The use of pseudo-equilibrium constant affords improved QSAR models of human plasma protein binding

Multitask Deep Learning Models of Combined Industrial Absorption, Distribution, Metabolism, and Excretion Datasets to Improve Generalization

The optimization of absorption, distribution, metabolism, and excretion (ADME) profiles of compounds is critical to the drug discovery process. As such, machine learning (ML) models for ADME are widely used for prioritizing the design and synthesis of compounds. The effectiveness of ML models for ADME depends on the availability of high-quality experimental data for a diverse set of compounds that is relevant to the emerging chemical space being explored by the drug discovery teams. To that end, ADME data sets from Genentech and Roche were combined to evaluate the impact of expanding the chemical space on the performance of ML models, a first experiment of its kind for large-scale, historical ADME data sets. The combined ADME data set consisted of over 1 million individual measurements distributed across 11 assay end points. We utilized a multitask (MT) neural network architecture that enables the modeling of multiple end points simultaneously and thereby exploits information transfer between interconnected ADME end points. Both single- and cross-site MT models were trained and compared against single-site, single-task baseline models. Given the differences in assay protocols across the two sites, the data for corresponding end points across sites were modeled as separate tasks. Models were evaluated against test sets representing varying degrees of extrapolation difficulty, including cluster-based, temporal, and external test sets. We found that cross-site MT models appeared to provide a greater generalization capacity compared to single-site models. The performance improvement of the cross-site MT models was more pronounced for the relatively "distant" external and temporal test sets, suggesting an expanded applicability domain. The data exchange exercise described here demonstrates the value of expanding the learning from ADME data from multiple sources without the need to aggregate such data when the experimental methods are disparate.

Refined ADME Profiles for ATC Drug Classes

Background: Modern generative chemistry initiatives aim to produce potent and selective novel synthetically feasible molecules with suitable pharmacokinetic properties. General ranges of physicochemical properties relevant for the absorption, distribution, metabolism, and excretion (ADME) of drugs have been used for decades. However, the therapeutic indication, dosing route, and pharmacodynamic response of the individual drug discovery program may ultimately define a distinct desired property profile. Methods: A methodological pipeline to build and validate machine learning (ML) models on physicochemical and ADME properties of small molecules is introduced. Results: The analysis of publicly available data on several ADME properties presented in this work reveals significant differences in the property value distributions across the various levels of the anatomical, therapeutic, and chemical (ATC) drug classification. For most properties, the predicted data distributions agree well with the corresponding distributions derived from experimental data across fourteen drug classes. Conclusions: The refined ADME profiles for ATC drug classes should be useful to guide the de novo generation of advanced lead structures directed toward specific therapeutic indications.

The state-of-the-art machine learning model for plasma protein binding prediction: Computational modeling with OCHEM and experimental validation

Plasma protein binding (PPB) is closely related to pharmacokinetics, pharmacodynamics and drug toxicity. Existing models for predicting PPB often suffer from low prediction accuracy and poor interpretability, especially for high PPB compounds, and are most often not experimentally validated. Here, we carried out a strict data curation protocol, and applied consensus modeling to obtain a model with a coefficient of determination of 0.90 and 0.91 on the training set and the test set, respectively. This model (available on the OCHEM platform https://ochem.eu/article/29) was further retrospectively validated for a set of 63 poly-fluorinated molecules and prospectively validated for a set of 25 highly diverse compounds, and its performance for both these sets was superior to that of the other previously reported models. Furthermore, we identified the physicochemical and structural characteristics of high and low PPB molecules for further structural optimization. Finally, we provide practical and detailed recommendations for structural optimization to decrease PPB binding of lead compounds.

Development of machine learning-based quantitative structure-activity relationship models for predicting plasma half-lives of drugs in six common food animal species

Plasma half-life is a crucial pharmacokinetic parameter for estimating extralabel withdrawal intervals of drugs to ensure the safety of food products derived from animals. This study focuses on developing a quantitative structure-activity relationship (QSAR) model incorporating multiple machine learning and artificial intelligence algorithms, and aims to predict the plasma half-lives of drugs in 6 food animals, including cattle, chickens, goats, sheep, swine, and turkeys. By integrating 4 machine learning algorithms with 5 molecular descriptor types, 20 QSAR models were developed using data from the Food Animal Residue Avoidance Databank (FARAD) Comparative Pharmacokinetic Database. The deep neural network (DNN) algorithm demonstrated the best prediction ability of plasma half-lives. The DNN model with all descriptors achieved superior performance with a high coefficient of determination (R2) of 0.82 ± 0.19 in 5-fold cross-validation on the training sets and an R2 of 0.67 on the independent test set, indicating accurate predictions and good generalizability. The final model was converted to a user-friendly web dashboard to facilitate its wide application by the scientific community. This machine learning-based QSAR model serves as a valuable tool for predicting drug plasma half-lives and extralabel withdrawal intervals in 6 common food animals based on physicochemical properties. It also provides a foundation to develop more advanced models to predict the tissue half-life of drugs in food animals.

Comprehensive benchmarking of computational tools for predicting toxicokinetic and physicochemical properties of chemicals

Ensuring the safety of chemicals for environmental and human health involves assessing physicochemical (PC) and toxicokinetic (TK) properties, which are crucial for absorption, distribution, metabolism, excretion, and toxicity (ADMET). Computational methods play a vital role in predicting these properties, given the current trends in reducing experimental approaches, especially those that involve animal experimentation. In the present manuscript, twelve software tools implementing Quantitative Structure-Activity Relationship (QSAR) models were selected for the prediction of 17 relevant PC and TK properties. A total of 41 validation datasets were collected from the literature, curated and used for assessing the models' external predictivity, emphasizing the performance of the models inside the applicability domain. Overall, the results confirmed the adequate predictive performance of the majority of the selected tools, with models for PC properties (R2 average = 0.717) generally outperforming those for TK properties (R2 average = 0.639 for regression, average balanced accuracy = 0.780 for classification). Notably, several of the tools evaluated exhibited good predictivity across different properties and were identified as recurring optimal choices. Moreover, a systematic analysis of the chemical space covered by the external validation datasets confirmed the validity of the collected results for relevant chemical categories (e.g., drugs and industrial chemicals), further increasing the confidence in the overall evaluation. The best performing models were ultimately suggested for each investigated property and proposed as robust computational tools for high-throughput assessment of highly relevant chemical properties. SCIENTIFIC CONTRIBUTION: The present manuscript provides an overview of the state-of-the-art available computational tools for predicting the PC and TK properties of chemicals. The results here offer valuable guidance to researchers, regulatory authorities, and the industry in identifying robust computational tools suitable for predicting relevant chemical properties in the context of chemical design, toxicity and environmental fate assessment.

