Computational approaches to predict the toxicity of bioactive natural products: a mini review of methodologies

Abstract

Despite the increasing global demand for functional foods, the challenges associated with bioactive natural food products due to their complex composition remain. Bioactive natural products can potentially interfere with physiological activity regulation and lead to undesired side effects. This finding emphasizes the need for machine learning (ML)-based food safety predictions focused on intrinsic toxicity. This review explores various strategies involved in current methods of model selection and validation techniques used in predictive analysis, highlighting their strengths, limitations, and progress. Future studies should focus on testing compound combinations using top-down or bottom-up approaches with appropriate models to advance in silico toxicity modeling of bioactive natural products.

Introduction

The diversity of bioactive materials implemented in functional foods and health-related products is expanding globally, followed by consumer demand not only for nutritional value but also for additional health benefits (Gutiérrez-Grijalva et al., 2024). However, due to the complex composition of bioactive natural food products consisting of numerous phytochemicals, there is a possibility of interference with targeted physiological activity regulation or the occurrence of undesired side effects, despite their positive impact on human health. These risks are particularly frequent when these materials are combined with other food ingredients or medications (Lila and Raskin, 2005). The efficacy and safety of natural products may be reported, but unexpected side effects remain a concern. New materials such as novel botanical species or byproducts might lack a history of human use and could contain toxic constituents (Kroes and Walker, 2004). Scientific evidence is crucial for the use of these materials in functional foods, ensuring both efficacy and safety. (Díaz et al., 2020). Although there are conventional scientific studies for safety assessment, such as in vitro tests, animal toxicology studies, and clinical trials (Kruger and Mann, 2003), these studies are expensive and time-consuming due to extensive trial and error processes and have difficulties in reproducibility (Manful et al., 2023). Thus, new methods are required for evaluating the toxicity of food ingredients. The application of machine learning (ML) based on big data has significantly transformed industries such as engineering, pharmacy, and medicine, offering innovative solutions to complex problems. Due to the contributions of the fourth industrial revolution to the food industry, big data and ML are widely used in food safety applications, including the detection of foodborne pathogens, food safety management, and the prediction of food safety risk (Kim and Kim, 2022). Several computational programs, such as the QSAR Toolbox, LiverTox, DILI-Rank, ToxCast, and Tox21, have been developed by the OECD, NIH, FDA, and US EPA to assist in food toxicity prediction (Chen et al., 2016; Dimitrov et al., 2016; Dix et al., 2007; NIDDK, 2012; Thomas et al., 2018).

However, current ML-based food safety predictions are focused on extrinsic toxicity and hazard analysis, still research based on the components of natural substances for toxicity is necessary (Hudson et al., 2018; Tralau et al., 2021). Furthermore, even though there have been many efforts to minimize drug failures due to toxicity development by the development of in silico toxicity prediction models for drug safety, (Dearden, 2003; Rim, 2020) the development of in silico toxicity predictions of bioactive food ingredients is still limited by the ambiguity and complexity of natural ingredients.

To achieve reliable in silico prediction results, several steps, such as proper data collection, toxicity endpoint selection, model selection, data preparation, model development, and model performance assessment, are necessary (Rim, 2020). This review focused on currently employed methods of toxicity prediction models and validation in predictive analysis, highlighting their strengths, limitations, and progress for the further development of in silico bioactive food ingredient toxicity prediction models.

Criteria to classified two modeling approaches: top-down and bottom-up

To apply in silico methods on the biologically active natural substance toxicity prediction, establishing the proper ML models that can be used in two primary computational strategies in this field is required: top-down approaches and bottom-up approaches (Tsamandouras et al., 2015). Top-down approaches involve the utilization of existing knowledge or databases to predict toxicity, and the aim is to use these established correlations to make predictions or understand outcomes. These methods typically rely on established relationships between chemical structures and toxicity endpoints, often leveraging statistical models or machine learning algorithms trained on large datasets of experimental toxicity data (Wang and Maranas, 2018). By extrapolating patterns from known toxic compounds, top-down approaches can rapidly screen and prioritize natural products for further evaluation, offering valuable insights into potential hazards early in the drug discovery process. Conversely, bottom-up approaches start at a more granular level, focus on understanding the underlying molecular mechanisms, emphasis is on building a model based on fundamental details of toxicity from first principles. These methods involve the computational simulation of molecular interactions and biological pathways to elucidate how natural products may interact with cellular components or physiological processes to induce toxicity (Bausch and Kroy, 2006). By integrating principles from chemistry, biology, and physics, bottom-up approaches provide detailed mechanistic insights into toxicity mechanisms, enabling the identification of specific molecular targets or pathways affected by natural product exposure. Both top-down and bottom-up approaches summarized in Fig. 1 offer unique advantages and challenges in predicting intrinsic natural product toxicity (Palsson, 2002). The principles, methodologies, and applications of top-down and bottom-up approaches in predicting intrinsic natural product toxicity using computational models are summarized in Table 1. The following sections will describe two different strategies for constructing and analyzing natural product toxicity prediction models.

