Unlocking the Potential of Clustering and Classification Approaches: Navigating Supervised and Unsupervised Chemical Similarity

Abstract

Background:

The field of toxicology has witnessed substantial advancements in recent years, particularly with the adoption of new approach methodologies (NAMs) to understand and predict chemical toxicity. Class-based methods such as clustering and classification are key to NAMs development and application, aiding the understanding of hazard and risk concerns associated with groups of chemicals without additional laboratory work. Advances in computational chemistry, data generation and availability, and machine learning algorithms represent important opportunities for continued improvement of these techniques to optimize their utility for specific regulatory and research purposes. However, due to their intricacy, deep understanding and careful selection are imperative to align the adequate methods with their intended applications.

Objectives:

This commentary aims to deepen the understanding of class-based approaches by elucidating the pivotal role of chemical similarity (structural and biological) in clustering and classification approaches (CCAs). It addresses the dichotomy between general end point–agnostic similarity, often entailing unsupervised analysis, and end point–specific similarity necessitating supervised learning. The goal is to highlight the nuances of these approaches, their applications, and common misuses.

Discussion:

Understanding similarity is pivotal in toxicological research involving CCAs. The effectiveness of these approaches depends on the right definition and measure of similarity, which varies based on context and objectives of the study. This choice is influenced by how chemical structures are represented and the respective labels indicating biological activity, if applicable. The distinction between unsupervised clustering and supervised classification methods is vital, requiring the use of end point–agnostic vs. end point–specific similarity definition. Separate use or combination of these methods requires careful consideration to prevent bias and ensure relevance for the goal of the study. Unsupervised methods use end point–agnostic similarity measures to uncover general structural patterns and relationships, aiding hypothesis generation and facilitating exploration of datasets without the need for predefined labels or explicit guidance. Conversely, supervised techniques demand end point–specific similarity to group chemicals into predefined classes or to train classification models, allowing accurate predictions for new chemicals. Misuse can arise when unsupervised methods are applied to end point–specific contexts, like analog selection in read-across, leading to erroneous conclusions. This commentary provides insights into the significance of similarity and its role in supervised classification and unsupervised clustering approaches. https://doi.org/10.1289/EHP14001

Introduction

Chemical hazard and risk assessments aim to estimate toxicological outcomes and evaluate the potential risks that chemicals may pose to human health and the environment. However, studying chemicals one-by-one is expensive, time-consuming, and can contribute to harmful chemicals being replaced by new chemicals that are structurally similar (to perform the same function in manufacturing or final product as the harmful chemicals did) but lack toxicological information.

Therefore, there has been an increasing shift in focus from individual chemical analyses to assessing groups of chemicals or agents. Class-based assessment approaches often employ high-throughput methods to forecast toxicity by drawing on the similarities of previously characterized compounds. Historical class-based efforts have typically focused on categorizing chemicals solely based on their observed or overall structural similarities.^1,2 This approach often overlooks the nuanced aspects of chemical similarity which, depending on how structures are represented and described, might have big implications on relating it to biological activity.^3,4 While chemical structure is the primary criterion for grouping substances together, other factors, such as biological activity, internal exposure, and/or pathways, should also be considered for classification studies.^2,5,6 Thanks to advances in information technology, including machine learning, various computational techniques can now be deployed to develop more efficient methods for toxicology and hazard assessment by incorporating structural and functional information as well as phenotypic outcomes and mechanistic data. These computational technique advancements update the standard class-based approach by looking beyond general chemical structure properties and utilizing multiple streams of information on select chemicals, allowing for a more specific grouping of the data, to better identify potential concerns.^7–9 Nevertheless, there remains a pressing need to understand the circumstances and techniques under which such approaches can be efficiently and effectively used to evaluate the hazards of chemical exposure on human health.

The aim of this commentary was to provide an overview of clustering and classification approaches (CCAs). The key concepts for defining similarity, considerations around supervised and unsupervised approaches, and current applications in regulatory contexts are discussed. This work also highlights challenges and potential solutions identified, which inspired the development of a workflow applying CCAs,^10,11 and proposed future collaborations to further develop best practices and increase the use of CCAs in toxicity research and risk assessment. These topics were discussed in depth at a workshop held by the Division of Translational Toxicology (DTT), National Institute of Environmental Health Sciences (NIEHS). Detailed information about the organization of the event and the recordings of the different sessions, which informed the current commentary, are available at https://www.niehs.nih.gov/news/events/pastmtg/2022/nams2022/index.cfm.