Exploring molecular fragments for fraction unbound in human plasma of chemicals: a fragment-based cheminformatics approach

Fraction unbound in plasma (fu,p) of drugs is an significant factor for drug delivery and other biological incidences related to the pharmacokinetic behaviours of drugs. Exploration of different molecular fragments for fu,p of different small molecules/agents can facilitate in identification of suitable candidates in the preliminary stage of drug discovery. Different researchers have implemented strategies to build several prediction models for fu,p of different drugs. However, these studies did not focus on the identification of responsible molecular fragments to determine the fraction unbound in plasma. In the current work, we tried to focus on the development of robust classification-based QSAR models and evaluated these models with multiple statistical metrics to identify essential molecular fragments/structural attributes for fractions unbound in plasma. The study unequivocally suggests various N-containing aromatic rings and aliphatic groups have positive influences and sulphur-containing thiadiazole rings have negative influences for the fu,p values. The molecular fragments may help for the assessment of the fu,p values of different small molecules/drugs in a speedy way in comparison to experiment-based in vivo and in vitro studies.

Systematic evaluation of high-throughput PBK modelling strategies for the prediction of intravenous and oral pharmacokinetics in humans

Physiologically based kinetic (PBK) modelling offers a mechanistic basis for predicting the pharmaco-/toxicokinetics of compounds and thereby provides critical information for integrating toxicity and exposure data to replace animal testing with in vitro or in silico methods. However, traditional PBK modelling depends on animal and human data, which limits its usefulness for non-animal methods. To address this limitation, high-throughput PBK modelling aims to rely exclusively on in vitro and in silico data for model generation. Here, we evaluate a variety of in silico tools and different strategies to parameterise PBK models with input values from various sources in a high-throughput manner. We gather 2000 + publicly available human in vivo concentration-time profiles of 200 + compounds (IV and oral administration), as well as in silico, in vitro and in vivo determined compound-specific parameters required for the PBK modelling of these compounds. Then, we systematically evaluate all possible PBK model parametrisation strategies in PK-Sim and quantify their prediction accuracy against the collected in vivo concentration-time profiles. Our results show that even simple, generic high-throughput PBK modelling can provide accurate predictions of the pharmacokinetics of most compounds (87% of Cmax and 84% of AUC within tenfold). Nevertheless, we also observe major differences in prediction accuracies between the different parameterisation strategies, as well as between different compounds. Finally, we outline a strategy for high-throughput PBK modelling that relies exclusively on freely available tools. Our findings contribute to a more robust understanding of the reliability of high-throughput PBK modelling, which is essential to establish the confidence necessary for its utilisation in Next-Generation Risk Assessment.

Advances in Modeling Approaches for Oral Drug Delivery: Artificial Intelligence, Physiologically-Based Pharmacokinetics, and First-Principles Models

Oral drug absorption is the primary route for drug administration. However, this process hinges on multiple factors, including the drug's physicochemical properties, formulation characteristics, and gastrointestinal physiology. Given its intricacy and the exorbitant costs associated with experimentation, the trial-and-error method proves prohibitively expensive. Theoretical models have emerged as a cost-effective alternative by assimilating data from diverse experiments and theoretical considerations. These models fall into three categories: (i) data-driven models, encompassing classical pharmacokinetics, quantitative-structure models (QSAR), and machine/deep learning; (ii) mechanism-based models, which include quasi-equilibrium, steady-state, and physiologically-based pharmacokinetics models; and (iii) first principles models, including molecular dynamics and continuum models. This review provides an overview of recent modeling endeavors across these categories while evaluating their respective advantages and limitations. Additionally, a primer on partial differential equations and their numerical solutions is included in the appendix, recognizing their utility in modeling physiological systems despite their mathematical complexity limiting widespread application in this field.

The Trends and Future Prospective of In Silico Models from the Viewpoint of ADME Evaluation in Drug Discovery

Drug discovery and development are aimed at identifying new chemical molecular entities (NCEs) with desirable pharmacokinetic profiles for high therapeutic efficacy. The plasma concentrations of NCEs are a biomarker of their efficacy and are governed by pharmacokinetic processes such as absorption, distribution, metabolism, and excretion (ADME). Poor ADME properties of NCEs are a major cause of attrition in drug development. ADME screening is used to identify and optimize lead compounds in the drug discovery process. Computational models predicting ADME properties have been developed with evolving model-building technologies from a simplified relationship between ADME endpoints and physicochemical properties to machine learning, including support vector machines, random forests, and convolution neural networks. Recently, in the field of in silico ADME research, there has been a shift toward evaluating the in vivo parameters or plasma concentrations of NCEs instead of using predictive results to guide chemical structure design. Another research hotspot is the establishment of a computational prediction platform to strengthen academic drug discovery. Bioinformatics projects have produced a series of in silico ADME models using free software and open-access databases. In this review, we introduce prediction models for various ADME parameters and discuss the currently available academic drug discovery platforms.

Descriptor-Free Deep Learning QSAR Model for the Fraction Unbound in Human Plasma

Chemical-specific parameters are either measured in vitro or estimated using quantitative structure-activity relationship (QSAR) models. The existing body of QSAR work relies on extracting a set of descriptors or fingerprints, subset selection, and training a machine learning model. In this work, we used a state-of-the-art natural language processing model, Bidirectional Encoder Representations from Transformers, which allowed us to circumvent the need for calculation of these chemical descriptors. In this approach, simplified molecular-input line-entry system (SMILES) strings were embedded in a high-dimensional space using a two-stage training approach. The model was first pre-trained on a masked SMILES token task and then fine-tuned on a QSAR prediction task. The pre-training task learned meaningful high-dimensional embeddings based upon the relationships between the chemical tokens in the SMILES strings derived from the "in-stock" portion of the ZINC 15 dataset─a large dataset of commercially available chemicals. The fine-tuning task then perturbed the pre-trained embeddings to facilitate prediction of a specific QSAR endpoint of interest. The power of this model stems from the ability to reuse the pre-trained model for multiple different fine-tuning tasks, reducing the computational burden of developing multiple models for different endpoints. We used our framework to develop a predictive model for fraction unbound in human plasma (fu,p). This approach is flexible, requires minimum domain expertise, and can be generalized for other parameters of interest for rapid and accurate estimation of absorption, distribution, metabolism, excretion, and toxicity.

