Davies, A. Hersey, Y. Light, S. McGlinchey, D. Michalovich, B. Al-Lazikani, J.P. Overington. ChEMBL: A large-scale bioactivity database for drug discovery. Nucleic Acids Research 40 (2012) D1100-D1107. https://doi.org/10.1093/nar/gkr777 [25] Y. Zhou, Y. Zhang, D. Zhao, X. Yu, X. Shen, Y. Zhou, S. Wang, Y. Qiu, Y. Chen, F. Zhu. T TD: Ther apeutic Targ et D atabase describing tar get drugg ability inf ormation. Nucleic Acids Research 52 (2024) D1465-D1477. https://doi.org/10.1093/nar/gkad751 [26] R.R. Sayre, J.F. Wambaugh, C.M. Grulke. Database of pharmacokinetic time-series data and parameters for 144 environmental chemicals. Scientific Data 7 (2020) 122. https://doi.org/10.1038/s41597-020-0455-1 [27] S. Kim, J. Chen, T. Cheng, A. Gindulyte, J. He, S. He, Q. Li, B.A. Shoemaker, P.A. Thiessen, B. Yu, L. Zaslavsky, J. Zhang, E.E. Bolton. PubChem 2019 update: Improved access to chemical data. Nucleic Acids Research 47 (2019) D1102-D1109. https://doi.org/10.1093/nar/gky1033 [28] T. Sterling, J.J. Irwin. ZINC 15 - Ligand Discovery for Everyone. Journal of Chemical Information and Modeling 55 (2015) 2324-2337. https://doi.org/10.1021/acs.jcim.5b00559 [29] S. Preissner, K. Kroll, M. Dunkel, C. Senger, G. Goldsobel, D. Kuzman, S. Guenther, R. Winnenburg, M. Schroeder, R. Preissner. SuperCYP: A comprehensive database on Cytochrome P450 enzymes including a tool for analysis of CYP-drug interactions. Nucleic Acids Research 38 (2009) D237-D243. https://doi.org/10.1093/nar/gkp970 [30] Y. Xu, X. Liu, W. Xia, J. Ge, C.W. Ju, H. Zhang, J.Z.H. Zhang. ChemXTree: A Feature-Enhanced Graph Neural Network-Neural Decision Tree Framework for ADMET Prediction. Journal of Chemical Information and Modeling 64(22) (2024) 8440-8452. https://doi.org/10.1021/acs.jcim.4c01186 [31] E.N. Feinberg, E. Joshi, V.S. Pande, A.C. Cheng. Improvement in ADMET Prediction with Multitask Deep Featurization. Journal of Medicinal Chemistry 63 (2020) 8835-8848. https://doi.org/10.1021/acs.jmedchem.9b02187 [32] M. Malekipirbazari, V. Aksakalli, W. Shafqat, A. Eberhard. Performance comparison of feature selection and extraction methods with random instance selection. Expert Systems with Applications 179 (2021) 115072. https://doi.org/10.1016/j.eswa.2021.115072 [33] B. Venkatesh, J. Anuradha. A review of Feature Selection and its methods. Cybernetics and Information Technologies 19 (2019) 3-26. https://doi.org/10.2478/CAIT-2019-0001 [34] N. Sánchez-Maroño, A. Alonso-Betanzos, M. Tombilla-Sanromán. Filter methods for feature selection - A comparative study. Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 4881 LNCS (2007) 178-187. https://doi.org/10.1007/978-3-540-77226-2_19 [35] S.S.S.J. Ahmed, V. Ramakrishnan. Systems biological approach of molecular descriptors connectivity: Optimal descriptors for oral bioavailability prediction. PLoS ONE 7 (2012) e40654. https://doi.org/10.1371/journal.pone.0040654 [36] I. Tsamardinos, G. Borboudakis, P. Katsogridakis, P. Pratikakis, V. Christophides. A greedy feature selection algorithm for Big Data of high dimensionality. Machine Learning 108 (2019) 149-202. https://doi.org/10.1007/s10994-018-5748-7 [37] H. Liu, M. Zhou, Q. Liu. An embedded feature selection method for imbalanced data classification. IEEE/CAA Journal of Automatica Sinica 6 (2019) 703-715. https://doi.org/10.1109/JAS.2019.1911447 [38] H.H. Hsu, C.W. Hsieh, M. Da Lu. Hybrid feature selection by combining filters and wrappers. Expert Systems with Applications 38 (2011) 8144-8150. https://doi.org/10.1016/j.eswa.2010.12.156 https://doi.org/10.1093/nar/gkx1037 https://doi.org/10.1093/nar/gkr777 https://doi.org/10.1093/nar/gkad751 https://doi.org/10.1038/s41597-020-0455-1 https://doi.org/10.1093/nar/gky1033 https://doi.org/10.1021/acs.jcim.5b00559 https://doi.org/10.1093/nar/gkp970 https://doi.org/10.1021/acs.jcim.4c01186 https://doi.org/10.1021/acs.jmedchem.9b02187 https://doi.org/10.1016/j.eswa.2021.115072 https://doi.org/10.2478/CAIT-2019-0001 https://doi.org/10.1007/978-3-540-77226-2_19 https://doi.org/10.1371/journal.pone.0040654 https://doi.org/10.1007/s10994-018-5748-7 https://doi.org/10.1109/JAS.2019.1911447 https://doi.org/10.1016/j.eswa.2010.12.156 ADMET & DMPK 13(3) (2025) 2772 Machine learning models for ADMET prediction in drug development doi: https://doi.org/10.5599/admet.2772 27 [39] P. Carracedo-Reboredo, J. Liñares-Blanco, N. Rodríguez-Fernández, F. Cedrón, F.J. Novoa, A. Carballal, V. Maojo, A. Pazos, C. Fernandez-Lozano. A review on machine learning approaches and trends in drug discovery. Computational and Structural Biotechnology Journal 19 (2021) 4538-4558. https://doi.org/10.1016/j.csbj.2021.08.011 [40] F. Grisoni, D. Ballabio, R. Todeschini, V. Consonni. Molecular descriptors for structure-activity applications: A hands-on approach. Methods in Molecular Biology 1800 (2018) 3-53. https://doi.org/10.1007/978-1-4939-7899-1_1 [41] A. Mauri, V. Consonni, R. Todeschini. Molecular descriptors. Handbook of Computational Chemistry (2017) 2065-2093. https://doi.org/10.1007/978-3-319-27282-5_51 [42] T. Mueller, A.G. Kusne, R. Ramprasad. Machine Learning in Materials Science: Recent Progress and Emerging Applications. Reviews in Computational Chemistry 29 (2016) 186-273. https://doi.org/10.1002/9781119148739.ch4 [43] R. Todeschini, V. Consonni. Molecular Descriptors for Chemoinformatics. Molecular Descriptors for Chemoinformatics 2 (2010) 1-252. https://doi.org/10.1002/9783527628766 [44] A.I. Odugbemi, C. Nyirenda, A. Christoffels, S.A. Egieyeh. Artificial intelligence in antidiabetic drug discovery: The advances in QSAR and the prediction of α-glucosidase inhibitors. Computational and Structural Biotechnology Journal 23 (2024) 2964-2977. https://doi.org/10.1016/j.csbj.2024.07.003 [45] J. Deng, Z. Yang, H. Wang, I. Ojima, D. Samaras, F. Wang. A systematic study of key elements underlying molecular property prediction. Nature Communications 14 (2023) 6395. https://doi.org/10.1038/s41467-023-41948-6 [46] B. Chandrasekaran, S.N. Abed, O. Al-Attraqchi, K. Kuche, R.K. Tekade. Computer-Aided Prediction of Pharmacokinetic (ADMET) Properties. Dosage Form Design Parameters 2 (2018) 731-755. https://doi.org/10.1016/B978-0-12-814421-3.00021-X [47] D. Wan, J. Yang, T. Zhang, Y. Xiong. A novel method to identify influential nodes based on hybrid topology structure. Physical Communication 58 (2023) 102046. https://doi.org/10.1016/j.phycom.2023.102046 [48] D.C. Kombo, K. Tallapragada, R. Jain, J. Chewning, A.A. Mazurov, J.D. Speake, T.A. Hauser, S. Toler. 3D molecular descriptors important for clinical success. Journal of Chemical Information and Modeling 53 (2013) 327-342. https://doi.org/10.1021/ci300445e [49] D. Fourches, J. Ash. 4D- quantitative structure-activity relationship modeling: making a comeback. Expert Opinion on Drug Discovery 14 (2019) 1227-1235. https://doi.org/10.1080/17460441.2019.1664467 [50] E. Jung, J. Kim, M. Kim, D.H. Jung, H. Rhee, J.M. Shin, K. Choi, S.K. Kang, M.K. Kim, C.H. Yun, Y.J. Choi, S.H. Choi. Artificial neural network models for prediction of intestinal permeability of oligopeptides. BMC Bioinformatics 8 (2007) 245. https://doi.org/10.1186/1471-2105-8-245 [51] R. Kumar, A. Sharma, M.H. Siddiqui, R.K. Tiwari. Prediction of Human Intestinal Absorption of Compounds Using Artificial Intelligence Techniques. Current Drug Discovery Technologies 14 (2017) 244 - 254. https://doi.org/10.2174/1570163814666170404160911 [52] V. Acuña-Guzman, M.E. Montoya-Alfaro, L.P. Negrón-Ballarte, C. Solis-Calero. A Machine Learning Approach for Predicting Caco-2 Cell Permeability in Natural Products from the Biodiversity in Peru. Pharmaceuticals 17(6) (2024) 750. https://doi.org/10.3390/ph17060750 [53] P. Stenberg, U. Norinder, K. Luthman, P. Artursson. Experimental and computational screening models for the prediction of intestinal drug absorption. Journal of Medicinal Chemistry 44 (2001) 1927-1937. https://doi.org/10.1021/jm001101a [54] T. Hou, J. Wang, W. Zhang, X. Xu. ADME evaluation in drug discovery. 