Fig. 1.

A schematic diagram for the top-down and bottom-up approaches. TM text mining, ARM association rule mining, SVM support vector machines, PBPK physiologically based pharmacokinetic, RWR random walk with restart, MD molecular docking

Table 1.

Methods applicable to predict the toxicity of bioactive natural products

Method	Classification	Algorithm	Description	Reference
Text Mining (TM)	Top-down	Latent Dirichlet Allocation (LDA) Named Entity Recognition (NER)	Text mining (TM) methods can play a valuable role in predicting food toxicity by extracting and analyzing relevant information from textual sources to identify potential risks, hazards, and safety concerns associated with food products and ingredients. TM provides a powerful tool for analyzing large volumes of textual data and uncovering valuable insights to inform food safety and risk assessment efforts	(Fang et al., 2008) (Zhou et al., 2010) (Sharma et al., 2020)
Association Rule Mining (ARM)	Top-down	Dijkstra’s algorithm FP-Growth algorithm Apriori algorithm	Association rule mining (ARM) is focused on identifying correlations between natural product components and toxicity outcomes without necessarily delving into the detailed molecular mechanisms underlying toxicity	(Yoo et al., 2018) (Pu et al., 2019)
Support Vector Machines (SVM)	Top-down	SVM algorithm Kernel Trick algorithm	The support vector machine (SVM) model is trained on a dataset containing chemical descriptors and corresponding toxicity labels. The SVM model learns to classify compounds as toxic or nontoxic based on the patterns present in the training data	(Li et al., 2016)
Quantitative Structure Activity Relationship (QSAR)	Top-down	SVM Random Forest (RF) Artificial Neural Networks (ANNs)	Quantitative structure–activity relationship (QSAR) models can be developed specifically for food-related chemicals to predict their toxicity based on chemical structure and properties. These models correlate structural features of chemicals with their biological activity or toxicity endpoints	(Yang et al., 2015) (Huang et al., 2015) (Zhao et al., 2017)
Random walk with restart (RWR)	Bottom-up	RWR heNetRW algorithm	The random walk with restart (RWR) model operates by simulating a random walk process on a network representing the relationships between different compounds, targets, biological pathways, and toxicity outcomes. The model assigns probabilities to each node in the network, indicating the likelihood of a random walker (representing a toxicological event) transitioning from one node to another	(Yang et al., 2015) (Lee and Nam, 2018) (Yang et al., 2018)
Physiologically Based Pharmacokinetic (PBPK)	Bottom-up	Nonlinear Mixed-Effects Modeling (NONMEM) Markov Chain Monte Carlo (MCMC)	Physiologically based pharmacokinetic (PBPK) models are mathematical representations used to predict the ADME of substances in the body, including drugs, chemicals, and natural products, based on their physicochemical properties and biological processes	(Rivero-Segura and Gomez-Verjan, 2021) (Li et al., 2021)
Molecular Docking (MD)	Bottom-up	Rigid-body Docking algorithm Flexible Docking algorithm	Molecular docking (MD) involves predicting the preferred orientation and conformation of a ligand (compound) within the binding site of a protein target. Docking algorithms calculate the binding energy or affinity between the ligand and the target, helping to prioritize compounds for further testing based on their predicted binding strength	(An et al., 2022)

Top-down approach models in predicting food ingredients toxicity

While top-down approaches excel in their ability to rapidly screen large chemical libraries and prioritize compounds for experimental testing, they may lack mechanistic detail and struggle with limited or biased training data. Top-down approaches are more applicable to datasets with direct information on toxicity determination in natural food products (Kenny et al., 2022). Text mining (TM) techniques can be used for topic modeling, sentiment analysis, or named entity recognition to extract relevant information from textual sources via natural language processing. The application of TM might involve analyzing large corpora of text data, such as scientific literature, regulatory documents, or clinical reports, to identify trends, patterns, or associations related to toxicity. In previous research, Fang et al. associated traditional Chinese medicine (TCM), gene and disease information using the TM method via a top-down approach named TCMGeneDIT (Fang et al., 2008). Knowledge discovery from traditional literature and herbal adverse events from reports are also utilized for machine learning through TM (Sharma et al., 2020; Zhou et al., 2010).

