Direct Comparison of Total Clearance Prediction: Computational Machine Learning Model versus Bottom-Up Approach Using In Vitro Assay

Prediction of Human Clearance Using In Silico Models with Reduced Bias

Predicting human clearance with high accuracy from in silico-derived parameters alone is highly desirable, as it is fast, saves in vitro resources, and is animal-sparing. We derived random forest (RF) models from 1340 compounds with human intravenous pharmacokinetic (PK) data, the largest data set publicly available today. To assess the general applicability of the RF models, we systematically removed structural-therapeutic class analogues and other compounds with structural similarity from the training sets. For a quasi-prospective test set of 343 compounds, we show that RF models devoid of structurally similar compounds in the training set predict human clearance with a geometric mean fold error (GMFE) of 3.3. While the observed GMFE illustrates how difficult it is to generate a useful model that is broadly applicable, we posit that our RF models yield a more realistic assessment of how well human clearance can be predicted prospectively. We deployed the conformal prediction formalism to assess the model applicability and to determine the prediction confidence intervals for each prediction. We observed that clearance can be predicted better for renally cleared compounds than for other clearance mechanisms. We show that applying a classification model for predicting renal clearance identifies a subset of compounds for which clearance can be predicted with higher accuracy, yielding a GMFE of 2.3. In addition, our in silico RF human clearance models compared well to models derived from scaling human hepatocytes or preclinical in vivo data.

The Comparison of Machine Learning and Mechanistic In Vitro-In Vivo Extrapolation Models for the Prediction of Human Intrinsic Clearance

Accurate prediction of human pharmacokinetics (PK) remains one of the key objectives of drug metabolism and PK (DMPK) scientists in drug discovery projects. This is typically performed by using in vitro-in vivo extrapolation (IVIVE) based on mechanistic PK models. In recent years, machine learning (ML), with its ability to harness patterns from previous outcomes to predict future events, has gained increased popularity in application to absorption, distribution, metabolism, and excretion (ADME) sciences. This study compares the performance of various ML and mechanistic models for the prediction of human IV clearance for a large (645) set of diverse compounds with literature human IV PK data, as well as measured relevant in vitro end points. ML models were built using multiple approaches for the descriptors: (1) calculated physical properties and structural descriptors based on chemical structure alone (classical QSAR/QSPR); (2) in vitro measured inputs only with no structure-based descriptors (ML IVIVE); and (3) in silico ML IVIVE using in silico model predictions for the in vitro inputs. For the mechanistic models, well-stirred and parallel-tube liver models were considered with and without the use of empirical scaling factors and with and without renal clearance. The best ML model for the prediction of in vivo human intrinsic clearance (CLint) was an in vitro ML IVIVE model using only six in vitro inputs with an average absolute fold error (AAFE) of 2.5. The best mechanistic model used the parallel-tube liver model, with empirical scaling factors resulting in an AAFE of 2.8. The corresponding mechanistic model with full in silico inputs achieved an AAFE of 3.3. These relative performances of the models were confirmed with the prediction of 16 Pfizer drug candidates that were not part of the original data set. Results show that ML IVIVE models are comparable to or superior to their best mechanistic counterparts. We also show that ML IVIVE models can be used to derive insights into factors for the improvement of mechanistic PK prediction.

A Machine Learning Framework to Improve Rat Clearance Predictions and Inform Physiologically Based Pharmacokinetic Modeling

During drug discovery and development, achieving appropriate pharmacokinetics is key to establishment of the efficacy and safety of new drugs. Physiologically based pharmacokinetic (PBPK) models integrating in vitro-to-in vivo extrapolation have become an essential in silico tool to achieve this goal. In this context, the most important and probably most challenging pharmacokinetic parameter to estimate is the clearance. Recent work on high-throughput PBPK modeling during drug discovery has shown that a good estimate of the unbound intrinsic clearance (CLint,u,) is the key factor for useful PBPK application. In this work, three different machine learning-based strategies were explored to predict the rat CLint,u as the input into PBPK. Therefore, in vivo and in vitro data was collected for a total of 2639 proprietary compounds. The strategies were compared to the standard in vitro bottom-up approach. Using the well-stirred liver model to back-calculate in vivo CLint,u from in vivo rat clearance and then training a machine learning model on this CLint,u led to more accurate clearance predictions (absolute average fold error (AAFE) 3.1 in temporal cross-validation) than the bottom-up approach (AAFE 3.6-16, depending on the scaling method) and has the advantage that no experimental in vitro data is needed. However, building a machine learning model on the bias between the back-calculated in vivo CLint,u and the bottom-up scaled in vitro CLint,u also performed well. For example, using unbound hepatocyte scaling, adding the bias prediction improved the AAFE in the temporal cross-validation from 16 for bottom-up to 2.9 together with the bias prediction. Similarly, the log Pearson r2 improved from 0.1 to 0.29. Although it would still require in vitro measurement of CLint,u., using unbound scaling for the bottom-up approach, the need for correction of the fu,inc by fu,p data is circumvented. While the above-described ML models were built on all data points available per approach, it is discussed that evaluation comparison across all approaches could only be performed on a subset because ca. 75% of the molecules had missing or unquantifiable measurements of the fraction unbound in plasma or in vitro unbound intrinsic clearance, or they dropped out due to the blood-flow limitation assumed by the well-stirred model. Advantageously, by predicting CLint,u as the input into PBPK, existing workflows can be reused and the prediction of the in vivo clearance and other PK parameters can be improved.