7. prediction of oral absorption by correlation and classification. Journal of Chemical Information and Modeling 47 (2007) 208-218. https://doi.org/10.1021/ci600343x [55] E. Deconinck, H. Ates, N. Callebaut, E. Van Gyseghem, Y. Vander Heyden. Evaluation of chromatographic descriptors for the prediction of gastro-intestinal absorption of drugs. Journal of Chromatography A 1138 (2007) 190-202. https://doi.org/10.1016/j.chroma.2006.10.068 https://doi.org/10.5599/admet.2772 https://doi.org/10.1016/j.csbj.2021.08.011 https://doi.org/10.1007/978-1-4939-7899-1_1 https://doi.org/10.1007/978-3-319-27282-5_51 https://doi.org/10.1002/9781119148739.ch4 https://doi.org/10.1002/9783527628766 https://doi.org/10.1016/j.csbj.2024.07.003 https://doi.org/10.1038/s41467-023-41948-6 https://doi.org/10.1016/B978-0-12-814421-3.00021-X https://doi.org/10.1016/j.phycom.2023.102046 https://doi.org/10.1021/ci300445e https://doi.org/10.1080/17460441.2019.1664467 https://doi.org/10.1186/1471-2105-8-245 https://doi.org/10.2174/1570163814666170404160911 https://doi.org/10.3390/ph17060750 https://doi.org/10.1021/jm001101a https://doi.org/10.1021/ci600343x https://doi.org/10.1016/j.chroma.2006.10.068 M. Venkataraman et al. ADMET & DMPK 13(3) (2025) 2772 28 [56] A. Talevi, M. Goodarzi, E. V. Ortiz, P.R. Duchowicz, C.L. Bellera, G. Pesce, E.A. Castro, L.E. Bruno- Blanch. Prediction of drug intestinal absorption by new linear and non-linear QSPR. European Journal of Medicinal Chemistry 46 (2011) 218-228. https://doi.org/10.1016/j.ejmech.2010.11.005 [57] A. Yan, Z. Wang, Z. Cai. Prediction of human intestinal absorption by GA feature selection and support vector machine regression. International Journal of Molecular Sciences 9 (2008) 1961-1976. https://doi.org/10.3390/ijms9101961 [58] J. Shen, F. Cheng, Y. Xu, W. Li, Y. Tang. Estimation of ADME properties with substructure pattern recognition. Journal of Chemical Information and Modeling 50 (2010) 1034-1041. https://doi.org/10.1021/ci100104j [59] R. Bei, J. Thomas, S. Kapur, M. Woldeyes, A. Rauk, J. Robarge, J. Feng, K. Abbou Oucherif. Predicting the clinical subcutaneous absorption rate constant of monoclonal antibodies using only the primary sequence: a machine learning approach. MAbs 16 (2024) 2352887. https://doi.org/10.1080/19420862.2024.2352887 [60] Y. Kamiya, K. Handa, T. Miura, J. Ohori, A. Kato, M. Shimizu, M. Kitajima, H. Yamazaki. Machine Learning Prediction of the Three Main Input Parameters of a Simplified Physiologically Based Pharmacokinetic Model Subsequently Used to Generate Time-Dependent Plasma Concentration Data in Humans after Oral Doses of 212 Disparate Chemicals. Biological and Pharmaceutical Bulletin 45 (2022) 124-128. https://doi.org/10.1248/bpb.b21-00769 [61] V.D. Karalis. Machine Learning in Bioequivalence: Towards Identifying an Appropriate Measure of Absorption Rate. Applied Sciences (Switzerland) 13(1) (2023) 418. https://doi.org/10.3390/app13010418 [62] K. Kumar, V. Chupakhin, A. Vos, D. Morrison, D. Rassokhin, M.J. Dellwo, K. McCormick, E. Paternoster, H. Ceulemans, R.L. Desjarlais. Development and implementation of an enterprise-wide predictive model for early absorption, distribution, metabolism and excretion properties. Future Medicinal Chemistry 13 (2021) 1639-1654. https://doi.org/10.4155/fmc-2021-0138 [63] U. Fagerholm, S. Hellberg, O. Spjuth. Article advances in predictions of oral bioavailability of candidate drugs in man with new machine learning methodology. Molecules 26(9) (2021) 2572. https://doi.org/10.3390/molecules26092572 [64] H. Bennett-Lenane, B.T. Griffin, J.P. O’Shea. Machine learning methods for prediction of food effects on bioavailability: A comparison of support vector machines and artificial neural networks. European Journal of Pharmaceutical Sciences 168 (2022) 106018. https://doi.org/10.1016/j.ejps.2021.106018 [65] S.S.S. Ng, Y. Lu. Evaluating the Use of Graph Neural Networks and Transfer Learning for Oral Bioavailability Prediction. Journal of Chemical Information and Modeling 63 (2023) 5035-5044. https://doi.org/10.1021/acs.jcim.3c00554 [66] K. Holt, S. Nagar, K. Korzekwa. Methods to Predict Volume of Distribution. Current Pharmacology Reports 5 (2019) 391-399. https://doi.org/10.1007/s40495-019-00186-5 [67] X. Liu, B.J. Smith, C. Chen, E. Callegari, S.L. Becker, X. Chen, J. Cianfrogna, A.C. Doran, S.D. Doran, J.P. Gibbs, N. Hosea, J. Liu, F.R. Nelson, M.A. Szewc, J. Van Deusen. Use of a physiologically based pharmacokinetic model to study the time to reach brain equilibrium: An experimental analysis of the role of blood-brain barrier permeability, plasma protein binding, and brain tissue binding. Journal of Pharmacology and Experimental Therapeutics 313 (2005) 1254-1262. https://doi.org/10.1124/jpet.104.079319 [68] Q.R. Smith, R. Samala. In situ and in vivo animal models. AAPS Advances in the Pharmaceutical Sciences Series 10 (2014) 199-211. https://doi.org/10.1007/978-1-4614-9105-7_7 [69] N.N. Wang, C. Huang, J. Dong, Z.J. Yao, M.F. Zhu, Z.K. Deng, B. Lv, A.P. Lu, A.F. Chen, D.S. Cao. Predicting human intestinal absorption with modified random forest approach: a comprehensive evaluation of molecular representation, unbalanced data, and applicability domain issues. RSC Advances 7 (2017) 19007-19018. https://doi.org/10.1039/C6RA28442F https://doi.org/10.1016/j.ejmech.2010.11.005 https://doi.org/10.3390/ijms9101961 https://doi.org/10.1021/ci100104j https://doi.org/10.1080/19420862.2024.2352887 https://doi.org/10.1248/bpb.b21-00769 https://doi.org/10.3390/app13010418 https://doi.org/10.4155/fmc-2021-0138 https://doi.org/10.3390/molecules26092572 https://doi.org/10.1016/j.ejps.2021.106018 https://doi.org/10.1021/acs.jcim.3c00554 https://doi.org/10.1007/s40495-019-00186-5 https://doi.org/10.1124/jpet.104.079319 https://doi.org/10.1007/978-1-4614-9105-7_7 https://doi.org/10.1039/C6RA28442F ADMET & DMPK 13(3) (2025) 2772 Machine learning models for ADMET prediction in drug development doi: https://doi.org/10.5599/admet.2772 29 [70] E. V. Radchenko, A.S. Dyabina, V.A. Palyulin. Towards Deep Neural Network Models for the Prediction of the Blood-Brain Barrier Permeability for Diverse Organic Compounds. Molecules 25(24) (2020) 5901. https://doi.org/10.3390/MOLECULES25245901 [71] V. Antontsev, A. Jagarapu, Y. Bundey, H. Hou, M. Khotimchenko, J. Walsh, J. Varshney. A hybrid modeling approach for assessing mechanistic models of small molecule partitioning in vivo using a machine learning-integrated modeling platform. Scientific Reports 11 (2021) 11143. https://doi.org/10.1038/s41598-021-90637-1 [72] Y. Yuan, S. Chang, Z. Zhang, Z. Li, S. Li, P. Xie, W.P. Yau, H. Lin, W. Cai, Y. Zhang, X. Xiang. A novel strategy for prediction of human plasma protein binding using machine learning techniques. Chemometrics and Intelligent Laboratory Systems 199 (2020) 103962. https://doi.org/10.1016/j.chemolab.2020.103962 [73] A. Khaouane, S. Ferhat, S. Hanini. A Novel Methodology for Human Plasma Protein Binding: Prediction, Validation, and Applicability Domain. Pharmaceutical and Biomedical Research 8 (2022) 311-322. https://doi.org/10.32598/pbr.8.4.1086.1 [74] I.F. Martins, A.L. Teixeira, L. Pinheiro, A.O. Falcao. A Bayesian approach to in Silico blood-brain barrier penetration modeling. Journal of Chemical Information and Modeling 52 (2012) 1686-1697. https://doi.org/10.1021/ci300124c [75] H. Golmohammadi, Z. Dashtbozorgi, W.E. Acree. Quantitative structure-activity relationship prediction of blood-to-brain partitioning behavior using support vector machine. European Journal of Pharmaceutical Sciences 47 (2012) 421-429. https://doi.org/10.1016/j.ejps.2012.06.021 [76] H. Iwata, T. Matsuo, H. Mamada, T. Motomura, M. Matsushita, T. Fujiwara, K. Maeda, K. Handa. Predicting Total Drug Clearance and Volumes of Distribution Using the Machine Learning-Mediated Multimodal Method through the Imputation of Various Nonclinical Data. Journal of Chemical Information and Modeling 62 (2022) 4057-4065. https://doi.org/10.1021/acs.jcim.2c00318 [77] N. Parrott, N. Manevski, A. Olivares-Morales. Can We Predict Clinical Pharmacokinetics of Highly Lipophilic Compounds by Integration of Machine Learning or in Vitro Data into Physiologically Based Models? A Feasibility Study Based on 12 Development Compounds. Molecular Pharmaceutics 19 (2022) 3858-3868. https://doi.org/10.1021/acs.molpharmaceut.2c00350 [78] V. Mulpuru, N. Mishra. In Silico Prediction of Fraction Unbound in Human Plasma from Chemical Fingerprint Using Automated Machine Learning. ACS Omega 6 (2021) 6791-6797. https://doi.org/10.1021/acsomega.0c05846 [79] H. Cao, J. Peng, Z. Zhou, Z. Yang, L. Wang, Y. Sun, Y. Wang, Y. Liang. Investigation