Insights and Practical Applications

Background

The field of toxicology has evolved significantly in recent years, with increasing emphasis on the use of new approach methodologies (NAMs) to understand chemical toxicity.^12,13 CCAs are important tools for analyzing chemical data and play a key role in the development and application of NAMs.^1,9 However, these methods can be complex and require careful considerations when selecting the appropriate approach for a given application. A wide array of CCAs have been used to interpret toxicological implications of chemical similarity and have been discussed in various guidance documents^9,14–18 and scholarly publications.^1,19,20

However, extending these tools to prioritize chemicals for testing, making risk-based decisions with data-poor chemicals, predicting toxicological effects, and integrating information to understand chemical effects in biological systems was always a challenge. This is, in particular, due to conflation between general chemical features denoting end point–agnostic similarity and end point–specific biological effect–based similarity.^3,21

Similarity is a fundamental concept in cheminformatics and is essential for the accurate application of CCAs in chemical grouping as well as other in silico NAMs [i.e., quantitative structure–activity relationships (QSAR) and read-across]. The concept of similarity is contextual and varies based on the analysis’s objectives and the data type being evaluated. Therefore, there is no singular type of similarity or universal definition for it.^22,23 Varnek showed an example of nine different geometrical elements that can be grouped based on various types of similarity: shape, color, size, and pattern. Each element can be deemed similar to every other one in one of the four aspects.²¹ Thus, the selection of similarity here is arbitrary and can only fulfill one definition since opting for one aspect of similarity automatically disregards the other types.

The same concept of similarity applies to chemical structures, which can all exhibit similarities in various ways. Consequently, a list of chemicals can be categorized differently depending on the attributes and features considered crucial for defining the desired similarity. It is essential to distinguish between general end point–agnostic similarity, typically observed visually or arbitrarily and deducted through unsupervised methods, and end point–specific similarity, requiring supervised analysis.²¹ This stems from the fact that general similarity is not tied to a particular biological activity, whereas end point–specific similarity is determined by the structural characteristics governing a specific biological activity or toxicity. Hence, the selection of a similarity approach relies on the study’s context and goals, which, in turn, are influenced by how chemical structures are represented and encoded in the explanatory variables. Varnek showed a list of 16 chemicals that, while structurally diverse, all share a formyl functional group consisting of a carbonyl bonded to hydrogen.²¹ Thus, all of these chemicals can be grouped as aldehydes, representing one of the similarities that they share. Alternatively, these chemicals can be grouped based on other similarity aspects, as depicted in the same figure by Varnek.²¹ The chemicals in this example were initially grouped based on the shared common scaffold, representing the most visually similar grouping.^21,24–26 However, the same chemicals can also be grouped based on common functional groups.²¹ Although the resulting grouped chemicals might not visually resemble each other, they share significant functional similarities that could be more relevant for certain biological activities compared to the first grouping.²¹ Another example where the type of similarity is not visible is constitutional isomers that share the same molecular formula (and weight) but look different due to their diverse atom connectivity (Figure 1).²⁷

Figure 1.

Constitutional isomers sharing the same molecular formula but differ in their connectivity of atoms.

To enable mathematical or statistical computation of similarity levels, chemical characteristics, often referred to as explanatory variables, features, or attributes, must be numerically represented. These variables can be generated directly from structural features and representations, experimental and predicted outcomes (in silico, in vitro, in vivo, physicochemical properties, genomics…), or a combination thereof. Explanatory variables generated based on structure are commonly known as molecular descriptors and fingerprints.

A molecular descriptor is a number or result derived from a logical and mathematical process that turns the symbolic representation of a molecule into useful information for analysis.²⁸ The molecular descriptors encode various types of information, ranging from 0D (e.g., atom counts), 1D (e.g., substructure representation), 2D (e.g., topological representation), 3D (e.g., spatial configuration), to 4D (e.g., stereo-electronic representation).^28–30 This range of information levels, as depicted in Figure 2, offers significant flexibility in analyzing chemical data and defining similarity for various grouping purposes, including end point–specific applications.^28,31,32

Figure 2.

Five levels of information encoded in different types of molecular descriptors. Note: QSAR, quantitative structure–activity relationships.

A fingerprint, conversely, is a string of binary bits (also known as 1D descriptors) that represent the existence or absence of predefined fragments, scaffolds, substructures, functional groups, atom pairs, or a combination thereof (Figure 3).^28,33 Examples of commonly used fingerprints include MACCS (Molecular ACCess System), PUBCHEM, and Morgan.^33–35 Fingerprints are of fixed length and provide a general description of chemicals that is most used for exploratory analysis of database search.^35–37 Thus, the importance of selecting the type of explanatory variables is based on the type of similarity considered for the grouping.

Figure 3.

Illustration of a fingerprint as a binary string indicating the presence or absence of specific substructures.

Similarity assessment can involve utilizing general unfiltered chemical and/or biological response data (end point–agnostic similarity) or a curated subset specifically related to a particular outcome of interest (end point–specific similarity). These scenarios correspond to unsupervised and supervised approaches, respectively.