Rapid equilibrium dialysis, ultrafiltration or ultracentrifugation? Evaluation of methods to quantify the unbound fraction of substances in plasma

In pharmacokinetics plasma protein binding (PPB) is a well-established parameter impacting drug disposition. The unbound fraction (f_u) is arguably regarded the effective concentration at the target site. Pharmacology and toxicology, increasingly use in vitro models. The translation of in vitro concentrations to in vivo doses can be supported by toxicokinetic modelling, e.g. physiologically based toxicokinetic models (PBTK). PPB of a test substance is an input parameter for PBTK. We compared three methods to quantify f_u: rapid equilibrium dialysis (RED), ultrafiltration (UF) and ultracentrifugation (UC) using twelve substances covering a wide range of Log P_ow (-0.1 to 6.8) and molecular weights (151 and 531 g/mol): Acetaminophen, Bisphenol A, Caffeine, Colchicine, Fenarimol, Flutamide, Genistein, Ketoconazole, α-Methyltestosterone, Tamoxifen, Trenbolone and Warfarin. After RED and UF separation, three polar substances (Log P_ow < 2) were largely unbound (f_u > 70%), while more lipophilic substances were largely bound (f_u < 33%). Compared to RED or UF, UC resulted in a generally higher f_u of lipophilic substances. f_u obtained after RED and UF were more consistent with published data. For half of the substances, UC resulted in f_u higher than the reference data. UF, RED and both UF and UC, resulted in lower f_u of Flutamide, Ketoconazole and Colchicine, respectively. For f_u quantifications, the separation method should be selected according to the test substance's properties. Based on our data, RED is suitable for a broader range of substances while UC and UF are suitable for polar substances.

A Machine Learning Model to Estimate Toxicokinetic Half-Lives of Per- and Polyfluoro-Alkyl Substances (PFAS) in Multiple Species

Per- and polyfluoroalkyl substances (PFAS) are a diverse group of man-made chemicals that are commonly found in body tissues. The toxicokinetics of most PFAS are currently uncharacterized, but long half-lives (t½) have been observed in some cases. Knowledge of chemical-specific t½ is necessary for exposure reconstruction and extrapolation from toxicological studies. We used an ensemble machine learning method, random forest, to model the existing in vivo measured t½ across four species (human, monkey, rat, mouse) and eleven PFAS. Mechanistically motivated descriptors were examined, including two types of surrogates for renal transporters: (1) physiological descriptors, including kidney geometry, for renal transporter expression and (2) structural similarity of defluorinated PFAS to endogenous chemicals for transporter affinity. We developed a classification model for t½ (Bin 1: <12 h; Bin 2: <1 week; Bin 3: <2 months; Bin 4: >2 months). The model had an accuracy of 86.1% in contrast to 32.2% for a y-randomized null model. A total of 3890 compounds were within domain of the model, and t½ was predicted using the bin medians: 4.9 h, 2.2 days, 33 days, and 3.3 years. For human t½, 56% of PFAS were classified in Bin 4, 7% were classified in Bin 3, and 37% were classified in Bin 2. This model synthesizes the limited available data to allow tentative extrapolation and prioritization.

Recent Studies of Artificial Intelligence on In Silico Drug Distribution Prediction

Drug distribution is an important process in pharmacokinetics because it has the potential to influence both the amount of medicine reaching the active sites and the effectiveness as well as safety of the drug. The main causes of 90% of drug failures in clinical development are lack of efficacy and uncontrolled toxicity. In recent years, several advances and promising developments in drug distribution property prediction have been achieved, especially in silico, which helped to drastically reduce the time and expense of screening undesired drug candidates. In this study, we provide comprehensive knowledge of drug distribution background, influencing factors, and artificial intelligence-based distribution property prediction models from 2019 to the present. Additionally, we gathered and analyzed public databases and datasets commonly utilized by the scientific community for distribution prediction. The distribution property prediction performance of five large ADMET prediction tools is mentioned as a benchmark for future research. On this basis, we also offer future challenges in drug distribution prediction and research directions. We hope that this review will provide researchers with helpful insight into distribution prediction, thus facilitating the development of innovative approaches for drug discovery.

Towards best use and regulatory acceptance of generic physiologically based kinetic (PBK) models for in vitro-to-in vivo extrapolation (IVIVE) in chemical risk assessment

With an increasing need to incorporate new approach methodologies (NAMs) in chemical risk assessment and the concomitant need to phase out animal testing, the interpretation of in vitro assay readouts for quantitative hazard characterisation becomes more important. Physiologically based kinetic (PBK) models, which simulate the fate of chemicals in tissues of the body, play an essential role in extrapolating in vitro effect concentrations to in vivo bioequivalent exposures. As PBK-based testing approaches evolve, it will become essential to standardise PBK modelling approaches towards a consensus approach that can be used in quantitative in vitro-to-in vivo extrapolation (QIVIVE) studies for regulatory chemical risk assessment based on in vitro assays. Based on results of an ECETOC expert workshop, steps are recommended that can improve regulatory adoption: (1) define context and implementation, taking into consideration model complexity for building fit-for-purpose PBK models, (2) harmonise physiological input parameters and their distribution and define criteria for quality chemical-specific parameters, especially in the absence of in vivo data, (3) apply Good Modelling Practices (GMP) to achieve transparency and design a stepwise approach for PBK model development for risk assessors, (4) evaluate model predictions using alternatives to in vivo PK data including read-across approaches, (5) use case studies to facilitate discussions between modellers and regulators of chemical risk assessment. Proof-of-concepts of generic PBK modelling approaches are published in the scientific literature at an increasing rate. Working on the previously proposed steps is, therefore, needed to gain confidence in PBK modelling approaches for regulatory use.

IDL-PPBopt: A Strategy for Prediction and Optimization of Human Plasma Protein Binding of Compounds via an Interpretable Deep Learning Method

The prediction and optimization of pharmacokinetic properties are essential in lead optimization. Traditional strategies mainly depend on the empirical chemical rules from medicinal chemists. However, with the rising amount of data, it is getting more difficult to manually extract useful medicinal chemistry knowledge. To this end, we introduced IDL-PPBopt, a computational strategy for predicting and optimizing the plasma protein binding (PPB) property based on an interpretable deep learning method. At first, a curated PPB data set was used to construct an interpretable deep learning model, which showed excellent predictive performance with a root mean squared error of 0.112 for the entire test set. Then, we designed a detection protocol based on the model and Wilcoxon test to identify the PPB-related substructures (named privileged substructures, PSubs) for each molecule. In total, 22 general privileged substructures (GPSubs) were identified, which shared some common features such as nitrogen-containing groups, diamines with two carbon units, and azetidine. Furthermore, a series of second-level chemical rules for each GPSub were derived through a statistical test and then summarized into substructure pairs. We demonstrated that these substructure pairs were equally applicable outside the training set and accordingly customized the structural modification schemes for each GPSub, which provided alternatives for the optimization of the PPB property. Therefore, IDL-PPBopt provides a promising scheme for the prediction and optimization of the PPB property and would be helpful for lead optimization of other pharmacokinetic properties.