of the Binding Fraction of PFAS in Human Plasma and Underlying Mechanisms Based on Machine Learning and Molecular Dynamics Simulation. Environmental Science and Technology 57 (2023) 17762-17773. https://doi.org/10.1021/acs.est.2c04400 [80] M. Riedl, S. Mukherjee, M. Gauthier. Descriptor-Free Deep Learning QSAR Model for the Fraction Unbound in Human Plasma. Molecular Pharmaceutics 20 (2023) 4984-4993. https://doi.org/10.1021/acs.molpharmaceut.3c00129 [81] F. Hammann, H. Gutmann, U. Baumann, C. Helma, J. Drewe. Classification of cytochrome P450 activities using machine learning methods. Molecular Pharmaceutics 6 (2009) 1920-1926. https://doi.org/10.1021/mp900217x [82] T. Fox, J. Kriegl. Machine Learning Techniques for In Silico Modeling of Drug Metabolism. Current Topics in Medicinal Chemistry 6 (2006) 1579-1591. https://doi.org/10.2174/156802606778108915 [83] H.X. Liu, X.J. Yao, R.S. Zhang, M.C. Liu, Z.D. Hu, B.T. Fan. Prediction of the tissue/blood partition coefficients of organic compounds based on the molecular structure using least-squares support vector machines. Journal of Computer-Aided Molecular Design 19 (2005) 499-508. https://doi.org/10.1007/s10822-005-9003-5 [84] J.Y. Ryu, J.H. Lee, B.H. Lee, J.S. Song, S. Ahn, K.S. Oh. PredMS: a random forest model for predicting metabolic stability of drug candidates in human liver microsomes. Bioinformatics 38 (2022) 364- 368. https://doi.org/10.1093/bioinformatics/btab547 https://doi.org/10.5599/admet.2772 https://doi.org/10.3390/MOLECULES25245901 https://doi.org/10.1038/s41598-021-90637-1 https://doi.org/10.1016/j.chemolab.2020.103962 https://doi.org/10.32598/pbr.8.4.1086.1 https://doi.org/10.1021/ci300124c https://doi.org/10.1016/j.ejps.2012.06.021 https://doi.org/10.1021/acs.jcim.2c00318 https://doi.org/10.1021/acs.molpharmaceut.2c00350 https://doi.org/10.1021/acsomega.0c05846 https://doi.org/10.1021/acs.est.2c04400 https://doi.org/10.1021/acs.molpharmaceut.3c00129 https://doi.org/10.1021/mp900217x https://doi.org/10.2174/156802606778108915 https://doi.org/10.1007/s10822-005-9003-5 https://doi.org/10.1093/bioinformatics/btab547 M. Venkataraman et al. ADMET & DMPK 13(3) (2025) 2772 30 [85] M. Baranwal, A. Magner, P. Elvati, J. Saldinger, A. Violi, A. Violi, A.O. Hero. A deep learning architecture for metabolic pathway prediction. Bioinformatics 36 (2020) 2547-2553. https://doi.org/10.1093/bioinformatics/btz954 [86] N.N. Wang, X.G. Wang, G.L. Xiong, Z.Y. Yang, A.P. Lu, X. Chen, S. Liu, T.J. Hou, D.S. Cao. Machine learning to predict metabolic drug interactions related to cytochrome P450 isozymes. Journal of Cheminformatics 14 (2022) 23. https://doi.org/10.1186/s13321-022-00602-x [87] H. Mamada, Y. Nomura, Y. Uesawa. Prediction Model of Clearance by a Novel Quantitative Structure-Activity Relationship Approach, Combination DeepSnap-Deep Learning and Conventional Machine Learning. ACS Omega 6 (2021) 23570-23577. https://doi.org/10.1021/acsomega.1c03689 [88] C.E. Keefer, G. Chang, L. Di, N.A. Woody, D.A. Tess, S.M. Osgood, B. Kapinos, J. Racich, A.A. Carlo, A. Balesano, N. Ferguson, C. Orozco, L. Zueva, L. Luo. The Comparison of Machine Learning and Mechanistic In Vitro-In Vivo Extrapolation Models for the Prediction of Human Intrinsic Clearance. Molecular Pharmaceutics 20 (2023) 5616-5630. https://doi.org/10.1021/acs.molpharmaceut.3c00502 [89] R. Rodríguez-Pérez, M. Trunzer, N. Schneider, B. Faller, G. Gerebtzoff. Multispecies Machine Learning Predictions of in Vitro Intrinsic Clearance with Uncertainty Quantification Analyses. Molecular Pharmaceutics 20 (2023) 383-394. https://doi.org/10.1021/acs.molpharmaceut.2c00680 [90] A. Andrews-Morger, M. Reutlinger, N. Parrott, A. Olivares-Morales. A Machine Learning Framework to Improve Rat Clearance Predictions and Inform Physiologically Based Pharmacokinetic Modeling. Molecular Pharmaceutics 20 (2023) 5052-5065. https://doi.org/10.1021/acs.molpharmaceut.3c00374 [91] T.T. Van Tran, H. Tayara, K.T. Chong. Artificial Intelligence in Drug Metabolism and Excretion Prediction: Recent Advances, Challenges, and Future Perspectives. Pharmaceutics 15(4) (2023) 1260. https://doi.org/10.3390/pharmaceutics15041260 [92] R. Watanabe, R. Ohashi, T. Esaki, H. Kawashima, Y. Natsume-Kitatani, C. Nagao, K. Mizuguchi. Development of an in silico prediction system of human renal excretion and clearance from chemical structure information incorporating fraction unbound in plasma as a descriptor. Scientific Reports 9 (2019) 18782. https://doi.org/10.1038/s41598-019-55325-1 [93] D. Bassani, N.J. Parrott, N. Manevski, J.D. Zhang. Another string to your bow: machine learning prediction of the pharmacokinetic properties of small molecules. Expert Opinion on Drug Discovery 19 (2024) 683-698. https://doi.org/10.1080/17460441.2024.2348157 [94] S. Seal, M.-A. Trapotsi, V. Subramanian, O. Spjuth, N. Greene, A. Bender. PKSmart: An Open-Source Computational Model to Predict in vivo Pharmacokinetics of Small Molecules. BioRxiv (2024). https://doi.org/10.1101/2024.02.02.578658 [95] J. Fan, S. Shi, H. Xiang, L. Fu, Y. Duan, D. Cao, H. Lu. Predicting Elimination of Small-Molecule Drug Half-Life in Pharmacokinetics Using Ensemble and Consensus Machine Learning Methods. Journal of Chemical Information and Modeling 64 (2024) 3080-3092. https://doi.org/10.1021/acs.jcim.3c02030 [96] Y.W. Hsiao, U. Fagerholm, U. Norinder. In silico categorization of in vivo intrinsic clearance using machine learning. Molecular Pharmaceutics 10 (2013) 1318-1321. https://doi.org/10.1021/mp300484r [97] H. Iwata, T. Matsuo, H. Mamada, T. Motomura, M. Matsushita, T. Fujiwara, M. Kazuya, K. Handa. Prediction of Total Drug Clearance in Humans Using Animal Data: Proposal of a Multimodal Learning Method Based on Deep Learning. Journal of Pharmaceutical Sciences 110 (2021) 1834- 1841. https://doi.org/10.1016/j.xphs.2021.01.020 [98] Y. Kosugi, N. Hosea. Direct Comparison of Total Clearance Prediction: Computational Machine Learning Model versus Bottom-Up Approach Using in Vitro Assay. Molecular Pharmaceutics 17 (2020) 2299-2309. https://doi.org/10.1021/acs.molpharmaceut.9b01294 [99] P. Paixão, L.F. Gouveia, J.A.G. Morais. Prediction of the in vitro intrinsic clearance determined in suspensions of human hepatocytes by using artificial neural networks. European Journal of Pharmaceutical Sciences 39 (2010) 310-321. https://doi.org/10.1016/j.ejps.2009.12.007 https://doi.org/10.1093/bioinformatics/btz954 https://doi.org/10.1186/s13321-022-00602-x https://doi.org/10.1021/acsomega.1c03689 https://doi.org/10.1021/acs.molpharmaceut.3c00502 https://doi.org/10.1021/acs.molpharmaceut.2c00680 https://doi.org/10.1021/acs.molpharmaceut.3c00374 https://doi.org/10.3390/pharmaceutics15041260 https://doi.org/10.1038/s41598-019-55325-1 https://doi.org/10.1080/17460441.2024.2348157 https://doi.org/10.1101/2024.02.02.578658 https://doi.org/10.1021/acs.jcim.3c02030 https://doi.org/10.1021/mp300484r https://doi.org/10.1016/j.xphs.2021.01.020 https://doi.org/10.1021/acs.molpharmaceut.9b01294 https://doi.org/10.1016/j.ejps.2009.12.007 ADMET & DMPK 13(3) (2025) 2772 Machine learning models for ADMET prediction in drug development doi: https://doi.org/10.5599/admet.2772 31 [100] S.W. Paine, P. Barton, J. Bird, R. Denton, K. Menochet, A. Smith, N.P. Tomkinson, K.K. Chohan. A rapid computational filter for predicting the rate of human renal clearance. Journal of Molecular Graphics and Modelling 29 (2010) 529-537. https://doi.org/10.1016/j.jmgm.2010.10.003 [101] Y. Wang, H. Liu, Y. Fan, X. Chen, Y. Yang, L. Zhu, J. Zhao, Y. Chen, Y. Zhang. In Silico Prediction of Human Intravenous Pharmacokinetic Parameters with Improved Accuracy. Journal of Chemical Information and Modeling 59 (2019) 3968-3980. https://doi.org/10.1021/acs.jcim.9b00300 [102] W. Guo, J. Liu, F. Dong, M. Song, Z. Li, M.K.H. Khan, T.A. Patterson, H. Hong. Review of machine learning and deep learning models for toxicity prediction. Experimental Biology and Medicine 248 (2023) 1952-1973. https://doi.org/10.1177/15353702231209421 [103] P. Rana, S. Kogut, X. Wen, F. Akhlaghi, M.D. Aleo. Most Influential Physicochemical and in Vitro Assay Descriptors for Hepatotoxicity and Nephrotoxicity Prediction. Chemical Research in Toxicology 33 (2020) 1780-1790. https://doi.org/10.1021/acs.chemrestox.0c00040 [104] M.Z.I. Khan, J.N. Ren, C. Cao, H.Y.X. Ye, H. Wang, Y.M. Guo, J.R. Yang, J.Z. Chen. Comprehensive hepatotoxicity prediction: ensemble model integrating machine learning and deep learning. Frontiers in Pharmacology 15 (2024) 1441587. https://doi.org/10.3389/fphar.2024.1441587 [105] R. Ancuceanu, M.V. Hovanet, A.I. Anghel, F. Furtunescu, M. Neagu, C. Constantin, M. Dinu. Computational models using multiple machine learning algorithms for predicting drug hepatotoxicity with the dilirank dataset. International Journal of Molecular Sciences 21(6) (2020) 2114. https://doi.org/10.3390/ijms21062114 [106] Y. Lu, L. Liu, D. Lu, Y. Cai, M. Zheng, X. Luo, H. Jiang, K. Chen. Predicting Hepatotoxicity of Drug Metabolites Via an Ensemble Approach Based on Support Vector Machine. Combinatorial Chemistry & High Throughput Screening 20 (2017) 839-849. https://doi.org/10.2174/1386207320666171121113255 [107] L.W. Chiu, Y.E. Ku, F.Y. Chan, W.N. Lie, H.J. Chao, S.Y. Wang, W.C. Shen, H.Y. Chen. Machine learning algorithms to predict colistin-induced nephrotoxicity from electronic health records in patients with multidrug-resistant gram-negative infection. International Journal of Antimicrobial Agents 64 (2024) 107175. https://doi.org/10.1016/j.ijantimicag.2024.107175 [108] J.Y. Ryu, W.D. Jang, J. Jang, K.S. Oh. PredAOT: a computational framework for prediction of acute oral toxicity based on multiple random forest models. BMC Bioinformatics 24 (2023) 66. https://doi.org/10.1186/s12859-023-05176-5 [109] F. Mostafa, V. Howle, M. Chen. Machine Learning to Predict Drug-Induced Liver Injury and Its Validation on Failed Drug Candidates in Development †. Toxics 12(6) (2024) 385. https://doi.org/10.3390/toxics12060385 [110] C. Cai, P. Guo, Y. Zhou, J. Zhou, Q. Wang, F. Zhang, J. Fang, F. Cheng. Deep Learning-Based Prediction of Drug-Induced Cardiotoxicity. Journal of Chemical Information and Modeling 59 (2019) 1073-1084. https://doi.org/10.1021/acs.jcim.8b00769 [111] D. Fan, H. Yang, F. Li, L. Sun, P. Di, W. Li, Y. Tang, G. Liu. In silico prediction of chemical genotoxicity using machine learning methods and structural alerts. Toxicology Research 7 (2018) 211-220. https://doi.org/10.1039/c7tx00259a [112] N.K. Shinada, N. Koyama, M. Ikemori, T. Nishioka, S. Hitaoka, A. Hakura, S. Asakura, Y. Matsuoka, S.K. Palaniappan. Optimizing machine-learning models for mutagenicity prediction through better feature selection. Mutagenesis 37 (2022) 191-202. https://doi.org/10.1093/mutage/geac010 [113] T. Li, W. Tong, R. Roberts, Z. Liu, S. Thakkar. DeepCarc: Deep Learning-Powered Carcinogenicity Prediction Using Model-Level Representation. Frontiers in Artificial Intelligence 4 (2021) 757780. https://doi.org/10.3389/frai.2021.757780 [114] A. Talebi, A. Bitarafan-Rajabi, A. Alizadeh-asl, P. Seilani, B. Khajetash, G. Hajianfar, M. Tavakoli. Machine learning based radiomics model to predict radiotherapy induced cardiotoxicity in breast cancer. Journal of Applied Clinical Medical Physics 26(4) (2024) e14614. https://doi.org/10.1002/acm2.14614 https://doi.org/10.5599/admet.2772 https://doi.org/10.1016/j.jmgm.2010.10.003 https://doi.org/10.1021/acs.jcim.9b00300 https://doi.org/10.1177/15353702231209421 https://doi.org/10.1021/acs.chemrestox.0c00040 https://doi.org/10.3389/fphar.2024.1441587 https://doi.org/10.3390/ijms21062114 https://doi.org/10.2174/1386207320666171121113255 https://doi.org/10.1016/j.ijantimicag.2024.107175 https://doi.org/10.1186/s12859-023-05176-5 https://doi.org/10.3390/toxics12060385 https://doi.org/10.1021/acs.jcim.8b00769 https://doi.org/10.1039/c7tx00259a https://doi.org/10.1093/mutage/geac010 https://doi.org/10.3389/frai.2021.757780 https://doi.org/10.1002/acm2.14614 M. Venkataraman et al. ADMET & DMPK 13(3) (2025) 2772 32 [115] S. Iftkhar, A.G.C. De Sá, J.P.L. Velloso, R. Aljarf, D.E.V. Pires, D.B. Ascher. CardioToxCSM: A Web Server for Predicting Cardiotoxicity of Small Molecules. Journal of Chemical Information and Modeling 62 (2022) 4827-4836. https://doi.org/10.1021/acs.jcim.2c00822 [116] P. Trairatphisan, L. Dorsheimer, P. Monecke, J. Wenzel, R. James, A. Czich, Y. Dietz-Baum, F. Schmidt. Machine learning enhances genotoxicity assessment using MultiFlow® DNA damage assay. Environmental and Molecular Mutagenesis 66 (2024) 45-57. https://doi.org/10.1002/em.22648 [117] C.N. Cavasotto, V. Scardino. Machine Learning Toxicity Prediction: Latest Advances by Toxicity End Point. ACS Omega 7 (2022) 47536-47546. https://doi.org/10.1021/acsomega.2c05693 [118] X. Ma, R. Wang, Y. Xue, Z. Li, S. Yang, Y. Wei, Y. Chen. Advances in Machine Learning Prediction of Toxicological Properties and Adverse Drug Reactions of Pharmaceutical Agents. Current Drug Safety 3 (2008) 100-114. https://doi.org/10.2174/157488608784529224 [119] N. Madhukar, K. Gayvert, C. Gilvary, O. Elemento. A Machine Learning Approach Predicts Tissue- Specific Drug Adverse Events. BioRxiv (2018). https://doi.org/10.1101/288332 [120] D. Naga, W. Muster, E. Musvasva, G.F. Ecker. Off-targetP ML: an open source machine learning framework for off-target panel safety assessment of small molecules. Journal of Cheminformatics 14 (2022) 27. https://doi.org/10.1186/s13321-022-00603-w [121] L. Tonoyan, A.G. Siraki. Machine learning in toxicological sciences: opportunities for assessing drug toxicity. Frontiers in Drug Discovery 4 (2024) 1336025. https://doi.org/10.3389/fddsv.2024.1336025 [122] S. Okechukwuyem Ojji. Emerging Technology Integration - Artificial Intelligence (AI) and Machine Learning (ML) for Predictive Analysis for Safety and Toxicity Assessment in Environmental Toxicology. International Journal of Scientific Research and Management (IJSRM) 12 (2024) 1182- 1195. https://doi.org/10.18535/ijsrm/v12i05.ec03 [123] V. Blay, X. Li, J. Gerlach, F. Urbina, S. Ekins. Combining DELs and machine learning for toxicology prediction. Drug Discovery Today 27 (2022) 103351. https://doi.org/10.1016/j.drudis.2022.103351 [124] A. Lavecchia. Deep learning in drug discovery: opportunities, challenges and future prospects. Drug Discovery Today 24 (2019) 2017-2032. https://doi.org/10.1016/j.drudis.2019.07.006 [125] H. Chen, O. Engkvist, Y. Wang, M. Olivecrona, T. Blaschke. The rise of deep learning in drug discovery. Drug Discovery Today 23 (2018) 1241-1250. https://doi.org/10.1016/j.drudis.2018.01.039 [126] S. Min, B. Lee, S. Yoon. Deep learning in bioinformatics. Briefings in Bioinformatics 18 (2017) 851- 869. https://doi.org/10.1093/bib/bbw068 [127] H. Belyadi, A. Haghighat. Supervised learning. Machine Learning Guide for Oil and Gas Using Python (2021) 169-295. https://doi.org/10.1016/b978-0-12-821929-4.00004-4 [128] I.V.D. Srihith, P.V. Lakshmi, A.D. Donald, T. Aditya, T.A.S. Srinivas, G. Thippanna. A Forest of Possibilities: Decision Trees and Beyond. HARB Publication 6 (2023) 29-37. http://dx.doi.org/10.5281/zenodo.8372196 [129] D.S. Palmer, N.M. O’Boyle, R.C. Glen, J.B.O. Mitchell. Random forest models to predict aqueous solubility. Journal of Chemical Information and Modeling 47 (2007) 150-158. https://doi.org/10.1021/ci060164k [130] D.S. Cao, Y.N. Yang, J.C. Zhao, J. Yan, S. Liu, Q.N. Hu, Q.S. Xu, Y.Z. Liang. Computer-aided prediction of toxicity with substructure pattern and random forest. Journal of Chemometrics 26 (2012) 7-15. https://doi.org/10.1002/cem.1416 [131] D.S. Cao, Q.N. Hu, Q.S. Xu, Y.N. Yang, J.C. Zhao, H.M. Lu, L.X. Zhang, Y.Z. Liang. In silico classification of human maximum recommended daily dose based on modified random forest and substructure fingerprint. Analytica Chimica Acta 692 (2011) 50-56. https://doi.org/10.1016/j.aca.2011.02.010 [132] Y. Uesawa. Rigorous selection of random forest models for identifying compounds that activate toxicity-related pathways. Frontiers in Environmental Science 4 (2016) 9. https://doi.org/10.3389/fenvs.2016.00009 [133] M.M. Suprijono, H. Sujuti, D. Kurnia, S.B. Widjanarko. Absorption, distribution, metabolism, excretion, and toxicity evaluation of Papua red fruit flavonoids through a computational study. in: https://doi.org/10.1021/acs.jcim.2c00822 https://doi.org/10.1002/em.22648 https://doi.org/10.1021/acsomega.2c05693 https://doi.org/10.2174/157488608784529224 https://doi.org/10.1101/288332 https://doi.org/10.1186/s13321-022-00603-w https://doi.org/10.3389/fddsv.2024.1336025 https://doi.org/10.18535/ijsrm/v12i05.ec03 