Unsupervised learning methods, such as clustering, rely on general similarity measures (e.g., chemical features, biological response data from high-throughput assays) to identify patterns and relationships among chemicals without any prior knowledge of the end point of interest (unlabeled data).^38,39 These methods are useful in identifying groups of similar chemicals (i.e., chemicals with the most similar features).⁴⁰ Unsupervised methods are useful for hypothesis generation. However, unsupervised methods may not be suitable for end point–specific grouping of chemicals (e.g., similar structures based on an arbitrary list of features may not produce similar adverse health effects), and their results need to be interpreted in that regard.^38,39,41

In contrast, supervised learning methods, such as classification, require end point–specific similarity measures to analyze labeled data for specific biological or functional end points study.^5,42 These methods use training data labels that contain information about the adverse findings or end points of interest to learn the relationship between the chemicals’ features and the biological activity. Supervised learning methods are suitable for developing end point–specific hypotheses and building predictive models to assess new chemicals.^43,44

While both unsupervised and supervised techniques provide a wealth of information, we argue that it is important to exercise caution to ensure each approach is used for its intended purpose. For example, a common misuse involves the application of unsupervised learning methods as part of a process where the final goal is end point specific.⁴⁵ An example of this is the unsupervised selection of an initial set of analogues in read-across that targets an end point–specific prediction. For this approach, the analogues selected depend on the first step (i.e., identifying patterns and relationships of the analogues of interest). Essentially, applying an unsupervised method (such as an arbitrarily selected fingerprint with a Tanimoto score) that relies on end point–agnostic similarities can only lead to analogues sharing general features not specific to the target end point. This poses a risk that the analogues selected will be used for different end points, even though the secondary layers might depend on the availability of data. Visual similarity, or end point–agnostic chemical structural similarity based on general arbitrarily selected descriptors or fingerprints (a commonly noticed issue during the selection of initial analogues for read-across), does not necessarily lead to end point–specific similarity, and using the same chemicals for different end point studies can result in erroneous interpretations.³⁹ Similarly, when a supervised approach is used, any attempt to derive a general interpretation will be biased by the initial end point–specific dataset on which the algorithm was trained.³⁸

The distinction between using supervised and unsupervised similarity is crucial for developing end point–specific models and making predictions for new chemicals.^39,41,42 Unsupervised methods use an arbitrary choice of the initial explanatory variables and lead to general similarity while supervised methods employ machine learning and/or expert judgement to select the most influential variables for a specific end point leading to a more targeted aspect of similarity. The combination of supervised and unsupervised methods can be relevant at times.³⁸ However, it requires careful consideration to avoid bias and ensure that the similarity aspect in selecting analogues, for example, is appropriate for the end point of interest.⁴⁶

Chemical Similarity: General and End Point Specific

Understanding similarity is essential for utilizing CCAs; similarity encompasses everything from chemical structure, biological activity, and phenotypic outcomes to the inclusion of mechanistic data and is dependent on the context and goal of research or study. Thus, the importance of distinguishing between end point–specific (supervised) and general/end point–agnostic (unsupervised) similarity cannot be overstated for choosing the appropriate grouping method.^38,43

In class-based assessments, we assume that “similar” chemicals belong to one class, where similarity can be defined using different information. Chemical similarity can be based on structural attributes (e.g., bonds, number of aromatic rings, common functional group), property (e.g., physicochemical properties, chemical reactivity), biological and toxicological effect (e.g., biological response, specific toxicity), toxicokinetic (e.g., skin absorption level, same precursor or metabolite), mechanistic [e.g., how a toxicity is induced, mode of action, adverse outcome pathway (AOP), or mechanism of toxicity], or a combination thereof. The word similarity applies to having a consistent trend (e.g., for molecules with a given functional group, boiling point increases with molecular weight) or following a pattern.^47,48 However, incorporating more parameters and descriptors does not always lead to better grouping of chemicals.^49–51 Consonni et al.⁵² and Varmuza et al.,⁵³ among others, have showcased that exceeding an optimal number of variables in a QSAR model invariably results in overfitting and diminished predictive capability.