Comparative Study for Prediction of Low and High Plasma Protein Binding Drugs by Various Machine Learning-Based Classification Algorithms

<p>In the drug discovery path, most drug candidates failed at the early stages due to their pharmacokinetic behavior in the system. Early prediction of pharmacokinetic properties and screening methods can reduce the time and investment for lead discoveries. Plasma protein binding is one of these properties which has a vital role in drug discovery and development. The focus of the current study is to develop a computational model for the classification of Low Plasma Protein Binding (LPPB) and High Plasma Protein Binding (HPPB) drugs using machine learning methods for early screening of molecules through WEKA. Plasma protein binding drugs data was collated from the Drug Bank database where 617 drug candidates were found to interact with plasma proteins, out of which an equal proportion of high and low plasma protein binding drugs were extracted to build a training set of ~300 drugs. The machine learning algorithms were trained with a training set and evaluated by a test set. We also compared various machine learning-based classification algorithms i.e., the Naïve Bayes algorithm, Instance-Based Learner (IBK), multilayer perceptron, and random forest to determine the best model based on accuracy. It was observed that the random forest algorithm-based model outperforms with an accuracy of 99.67% and 0.9933 kappa value on training set and on test set as compared to other classification methods and can predict drug plasma binding capacity in the given data set using the WEKA tool.</p>

Plasma protein binding prediction focusing on residue-level features and circularity of cyclic peptides by deep learning

MOTIVATION

In recent years, cyclic peptide drugs have been receiving increasing attention because they can target proteins that are difficult to be tackled by conventional small-molecule drugs or antibody drugs. Plasma protein binding rate (%PPB) is a significant pharmacokinetic property of a compound in drug discovery and design. However, due to structural differences, previous computational prediction methods developed for small-molecule compounds cannot be successfully applied to cyclic peptides, and methods for predicting the PPB rate of cyclic peptides with high accuracy are not yet available.

RESULTS

Cyclic peptides are larger than small molecules, and their local structures have a considerable impact on PPB; thus, molecular descriptors expressing residue-level local features of cyclic peptides, instead of those expressing the entire molecule, as well as the circularity of the cyclic peptides should be considered. Therefore, we developed a prediction method named CycPeptPPB using deep learning that considers both factors. First, the macrocycle ring of cyclic peptides was decomposed residue by residue. The residue-based descriptors were arranged according to the sequence information of the cyclic peptide. Furthermore, the circular data augmentation method was used, and the circular convolution method CyclicConv was devised to express the cyclic structure. CycPeptPPB exhibited excellent performance, with mean absolute error (MAE) of 4.79% and correlation coefficient (R) of 0.92 for the public drug dataset, compared to the prediction performance of the existing PPB rate prediction software (MAE=15.08%, R=0.63).

AVAILABILITY AND IMPLEMENTATION

The data underlying this article are available in the online supplementary material. The source code of CycPeptPPB is available at https://github.com/akiyamalab/cycpeptppb.

SUPPLEMENTARY INFORMATION

Supplementary data are available at Bioinformatics online.

Designing QSARs for Parameters of High-Throughput Toxicokinetic Models Using Open-Source Descriptors

The intrinsic metabolic clearance rate (Cl_int) and the fraction of the chemical unbound in plasma (f_up) serve as important parameters for high-throughput toxicokinetic (TK) models, but experimental data are limited for many chemicals. Open-source quantitative structure-activity relationship (QSAR) models for both parameters were developed to offer reliable in silico predictions for a diverse set of chemicals regulated under the U.S. law, including pharmaceuticals, pesticides, and industrial chemicals. As a case study to demonstrate their utility, model predictions served as inputs to the TK component of a risk-based prioritization approach based on bioactivity/exposure ratios (BERs), in which a BER < 1 indicates that exposures are predicted to exceed a biological activity threshold. When applied to a subset of the Tox21 screening library (6484 chemicals), we found that the proportion of chemicals with BER <1 was similar using either in silico (1133/6484; 17.5%) or in vitro (148/848; 17.5%) parameters. Further, when considering only the chemicals in the Tox21 set with in vitro data, there was a high concordance of chemicals classified with either BER <1 or >1 using either in silico or in vitro parameters (767/848, 90.4%). Thus, the presented QSARs may be suitable for prioritizing the risk posed by many chemicals for which measured in vitro TK data are lacking.

In Silico Prediction of Fraction Unbound in Human Plasma from Chemical Fingerprint Using Automated Machine Learning

Predicting the fraction unbound of a drug in plasma plays a significant role in understanding its pharmacokinetic properties during in vitro studies of drug design and discovery. Owing to the gaining reliability of machine learning in biological predictive models and development of automated machine learning techniques for the ease of nonexperts of machine learning to optimize and maximize the reliability of the model, in this experiment, we built an in silico prediction model of a fraction unbound drug in human plasma using a chemical fingerprint and a freely available AutoML framework. The predictive model was trained on one of the largest data sets ever of 5471 experimental values using four different AutoML frameworks to compare their performance on this problem and to choose the most significant one. With a coefficient of determination of 0.85 on the test data set, our best prediction model showed better performance than other previously published models, giving our model significant importance in pharmacokinetic modeling.

Machine learning in pharmacometrics: Opportunities and challenges

The explosive growth in medical devices, imaging and diagnostics, computing, and communication and information technologies in drug development and healthcare has created an ever-expanding data landscape that the pharmacometrics (PMX) research community must now traverse. The tools of machine learning (ML) have emerged as a powerful computational approach in other data-rich disciplines but its effective utilization in the pharmaceutical sciences and PMX modelling is in its infancy. ML-based methods can complement PMX modelling by enabling the information in diverse sources of big data, e.g. population-based public databases and disease-specific clinical registries, to be harnessed because they are capable of efficiently identifying salient variables associated with outcomes and delineating their interdependencies. ML algorithms are computationally efficient, have strong predictive capabilities and can enable learning in the big data setting. ML algorithms can be viewed as providing a computational bridge from big data to complement PMX modelling. This review provides an overview of the strengths and weaknesses of ML approaches vis-à-vis population methods, assesses current research into ML applications in the pharmaceutical sciences and provides perspective for potential opportunities and strategies for the successful integration and utilization of ML in PMX.