https://doi.org/10.1016/j.drudis.2022.103351 https://doi.org/10.1016/j.drudis.2019.07.006 https://doi.org/10.1016/j.drudis.2018.01.039 https://doi.org/10.1093/bib/bbw068 https://doi.org/10.1016/b978-0-12-821929-4.00004-4 http://dx.doi.org/10.5281/zenodo.8372196 https://doi.org/10.1021/ci060164k https://doi.org/10.1002/cem.1416 https://doi.org/10.1016/j.aca.2011.02.010 https://doi.org/10.3389/fenvs.2016.00009 ADMET & DMPK 13(3) (2025) 2772 Machine learning models for ADMET prediction in drug development doi: https://doi.org/10.5599/admet.2772 33 IOP Conf. Ser. Earth Environ. Sci., IOP Science, Malang East Java Indonesia, (2020) https://doi.org/10.1088/1755-1315/475/1/012078 [134] M.N. Murty, R. Raghava. Kernel-based SVM. SpringerBriefs in Computer Science 0 (2016) 57-67. https://doi.org/10.1007/978-3-319-41063-0_5 [135] M.W.B. Trotter, S.B. Holden. Support vector machines for ADME property classification. QSAR and Combinatorial Science 22 (2003) 533-548. https://doi.org/10.1002/qsar.200310006 [136] J. Gola, O. Obrezanova, E. Champness, M. Segall. ADMET property prediction: The state of the art and current challenges. QSAR and Combinatorial Science 25 (2006) 1172-1180. https://doi.org/10.1002/qsar.200610093 [137] Y. Shi, K. Yang, Z. Yang, Y. Zhou. Primer on artificial intelligence. Mobile Edge Artificial Intelligence (2022) 7-36. https://doi.org/10.1016/b978-0-12-823817-2.00011-5 [138] D. Gadaleta, F. Pizzo, A. Lombardo, A. Carotti, S.E. Escher, O. Nicolotti, E. Benfenati. A k-NN algorithm for predicting oral sub-chronic toxicity in the rat. Altex 31 (2014) 423-432. https://doi.org/10.14573/altex.1405091 [139] F. Como, E. Carnesecchi, S. Volani, J.L. Dorne, J. Richardson, A. Bassan, M. Pavan, E. Benfenati. Predicting acute contact toxicity of pesticides in honeybees (Apis mellifera) through a k-nearest neighbor model. Chemosphere 166 (2017) 438-444. https://doi.org/10.1016/j.chemosphere.2016.09.092 [140] S. Chavan, R. Friedman, I.A. Nicholls. Acute toxicity-supported chronic toxicity prediction: A k- nearest neighbor coupled read-across strategy. International Journal of Molecular Sciences 16 (2015) 11659-11677. https://doi.org/10.3390/ijms160511659 [141] H. Tian, R. Ketkar, P. Tao. ADMETboost: a web server for accurate ADMET prediction. Journal of Molecular Modeling 28 (2022) 408. https://doi.org/10.1007/s00894-022-05373-8 [142] T. An, Y. Chen, Y. Chen, L. Ma, J. Wang, J. Zhao. A machine learning-based approach to ERα bioactivity and drug ADMET prediction. Frontiers in Genetics 13 (2023) 1087273. https://doi.org/10.3389/fgene.2022.1087273 [143] X. Li, L. Tang, Z. Li, D. Qiu, Z. Yang, B. Li. Prediction of ADMET Properties of Anti-Breast Cancer Compounds Using Three Machine Learning Algorithms. Molecules 28 (2023) 2326. https://doi.org/10.3390/molecules28052326 [144] P.M. Vassiliev, A. V. Golubeva, A.R. Koroleva, M.A. Perfilev, A.N. Kochetkov. In Silico Prediction of Toxicological and Pharmacokinetic Characteristics of Medicinal Compounds. Safety and Risk of Pharmacotherapy 11 (2023) 390-408. https://doi.org/10.30895/2312-7821-2023-11-4-390-408 [145] J. Wenzel, H. Matter, F. Schmidt. Predictive Multitask Deep Neural Network Models for ADME-Tox Properties: Learning from Large Data Sets. Journal of Chemical Information and Modeling 59 (2019) 1253-1268. https://doi.org/10.1021/acs.jcim.8b00785 [146] D.A. Winkler. Neural networks in ADME and toxicity prediction. Drugs of the Future 29 (2004) 1043- 1057. https://doi.org/10.1358/dof.2004.029.10.863395 [147] I. Pantic, J. Paunovic, J. Cumic, S. Valjarevic, G.A. Petroianu, P.R. Corridon. Artificial neural networks in contemporary toxicology research. Chemico-Biological Interactions 369 (2023) 110269. https://doi.org/10.1016/j.cbi.2022.110269 [148] N. Schapin, M. Majewski, A. Varela-Rial, C. Arroniz, G. De Fabritiis. Machine learning small molecule properties in drug discovery. Artificial Intelligence Chemistry 1 (2023) 100020. https://doi.org/10.1016/j.aichem.2023.100020 [149] P. Schyman, R. Liu, V. Desai, A. Wallqvist. vNN web server for ADMET predictions. Frontiers in Pharmacology 8 (2017) 889. https://doi.org/10.3389/fphar.2017.00889 [150] H. González-Díaz. ADMET-Multi-Output Cheminformatics Models for Drug Delivery, Interactomics, and Nanotoxicology. Current Drug Delivery 13 (2016) 1. https://pubmed.ncbi.nlm.nih.gov/27417300/ https://doi.org/10.5599/admet.2772 https://doi.org/10.1088/1755-1315/475/1/012078 https://doi.org/10.1007/978-3-319-41063-0_5 https://doi.org/10.1002/qsar.200310006 https://doi.org/10.1002/qsar.200610093 https://doi.org/10.1016/b978-0-12-823817-2.00011-5 https://doi.org/10.14573/altex.1405091 https://doi.org/10.1016/j.chemosphere.2016.09.092 https://doi.org/10.3390/ijms160511659 https://doi.org/10.1007/s00894-022-05373-8 https://doi.org/10.3389/fgene.2022.1087273 https://doi.org/10.3390/molecules28052326 https://doi.org/10.30895/2312-7821-2023-11-4-390-408 https://doi.org/10.1021/acs.jcim.8b00785 https://doi.org/10.1358/dof.2004.029.10.863395 https://doi.org/10.1016/j.cbi.2022.110269 https://doi.org/10.1016/j.aichem.2023.100020 https://doi.org/10.3389/fphar.2017.00889 https://pubmed.ncbi.nlm.nih.gov/27417300/ M. Venkataraman et al. ADMET & DMPK 13(3) (2025) 2772 34 [151] A.T. Müller, J.A. Hiss, G. Schneider. Recurrent Neural Network Model for Constructive Peptide Design. Journal of Chemical Information and Modeling 58 (2018) 472-479. https://doi.org/10.1021/acs.jcim.7b00414 [152] A. De Carlo, D. Ronchi, M. Piastra, E.M. Tosca, P. Magni. Predicting ADMET Properties from Molecule SMILE: A Bottom-Up Approach Using Attention-Based Graph Neural Networks. Pharmaceutics 16 (2024) 776. https://doi.org/10.3390/pharmaceutics16060776 [153] M. Aburidi, R. Marcia. Wasserstein Distance-Based Graph Kernel for Enhancing Drug Safety and Efficacy Prediction*. Proceedings - 2024 IEEE 1st International Conference on Artificial Intelligence for Medicine, Health and Care, AIMHC (2024) 113-119. https://doi.org/10.1109/AIMHC59811.2024.00029 [154] L. Mescheder, S. Nowozin, A. Geiger. Adversarial Variational Bayes: Unifying Variational Autoencoders and Generative Adversarial Networks. ArXiv (2017). http://dx.doi.org/10.48550/arXiv.1701.04722 [155] A. Genevay, G. Peyré, M. Cuturi. GAN and VAE from an Optimal Transport Point of View. (2017). http://arxiv.org/abs/1706.01807 [156] K. Tsaioun, B.J. Blaauboer, T. Hartung. Evidence-based absorption, distribution, metabolism, excretion (ADME) and its interplay with alternative toxicity methods. Altex 33 (2016) 343-358. https://doi.org/10.14573/altex.1610101 [157] L.K. Vora, A.D. Gholap, K. Jetha, R.R.S. Thakur, H.K. Solanki, V.P. Chavda. Artificial Intelligence in Pharmaceutical Technology and Drug Delivery Design. Pharmaceutics 15 (2023) 1916. https://doi.org/10.3390/pharmaceutics15071916 [158] J. Kim, W. Chang, H. Ji, I.S. Joung. Quantum-Informed Molecular Representation Learning Enhancing ADMET Property Prediction. Journal of Chemical Information and Modeling 64 (2024) 5028-5040. https://doi.org/10.1021/acs.jcim.4c00772 [159] C. Hu, K. V. Saboo, A.H. Ali, B.D. Juran, K.N. Lazaridis, R.K. Iyer. REMEDI: REinforcement learning- driven adaptive MEtabolism modeling of primary sclerosing cholangitis DIsease progression. Proceedings of Machine Learning Research 225 (2023) 157-189. https://doi.org/10.48550/arXiv.2310.01426 [160] R.K. Tan, Y. Liu, L. Xie. Reinforcement learning for systems pharmacology-oriented and personalized drug design. Expert Opinion on Drug Discovery 17 (2022) 849-863. https://doi.org/10.1080/17460441.2022.2072288 [161] S. Park, Y.H. Ko, B. Lee, B. Shin, B.R. Beck. Abstract 35: Molecular optimization of phase III trial failed anticancer drugs using target affinity and toxicity-centered multiple properties reinforcement learning. Clinical Cancer Research 26 (2020) 35-35. https://doi.org/10.1158/1557- 3265.advprecmed20-35 [162] D. Yuan, H. He, Y. Wu, J. Fan, Y. Cao. Physiologically Based Pharmacokinetic Modeling of Nanoparticles. Journal of Pharmaceutical Sciences 108 (2019) 58-72. https://doi.org/10.1016/j.xphs.2018.10.037 [163] D. Deepika, V. Kumar. The Role of “Physiologically Based Pharmacokinetic Model (PBPK)” New Approach Methodology (NAM) in Pharmaceuticals and Environmental Chemical Risk Assessment. International Journal of Environmental Research and Public Health 