We posit that exploration of chemical similarity necessitates the characterization of the molecular structures of chemicals. Molecular descriptors and fingerprints provide avenues for representing molecular structures and their features.³⁵ Through the establishment of a shared representation framework, these tools facilitate the encoding of both detailed and overarching characteristics inherent to molecular architectures.²⁸ This process involves the conversion of the molecular structure into numerical representations, which can be used in machine learning approaches for tasks requiring similarity analysis.²⁸ When selecting appropriate molecular descriptors and fingerprints from the array of available types, one should consider the context and objectives of the study, emphasizing the identification of the most pertinent facets of the structural properties (i.e., the selection of analogues for read-across, the arbitrary choice of different fingerprints lead to different analogues).^3,54

Calculating similarity between chemicals is vital for both supervised and unsupervised approaches.^2,55 One way to do this is by utilizing continuous molecular descriptors, fingerprints, and toxicological alerts.^23,28,35 Similarity between chemicals is calculated by comparing these numerical representations from the chemical pairs. To ensure meaningful results, it is crucial to select descriptors relevant to the specific end point under consideration. When dealing with continuous descriptors, distance metrics like Euclidean or cosine distance can be employed.^56–59 Binary descriptors, like fingerprints or structural alerts, require specialized measures such as Jaccard–Tanimoto similarity coefficient.^{33,36,60–64} Comparisons of available similarity coefficients can inform selection of appropriate tools.⁶⁵

In the context of end point–specific studies, similarity determined by supervised approaches is often impacted by activity cliffs.^66–69 Activity cliffs are defined as cases where chemicals with structurally similar features display large differences in potency for a common biological target.⁷⁰ The successes and failures of the conventional “similar structures, similar properties” paradigm can be attributed to the intricacies encompassing activity cliffs.^67,68 Even a minor difference in structure can significantly impact biological activity. Furthermore, employing arbitrarily chosen and unfiltered general descriptors, such a fixed-length fingerprint, might not be optimal for end point–specific similarity analysis. For instance, using a fingerprint like MACCS along with a similarity index (such as Jaccard–Tanimoto) might consider certain chemicals highly similar, despite their substantial difference in biological activity.²¹ Therefore, we argue that it is crucial to establish a foundation for quantifying structural similarity and to recognize the implications of this quantification when exploring structure–activity relationships. This entails carefully selecting the most relevant descriptors to the specific end point under investigation.⁷¹ This approach selectively encodes the impactful structural features for a particular end point, effectively shaping the supervised similarity, and in turn minimizing the activity cliffs effect.^66,72 The concept of activity cliffs can be extended to structural cliffs, wherein structurally dissimilar compounds exhibit similar biological activities.⁷³

When assessing molecular similarity between bioactive substances from a structural bioinformatics perspective, two common representations are employed: 2D structures, which focus on the connectivity of atoms, and 3D structures, which consider the configuration of atoms in three-dimensional space.⁷⁴ Both 2D and 3D descriptors capture different aspects of molecular properties.^{29,30,68,75,76} In the case of receptor-targeting drugs, for example, utilizing structural similarity can provide valuable insights.⁷⁷ By comparing structures of different compounds, researchers can discern patterns that distinguish molecules with binding activity from those without. This is especially important when categorizing agonists, which activate the receptor, and antagonists, which inhibit the receptor activity. Through such analyses, researchers can identify key structural features that are consistently present in binders and absent in nonbinders. This information is invaluable in drug discovery, as it allows for the design of new compounds with optimized binding characteristics, thereby contributing to the development of more effective and targeted pharmaceuticals.^3,78

In conclusion, we assert that an investigation of chemical similarity should emphasize its context-dependent nature, raising questions about defining similarity using diverse tools and differentiating end point–agnostic from end point–specific likeness. Defining end point–specific similarity entails selecting the most important among features (structural descriptors, physicochemical properties, biological aspects, etc.) that either separately and/or synergistically govern the studied end point.^32,79 The applications and interpretations of similarity vary among stakeholders, including medicinal chemists, QSAR model developers, and regulators. Amid these applications, a consensus emerges that similarity definitions should adapt to specific goals and contexts.³⁸ This extends to integrating diverse data, like high-throughput screening results^80,81 and molecular initiating events, to bridge gaps and establish similarity. Although end point–specific frameworks are available for end points including phototoxicity, repeated dose toxicity, and reproductive toxicity, experts in the field collectively acknowledge that a universal method for defining end point–specific similarity is lacking.⁸² Known good practices include broadly considering available data, adopting precise end point definitions, and collaborating across scientific disciplines. This consensus highlights the importance of a more wholistic approach to navigate the intricate concept of chemical structure similarity.

Approaches: Supervised and Unsupervised

Molecular similarity, while theoretically customizable by attribute, gains practical utility when considered in the context of a specific problem. Detaching similarity from specific case scenarios renders it less practical and often devoid of significance.⁵⁴ Thus, customizing similarity to suit a given problem is imperative to confer practical value.³ This involves adopting computational methods to represent molecules before employing supervised and unsupervised techniques tailored to the context.