Ensemble Machine Learning Approaches Based on Molecular Descriptors and Graph Convolutional Networks for Predicting the Efflux Activities of MDR1 and BCRP Transporters

Multidrug resistance (MDR1) and breast cancer resistance protein (BCRP) play important roles in drug absorption and distribution. Computational prediction of substrates for both transporters can help reduce time in drug discovery. This study aimed to predict the efflux activity of MDR1 and BCRP using multiple machine learning approaches with molecular descriptors and graph convolutional networks (GCNs). In vitro efflux activity was determined using MDR1- and BCRP-expressing cells. Predictive performance was assessed using an in-house dataset with a chronological split and an external dataset. CatBoost and support vector regression showed the best predictive performance for MDR1 and BCRP efflux activities, respectively, of the 25 descriptor-based machine learning methods based on the coefficient of determination (R2). The single-task GCN showed a slightly lower performance than descriptor-based prediction in the in-house dataset. In both approaches, the percentage of compounds predicted within twofold of the observed values in the external dataset was lower than that in the in-house dataset. Multi-task GCN did not show any improvements, whereas multimodal GCN increased the predictive performance of BCRP efflux activity compared with single-task GCN. Furthermore, the ensemble approach of descriptor-based machine learning and GCN achieved the highest predictive performance with R2 values of 0.706 and 0.587 in MDR1 and BCRP, respectively, in time-split test sets. This result suggests that two different approaches to represent molecular structures complement each other in terms of molecular characteristics. Our study demonstrated that predictive models using advanced machine learning approaches are beneficial for identifying potential substrate liability of both MDR1 and BCRP.

Human Brain Penetration Prediction Using Scaling Approach from Animal Machine Learning Models

Machine learning (ML) approaches have been applied to predicting drug pharmacokinetic properties. Previously, we predicted rat unbound brain-to-plasma ratio (Kpuu,brain) by ML models. In this study, we aimed to predict human Kpuu,brain through animal ML models. First, we re-evaluated ML models for rat Kpuu,brain prediction by using trendy open-source packages. We then developed ML models for monkey Kpuu,brain prediction. Leave-one-out cross validation was utilized to rationally build models using a relatively small dataset. After establishing the monkey and rat ML models, human Kpuu,brain prediction was achieved by implementing the animal models considering appropriate scaling methods. Mechanistic NeuroPK models for the identical monkey and human dataset were treated as the criteria for comparison. Results showed that rat Kpuu,brain predictivity was successfully replicated. The optimal ML model for monkey Kpuu,brain prediction was superior to the NeuroPK model, where accuracy within 2-fold error was 78% (R2 = 0.76). For human Kpuu,brain prediction, rat model using relative expression factor (REF), scaled transporter efflux ratios (ERs), and monkey model using in vitro ERs can provide comparable predictivity to the NeuroPK model, where accuracy within 2-fold error was 71% and 64% (R2 = 0.30 and 0.52), respectively. We demonstrated that ML models can deliver promising Kpuu,brain prediction with several advantages: (1) predict reasonable animal Kpuu,brain; (2) prospectively predict human Kpuu,brain from animal models; and (3) can skip expensive monkey studies for human prediction by using the rat model. As a result, ML models can be a powerful tool for drug Kpuu,brain prediction in the discovery stage.

An artificial intelligence-assisted physiologically-based pharmacokinetic model to predict nanoparticle delivery to tumors in mice

The critical barrier for clinical translation of cancer nanomedicine stems from the inefficient delivery of nanoparticles (NPs) to target solid tumors. Rapid growth of computational power, new machine learning and artificial intelligence (AI) approaches provide new tools to address this challenge. In this study, we established an AI-assisted physiologically based pharmacokinetic (PBPK) model by integrating an AI-based quantitative structure-activity relationship (QSAR) model with a PBPK model to simulate tumor-targeted delivery efficiency (DE) and biodistribution of various NPs. The AI-based QSAR model was developed using machine learning and deep neural network algorithms that were trained with datasets from a published "Nano-Tumor Database" to predict critical input parameters of the PBPK model. The PBPK model with optimized NP cellular uptake kinetic parameters was used to predict the maximum delivery efficiency (DEmax) and DE at 24 (DE24) and 168 h (DE168) of different NPs in the tumor after intravenous injection and achieved a determination coefficient of R2 = 0.83 [root mean squared error (RMSE) = 3.01] for DE24, R2 = 0.56 (RMSE = 2.27) for DE168, and R2 = 0.82 (RMSE = 3.51) for DEmax. The AI-PBPK model predictions correlated well with available experimentally-measured pharmacokinetic profiles of different NPs in tumors after intravenous injection (R2 ≥ 0.70 for 133 out of 288 datasets). This AI-based PBPK model provides an efficient screening tool to rapidly predict delivery efficiency of a NP based on its physicochemical properties without relying on an animal training dataset.