The information from the mined database contains identified correlations. For example, if the taxonomic information of one herb is associated with a specific toxicity and unique compound, other herbal compounds commonly found in natural substances with the same toxicity are more likely to be predicted to be toxic. Association rule mining (ARM) helps identify correlations between natural product components and toxicity outcomes. This data-driven top-down approach was applied to identify medicinal combinations of natural products to predict therapeutically adverse effects in the Compound Combination-Oriented Natural Product Database with Unified Terminology (COCONUT) (Yoo et al., 2018). Not only the classification of toxic natural products but also the classification of chemical descriptors and corresponding toxicity labels can be applied in top-down approaches. Support vector machines (SVMs), as a top-down approach, enable the classification of compounds as toxic or nontoxic based on the patterns present in the training data with forbidden features such as their structures and physicochemical properties (Jakkula, 2006). Another data-driven toxicity application database is The Hazardous Substances Data Bank (HSDB), constructed by the National Library of Medicine, is a service that provides a file of chemical and substance information with one record for each specific or for a category of chemicals or substances. (Fonger et al., 2014). Based on HSDB, Pu et al. estimated the toxicity of drug candidates by employing SVM algorithm (Pu et al., 2019). These features are converted to a multidimensional vector, and the data points are separated into different classes by finding the hyperplane that maximizes the margin between the toxic or nontoxic classes. To predict the cardiotoxicity of natural product combinations, Li et al. applied an SVM model in conjunction with metabonomics via UPLC/Q–TOF–MS technology to develop a rapid, sensitive, and specific prediction method (Li et al., 2016).

The quantitative structure–activity relationship (QSAR) model starts with the observed biological activity or toxicity data of various compounds and seek to understand and predict this activity based on the chemical structure of the compounds. The aim is to derive a mathematical relationship between the chemical structure as an input and the biological activity as an output, which reflects a high-level objective. The QSAR model, which is built on chemical structure and knowledge-based properties, is trained by a prediction algorithm to correlate biological activities or toxicity endpoints in the training set. SVM, random forest (RF), and artificial neural networks (ANNs) can be applied to ML algorithms to determine correlations between chemical properties and toxic pathways (Mozafari et al., 2020). Several studies have investigated the application of QSAR modeling in predicting the toxicity of natural products. Yang et al. applied QSAR models for predicting the interaction between TCM-oriented compounds and p-glycoprotein inhibitors, which was reported to be one of the reasons for unexpected toxic reactions or drug–herb interactions in humans (Yang et al., 2015). Huang et al. developed a QSAR model for hepatotoxicity prediction in natural products. Briefly, drugs associated with drug-induced liver injury were characterized by calculating molecular descriptors and fingerprints and classified as hepatotoxic compounds. The features were trained by the RF algorithm and validated. Finally, natural product compounds were tested by the QSAR model (Huang et al., 2015).

Bottom-up approach models for predicting food ingredients toxicity

In contrast to top-down approaches, bottom-up approaches start with the smallest components or individual entities and build to understand the behavior of the system. Bottom-up approaches provide a deeper understanding of toxicity mechanisms by integrating data from molecular cells and tissues into mathematical models, providing insights into the functioning and dynamic behavior of cells, organs, and organisms, such as herb–compound–gene–pathway networks, pharmacokinetics and pharmacodynamics tracking, and molecular docking (Oulas et al., 2017). Bottom-up approaches can find new derivations that were not reported in the literature but often require extensive computational resources and expertise for accurate implementation.

To complement the detailed insights provided by bottom-up methods, integrated network-based approaches can be employed. By integrating both bottom-up and network-based approaches, a more comprehensive understanding of food ingredient toxicity can be achieved, leveraging detailed molecular data and broad systemic interactions. The Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway database can be one of the examples of molecular pathway of the mechanism (Kanehisa, 2002). As bioactive natural products consist of multiple combinations of compounds, they systematically disrupt multiple targets rather than a single target, causing specific toxicity. A topological analysis is required by constructing an herb–compound–target network, and interactions should be associated with protein–protein interactions. After establishing the comprehensive network, random walk with restart (RWR) models can be used to highlight key toxic pathways and nodes as one of the most effective topological analyses (Le, 2017). The RWR model simulates a random walk process on a network, mapping relationships among compounds, targets, pathways, and toxicity outcomes. It assigns probabilities to nodes, indicating the likelihood of transitioning between them, providing insight into toxicological events (Luo et al., 2016). The RWR considers not only directly connected nodes but also potentially toxic interactions in target‒target networks. This process reduces the number of cases where unknown interaction information is missing (Lee and Nam, 2018). The heNetRW algorithm utilizes a random walk approach on a heterogeneous herb–target network to identify protein targets of herbs. By constructing a network comprising herbs, targets, and their interactions and simulating random walk algorithms, candidate targets for a given herb can be predicted (Yang et al., 2018).