Supervised and unsupervised learning approaches have shown promise in chemical similarity analysis.^5,38,43 The distinction between supervised and unsupervised approaches highlights the different ways in which structural similarity analysis can be performed, depending on the availability of labeled data and the specific goals of the study. End point–specific methods leverage prior knowledge about the end point to guide the analysis, whereas end point–agnostic similarity methods explore the overall structural landscape of chemical compounds in an unbiased manner.

While both supervised and unsupervised approaches can start with the same features and use similar visualization tools, supervised learning tries to predict a specific property, while unsupervised learning explores general similarities without a specific outcome in mind. Unsupervised clustering can help identify important features in each cluster, which can be used for supervised classification. Supervised learning requires applying algorithms for variable selection such as genetic algorithms and simulated annealing.^79,83–86 Unsupervised methods can be useful for exploring the structural landscape, identifying outliers, and getting a sense of the data space before applying supervised methods.^38,87 Unlike supervised methods where variable selection is applied for end point–specific similarity, for unsupervised methods only dimensionality reduction is necessary since high numbers of descriptors and fingerprints can make the interpretation of results overwhelming.^39,88,89

Supervised and unsupervised approaches usually employ univariate or multivariate analysis.^90,91 However, multivariate calculations have advantages over univariate ones.⁹² Univariate approaches consider properties independently, which might overlook complex interactions.⁹³ Multivariate methods, like principal component analysis (PCA) and hierarchical clustering, capture these interactions, yielding more comprehensive insights.^{90,92,94–98}

Integrating both supervised and unsupervised approaches within the same study is valuable, as they can complement each other effectively.³⁸ When combining different techniques, various synergies might emerge and contribute to a richer understanding of the dataset. In the realm of machine learning, the selection of a particular approach is not unidirectional. Instead, it is a nuanced decision influenced by dataset characteristics and the study objectives. Employing multiple approaches in parallel can yield outcomes that are more interpretable and reliable and enhance the modeling process. When it comes to selecting appropriate analogues and surrogates, context and study goals take center stage. Choosing analogues and surrogates that align with the specific objectives of the study is imperative.⁸² For example, to identify and justify analogues for systemic toxicity predictions, a combination of metabolic profiling, physicochemical properties, and structural features indicating reactivity potential can be used.⁹⁹ Robust feature selection techniques play a critical role in ensuring the relevance of the similarity aspect under examination.^49,79 In essence, combining supervised and unsupervised techniques enhances the depth of insights acquired. The flexibility to use multiple approaches and the careful selection of analogues, surrogates, and features can uphold the integrity and relevance of the study’s findings, ultimately enhancing the quality of the results.

Considerations in Applying Different Approaches

After understanding the concept of similarity (end point–agnostic and end point–specific) and its uses in supervised and unsupervised learning approaches, we argue that users need to explore the various ways of applying these techniques to conduct clustering and classification studies. We feel that interpreting the outcomes of these approaches, whether used separately or in combination, is key as is implementing best practices.

While class-based assessment is built upon the assumption that similar chemicals share similar properties, not all similarities have equal influence on the studied end point (biological activity, toxicity, physicochemical property, etc.). For instance, a functional group might not contribute to the chemical reactivity because it is hindered or irrelevant. One of the common pitfalls in carrying out read-across is misapplying unsupervised approaches for end point–specific studies.^43,82 Conversely, machine learning leverages comprehensive chemical descriptors for toxicity prediction through supervised learning. By comparing and contrasting these approaches, a pathway emerges for their integration. The goal is to enhance prediction accuracy while retaining the transparency inherent in model interpretation. Integrating the strengths of both methods can yield a powerful predictive model that takes advantage of the unsupervised clustering’s interpretability and the supervised classification’s accuracy, contributing to more effective and reliable applications.³⁸ Figure 4 provides an overview of clustering and classification methods, offering recommendations for their appropriate applications and examples of algorithms commonly employed.

Figure 4.

Explanation of clustering and classification methods, recommendations for their applications, and examples of used algorithms.

We propose that navigating the realm of supervised and unsupervised clustering and classification necessitates adhering to best practices for optimal outcomes. It is crucial to avoid a common pitfall: constructing models (e.g., using random forest or nearest neighbor approaches) without refining the molecular descriptors for the specific dataset.^49,100–103 Instead, an iterative supervised feature selection process emerges as a prudent choice. By strategically choosing a subset of descriptors that enhance the performance of a QSAR model within internal cross-validation, the predictive capabilities for new chemicals can be maximized.^52,100 Beyond this, establishing best practices for supervised and unsupervised learning in the context of toxicity and physicochemical property datasets is essential.¹⁰⁴ Another aspect of this exploration entails assessing how supervised and unsupervised feature selection impact external prediction performance across various QSAR methods like nearest analog, k-nearest neighbors, random forest, and support vector machines. Embracing these practices can improve the clarity and effectiveness of these techniques, enabling insightful analysis and robust predictions.