Prediction of Compound Plasma Concentration-Time Profiles in Mice Using Random Forest

Pharmacokinetic (PK) parameters such as clearance (CL) and volume of distribution (Vd) have been the subject of previous in silico predictive models. However, having information of the concentration over time profile explicitly can provide additional value like time above MIC or AUC, etc., to understand both the efficacy and safety-related aspects of a compound. In this work, we developed machine learning models for plasma concentration-time profiles after both i.v. and p.o. dosing for a series of 17 in-house projects. For explanatory variables, MACCS Keys chemical descriptors as well as in silico and experimental in vitro PK parameters were used. The predictive accuracy of random forest (RF), message passing neural network, 2-compartment models using estimated CL and Vdss, and an average model (as a control experiment) was investigated using 5-fold cross-validation (5-fold CV) and leave-one-project-out validation (LOPO-V). The predictive accuracy of RF in 5-fold CV for i.v. and p.o. plasma concentration-time profiles was the best among the models studied, with an RMSE for i.v. dosing at 0.08, 1, and 8 h of 0.245, 0.474, and 0.462, respectively, and an RMSE for p.o. dosing at 0.25, 1, and 8 h of 0.500, 0.612, and 0.509, respectively. Furthermore, by investigating the importance of the in vitro PK parameters using the Gini index, we observed that the general prior knowledge in ADME research was reflected well in the respective feature importance of in vitro parameters such as predicted human Vd (hVd) for the initial distribution, mouse intrinsic CL and unbound fraction of mouse plasma for the elimination process, and Caco2 permeability for the absorption process. Also, this model is the first model that can predict twin peaks in the concentration-time profile much better than a baseline compartment model. Because of its combination of sufficient accuracy and speed of prediction, we found the model to be fit-for-purpose for practical lead optimization.

Artificial Intelligence in Drug Metabolism and Excretion Prediction: Recent Advances, Challenges, and Future Perspectives

Drug metabolism and excretion play crucial roles in determining the efficacy and safety of drug candidates, and predicting these processes is an essential part of drug discovery and development. In recent years, artificial intelligence (AI) has emerged as a powerful tool for predicting drug metabolism and excretion, offering the potential to speed up drug development and improve clinical success rates. This review highlights recent advances in AI-based drug metabolism and excretion prediction, including deep learning and machine learning algorithms. We provide a list of public data sources and free prediction tools for the research community. We also discuss the challenges associated with the development of AI models for drug metabolism and excretion prediction and explore future perspectives in the field. We hope this will be a helpful resource for anyone who is researching in silico drug metabolism, excretion, and pharmacokinetic properties.

Artificial intelligence for compound pharmacokinetics prediction

Optimisation of compound pharmacokinetics (PK) is an integral part of drug discovery and development. Animal in vivo PK data as well as human and animal in vitro systems are routinely utilised to evaluate PK in humans. In recent years machine learning and artificial intelligence (AI) emerged as a major tool for modelling of in vivo animal and human PK, enabling prediction from chemical structure early in drug discovery, and therefore offering opportunities to guide the design and prioritisation of molecules based on relevant in vivo properties and, ultimately, predicting human PK at the point of design. This review presents recent advances in machine learning and AI models for in vivo animal and human PK for small-molecule compounds as well as some examples for antibody therapeutics.

Opportunities and challenges in application of artificial intelligence in pharmacology

Artificial intelligence (AI) is a machine science that can mimic human behaviour like intelligent analysis of data. AI functions with specialized algorithms and integrates with deep and machine learning. Living in the digital world can generate a huge amount of medical data every day. Therefore, we need an automated and reliable evaluation tool that can make decisions more accurately and faster. Machine learning has the potential to learn, understand and analyse the data used in healthcare systems. In the last few years, AI is known to be employed in various fields in pharmaceutical science especially in pharmacological research. It helps in the analysis of preclinical (laboratory animals) and clinical (in human) trial data. AI also plays important role in various processes such as drug discovery/manufacturing, diagnosis of big data for disease identification, personalized treatment, clinical trial research, radiotherapy, surgical robotics, smart electronic health records, and epidemic outbreak prediction. Moreover, AI has been used in the evaluation of biomarkers and diseases. In this review, we explain various models and general processes of machine learning and their role in pharmacological science. Therefore, AI with deep learning and machine learning could be relevant in pharmacological research.