20 (2023) 3473. https://doi.org/10.3390/ijerph20043473 [164] Y. Kamiya, K. Handa, T. Miura, M. Yanagi, K. Shigeta, S. Hina, M. Shimizu, M. Kitajima, F. Shono, K. Funatsu, H. Yamazaki. In Silico Prediction of Input Parameters for Simplified Physiologically Based Pharmacokinetic Models for Estimating Plasma, Liver, and Kidney Exposures in Rats after Oral Doses of 246 Disparate Chemicals. Chemical Research in Toxicology 34 (2021) 507-513. https://doi.org/10.1021/acs.chemrestox.0c00336 [165] W.C. Chou, Z. Lin. Machine learning and artificial intelligence in physiologically based pharmacokinetic modeling. Toxicological Sciences 191 (2023) 1-14. https://doi.org/10.1093/toxsci/kfac101 https://doi.org/10.1021/acs.jcim.7b00414 https://doi.org/10.3390/pharmaceutics16060776 https://doi.org/10.1109/AIMHC59811.2024.00029 http://dx.doi.org/10.48550/arXiv.1701.04722 http://arxiv.org/abs/1706.01807 https://doi.org/10.14573/altex.1610101 https://doi.org/10.3390/pharmaceutics15071916 https://doi.org/10.1021/acs.jcim.4c00772 https://doi.org/10.48550/arXiv.2310.01426 https://doi.org/10.1080/17460441.2022.2072288 https://doi.org/10.1158/1557-3265.advprecmed20-35 https://doi.org/10.1158/1557-3265.advprecmed20-35 https://doi.org/10.1016/j.xphs.2018.10.037 https://doi.org/10.3390/ijerph20043473 https://doi.org/10.1021/acs.chemrestox.0c00336 https://doi.org/10.1093/toxsci/kfac101 ADMET & DMPK 13(3) (2025) 2772 Machine learning models for ADMET prediction in drug development doi: https://doi.org/10.5599/admet.2772 35 [166] Y. Li, Z. Wang, Y. Li, J. Du, X. Gao, Y. Li, L. Lai. A Combination of Machine Learning and PBPK Modeling Approach for Pharmacokinetics Prediction of Small Molecules in Humans. Pharmaceutical Research 41 (2024) 1369-1379. https://doi.org/10.1007/s11095-024-03725-y [167] D. Naga, N. Parrott, G.F. Ecker, A. Olivares-Morales. Evaluation of the Success of High-Throughput Physiologically Based Pharmacokinetic (HT-PBPK) Modeling Predictions to Inform Early Drug Discovery. Molecular Pharmaceutics 19 (2022) 2203-2216. https://doi.org/10.1021/ACS.MOLPHARMACEUT.2C00040 [168] S. Habiballah, B. Reisfeld. Adapting physiologically-based pharmacokinetic models for machine learning applications. Scientific Reports 13 (2023) 14934. https://doi.org/10.1038/s41598-023- 42165-3 [169] Y. Zhang, Q. Yang. An overview of multi-task learning. National Science Review 5 (2018) 30-43. https://doi.org/10.1093/nsr/nwx105 [170] S. Zhang, X. Luo, B. Mai. Multi-task machine learning models for simultaneous prediction of tissue- to-blood partition coefficients of chemicals in mammals. Environmental Research 241 (2024) 117603. https://doi.org/10.1016/j.envres.2023.117603 [171] M. Walter, J.M. Borghardt, L. Humbeck, M. Skalic. Multi-task ADME/PK Prediction at Industrial Scale: Leveraging Large and Diverse Experimental Datasets. ChemRxiv (2024) e202400079. https://doi.org/10.1002/minf.202400079 [172] Z. Zhao, L. Alzubaidi, J. Zhang, Y. Duan, Y. Gu. A comparison review of transfer learning and self- supervised learning: Definitions, applications, advantages and limitations. Expert Systems with Applications 242 (2024) 122807. https://doi.org/10.1016/j.eswa.2023.122807 [173] Z. Ye, Y. Yang, X. Li, D. Cao, D. Ouyang. An Integrated Transfer Learning and Multitask Learning Approach for Pharmacokinetic Parameter Prediction. Molecular Pharmaceutics 16 (2019) 533-541. https://doi.org/10.1021/acs.molpharmaceut.8b00816 [174] K. Abbasi, A. Poso, J. Ghasemi, M. Amanlou, A. Masoudi-Nejad. Deep transferable compound representation across domains and tasks for low data drug discovery. Journal of Chemical Information and Modeling (2019) 4528-4539. https://doi.org/10.1021/acs.jcim.9b00626 [175] S. Wang, Y. Guo, Y. Wang, H. Sun, J. Huang. Smiles-Bert: Large scale unsupervised pre-training for molecular property prediction. ACM-BCB 2019 - Proceedings of the 10th ACM International Conference on Bioinformatics, Computational Biology and Health Informatics (2019) 429-436. https://doi.org/10.1145/3307339.3342186 [176] X. Li, D. Fourches. Inductive transfer learning for molecular activity prediction: Next-Gen QSAR Models with MolPMoFiT. Journal of Cheminformatics 12 (2020) 27. https://doi.org/10.1186/s13321-020-00430-x [177] S. Zhang, Z. Yan, Y. Huang, L. Liu, D. He, W. Wang, X. Fang, X. Zhang, F. Wang, H. Wu, H. Wang. HelixADMET: A robust and endpoint extensible ADMET system incorporating self-supervised knowledge transfer. Bioinformatics 38 (2022) 3444-3453. https://doi.org/10.1093/bioinformatics/btac342 [178] W. Jung, S. Goo, T. Hwang, H. Lee, Y.K. Kim, J.W. Chae, H.Y. Yun, S. Jung. Absorption Distribution Metabolism Excretion and Toxicity Property Prediction Utilizing a Pre-Trained Natural Language Processing Model and Its Applications in Early-Stage Drug Development. Pharmaceuticals 17 (2024) 382. https://doi.org/10.3390/ph17030382 [179] J. Wenzel, H. Matter, F. Schmidt. Predictive Multitask Deep Neural Network Models for ADME-Tox Properties: Learning from Large Data Sets. Journal of Chemical Information and Modeling 59 (2019) 1253-1268. https://doi.org/10.4155/fmc-2021-0138 [180] W. Saeed, C. Omlin. Explainable AI (XAI): A systematic meta-survey of current challenges and future opportunities. Knowledge-Based Systems 263 (2023) 110273. https://doi.org/10.1016/j.knosys.2023.110273 [181] R. Alizadehsani, S.S. Oyelere, S. Hussain, S.K. Jagatheesaperumal, R.R. Calixto, M. Rahouti, M. Roshanzamir, V.H.C. De Albuquerque. Explainable Artificial Intelligence for Drug Discovery and https://doi.org/10.5599/admet.2772 https://doi.org/10.1007/s11095-024-03725-y https://doi.org/10.1021/ACS.MOLPHARMACEUT.2C00040 https://doi.org/10.1038/s41598-023-42165-3 https://doi.org/10.1038/s41598-023-42165-3 https://doi.org/10.1093/nsr/nwx105 https://doi.org/10.1016/j.envres.2023.117603 https://doi.org/10.1002/minf.202400079 https://doi.org/10.1016/j.eswa.2023.122807 https://doi.org/10.1021/acs.molpharmaceut.8b00816 https://doi.org/10.1021/acs.jcim.9b00626 https://doi.org/10.1145/3307339.3342186 https://doi.org/10.1186/s13321-020-00430-x https://doi.org/10.1093/bioinformatics/btac342 https://doi.org/10.3390/ph17030382 https://doi.org/10.4155/fmc-2021-0138 https://doi.org/10.1016/j.knosys.2023.110273 M. Venkataraman et al. ADMET & DMPK 13(3) (2025) 2772 36 Development: A Comprehensive Survey. IEEE Access 12 (2024) 35796-35812. https://doi.org/10.1109/ACCESS.2024.3373195 [182] A. McGrath. What is AI interpretability? IBM (2024). https://www.ibm.com/think/topics/interpretability [183] V. Vimbi, N. Shaffi, M. Mahmud. Interpreting artificial intelligence models: a systematic review on the application of LIME and SHAP in Alzheimer’s disease detection. Brain Informatics 11 (2024) 10. https://doi.org/10.1186/s40708-024-00222-1 [184] J. Patterson, N. Tatonetti. KG-LIME: predicting individualized risk of adverse drug events for multiple sclerosis disease-modifying therapy. Journal of the American Medical Informatics Association 31 (2024) 1693-1703. https://doi.org/10.1093/jamia/ocae155 [185] F. Gabbay, S. Bar-Lev, O. Montano, N. Hadad. A lime-based explainable machine learning model for predicting the severity level of covid-19 diagnosed patients. Applied Sciences (Switzerland) 11 (2021) 10417. https://doi.org/10.3390/app112110417 [186] S. Singh, R. Kumar, S. Payra, S.K. Singh. Artificial Intelligence and Machine Learning in Pharmacological Research: Bridging the Gap Between Data and Drug Discovery. Cureus 15(8) (2023) e44359. https://doi.org/10.7759/cureus.44359 [187] T.Z. Long, D.J. Jiang, S.H. Shi, Y.C. Deng, W.X. Wang, D.S. Cao. Enhancing Multi-species Liver Microsomal Stability Prediction through Artificial Intelligence. Journal of Chemical Information and Modeling 64 (2024) 3222-3236. https://doi.org/10.1021/acs.jcim.4c00159 [188] C. König, A. Vellido. Understanding predictions of drug profiles using explainable machine learning models. BioData Mining 17 (2024) 25. https://doi.org/10.1186/s13040-024-00378-w [189] R. Rodríguez-Pérez, J. Bajorath. Interpretation of machine learning models using shapley values: application to compound potency and multi-target activity predictions. Journal of Computer-Aided Molecular Design 34 (2020) 1013-1026. https://doi.org/10.1007/s10822-020-00314-0 [190] S. Marshall, R. Madabushi, E. Manolis, K. Krudys, A. Staab, K. Dykstra, S.A.G. Visser. Model-Informed