More broadly, many themes converge when it comes to best practices for using clustering and classification techniques for toxicity assessment of chemicals. One prominent theme involves establishing the precise objective of a study and subsequently selecting the most suitable approach. Throughout this process, we argue that it is essential to carefully consider factors such as end point type, data availability, and the depth of decision-making required. Additionally, it is pivotal to consider the synergy between supervised and unsupervised methods, emphasizing their combined utility. In our opinion, an essential aspect of this story lies in acknowledging the degree of complexity of the end point under examination and understanding its influence on modeling accuracy. In this confluence of themes, we believe the importance of tailored approaches for robust and accurate toxicity assessment emerges.

Another significant theme is the importance of tailoring fundamental concepts for inexperienced users. Examples of this are considering result accessibility and interpretability, along with the requisite training and knowledge accumulation. Amid this, a challenge emerges in assessing the efficacy and applicability of predictive models.^58,105,106 Experts propose a collaborative model development process, wherein developers engage users from the outset.¹⁰⁷ This proactive approach ensures that end products align closely with users’ most crucial requirements, rather than presenting solutions after the fact. This theme underscores the value of user-centric design and the shared effort to create impactful and user-friendly tools.

A third notable theme is the significance of employing generic features to predict end points and the necessity of factoring in the quality^108,109 and quantity of training datasets.¹¹⁰ This theme centers around input considerations, specifically the criteria guiding method and feature selection. We argue that machine learning’s role in harmonizing data across molecules to facilitate meaningful comparisons is also a crucial aspect within this theme. In essence, this theme underscores the careful deliberation required in determining input parameters, which includes selecting appropriate methods and features and ensuring data quality, ultimately contributing to accurate and reliable results.

In summary, a pivotal takeaway is the paramount significance of collaboration and knowledge exchange between tool developers and users of these approaches. Equally important is the recognition of tailoring machine learning and class-based approaches to the distinct objectives and limitations of each study. We believe that this collective approach ensures that the resulting interpretations are not only accurate and effective but also aligned with the real-world needs and constraints of the users.

Applications and Examples

CCAs find diverse applications in chemical assessment, spanning prioritization of toxicological research to risk assessment, and high adaptability in various contexts. A notable instance lies in the European Union’s Registration, Evaluation, Authorization and Restriction of Chemicals (REACH) regulation, where the use of read-across (by showing chemical similarity and a read-across hypothesis) is encouraged for substance registration.¹¹¹ Here, the integration of both structural and mechanistic similarity is pivotal for successful read-across submissions by REACH registrants.^9,112 Delving further into the read-across approaches, the process of analogues selection can be a combination of supervised and unsupervised selection methods. A practical example is the integration of QSAR and read-across.¹¹³ The effectiveness of this integration is shaped by the distribution of analogous compounds within the small chemical space around the target.

This structural similarity also proves valuable in the European Chemicals Agency’s (ECHA) grouping of substances for screening and prioritization objectives.¹¹¹ An exemplar of this functionality is present in the OECD (Organization for Economic Co-operation and Development) QSAR Toolbox (https://qsartoolbox.org/), a software designed for substance grouping, freely accessible through collaboration between ECHA and the OECD. These international bodies have recently published an updated QSAR Assessment Framework, which is intended to support transparent model reporting, verification, validation, and regulatory implementation.¹¹⁴ These practical applications underscore the versatility and relevance of clustering and classification in advancing chemical assessments.

Similarly, an examination of the US Environmental Protection Agency (EPA)’s historical applications and upcoming requirements for prioritization using NAMs reveals their reliance on chemical similarity.¹¹⁵ This involves assessing structural similarity encompassing both two-dimensional (2D) and three-dimensional (3D) aspects, as well as considering bioactivity viewpoints. Such analyses aid in establishing the significance of these approaches in the EPA’s efforts to prioritize chemicals effectively, both in the past and moving forward.^116,117

Another illustration involves employing class-based methodologies to assess exposure, hazard, and risk, utilizing subclasses of organohalogen flame retardants (OFRs) as a focal point.¹¹⁸ This comprehensive study encompasses multiple facets of OFR evaluation. It commences with scoping and evidence mapping to categorize existing studies on exposure, mechanisms, and toxicity. Additionally, this investigation delves into exposure, risk, and regulatory implications concerning distinct subclasses of OFRs. This example underscores the versatility of class-based approaches in dissecting multifaceted chemical assessments and regulatory considerations.

Class-based assessments offer a sustainable path forward in different NAMs applications. Effective implementation requires open dialogue among stakeholders, solid case studies, and a focus on education and data management.¹¹⁸ We believe the overarching message is that the collaboration between stakeholders and the continued refinement of class-based methodologies will ultimately shape the future of chemical assessment.