Coloring Molecules with Explainable Artificial Intelligence for Preclinical Relevance Assessment

Graph neural networks are able to solve certain drug discovery tasks such as molecular property prediction and de novo molecule generation. However, these models are considered "black-box" and "hard-to-debug". This study aimed to improve modeling transparency for rational molecular design by applying the integrated gradients explainable artificial intelligence (XAI) approach for graph neural network models. Models were trained for predicting plasma protein binding, hERG channel inhibition, passive permeability, and cytochrome P450 inhibition. The proposed methodology highlighted molecular features and structural elements that are in agreement with known pharmacophore motifs, correctly identified property cliffs, and provided insights into unspecific ligand-target interactions. The developed XAI approach is fully open-sourced and can be used by practitioners to train new models on other clinically relevant endpoints.

Evaluation of Quantitative Structure Property Relationship Algorithms for Predicting Plasma Protein Binding in Humans

The extent of plasma protein binding is an important compound-specific property that influences a compound's pharmacokinetic behavior and is a critical input parameter for predicting exposure in physiologically based pharmacokinetic (PBPK) modeling. When experimentally determined fraction unbound in plasma (fup) data are not available, quantitative structure-property relationship (QSPR) models can be used for prediction. Because available QSPR models were developed based on training sets containing pharmaceutical-like compounds, we compared their prediction accuracy for environmentally relevant and pharmaceutical compounds. Fup values were calculated using Ingle et al., Watanabe et al. and ADMET Predictor (Simulation Plus). The test set included 818 pharmaceutical and environmentally relevant compounds with fup values ranging from 0.01 to 1. Overall, the three QSPR models resulted in over-prediction of fup for highly binding compounds and under-prediction for low or moderately binding compounds. For highly binding compounds (0.01≤ fup ≤ 0.25), Watanabe et al. performed better with a lower mean absolute error (MAE) of 6.7% and a lower mean absolute relative prediction error (RPE) of 171.7 % than other methods. For low to moderately binding compounds, both Ingle et al. and ADMET Predictor performed better than Watanabe et al. with superior MAE and RPE values. The positive polar surface area, the number of basic functional groups and lipophilicity were the most important chemical descriptors for predicting fup. This study demonstrated that the prediction of fup was the most uncertain for highly binding compounds. This suggested that QSPR-predicted fup values should be used with caution in PBPK modeling.

Using Chemical Structure Information to Develop Predictive Models for In Vitro Toxicokinetic Parameters to Inform High-throughput Risk-assessment

The toxicokinetic (TK) parameters fraction of the chemical unbound to plasma proteins and metabolic clearance are critical for relating exposure and internal dose when building in vitro-based risk assessment models. However, experimental toxicokinetic studies have only been carried out on limited chemicals of environmental interest (~1000 chemicals with TK data relative to tens of thousands of chemicals of interest). This work evaluated the utility of chemical structure information to predict TK parameters in silico; development of cluster-based read-across and quantitative structure-activity relationship models of fraction unbound or fub (regression) and intrinsic clearance or Cl_int (classification and regression) using a dataset of 1487 chemicals; utilization of predicted TK parameters to estimate uncertainty in steady-state plasma concentration (C_ss); and subsequent in vitro-in vivo extrapolation analyses to derive bioactivity-exposure ratio (BER) plot to compare human oral equivalent doses and exposure predictions using androgen and estrogen receptor activity data for 233 chemicals as an example dataset. The results demonstrate that fub is structurally more predictable than Cl_int. The model with the highest observed performance for fub had an external test set RMSE/σ=0.62 and R²=0.61, for Cl_int classification had an external test set accuracy = 65.9%, and for intrinsic clearance regression had an external test set RMSE/σ=0.90 and R²=0.20. This relatively low performance is in part due to the large uncertainty in the underlying Cl_int data. We show that C_ss is relatively insensitive to uncertainty in Cl_int. The models were benchmarked against the ADMET Predictor software. Finally, the BER analysis allowed identification of 14 out of 136 chemicals for further risk assessment demonstrating the utility of these models in aiding risk-based chemical prioritization.

Development of Simplified in Vitro P-Glycoprotein Substrate Assay and in Silico Prediction Models To Evaluate Transport Potential of P-Glycoprotein

For efficient drug discovery and screening, it is necessary to simplify P-glycoprotein (P-gp) substrate assays and to provide in silico models that predict the transport potential of P-gp. In this study, we developed a simplified in vitro screening method to evaluate P-gp substrates by unidirectional membrane transport in P-gp-overexpressing cells. The unidirectional flux ratio positively correlated with parameters of the conventional bidirectional P-gp substrate assay ( R² = 0.941) and in vivo K_p,brain ratio (mdr1a/1b KO/WT) in mice ( R² = 0.800). Our in vitro P-gp substrate assay had high reproducibility and required approximately half the labor of the conventional method. We also constructed regression models to predict the value of P-gp-mediated flux and three-class classification models to predict P-gp substrate potential (low-, medium-, and high-potential) using 2397 data entries with the largest data set collected under the same experimental conditions. Most compounds in the test set fell within two- and three-fold errors in the random forest regression model (71.3 and 88.5%, respectively). Furthermore, the random forest three-class classification model showed a high balanced accuracy of 0.821 and precision of 0.761 for the low-potential classes in the test set. We concluded that the simplified in vitro P-gp substrate assay was suitable for compound screening in the early stages of drug discovery and that the in silico regression model and three-class classification model using only chemical structure information could identify the transport potential of compounds including P-gp-mediated flux ratios. Our proposed method is expected to be a practical tool to optimize effective central nervous system (CNS) drugs, to avoid CNS side effects, and to improve intestinal absorption.

An Integrated Transfer Learning and Multitask Learning Approach for Pharmacokinetic Parameter Prediction

BACKGROUND

Pharmacokinetic evaluation is one of the key processes in drug discovery and development. However, current absorption, distribution, metabolism, and excretion prediction models still have limited accuracy.

AIM

This study aims to construct an integrated transfer learning and multitask learning approach for developing quantitative structure-activity relationship models to predict four human pharmacokinetic parameters.

METHODS

A pharmacokinetic data set included 1104 U.S. FDA approved small molecule drugs. The data set included four human pharmacokinetic parameter subsets (oral bioavailability, plasma protein binding rate, apparent volume of distribution at steady-state, and elimination half-life). The pretrained model was trained on over 30 million bioactivity data entries. An integrated transfer learning and multitask learning approach was established to enhance the model generalization.

RESULTS

The pharmacokinetic data set was split into three parts (60:20:20) for training, validation, and testing by the improved maximum dissimilarity algorithm with the representative initial set selection algorithm and the weighted distance function. The multitask learning techniques enhanced the model predictive ability. The integrated transfer learning and multitask learning model demonstrated the best accuracies, because deep neural networks have the general feature extraction ability; transfer learning and multitask learning improve the model generalization.