Drug Discovery and Development: Current Industry Good Practice and Regulatory Expectations and Future Perspectives. CPT: Pharmacometrics and Systems Pharmacology 8 (2019) 87-96. https://doi.org/10.1002/psp4.12372 [191] N. Terranova, D. Renard, M.H. Shahin, S. Menon, Y. Cao, C.E.C.A. Hop, S. Hayes, K. Madrasi, S. Stodtmann, T. Tensfeldt, P. Vaddady, N. Ellinwood, J. Lu. Artificial Intelligence for Quantitative Modeling in Drug Discovery and Development: An Innovation and Quality Consortium Perspective on Use Cases and Best Practices. Clinical Pharmacology and Therapeutics 115 (2024) 658-672. https://doi.org/10.1002/cpt.3053 [192] S.H. Kang, M.R. Poynton, K.M. Kim, H. Lee, D.H. Kim, S.H. Lee, K.S. Bae, O. Linares, S.E. Kern, G.J. Noh. Population pharmacokinetic and pharmacodynamic models of remifentanil in healthy volunteers using artificial neural network analysis. British Journal of Clinical Pharmacology 64 (2007) 3-13. https://doi.org/10.1111/j.1365-2125.2007.02845.x [193] B. Canault, S. Bourg, P. Vayer, P. Bonnet. Comprehensive Network Map of ADME-Tox Databases. Molecular Informatics 36 (2017) 1700029. https://doi.org/10.1002/minf.201700029 [194] S. Ekins, A.J. Williams. Precompetitive preclinical ADME/Tox data: Set it free on the web to facilitate computational model building and assist drug development. Lab on a Chip 10 (2010) 13-22. https://doi.org/10.1039/b917760b [195] G. Pawar, J.C. Madden, D. Ebbrell, J.W. Firman, M.T.D. Cronin. In silico toxicology data resources to support read-across and (Q)SAR. Frontiers in Pharmacology 10 (2019) 561. https://doi.org/10.3389/fphar.2019.00561 [196] M.A. Miteva, S. Violas, M. Montes, D. Gomez, P. Tuffery, B.O. Villoutreix. FAF-Drugs: Free ADME/tox filtering of compound collections. Nucleic Acids Research 34 (2006) W738-W744. https://doi.org/10.1093/nar/gkl065 https://doi.org/10.1109/ACCESS.2024.3373195 https://www.ibm.com/think/topics/interpretability https://doi.org/10.1186/s40708-024-00222-1 https://doi.org/10.1093/jamia/ocae155 https://doi.org/10.3390/app112110417 https://doi.org/10.7759/cureus.44359 https://doi.org/10.1021/acs.jcim.4c00159 https://doi.org/10.1186/s13040-024-00378-w https://doi.org/10.1007/s10822-020-00314-0 https://doi.org/10.1002/psp4.12372 https://doi.org/10.1002/cpt.3053 https://doi.org/10.1111/j.1365-2125.2007.02845.x https://doi.org/10.1002/minf.201700029 https://doi.org/10.1039/b917760b https://doi.org/10.3389/fphar.2019.00561 https://doi.org/10.1093/nar/gkl065 ADMET & DMPK 13(3) (2025) 2772 Machine learning models for ADMET prediction in drug development doi: https://doi.org/10.5599/admet.2772 37 [197] D. Rao, V.N. Gudivada, V. V. Raghavan. Data quality issues in big data. Proceedings - 2015 IEEE International Conference on Big Data, IEEE Big Data (2015) 2654-2660. https://doi.org/10.1109/BigData.2015.7364065 [198] W. Fan, F. Geerts. Foundations of Data Quality Management, 1st ed., Springer Nature Link, (2012) https://doi.org/10.1007/978-3-031-01892-3 [199] K. Natarajan, J. Li, A. Koronios. Data mining techniques for data cleaning. Engineering Asset Lifecycle Management - Proceedings of the 4th World Congress on Engineering Asset Management, WCEAM (2009) 796-804. https://doi.org/10.1007/978-0-85729-320-6_91 [200] C. Batini, M. Scannapieco. Data Quality: Concepts, Methodologies and Techniques. Springer 1 (2006) 161-200. http://dx.doi.org/10.1007/3-540-33173-5 [201] M. Picard, M.P. Scott-Boyer, A. Bodein, O. Périn, A. Droit. Integration strategies of multi-omics data for machine learning analysis. Computational and Structural Biotechnology Journal 19 (2021) 3735- 3746. https://doi.org/10.1016/j.csbj.2021.06.030 [202] H. Bansal, H. Luthra, S.R. Raghuram. A Review on Machine Learning Aided Multi-omics Data Integration Techniques for Healthcare. Studies in Big Data 132 (2023) 211-239. https://doi.org/10.1007/978-3-031-38325-0_10 [203] M. Akhmedov, A. Arribas, R. Montemanni, F. Bertoni, K. Ivo. OmicsNet: Integration of Multi-Omics Data using Path Analysis in Multilayer Networks. BioRxiv (2017). https://doi.org/10.1101/238766 [204] I. Subramanian, S. Verma, S. Kumar, A. Jere, K. Anamika. Multi-omics Data Integration, Interpretation, and Its Application. Bioinformatics and Biology Insights 14 (2020) 16-21. https://doi.org/10.1177/1177932219899051 [205] P.K. Davis, A. O’Mahony, J. Pfautz. Social-behavioral modeling for complex systems. Social- Behavioral Modeling for Complex Systems (2019) 1-947. https://doi.org/10.1002/9781119485001 [206] A. Pardo, G. Siemens. Ethical and privacy principles for learning analytics. British Journal of Educational Technology 45 (2014) 438-450. https://doi.org/10.1111/bjet.12152 [207] J. Francom. Rfid: a Survey of Ethical and Privacy Concerns. Issues In Information Systems 8(2) (2007) 336-340. https://doi.org/10.48009/2_iis_2007_336-340 [208] P. Wisniewski, J. Vitak, X. Page, B. Knijnenburg, Y. Wang, C. Fiesler. In whose best interest? Exploring the real, potential, and imagined ethical concerns in privacy-focused agenda. CSCW 2017 - Companion of the 2017 ACM Conference on Computer Supported Cooperative Work and Social Computing (2017) 377-382. https://doi.org/10.1145/3022198.3022660 [209] A. Nayarisseri, R. Khandelwal, P. Tanwar, M. Madhavi, D. Sharma, G. Thakur, A. Speck-Planche, S.K. Singh. Artificial Intelligence, Big Data and Machine Learning Approaches in Precision Medicine & Drug Discovery. Current Drug Targets 22 (2021) 631-655. https://doi.org/10.2174/1389450122999210104205732 [210] V.G. Maltarollo, J.C. Gertrudes, P.R. Oliveira, K.M. Honorio. Applying machine learning techniques for ADME-Tox prediction: A review. Expert Opinion on Drug Metabolism and Toxicology 11 (2015) 259-271. https://doi.org/10.1517/17425255.2015.980814 [211] Y. Bidault. A flexible approach for optimising in silico ADME/Tox characterisation of lead candidates. Expert Opinion on Drug Metabolism and Toxicology 2 (2006) 157-168. https://doi.org/10.1517/17425255.2.1.157 [212] J.G.M. Bessems, L. Geraets. Proper knowledge on toxicokinetics improves human hazard testing and subsequent health risk characterisation. A case study approach. Regulatory Toxicology and Pharmacology 67 (2013) 325-334. https://doi.org/10.1016/j.yrtph.2013.08.010 [213] S.N. Deftereos, C. Andronis, E.J. Friedla, A. Persidis, A. Persidis. Drug repurposing and adverse event prediction using high-throughput literature analysis. Wiley Interdisciplinary Reviews: Systems Biology and Medicine 3 (2011) 323-334. https://doi.org/10.1002/wsbm.147 [214] L. Mucke. EC-03-04: Tau-Dependent Signaling and Neuronal Network Hyperexcitability. Alzheimer’s & Dementia 12 (2016) P269. https://doi.org/10.1016/j.jalz.2016.06.2382 https://doi.org/10.5599/admet.2772 https://doi.org/10.1109/BigData.2015.7364065 https://doi.org/10.1007/978-3-031-01892-3 https://doi.org/10.1007/978-0-85729-320-6_91 http://dx.doi.org/10.1007/3-540-33173-5 https://doi.org/10.1016/j.csbj.2021.06.030 https://doi.org/10.1007/978-3-031-38325-0_10 https://doi.org/10.1101/238766 https://doi.org/10.1177/1177932219899051 https://doi.org/10.1002/9781119485001 https://doi.org/10.1111/bjet.12152 https://doi.org/10.48009/2_iis_2007_336-340 https://doi.org/10.1145/3022198.3022660 https://doi.org/10.2174/1389450122999210104205732 https://doi.org/10.1517/17425255.2015.980814 https://doi.org/10.1517/17425255.2.1.157 https://doi.org/10.1016/j.yrtph.2013.08.010 https://doi.org/10.1002/wsbm.147 https://doi.org/10.1016/j.jalz.2016.06.2382 M. Venkataraman et al. ADMET & DMPK 13(3) (2025) 2772 38 [215] F. Farouk, R. Shamma. Chemical structure modifications and nano-technology applications for improving ADME-Tox properties, a review. Archiv Der Pharmazie 352 (2019) 1800213. https://doi.org/10.1002/ardp.201800213 [216] G. Szakács, A. Váradi, C. Özvegy-Laczka, B. Sarkadi. The role of ABC transporters in drug absorption, distribution, metabolism, excretion and toxicity (ADME-Tox). Drug Discovery Today 13 (2008) 379- 393. https://doi.org/10.1016/j.drudis.2007.12.010 [217] C. Kim, J. Jeong, J. Choi. Effects of Class Imbalance and Data Scarcity on the Performance of Binary Classification Machine Learning Models Developed Based on ToxCast/Tox21 Assay Data. Chemical Research in Toxicology 35 (2022) 2219-2226. https://doi.org/10.1021/acs.chemrestox.2c00189 [218] Y.O. Lee, Y.J. Kim. The Effect of Resampling