Discussion

Adapting class-based approaches for chemical regulations and hazard/risk evaluations is essential but comes with its set of challenges. Central to this endeavor is understanding the pivotal attributes of chemical structures for targeted similarity assessments. We posit that establishing flexible guidelines tailored to regulatory end points is imperative for navigating this complex landscape effectively. In this section, we delve into the intricacies of adapting class-based methods, focusing on three key takeaways that underscore the importance of robust case studies in evaluating hazards or risks.

How to Choose Between General and End Point–Specific Similarity?

In summary, end point–specific structural similarity is a concept used in cheminformatics and computational chemistry to assess the similarity of chemical compounds based on their structural features with respect to a particular biological or chemical end point. This end point could be the activity against a specific target (e.g., enzyme inhibition, receptor binding), toxicity, metabolic stability, or any other desired property.

The key difference between end point–specific structural similarity and general, end point–agnostic similarity lies in focused optimization vs. a broader consideration of structural characteristics. In end point–specific similarity, the structural features that are considered relevant to the end point of interest are prioritized, while irrelevant features are either minimized or ignored. Conversely, end point–agnostic similarity considers all structural features or general arbitrarily chosen fingerprints without necessarily prioritizing any specific end point.

For example, let’s consider two compounds, Compound A and Compound B, both designed as potential inhibitors of a particular enzyme. End point–specific structural similarity analysis between Compound A and Compound B would primarily focus on the key structural motifs or features known to influence enzyme inhibition. This analysis would disregard other structural aspects that might not be relevant to the inhibition of the enzyme. On the contrary, end point–agnostic similarity analysis between Compound A and Compound B would considers all structural features or fingerprints in general, regardless of their relevance to enzyme inhibition. This analysis provides a broader perspective on the overall structural similarity between the two compounds but may not directly address their similarity in terms of enzyme inhibition potency.

While both end point–specific and end point–agnostic general similarity analyses are valuable in cheminformatics and computational chemistry, their applications differ based on the specific goals of the study, with end point–specific similarity being more targeted toward a particular end point or property of interest, and end point–agnostic similarity providing a broader perspective on overall structural similarity. A comprehensive overview of the two types of similarity is provided in Table 1.

Table 1.

Comparative overview of end point–specific and general chemical structure similarity.

Category	End point–specific similarity	End point–agnostic similarity
Goals of the study	Drug Design: Helps identifying compounds with similar structural features to known active compounds for a specific target, facilitating drug discovery.119 Toxicity Prediction: Assists in predicting the toxicity of new compounds by comparing their structures to known toxic compounds for a specific end point.109 Metabolism Prediction: Aids in predicting the metabolic fate of compounds by comparing their structures to known substrates or inhibitors of specific metabolic enzymes.45	Compound Clustering: Allows grouping compounds into clusters based on their overall structural similarity, providing insights into chemical space exploration and diversity analysis.64 Virtual Screening: Helps fast screening of large compound libraries to identify potential hits for various biological targets without focusing on specific end points initially.25 Chemical Space Exploration: Aids in exploring the chemical space by identifying compounds with diverse structural features, to assist in lead optimization and scaffold hopping.120
Methodologies and tools	Pharmacophore Modeling: Identifies the essential features in a molecule that are crucial for binding to a specific target. Pharmacophore-based similarity assessment compares the pharmacophoric features of compounds rather than their entire structures.121 Tools such as LigandScout, MOE (Molecular Operating Environment), and Discovery Studio provide functionalities for pharmacophore-based similarity analysis.121–123 3D Structure Alignment: For compounds with known 3D structures, structural alignment algorithms can be used to overlay and compare the binding conformations of ligands in the active site of a target protein.124 Methods include maximum common substructure (MCS) alignment or flexible docking. Tools such as AutoDock, GOLD, and Schrödinger Suite offer 3D structure alignment and similarity analysis capabilities.125–127 QSAR: QSAR models can be utilized to predict the activity of compounds against a specific target based on their structural descriptors.23 Similarity can be assessed by comparing the predicted activities of compounds using QSAR models.63 Tools like QSAR Toolbox, OPERA, and MOE provide QSAR models based on end point–specific similarity.122,128	Fingerprint-based Similarity: Fingerprints encode the structural features of compounds into binary vectors.33 Similarity between compounds is assessed by comparing their fingerprints using metrics such as Tanimoto coefficient or Euclidean distance.62 Popular fingerprint types include MACCS keys, ECFP (Extended Connectivity Fingerprints), and Morgan fingerprints.34,129 Tools like RDKit, Open Babel, and ChemAxon offer functionalities for fingerprint-based similarity analysis.64,130 Graph-based Similarity: Compounds can be represented as graphs, where atoms are nodes and bonds are edges.27 Graph similarity algorithms compare the topology of these graphs to assess structural similarity between compounds.34 Methods include graph kernel methods, graph edit distance, and maximum common subgraph algorithms.60 Tools such as NetworkX, igraph, and MOE provide graph-based similarity analysis capabilities.122 Fragment-based Similarity: Compounds are cut into smaller chemical fragments, and similarity is assessed based on the presence or absence of these fragments.131 Fragment-based similarity approaches offer a balance between molecular complexity and computational efficiency.132 Tools providing fragment-based similarity include Fragment Network, RDKit, and ChemAxon.133
Type of learning	Supervised learning: End point–specific similarity approaches apply supervised learning because they are tailored to a specific task (e.g., predicting activity against a particular target) and utilize labeled data or prior knowledge to guide the analysis.87 Operating mechanism: The algorithm learns from labeled data, where each example is associated with a target value or class label. Similarly, in end point–specific similarity, the analysis is guided by knowledge of the specific end point or biological activity of interest. The structural features considered in end point–specific similarity are chosen based on their relevance to the end point.63 Example methods: Pharmacophore and QSAR modeling explicitly incorporate information about the end point into the analysis. These methods learn to differentiate between compounds based on their similarity to known active compounds or their predicted activity against the target.63,121	Unsupervised learning: General similarity approaches apply unsupervised learning, operating in a data-driven manner without relying on labels or knowledge about a specific end point. They aim to uncover underlying relationships in the data without predefined targets or classes.41 Operating mechanism: The algorithm autonomously explores the structure of the data without explicit guidance from labeled examples. Similarly, general similarity methods compare compounds based on their overall structural features without specific reference to a particular end point, allowing for a comprehensive exploration of chemical space and identification of underlying patterns.38 Example methods: Fingerprint-based similarity and clustering algorithms delve into the inherent structure of the chemical space to discern patterns or similarities among the studied compounds. This analysis facilitates a deeper understanding of the relationships among the chemicals of a dataset.39,64