CONCLUSIONS

The integrated transfer learning and multitask learning approach with the improved data set splitting algorithm was first introduced to predict the pharmacokinetic parameters. This method can be further employed in drug discovery and development.

Computational prediction of plasma protein binding of cyclic peptides from small molecule experimental data using sparse modeling techniques

BACKGROUND

Cyclic peptide-based drug discovery is attracting increasing interest owing to its potential to avoid target protein depletion. In drug discovery, it is important to maintain the biostability of a drug within the proper range. Plasma protein binding (PPB) is the most important index of biostability, and developing a computational method to predict PPB of drug candidate compounds contributes to the acceleration of drug discovery research. PPB prediction of small molecule drug compounds using machine learning has been conducted thus far; however, no study has investigated cyclic peptides because experimental information of cyclic peptides is scarce.

RESULTS

First, we adopted sparse modeling and small molecule information to construct a PPB prediction model for cyclic peptides. As cyclic peptide data are limited, applying multidimensional nonlinear models involves concerns regarding overfitting. However, models constructed by sparse modeling can avoid overfitting, offering high generalization performance and interpretability. More than 1000 PPB data of small molecules are available, and we used them to construct a prediction models with two enumeration methods: enumerating lasso solutions (ELS) and forward beam search (FBS). The accuracies of the prediction models constructed by ELS and FBS were equal to or better than those of conventional non-linear models (MAE = 0.167-0.174) on cross-validation of a small molecule compound dataset. Moreover, we showed that the prediction accuracies for cyclic peptides were close to those for small molecule compounds (MAE = 0.194-0.288). Such high accuracy could not be obtained by a simple method of learning from cyclic peptide data directly by lasso regression (MAE = 0.286-0.671) or ridge regression (MAE = 0.244-0.354).

CONCLUSION

In this study, we proposed a machine learning techniques that uses low-dimensional sparse modeling to predict the PPB value of cyclic peptides computationally. The low-dimensional sparse model not only exhibits excellent generalization performance but also improves interpretation of the prediction model. This can provide common an noteworthy knowledge for future cyclic peptide drug discovery studies.

QSAR Development for Plasma Protein Binding: Influence of the Ionization State

Purpose

This study explored several strategies to improve the performance of literature QSAR models for plasma protein binding (PPB), such as a suitable endpoint transformation, a correct representation of chemicals, more consistency in the dataset, and a reliable definition of the applicability domain.

Methods

We retrieved human fraction unbound (Fu) data for 670 compounds from the literature and carefully checked them for consistency. Descriptors were calculated taking account of the ionization state of molecules at physiological pH (7.4), in order to better estimate the affinity of molecules to blood proteins. We used different algorithms and chemical descriptors to explore the most suitable strategy for modeling the endpoint. SMILES (simplified molecular input line entry system)-based string descriptors were also tested with the CORAL software (CORelation And Logic). We did an outlier analysis to establish the models to use (or not to use) in case of well recognized families.

Results

Internal validation of the selected models returned Q² values close to 0.60. External validation also gave r² values always greater than 0.60. The CORAL descriptor based model for √fu was the best, with r² 0.74 in external validation.

Conclusions

Performance in prediction confirmed the robustness of all the derived models and their suitability for real-life purposes, i.e. screening chemicals for their ADMET profiling. Optimization of descriptors can be useful in order to obtain the correct results with a ionized molecule.

Electronic supplementary material

The online version of this article (10.1007/s11095-018-2561-8) contains supplementary material, which is available to authorized users.

Predicting Fraction Unbound in Human Plasma from Chemical Structure: Improved Accuracy in the Low Value Ranges

Predicting the fraction unbound in plasma provides a good understanding of the pharmacokinetic properties of a drug to assist candidate selection in the early stages of drug discovery. It is also an effective tool to mitigate the risk of late-stage attrition and to optimize further screening. In this study, we built in silico prediction models of fraction unbound in human plasma with freely available software, aiming specifically to improve the accuracy in the low value ranges. We employed several machine learning techniques and built prediction models trained on the largest ever data set of 2738 experimental values. The classification model showed a high true positive rate of 0.826 for the low fraction unbound class on the test set. The strongly biased distribution of the fraction unbound in plasma was mitigated by a logarithmic transformation in the regression model, leading to improved accuracy at lower values. Overall, our models showed better performance than those of previously published methods, including commercial software. Our prediction tool can be used on its own or integrated into other pharmacokinetic modeling systems.

Rapid experimental measurements of physicochemical properties to inform models and testing

The structures and physicochemical properties of chemicals are important for determining their potential toxicological effects, toxicokinetics, and route(s) of exposure. These data are needed to prioritize the risk for thousands of environmental chemicals, but experimental values are often lacking. In an attempt to efficiently fill data gaps in physicochemical property information, we generated new data for 200 structurally diverse compounds, which were rigorously selected from the USEPA ToxCast chemical library, and whose structures are available within the Distributed Structure-Searchable Toxicity Database (DSSTox). This pilot study evaluated rapid experimental methods to determine five physicochemical properties, including the log of the octanol:water partition coefficient (known as log(Kow) or logP), vapor pressure, water solubility, Henry's law constant, and the acid dissociation constant (pKa). For most compounds, experiments were successful for at least one property; log(Kow) yielded the largest return (176 values). It was determined that 77 ToxPrint structural features were enriched in chemicals with at least one measurement failure, indicating which features may have played a role in rapid method failures. To gauge consistency with traditional measurement methods, the new measurements were compared with previous measurements (where available). Since quantitative structure-activity/property relationship (QSAR/QSPR) models are used to fill gaps in physicochemical property information, 5 suites of QSPRs were evaluated for their predictive ability and chemical coverage or applicability domain of new experimental measurements. The ability to have accurate measurements of these properties will facilitate better exposure predictions in two ways: 1) direct input of these experimental measurements into exposure models; and 2) construction of QSPRs with a wider applicability domain, as their predicted physicochemical values can be used to parameterize exposure models in the absence of experimental data.

In vitro to in vivo extrapolation for high throughput prioritization and decision making

In vitro chemical safety testing methods offer the potential for efficient and economical tools to provide relevant assessments of human health risk. To realize this potential, methods are needed to relate in vitro effects to in vivo responses, i.e., in vitro to in vivo extrapolation (IVIVE). Currently available IVIVE approaches need to be refined before they can be utilized for regulatory decision-making. To explore the capabilities and limitations of IVIVE within this context, the U.S. Environmental Protection Agency Office of Research and Development and the National Toxicology Program Interagency Center for the Evaluation of Alternative Toxicological Methods co-organized a workshop and webinar series. Here, we integrate content from the webinars and workshop to discuss activities and resources that would promote inclusion of IVIVE in regulatory decision-making. We discuss properties of models that successfully generate predictions of in vivo doses from effective in vitro concentration, including the experimental systems that provide input parameters for these models, areas of success, and areas for improvement to reduce model uncertainty. Finally, we provide case studies on the uses of IVIVE in safety assessments, which highlight the respective differences, information requirements, and outcomes across various approaches when applied for decision-making.