Machine learning and artificial intelligence in physiologically based pharmacokinetic modeling

Physiologically based pharmacokinetic (PBPK) models are useful tools in drug development and risk assessment of environmental chemicals. PBPK model development requires the collection of species-specific physiological, and chemical-specific absorption, distribution, metabolism, and excretion (ADME) parameters, which can be a time-consuming and expensive process. This raises a need to create computational models capable of predicting input parameter values for PBPK models, especially for new compounds. In this review, we summarize an emerging paradigm for integrating PBPK modeling with machine learning (ML) or artificial intelligence (AI)-based computational methods. This paradigm includes 3 steps (1) obtain time-concentration PK data and/or ADME parameters from publicly available databases, (2) develop ML/AI-based approaches to predict ADME parameters, and (3) incorporate the ML/AI models into PBPK models to predict PK summary statistics (eg, area under the curve and maximum plasma concentration). We also discuss a neural network architecture "neural ordinary differential equation (Neural-ODE)" that is capable of providing better predictive capabilities than other ML methods when used to directly predict time-series PK profiles. In order to support applications of ML/AI methods for PBPK model development, several challenges should be addressed (1) as more data become available, it is important to expand the training set by including the structural diversity of compounds to improve the prediction accuracy of ML/AI models; (2) due to the black box nature of many ML models, lack of sufficient interpretability is a limitation; (3) Neural-ODE has great potential to be used to generate time-series PK profiles for new compounds with limited ADME information, but its application remains to be explored. Despite existing challenges, ML/AI approaches will continue to facilitate the efficient development of robust PBPK models for a large number of chemicals.

Computational Predictions of Nonclinical Pharmacokinetics at the Drug Design Stage

Although computational predictions of pharmacokinetics (PK) are desirable at the drug design stage, existing approaches are often limited by prediction accuracy and human interpretability. Using a discovery data set of mouse and rat PK studies at Roche (9,685 unique compounds), we performed a proof-of-concept study to predict key PK properties from chemical structure alone, including plasma clearance (CLp), volume of distribution at steady-state (Vss), and oral bioavailability (F). Ten machine learning (ML) models were evaluated, including Single-Task, Multitask, and transfer learning approaches (i.e., pretraining with in vitro data). In addition to prediction accuracy, we emphasized human interpretability of outcomes, especially the quantification of uncertainty, applicability domains, and explanations of predictions in terms of molecular features. Results show that intravenous (IV) PK properties (CLp and Vss) can be predicted with good precision (average absolute fold error, AAFE of 1.96-2.84 depending on data split) and low bias (average fold error, AFE of 0.98-1.36), with AutoGluon, Gaussian Process Regressor (GP), and ChemProp displaying the best performance. Driven by higher complexity of oral PK studies, predictions of F were more challenging, with the best AAFE values of 2.35-2.60 and higher overprediction bias (AFE of 1.45-1.62). Multi-Task approaches and pretraining of ChemProp neural networks with in vitro data showed similar precision to Single-Task models but helped reduce the bias and increase correlations between observations and predictions. A combination of GP-computed prediction variance, molecular clustering, and dimensionality-reduction provided valuable quantitative insights into prediction uncertainty and applicability domains. SHAPley Additive exPlanations (SHAPs) highlighted molecular features contributing to prediction outcomes of Vss, providing explanations that could aid drug design. Combined results show that computational predictions of PK are feasible at the drug design stage, with several ML technologies converging to successfully leverage historical PK data sets. Further studies are needed to unlock the full potential of this approach, especially with respect to data set sizes and quality, transfer learning between in vitro and in vivo data sets, model-independent quantification of uncertainty, and explainability of predictions.

Application of machine learning techniques to the analysis and prediction of drug pharmacokinetics

In this review, we describe the current status and challenges in applying machine-learning techniques to the analysis and prediction of pharmacokinetic data. The theory of pharmacokinetics has been developed over decades on the basis of physiology and reaction kinetics. Mathematical models allow the reduction of pharmacokinetic data to parameter values, giving insight and understanding into ADME processes and predicting the outcome of different dosing scenarios. However, much information hidden in the data is lost through conceptual simplification with models. It is difficult to use mechanistic models alone to predict diverse pharmacokinetic time profiles, including inter-drug and inter-individual differences, in a cross-sectional manner. Machine learning is a prediction platform that can handle complex phenomena through data-driven analysis. As a resule, machine learning has been successfully adopted in various fields, including image recognition and language processing, and has been used for over two decades in pharmacokinetic research, primarily in the area of quantitative structure-activity relationships for pharmacokinetic parameters. Machine-learning models are generally known to provide better predictive performance than conventional linear models. Owing to the recent success in deep learning, models with new structures are being consistently proposed. These models include transfer learning and generative adversarial networks, which contribute to the effective use of a limited amount of data by diverting existing similar models or generating pseudo-data. How to make such newly emerging machine learning technologies applicable to meet challenges in the pharmacokinetics/pharmacodynamics field is now the key issue.

Utilizing public and private sector data to build better machine learning models for the prediction of pharmacokinetic parameters

One solution to compensate for the shortage of publicly available data is to collect more quality-controlled data from the private sector through public-private partnerships. However, several issues must be resolved before implementing such a system. Here, we review the technical aspects of public-private partnerships using our initiative in Japan as an example. In particular, we focus on the procedure for collecting data from multiple private sector companies and building prediction models and discuss how merging public and private sector datasets will help to improve the chemical space coverage and prediction performance. Teaser: Japan's first public-private consortium in pharmacokinetics has incorporated data from multiple pharmaceutical companies to create useful predictive models.