Another basic principle that can be started from the bottom-up approach is the absorption, distribution, metabolism and elimination (ADME) approach based on pharmacokinetics, which can be explained by physiological calculations of the bioactive substances of natural products. Physiologically based pharmacokinetic (PBPK) models have been increasingly applied in the field of natural product toxicity to understand the pharmacokinetics and toxicokinetics of bioactive compounds derived from natural sources. In the context of natural product toxicity, PBPK models can provide valuable insights into how these compounds are absorbed, distributed, metabolized, and eliminated in the body and how these processes may influence their toxicity (Ruiz et al., 2020). By integrating information on the physicochemical properties of natural products, such as molecular weight, lipophilicity, and solubility, with physiological parameters such as blood flow rates, organ volumes, and metabolic enzyme activities, PBPK models can simulate the concentration–time profiles of natural products and their metabolites in different tissues and organs. Several studies have applied the PBPK model to predict the toxicity of natural products. Natural products isolated from Mexican herbal medicines against COVID-19. Rivero-Segura and Gomez-Verjan calculated the lipophilicity, hydrophilicity, drug-likeness, lead-likeness, molecular weight, and toxico-informatic properties of compounds to determine safe chemicals by in silico screening (Rivero-Segura and Gomez-Verjan, 2021). Li et al. predicted the oral hepatotoxic dose of natural products based on the PBPK model and SVM classifier to suggest an acceptable daily dose of TCM (Li et al., 2021).

In contrast to topological analysis, which considers multiple compound-multitarget pathways, the molecular docking (MD) model are focused on a single bioactive compound from a natural product. The MD model focused on the interaction between a bioactive compound and a large biological molecule at the atomic, transcriptomic, and proteomic levels involves predicting the preferred orientation and conformation of the compound within the affinity of the binding site of a protein target (Fan et al., 2019). Because MD correlates direct interactions with chemicals, target genes, or proteins, it can be a key tool in structural molecular biology and toxicology, but it needs to be combined with other models, such as network analysis, to detect correlations in biological pathways (Jiao et al., 2021). An et al. applied MD and network pharmacology for the prediction of neurotoxicity caused by fuzi (aconite, Radix Aconiti praeparata). As a result, the prediction of potential neurotoxic compounds of high affinity for ALB, AKT1 and CASP3 was performed, and network pharmacology analysis revealed that these compounds exhibit neurotoxicity (An et al., 2022).

Future suggestions for better toxicity prediction

Each database could provide taxonomical information, a basis for intake, a biological network, and chemical features that can be utilized in data-driven toxicity prediction of bioactive natural products. Researchers need a large, diverse, and high-quality dataset for training the model. The dataset should accurately represent the scenarios of intrinsic toxicity to natural products. The data were cleaned and preprocessed appropriately before they were input into the toxicity prediction model. This may involve handling missing values, removing outliers, scaling features, or encoding categorical variables. A well-preprocessed dataset can improve the model performance regarding toxicity. Additionally, based on these databases, various models have been introduced depending on the toxicity of interest. The in silico models are classified by two approaches, top-down and bottom-up approaches, which differ in their starting points and the level of detail initially considered. The choice between these approaches depends on the specific research question, available data, and level of understanding of the system being modeled. The selection of an appropriate model architecture or algorithm that is suitable for toxicity prediction is also suggested. Different models have different strengths and weaknesses, and selecting the right model can significantly impact the predictive accuracy. To compensate for the limitations of each model, it is necessary to consider the perspectives of both approaches. The bottom-up approach starts from the molecular level, needs detailed molecular characterization. Also, conducting detailed studies on the biochemical pathways and mechanisms through which compounds exert their effects, both beneficial and toxic is required (Di Giulio et al., 2020). For the top-down approach which begins at the level of observed toxicological effects in organisms and works backwards to understand the underlying causes, establishing robust systems for reporting and collecting data on adverse effects associated with natural product consumption in human populations is suggested (Devleesschauwer et al., 2015). Finally, Various models or ensemble methods should be explored to improve the performance. Predictions from multiple models should be combined to make a final prediction. Using top-down and bottom-up approaches separately are not sufficient (Ekins et al., 2005). Ensemble methods, such as averaging predictions or using more advanced techniques such as stacking or boosting, can often yield better results than using a single model. Continuously refining the model by experimenting with different techniques, architectures, and strategies will be required to improve the prediction of the toxicity of bioactive natural products in future studies.

In this review, the applicable ML models used for toxicity prediction research were reviewed, classified by two approaches, top-down and bottom-up approaches, and suggested their application to advance in silico toxicity modeling of bioactive natural products.

Declarations

Conflict of interest

The authors declare that they have no conflicts of interest.