Evaluation and calibration of high-throughput predictions of chemical distribution to tissues

Toxicokinetics (TK) provides critical information for integrating chemical toxicity and exposure assessments in order to determine potential chemical risk (i.e., the margin between toxic doses and plausible exposures). For thousands of chemicals that are present in our environment, in vivo TK data are lacking. The publicly available R package "httk" (version 1.8, named for "high throughput TK") draws from a database of in vitro data and physico-chemical properties in order to run physiologically-based TK (PBTK) models for 553 compounds. The PBTK model parameters include tissue:plasma partition coefficients (K_p) which the httk software predicts using the model of Schmitt (Toxicol In Vitro 22 (2):457-467, 2008). In this paper we evaluated and modified httk predictions, and quantified confidence using in vivo literature data. We used 964 rat K_p measured by in vivo experiments for 143 compounds. Initially, predicted K_p were significantly larger than measured K_p for many lipophilic compounds (log₁₀ octanol:water partition coefficient > 3). Hence the approach for predicting K_p was revised to account for possible deficiencies in the in vitro protein binding assay, and the method for predicting membrane affinity was revised. These changes yielded improvements ranging from a factor of 10 to nearly a factor of 10,000 for 83 K_p across 23 compounds with only 3 K_p worsening by more than a factor of 10. The vast majority (92%) of K_p were predicted within a factor of 10 of the measured value (overall root mean squared error of 0.59 on log₁₀-transformed scale). After applying the adjustments, regressions were performed to calibrate and evaluate the predictions for 12 tissues. Predictions for some tissues (e.g., spleen, bone, gut, lung) were observed to be better than predictions for other tissues (e.g., skin, brain, fat), indicating that confidence in the application of in silico tools to predict chemical partitioning varies depending upon the tissues involved. Our calibrated model was then evaluated using a second data set of human in vivo measurements of volume of distribution (V_ss) for 498 compounds reviewed by Obach et al. (Drug Metab Dispos 36(7):1385-1405, 2008). We found that calibration of the model improved performance: a regression of the measured values as a function of the predictions has a slope of 1.03, intercept of - 0.04, and R² of 0.43. Through careful evaluation of predictive methods for chemical partitioning into tissues, we have improved and calibrated these methods and quantified confidence for TK predictions in humans and rats.

Evaluation and calibration of high-throughput predictions of chemical distribution to tissues

Toxicokinetics (TK) provides critical information for integrating chemical toxicity and exposure assessments in order to determine potential chemical risk (i.e., the margin between toxic doses and plausible exposures). For thousands of chemicals that are present in our environment, in vivo TK data are lacking. The publicly available R package "httk" (version 1.8, named for "high throughput TK") draws from a database of in vitro data and physico-chemical properties in order to run physiologically-based TK (PBTK) models for 553 compounds. The PBTK model parameters include tissue:plasma partition coefficients (Kp) which the httk software predicts using the model of Schmitt (Toxicol In Vitro 22 (2):457-467, 2008). In this paper we evaluated and modified httk predictions, and quantified confidence using in vivo literature data. We used 964 rat Kp measured by in vivo experiments for 143 compounds. Initially, predicted Kp were significantly larger than measured Kp for many lipophilic compounds (log10 octanol:water partition coefficient > 3). Hence the approach for predicting Kp was revised to account for possible deficiencies in the in vitro protein binding assay, and the method for predicting membrane affinity was revised. These changes yielded improvements ranging from a factor of 10 to nearly a factor of 10,000 for 83 Kp across 23 compounds with only 3 Kp worsening by more than a factor of 10. The vast majority (92%) of Kp were predicted within a factor of 10 of the measured value (overall root mean squared error of 0.59 on log10-transformed scale). After applying the adjustments, regressions were performed to calibrate and evaluate the predictions for 12 tissues. Predictions for some tissues (e.g., spleen, bone, gut, lung) were observed to be better than predictions for other tissues (e.g., skin, brain, fat), indicating that confidence in the application of in silico tools to predict chemical partitioning varies depending upon the tissues involved. Our calibrated model was then evaluated using a second data set of human in vivo measurements of volume of distribution (Vss) for 498 compounds reviewed by Obach et al. (Drug Metab Dispos 36(7):1385-1405, 2008). We found that calibration of the model improved performance: a regression of the measured values as a function of the predictions has a slope of 1.03, intercept of - 0.04, and R2 of 0.43. Through careful evaluation of predictive methods for chemical partitioning into tissues, we have improved and calibrated these methods and quantified confidence for TK predictions in humans and rats.

Computer-Aided Drug Design Applied to Marine Drug Discovery: Meridianins as Alzheimer's Disease Therapeutic Agents

Computer-aided drug discovery/design (CADD) techniques allow the identification of natural products that are capable of modulating protein functions in pathogenesis-related pathways, constituting one of the most promising lines followed in drug discovery. In this paper, we computationally evaluated and reported the inhibitory activity found in meridianins A-G, a group of marine indole alkaloids isolated from the marine tunicate Aplidium, against various protein kinases involved in Alzheimer's disease (AD), a neurodegenerative pathology characterized by the presence of neurofibrillary tangles (NFT). Balance splitting between tau kinase and phosphate activities caused tau hyperphosphorylation and, thereby, its aggregation and NTF formation. Inhibition of specific kinases involved in its phosphorylation pathway could be one of the key strategies to reverse tau hyperphosphorylation and would represent an approach to develop drugs to palliate AD symptoms. Meridianins bind to the adenosine triphosphate (ATP) binding site of certain protein kinases, acting as ATP competitive inhibitors. These compounds show very promising scaffolds to design new drugs against AD, which could act over tau protein kinases Glycogen synthetase kinase-3 Beta (GSK3β) and Casein kinase 1 delta (CK1δ, CK1D or KC1D), and dual specificity kinases as dual specificity tyrosine phosphorylation regulated kinase 1 (DYRK1A) and cdc2-like kinases (CLK1). This work is aimed to highlight the role of CADD techniques in marine drug discovery and to provide precise information regarding the binding mode and strength of meridianins against several protein kinases that could help in the future development of anti-AD drugs.