Predicting Total Drug Clearance and Volumes of Distribution Using the Machine Learning-Mediated Multimodal Method through the Imputation of Various Nonclinical Data

Pharmacokinetic research plays an important role in the development of new drugs. Accurate predictions of human pharmacokinetic parameters are essential for the success of clinical trials. Clearance (CL) and volume of distribution (Vd) are important factors for evaluating pharmacokinetic properties, and many previous studies have attempted to use computational methods to extrapolate these values from nonclinical laboratory animal models to human subjects. However, it is difficult to obtain sufficient, comprehensive experimental data from these animal models, and many studies are missing critical values. This means that studies using nonclinical data as explanatory variables can only apply a small number of compounds to their model training. In this study, we perform missing-value imputation and feature selection on nonclinical data to increase the number of training compounds and nonclinical datasets available for these kinds of studies. We could obtain novel models for total body clearance (CL_tot) and steady-state Vd (Vd_ss) (CL_tot: geometric mean fold error [GMFE], 1.92; percentage within 2-fold error, 66.5%; Vd_ss: GMFE, 1.64; percentage within 2-fold error, 71.1%). These accuracies were comparable to the conventional animal scale-up models. Then, this method differs from animal scale-up methods because it does not require animal experiments, which continue to become more strictly regulated as time passes.

Non-specific binding of compounds in in vitro metabolism assays: a comparison of microsomal and hepatocyte binding in different species and an assessment of the accuracy of prediction models

Non-specific binding in in vitro metabolism systems leads to an underestimation of the true intrinsic metabolic clearance of compounds being studied. Therefore in vitro binding needs to be accounted for when extrapolating in vitro data to predict the in vivo metabolic clearance of a compound. While techniques exist for experimentally determining the fraction of a compound unbound in in vitro metabolism systems, early in drug discovery programmes computational approaches are often used to estimate the binding in the in vitro system.Experimental fraction unbound data (n = 60) were generated in liver microsomes (fu_mic) from five commonly used pre-clinical species (rat, mouse, dog, minipig, monkey) and humans. Unbound fraction in incubations with mouse, rat or human hepatocytes was determined for the same 60 compounds. These data were analysed to determine the relationship between experimentally determined binding in the different matrices and across different species. In hepatocytes there was a good correlation between fraction unbound in human and rat (r²=0.86) or mouse (r²=0.82) hepatocytes. Similar correlations were observed between binding in human liver microsomes and microsomes from rat, mouse, dog, Göttingen minipig or monkey liver microsomes (r² of >0.89, n = 51 - 52 measurements in different species). Physicochemical parameters (logP, pKa and logD) were predicted for all evaluated compounds. In addition, logP and/or logD were measured for a subset of compounds.Binding to human hepatocytes predicted using 5 different methods was compared to the measured data for a set of 59 compounds. The best methods evaluated used measured microsomal binding in human liver microsomes to predict hepatocyte binding. The collated physicochemical data were used to predict the human fu_mic using four different in silico models for a set of 53-60 compounds. The correlation (r²) and root mean square error between predicted and observed microsomal binding was 0.69 & 0.20, 0.47 & 0.23, 0.56 & 0.21 and 0.54 & 0.26 for the Turner-Simcyp, Austin, Hallifax-Houston and Poulin models, respectively. These analyses were extended to include measured literature values for binding in human liver microsomes for a larger set of compounds (n=697). For the larger dataset of compounds, microsomal binding was well predicted for neutral compounds (r²=0.67 - 0.70) using the Poulin, Austin, or Turner-Simcyp methods but not for acidic or basic compounds (r²<0.5) using any of the models. While the lipophilicity-based models can be used, the in vitro binding should be measured for compounds where more certainty is needed, using appropriately calibrated assays and possibly established weak, moderate, and strong binders as reference compounds to allow comparison across databases.

Novel QSAR Approach for a Regression Model of Clearance That Combines DeepSnap-Deep Learning and Conventional Machine Learning

The toxicity, absorption, distribution, metabolism, and excretion properties of some targets are difficult to predict by quantitative structure-activity relationship analysis. Therefore, there is a need for a new prediction method that performs well for these targets. The aim of this study was to develop a new regression model of rat clearance (CL). We constructed a regression model using 1545 in-house compounds for which we had rat CL data. Molecular descriptors were calculated using molecular operating environment, alvaDesc, and ADMET Predictor software. The classification model of DeepSnap and Deep Learning (DeepSnap-DL) with images of the three-dimensional chemical structures of compounds as features was constructed, and the prediction probabilities for each compound were calculated. For molecular descriptor-based methods that use molecular descriptors and conventional machine learning algorithms selected by DataRobot, the correlation coefficient (R ²) and root mean square error (RMSE) were 0.625-0.669 and 0.295-0.318, respectively. We combined molecular descriptors and prediction probability of DeepSnap-DL as features and developed a novel regression method we called the combination model. In the combination model with these two types of features and conventional algorithms selected by DataRobot, R ² and RMSE were 0.710-0.769 and 0.247-0.278, respectively. This finding shows that the combination model performed better than molecular descriptor-based methods. Our combination model will contribute to the design of more rational compounds for drug discovery. This method may be applicable not only to rat CL but also to other pharmacokinetic and pharmacological activity and toxicity parameters; therefore, applying it to other parameters may help to accelerate drug discovery.