on Data-imbalanced Conditions for Prediction towards Nuclear Receptor Profiling Using Deep Learning. Molecular Informatics 39 (2020) 1900131. https://doi.org/10.1002/minf.201900131 [219] L. Tao, P. Zhang, C. Qin, S.Y. Chen, C. Zhang, Z. Chen, F. Zhu, S.Y. Yang, Y.Q. Wei, Y.Z. Chen. Recent progresses in the exploration of machine learning methods as in-silico ADME prediction tools. Advanced Drug Delivery Reviews 86 (2015) 83-100. https://doi.org/10.1016/j.addr.2015.03.014 [220] X. Sun, J. Zhu, B. Chen, H. You, H. Xu. A feature transferring workflow between data-poor compounds in various tasks. PLoS ONE 17 (2022) e0266088. https://doi.org/10.1371/journal.pone.0266088 [221] L. Alzubaidi, J. Bai, A. Al-Sabaawi, J. Santamaría, A.S. Albahri, B.S.N. Al-dabbagh, M.A. Fadhel, M. Manoufali, J. Zhang, A.H. Al-Timemy, Y. Duan, A. Abdullah, L. Farhan, Y. Lu, A. Gupta, F. Albu, A. Abbosh, Y. Gu. A survey on deep learning tools dealing with data scarcity: definitions, challenges, solutions, tips, and applications. Journal of Big Data 10 (2023) 46. https://doi.org/10.1186/s40537- 023-00727-2 [222] D.A. Dablain, C. Bellinger, B. Krawczyk, D.W. Aha, N. V. Chawla. Interpretable ML for Imbalanced Data. ArXiv (2022). http://arxiv.org/abs/2212.07743 [223] Y. Chen, R. Calabrese, B. Martin-Barragan. Interpretable machine learning for imbalanced credit scoring datasets. European Journal of Operational Research 312 (2024) 357-372. https://doi.org/10.1016/j.ejor.2023.06.036 [224] C.H. Chang, S. Tan, B. Lengerich, A. Goldenberg, R. Caruana. How Interpretable and Trustworthy are GAMs? Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (2021) 95-105. https://doi.org/10.1145/3447548.3467453 [225] Y. Liu, X. Yu. Farthest Point Sampling in Property Designated Chemical Feature Space as a General Strategy for Enhancing the Machine Learning Model Performance for Small Scale Chemical Dataset. (2024). http://arxiv.org/abs/2404.11348 [226] N.C. Iovanac, B.M. Savoie. Improved Chemical Prediction from Scarce Data Sets via Latent Space Enrichment. Journal of Physical Chemistry A 123 (2019) 4295-4302. https://doi.org/10.1021/acs.jpca.9b01398 [227] D. Lemm, G.F. von Rudorff, O.A. von Lilienfeld. Improved decision making with similarity based machine learning. (2022). http://arxiv.org/abs/2205.05633 [228] D. Wade, A. Wilson, A. Reddy, R. Bharadwaj. Validating machine-learned diagnostic classifiers in safety critical applications with imbalanced populations. Proceedings of the Annual Conference of the Prognostics and Health Management Society, PHM (2018). https://doi.org/10.36001/phmconf.2018.v10i1.192 [229] J.D. Walker, L. Carlsen, J. Jaworska. Improving opportunities for regulatory acceptance of QSARs: The importance of model domain, uncertainty, validity and predictability. QSAR and Combinatorial Science 22 (2003) 346-350. https://doi.org/10.1002/qsar.200390024 [230] Y. Wang, K. Wang, C. Zhang. Applications of artificial intelligence/machine learning to high- performance composites. Composites Part B: Engineering 285 (2024) 111740. https://doi.org/10.1016/j.compositesb.2024.111740 https://doi.org/10.1002/ardp.201800213 https://doi.org/10.1016/j.drudis.2007.12.010 https://doi.org/10.1021/acs.chemrestox.2c00189 https://doi.org/10.1002/minf.201900131 https://doi.org/10.1016/j.addr.2015.03.014 https://doi.org/10.1371/journal.pone.0266088 https://doi.org/10.1186/s40537-023-00727-2 https://doi.org/10.1186/s40537-023-00727-2 http://arxiv.org/abs/2212.07743 https://doi.org/10.1016/j.ejor.2023.06.036 https://doi.org/10.1145/3447548.3467453 http://arxiv.org/abs/2404.11348 https://doi.org/10.1021/acs.jpca.9b01398 http://arxiv.org/abs/2205.05633 https://doi.org/10.36001/phmconf.2018.v10i1.192 https://doi.org/10.1002/qsar.200390024 https://doi.org/10.1016/j.compositesb.2024.111740 ADMET & DMPK 13(3) (2025) 2772 Machine learning models for ADMET prediction in drug development doi: https://doi.org/10.5599/admet.2772 39 [231] J.P. Janet, L. Mervin, O. Engkvist. Artificial intelligence in molecular de novo design: Integration with experiment. Current Opinion in Structural Biology 80 (2023). https://doi.org/10.1016/j.sbi.2023.102575 [232] S. Capponi, K.G. Daniels. Harnessing the power of artificial intelligence to advance cell therapy. Immunological Reviews 320 (2023) 147-165. https://doi.org/10.1111/imr.13236 [233] Y. Hu, W. Li, D. Wright, O. Aydin, D. Wilson, O. Maher, M. Raad. Artificial Intelligence Approaches. Geographic Information Science & Technology Body of Knowledge 2019 (2019) 102575. https://doi.org/10.22224/gistbok/2019.3.4 [234] J.E.T. Taylor, G.W. Taylor. Artificial cognition: How experimental psychology can help generate explainable artificial intelligence. Psychonomic Bulletin and Review 28 (2021) 454-475. https://doi.org/10.3758/s13423-020-01825-5 [235] F. Gossen, T. Margaria, B. Steffen. Formal Methods Boost Experimental Performance for Explainable AI. IT Professional 23 (2021) 8-12. https://doi.org/10.1109/MITP.2021.3123495 [236] D. Hemment, D. Murray-Rust, V. Belle, R. Aylett, ... Experiential AI: A transdisciplinary framework for legibility and agency in AI. ArXiv Preprint ArXiv … (2023). https://doi.org/10.48550/arXiv.2306.00635 [237] M.T. Hosain, J.R. Jim, M.F. Mridha, M.M. Kabir. Explainable AI approaches in deep learning: Advancements, applications and challenges. Computers and Electrical Engineering 117 (2024) 109246. https://doi.org/10.1016/j.compeleceng.2024.109246 [238] Y. Cao, J. Romero, A. Aspuru-Guzik. Potential of quantum computing for drug discovery. IBM Journal of Research and Development 62 (2018) 6:1-6:20. https://doi.org/10.1147/JRD.2018.2888987 [239] M. Avramouli, I.K. Savvas, A. Vasilaki, G. Garani. Unlocking the Potential of Quantum Machine Learning to Advance Drug Discovery. Electronics (Switzerland) 12 (2023) 2402. https://doi.org/10.3390/electronics12112402 [240] S.K. Kandula, N. Katam, P.R. Kangari, A. Hijmal, R. Gurrala, M. Mahmoud. Quantum Computing Potentials for Drug Discovery. Proceedings - 2023 International Conference on Computational Science and Computational Intelligence, CSCI 2023 (2023) 1467-1473. https://doi.org/10.1109/CSCI62032.2023.00240 [241] P.H. Wang, J.H. Chen, Y.Y. Yang, C. Lee, Y.J. Tseng. Recent Advances in Quantum Computing for Drug Discovery and Development. IEEE Nanotechnology Magazine 17 (2023) 26-30. https://doi.org/10.1109/MNANO.2023.3249499 [242] K. Louhichi, I. Abdelghani, H. Jdidi, Y. Boukhris, B. Roch, C. François, M. Toumi, A. Bakhutashvili. MSR8 Revolutionizing Drug Discovery and Preclinical Research Via Artificial Intelligence: A Targeted Literature Review. Value in Health 25 (2022) S518-S519. https://doi.org/10.1016/j.jval.2022.04.1215 [243] N. Stephenson, E. Shane, J. Chase, J. Rowland, D. Ries, N. Justice, J. Zhang, L. Chan, R. Cao. Survey of Machine Learning Techniques in Drug Discovery. Current Drug Metabolism 20 (2018) 185-193. https://doi.org/10.2174/1389200219666180820112457 [244] M. Baranwal, A. Magner, P. Elvati, J. Saldinger, A. Violi, A. Violi, A.O. Hero. A deep learning architecture for metabolic pathway prediction. Bioinformatics 36 (2020) 2547-2553. https://doi.org/10.1093/bioinformatics/btz954 [245] K. Sasahara, M. Shibata, H. Sasabe, T. Suzuki, K. Takeuchi, K. Umehara, E. Kashiyama. Predicting drug metabolism and pharmacokinetics features of in-house compounds by a hybrid machine- learning model. Drug Metabolism and Pharmacokinetics 39 (2021) 100395. https://doi.org/10.1016/j.dmpk.2021.100395 ©2025 by the authors; licensee IAPC, Zagreb, Croatia. This article is an open-access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/) https://doi.org/10.5599/admet.2772 https://doi.org/10.1016/j.sbi.2023.102575 https://doi.org/10.1111/imr.13236 https://doi.org/10.22224/gistbok/2019.3.4 https://doi.org/10.3758/s13423-020-01825-5 https://doi.org/10.1109/MITP.2021.3123495 https://doi.org/10.48550/arXiv.2306.00635 https://doi.org/10.1016/j.compeleceng.2024.109246 https://doi.org/10.1147/JRD.2018.2888987 https://doi.org/10.3390/electronics12112402 https://doi.org/10.1109/CSCI62032.2023.00240 https://doi.org/10.1109/MNANO.2023.3249499 https://doi.org/10.1016/j.jval.2022.04.1215 https://doi.org/10.2174/1389200219666180820112457 https://doi.org/10.1093/bioinformatics/btz954 https://doi.org/10.1016/j.dmpk.2021.100395 http://creativecommons.org/licenses/by/3.0/