Chembench: A Publicly Accessible, Integrated Cheminformatics Portal

The enormous increase in the amount of publicly available chemical genomics data and the growing emphasis on data sharing and open science mandates that cheminformaticians also make their models publicly available for broad use by the scientific community. Chembench is one of the first publicly accessible, integrated cheminformatics Web portals. It has been extensively used by researchers from different fields for curation, visualization, analysis, and modeling of chemogenomics data. Since its launch in 2008, Chembench has been accessed more than 1 million times by more than 5000 users from a total of 98 countries. We report on the recent updates and improvements that increase the simplicity of use, computational efficiency, accuracy, and accessibility of a broad range of tools and services for computer-assisted drug design and computational toxicology available on Chembench. Chembench remains freely accessible at https://chembench.mml.unc.edu.

Predicting binding affinities of diverse pharmaceutical chemicals to human serum plasma proteins using QSPR modelling approaches

The prediction of the plasma protein binding (PPB) affinity of chemicals is of paramount significance in the drug development process. In this study, ensemble machine learning-based QSPR models have been established for a four-category classification and PPB affinity prediction of diverse compounds using a large PPB dataset of 930 compounds and in accordance with the OECD guidelines. The structural diversity of the chemicals was tested by the Tanimoto similarity index. The external predictive power of the developed QSPR models was evaluated through internal and external validations. In the QSPR models, XLogP was the most important descriptor. In the test data, the classification QSPR models rendered an accuracy of >93%, while the regression QSPR models yielded r(2) of >0.920 between the measured and predicted PPB affinities, with the root mean squared error <9.77. Values of statistical coefficients derived for the test data were above their threshold limits, thus put a high confidence in this analysis. The QSPR models in this study performed better than any of the previous studies. The results suggest that the developed QSPR models are reliable for predicting the PPB affinity of structurally diverse chemicals. They can be useful for initial screening of candidate molecules in the drug development process.

In vitro, in silico and integrated strategies for the estimation of plasma protein binding. A review

Plasma protein binding (PPB) strongly affects drug distribution and pharmacokinetic behavior with consequences in overall pharmacological action. Extended plasma protein binding may be associated with drug safety issues and several adverse effects, like low clearance, low brain penetration, drug-drug interactions, loss of efficacy, while influencing the fate of enantiomers and diastereoisomers by stereoselective binding within the body. Therefore in holistic drug design approaches, where ADME(T) properties are considered in parallel with target affinity, considerable efforts are focused in early estimation of PPB mainly in regard to human serum albumin (HSA), which is the most abundant and most important plasma protein. The second critical serum protein α1-acid glycoprotein (AGP), although often underscored, plays also an important and complicated role in clinical therapy and thus the last years it has been studied thoroughly too. In the present review, after an overview of the principles of HSA and AGP binding as well as the structure topology of the proteins, the current trends and perspectives in the field of PPB predictions are presented and discussed considering both HSA and AGP binding. Since however for the latter protein systematic studies have started only the last years, the review focuses mainly to HSA. One part of the review highlights the challenge to develop rapid techniques for HSA and AGP binding simulation and their performance in assessment of PPB. The second part focuses on in silico approaches to predict HSA and AGP binding, analyzing and evaluating structure-based and ligand-based methods, as well as combination of both methods in the aim to exploit the different information and overcome the limitations of each individual approach. Ligand-based methods use the Quantitative Structure-Activity Relationships (QSAR) methodology to establish quantitate models for the prediction of binding constants from molecular descriptors, while they provide only indirect information on binding mechanism. Efforts for the establishment of global models, automated workflows and web-based platforms for PPB predictions are presented and discussed. Structure-based methods relying on the crystal structures of drug-protein complexes provide detailed information on the underlying mechanism but are usually restricted to specific compounds. They are useful to identify the specific binding site while they may be important in investigating drug-drug interactions, related to PPB. Moreover, chemometrics or structure-based modeling may be supported by experimental data a promising integrated alternative strategy for ADME(T) properties optimization. In the case of PPB the use of molecular modeling combined with bioanalytical techniques is frequently used for the investigation of AGP binding.

Recursive Random Forests Enable Better Predictive Performance and Model Interpretation than Variable Selection by LASSO

Variable selection is of crucial significance in QSAR modeling since it increases the model predictive ability and reduces noise. The selection of the right variables is far more complicated than the development of predictive models. In this study, eight continuous and categorical data sets were employed to explore the applicability of two distinct variable selection methods random forests (RF) and least absolute shrinkage and selection operator (LASSO). Variable selection was performed: (1) by using recursive random forests to rule out a quarter of the least important descriptors at each iteration and (2) by using LASSO modeling with 10-fold inner cross-validation to tune its penalty λ for each data set. Along with regular statistical parameters of model performance, we proposed the highest pairwise correlation rate, average pairwise Pearson's correlation coefficient, and Tanimoto coefficient to evaluate the optimal by RF and LASSO in an extensive way. Results showed that variable selection could allow a tremendous reduction of noisy descriptors (at most 96% with RF method in this study) and apparently enhance model's predictive performance as well. Furthermore, random forests showed property of gathering important predictors without restricting their pairwise correlation, which is contrary to LASSO. The mutual exclusion of highly correlated variables in LASSO modeling tends to skip important variables that are highly related to response endpoints and thus undermine the model's predictive performance. The optimal variables selected by RF share low similarity with those by LASSO (e.g., the Tanimoto coefficients were smaller than 0.20 in seven out of eight data sets). We found that the differences between RF and LASSO predictive performances mainly resulted from the variables selected by different strategies rather than the learning algorithms. Our study showed that the right selection of variables is more important than the learning algorithm for modeling. We hope that a standard procedure could be developed based on these proposed statistical metrics to select the truly important variables for model interpretation, as well as for further use to facilitate drug discovery and environmental toxicity assessment.

An updated review on drug-induced cholestasis: mechanisms and investigation of physicochemical properties and pharmacokinetic parameters

Drug-induced cholestasis is an important form of acquired liver disease and is associated with significant morbidity and mortality. Bile acids are key signaling molecules, but they can exert toxic responses when they accumulate in hepatocytes. This review focuses on the physiological mechanisms of drug-induced cholestasis associated with altered bile acid homeostasis due to direct (e.g., bile acid transporter inhibition) or indirect (e.g., activation of nuclear receptors, altered function/expression of bile acid transporters) processes. Mechanistic information about the effects of a drug on bile acid homeostasis is important when evaluating the cholestatic potential of a compound, but experimental data often are not available. The relationship between physicochemical properties, pharmacokinetic parameters, and inhibition of the bile salt export pump among 77 cholestatic drugs with different pathophysiological mechanisms of cholestasis (i.e., impaired formation of bile vs. physical obstruction of bile flow) was investigated. The utility of in silico models to obtain mechanistic information about the impact of compounds on bile acid homeostasis to aid in predicting the cholestatic potential of drugs is highlighted.