Artificial intelligence to bring nanomedicine to life

The technology of drug delivery systems (DDSs) has demonstrated an outstanding performance and effectiveness in production of pharmaceuticals, as it is proved by many FDA-approved nanomedicines that have an enhanced selectivity, manageable drug release kinetics and synergistic therapeutic actions. Nonetheless, to date, the rational design and high-throughput development of nanomaterial-based DDSs for specific purposes is far from a routine practice and is still in its infancy, mainly due to the limitations in scientists' capabilities to effectively acquire, analyze, manage, and comprehend complex and ever-growing sets of experimental data, which is vital to develop DDSs with a set of desired functionalities. At the same time, this task is feasible for the data-driven approaches, high throughput experimentation techniques, process automatization, artificial intelligence (AI) technology, and machine learning (ML) approaches, which is referred to as The Fourth Paradigm of scientific research. Therefore, an integration of these approaches with nanomedicine and nanotechnology can potentially accelerate the rational design and high-throughput development of highly efficient nanoformulated drugs and smart materials with pre-defined functionalities. In this Review, we survey the important results and milestones achieved to date in the application of data science, high throughput, as well as automatization approaches, combined with AI and ML to design and optimize DDSs and related nanomaterials. This manuscript mission is not only to reflect the state-of-art in data-driven nanomedicine, but also show how recent findings in the related fields can transform the nanomedicine's image. We discuss how all these results can be used to boost nanomedicine translation to the clinic, as well as highlight the future directions for the development, data-driven, high throughput experimentation-, and AI-assisted design, as well as the production of nanoformulated drugs and smart materials with pre-defined properties and behavior. This Review will be of high interest to the chemists involved in materials science, nanotechnology, and DDSs development for biomedical applications, although the general nature of the presented approaches enables knowledge translation to many other fields of science.

Prediction of In Vivo Pharmacokinetic Parameters and Time-Exposure Curves in Rats Using Machine Learning from the Chemical Structure

Animal pharmacokinetic (PK) data as well as human and animal in vitro systems are utilized in drug discovery to define the rate and route of drug elimination. Accurate prediction and mechanistic understanding of drug clearance and disposition in animals provide a degree of confidence for extrapolation to humans. In addition, prediction of in vivo properties can be used to improve design during drug discovery, help select compounds with better properties, and reduce the number of in vivo experiments. In this study, we generated machine learning models able to predict rat in vivo PK parameters and concentration-time PK profiles based on the molecular chemical structure and either measured or predicted in vitro parameters. The models were trained on internal in vivo rat PK data for over 3000 diverse compounds from multiple projects and therapeutic areas, and the predicted endpoints include clearance and oral bioavailability. We compared the performance of various traditional machine learning algorithms and deep learning approaches, including graph convolutional neural networks. The best models for PK parameters achieved R² = 0.63 [root mean squared error (RMSE) = 0.26] for clearance and R² = 0.55 (RMSE = 0.46) for bioavailability. The models provide a fast and cost-efficient way to guide the design of molecules with optimal PK profiles, to enable the prediction of virtual compounds at the point of design, and to drive prioritization of compounds for in vivo assays.

Machine-learning-guided early drug discovery of small molecules

Machine learning (ML) approaches have been widely adopted within the early stages of the drug discovery process, particularly within the context of small-molecule drug candidates. Despite this, the use of ML is still limited in the pharmacokinetic/pharmacodynamic (PK/PD) application space. Here, we describe recent progress and the role of ML used in preclinical drug discovery. We summarize the advances and current strategies used to predict ADME properties of small molecules based on their structures, and predict structures based on the desired properties for molecular screening and optimization. Finally, we discuss the use of ML to predict PK to rank the ability of drug candidates to achieve appropriate exposures and hence provide important insights into safety and efficacy. Teaser: In this short review we highlight key advances in the application of machine learning to early drug discovery for small molecules in the preclinical setting.

Development and Biological Characterization of a Novel Selective TrkA Agonist with Neuroprotective Properties against Amyloid Toxicity

Neurotrophins are growth factors that exert important neuroprotective effects by preventing neuronal death and synaptic loss. Nerve Growth Factor (NGF) acts through the activation of its high-affinity, pro-survival TrkA and low-affinity, pro-apoptotic p75^NTR receptors. NGF has been shown to slow or prevent neurodegenerative signals in Alzheimer’s Disease (AD) progression. However, its low bioavailability and its blood–brain-barrier impermeability limit the use of NGF as a potential therapeutic agent against AD. Based on our previous findings on synthetic dehydroepiandrosterone derivatives, we identified a novel NGF mimetic, named ENT-A013, which selectively activates TrkA and exerts neuroprotective, anti-amyloid-β actions. We now report the chemical synthesis, in silico modelling, metabolic stability, CYP-mediated reaction phenotyping and biological characterization of ENT-A013 under physiological and neurodegenerative conditions. We show that ENT-A013 selectively activates the TrkA receptor and its downstream kinases Akt and Erk1/2 in PC12 cells, protecting these cells from serum deprivation-induced cell death. Moreover, ENT-A013 promotes survival of primary Dorsal Root Ganglion (DRG) neurons upon NGF withdrawal and protects hippocampal neurons against Amyloid β-induced apoptosis and synaptic loss. Furthermore, this neurotrophin mimetic partially restores LTP impairment. In conclusion, ENT-A013 represents a promising new lead molecule for developing therapeutics against neurodegenerative disorders, such as Alzheimer’s Disease, selectively targeting TrkA-mediated pro-survival signals.

Comparison of Predictions by BCS, rDCS and Machine Learning for the Effect of Food on Oral Drug Absorption Based on Features Calculated In silico

In this study, observed food effects of 473 drugs were categorized into positive, negative, or no effects and compared with the predictions made by machine learning (ML), the Biopharmaceutics Classification System (BCS) and refined Developability Classification System (rDCS). All methods used primarily in silico estimates for prediction, and for ML, four algorithms were evaluated using nested cross-validation to select important information from 371 features calculated based on the chemical structure. Approximately 18 features, including estimated solubility in biorelevant media, were selected as important, and the random forest classifier was the best among four algorithms with 36.6% error rate (ER) and 10.8% opposite prediction rate (OPR). The prediction by rDCS utilizing solubility in a biorelevant medium was somewhat inferior, but not by much; 41.0% ER and 11.4% OPR. Compared with these two methods, the prediction by BCS was inferior; 54.5% ER and 21.4% OPR. ER was improved modestly by using measured features instead of in silico estimates when BCS was applied to a subset of 151 drugs (46.4% from 55.0%). ML and rDCS predicted the food effects of the same subset using in silico estimates with ERs of 37.7% and 42.4%, respectively, suggesting that the predictions by ML and rDCS using in silico features are similar or more accurate than those by BCS using measured features. These results suggest that ML was useful in revealing essential features from complex information and, together with rDCS, is effective in predicting food effects during drug development, including early drug discovery.

Machine Learning Models for Human In Vivo Pharmacokinetic Parameters with In-House Validation

Prior to clinical development, a comprehensive pharmacokinetic characterization of a novel drug is required to understand its exposure at the site of action and elimination. Accordingly, in vitro assays and animal pharmacokinetic studies are regularly employed to predict drug exposure in humans, which is often costly and time-consuming. For this reason, the prediction of human pharmacokinetics at the point of design would be of high value for drug discovery. Therefore, we have established a comprehensive data curation protocol that enables machine learning evaluation of 12 human in vivo pharmacokinetic parameters using only chemical structure information and available doses for 1001 unique compounds. These machine learning models were thoroughly investigated and validated using both an independent hold-out test set and AstraZeneca clinical data. In addition, the availability of preclinical predictions for a subset of internal clinical candidates allowed us to compare our in silico approach with state-of-the-art pharmacokinetic predictions. Based on this evaluation, three fit-for-purpose models for AUC PO (R_test² = 0.63; RMSE_test = 0.76), C_max PO (R_test² = 0.68; RMSE_test = 0.62), and Vd_ss IV (R_test² = 0.47; RMSE_test = 0.50) were identified. Based on the findings, our machine learning models have considerable potential for practical applications in drug discovery, such as influencing decision-making in drug discovery projects and progression of drug candidates toward the clinic.

Artificial Intelligence in Drug Safety and Metabolism

The use of artificial intelligence methods in drug safety began in the early 2000s with applications such as predicting bacterial mutagenicity and hERG inhibition. The field has been endlessly expanding ever since and the models have become more complex. These approaches are now integrated into molecule risk assessment processes along with in vitro and in vivo methods. Today, artificial intelligence can be used in every phase of drug discovery and development, from profiling chemical libraries in early discovery, to predicting off-target effects in the mid-discovery phase, to assessing potential mutagenic impurities in development and degradants as part of life cycle management. This chapter provides an overview of artificial intelligence in drug safety and describes its application throughout the entire discovery and development process.

Prediction of permeability across intestinal cell monolayers for 219 disparate chemicals using in vitro experimental coefficients in a pH gradient system and in silico analyses by trivariate linear regressions and machine learning

For medicines, the apparent membrane permeability coefficients (P_app) across human colorectal carcinoma cell line (Caco-2) monolayers under a pH gradient generally correlate with the fraction absorbed after oral intake. Furthermore, the in vitro P_app values of 29 industrial chemicals were found to have an inverse association with their reported no-observed effect levels for hepatotoxicity in rats. In the current study, we expanded our influx permeability predictions for the 90 previously investigated chemicals to both influx and efflux permeability predictions for 207 diverse primary compounds, along with those for 23 secondary compounds. Trivariate linear regression analysis found that the observed influx and efflux logP_app values determined by in vitro experiments significantly correlated with molecular weights and the octanol-water distribution coefficients at apical and basal pH levels (pH 6.0 and 7.4, respectively) (apical to basal, r = 0.76, n = 198; and basal to apical, r = 0.77, n = 202); the distribution coefficients were estimated in silico. Further, prediction accuracy was enhanced by applying a light gradient boosting machine learning system (LightGBM) to estimate influx and efflux logP_app values that incorporated 17 and 19 in silico chemical descriptors (r = 0.83-0.84, p < 0.001). The determination in vitro and/or prediction in silico of permeability coefficients across intestinal cell monolayers of a diverse range of industrial chemicals/food components/medicines could contribute to the safety evaluations of oral intakes of general chemicals in humans. Such new alternative methods could also reduce the need for animal testing during toxicity assessment.

Direct Comparison of the Prediction of the Unbound Brain-to-Plasma Partitioning Utilizing Machine Learning Approach and Mechanistic Neuropharmacokinetic Model

The mechanistic neuropharmacokinetic (neuroPK) model was established to predict unbound brain-to-plasma partitioning (K_p,uu,brain) by considering in vitro efflux activities of multiple drug resistance 1 (MDR1) and breast cancer resistance protein (BCRP). Herein, we directly compare this model to a computational machine learning approach utilizing physicochemical descriptors and efflux ratios of MDR1 and BCRP-expressing cells for predicting K_p,uu,brain in rats. Two different types of machine learning techniques, Gaussian processes (GP) and random forest regression (RF), were assessed by the time and cluster-split validation methods using 640 internal compounds. The predictivity of machine learning models based on only molecular descriptors in the time-split dataset performed worse than the cluster-split dataset, whereas the models incorporating MDR1 and BCRP efflux ratios showed similar predictivity between time and cluster-split datasets. The GP incorporating MDR1 and BCRP in the time-split dataset achieved the highest correlation (R² = 0.602). These results suggested that incorporation of MDR1 and BCRP in machine learning is beneficial for robust and accurate prediction. K_p,uu,brain prediction utilizing the neuroPK model was significantly worse compared to machine learning approaches for the same dataset. We also investigated the predictivity of K_p,uu,brain using an external independent test set of 34 marketed drugs. Compared to machine learning models, the neuroPK model showed better predictive performance with R² of 0.577. This work demonstrates that the machine learning model for K_p,uu,brain achieves maximum predictive performance within the chemical applicability domain, whereas the neuroPK model is applicable more widely beyond the chemical space covered in the training dataset.

Prediction of Oral Pharmacokinetics Using a Combination of In Silico Descriptors and In Vitro ADME Properties

Accurate prediction of oral pharmacokinetics remains challenging. This study investigated quantitative approaches for the prediction of the area under the plasma concentration-time curve after oral administration (AUC_p,oral) to rats using the in vitro-in vivo extrapolation (IVIVE), in silico model using machine learning approaches and the combination of the in silico model and in vitro data. A set of 595 structurally diverse compounds with determined AUC_p,oral at 1 mg/kg, in vitro intrinsic clearance (CL_int), an unbound fraction in plasma (f_u,p) in rats, and kinetic solubility at pH 6.8 was used for this assessment. Prediction models developed by two different types of machine learning techniques (i.e., random forest regression and Gaussian processes) were evaluated using three validation methods implementing the time and cluster-split training and test set and fivefold cross-validation. The developed machine learning models have a square of correlation coefficient (R²) in the range of 0.381-0.685 with 33-45% of the compounds being predicted within 2-fold of the observed AUC_p,oral value. The predictivity was improved by incorporating CL_int, f_u,p, and solubility as explanatory variables with R² = 0.554-0.743. In cases where extraction by the liver is the main elimination pathway and intestinal extraction is negligible, AUC_p,oral can be expressed by dose, CL_int, and f_u,p based on a well-stirred model. By using this conventional IVIVE approach, only 1.7-5.0% of compounds were predicted within the 2-fold error with R² = 0.354-0.487. Two empirical scaling factors (ESFs) determined by linear regression analysis and machine learning approaches improved the predictivity of AUC_p,oral with 33-44% predicted within twofold variability. The IVIVE using ESF predicted by random forest regression showed better predictivity of AUC_p,oral with R² = 0.471-0.618, while it still showed lower predictivity than machine learning approaches applied directly to AUC_p,oral prediction. This study demonstrated that the combination of in silico and in vitro parameters is useful to improve the predictivity of the machine learning model for rat AUC_p,oral and supports consideration for predicting AUC_p,oral for human and other non-clinical species in a similar manner.