Note: MACCS, Molecular ACCess System; QSAR, quantitative structure–activity relationships.

Why Can General Similarity Not Be Used for End Point–Specific Applications?

End point–agnostic similarity measures, such as arbitrarily chosen general fingerprint-based or graph-based comparisons, provide a broad assessment of structural similarity between chemical compounds. While they are valuable for tasks like compound clustering, virtual screening, and chemical space exploration, we argue that they are not suitable for end point–specific purposes due to several reasons:

• Relevance to End Point: End point–agnostic similarity methods do not prioritize the structural features that are specifically relevant to the end point of interest. They consider all structural aspects equally, regardless of their importance for the biological or chemical activity being studied. This can lead to a dilution of the signal relevant to the end point amid a myriad of other structural details.

• Lack of Specificity: General similarity metrics do not account for the nuanced structural requirements of a particular end point. They may identify compounds as similar even if they differ significantly in their ability to interact with the target or exhibit the desired biological activity. This lack of specificity can result in false positives or false negatives when predicting compound activity.

• Inability to Capture Pharmacophoric Features: General similarity methods often fail to capture the pharmacophoric features necessary for interaction with a specific target. These features, such as hydrogen bond donors/acceptors, aromatic rings, or hydrophobic regions, are crucial for binding affinity and biological activity. End point–specific similarity approaches, such as pharmacophore modeling, are designed to explicitly consider these features.

• Risk of Overlooking Important Structural Motifs: End point–specific activities often rely on specific structural motifs or functional groups known to be critical for interaction with the target. End point–agnostic similarity methods may overlook these motifs if they are not prevalent in the dataset or if they are overshadowed by other, more abundant structural features. End point–specific approaches ensure that these important motifs are given appropriate weight in the similarity assessment.

• Failure to Prioritize Key Interactions: In many cases, the biological activity of a compound is mediated by specific interactions with the target protein or receptor. General similarity measures may fail to prioritize compounds that share these key interactions, leading to inaccurate predictions of compound activity. End point–specific methods, such as 3D structure alignment or QSAR modeling, explicitly consider these interactions.

In summary, while end point–agnostic similarity methods are valuable for exploring overall structural relationships between compounds, we have shown that they are not suitable for end point–specific purposes due to their lack of specificity, inability to capture pharmacophoric features, and risk of overlooking important structural motifs and interactions. End point–specific similarity approaches are designed to address these limitations by focusing explicitly on the structural features relevant to the end point of interest.

How to Identify End Point–Specific Relevant Structural Features?

Identifying end point–specific relevant features for read-across and QSARs involves a systematic approach to selecting and prioritizing structural features that are most influential for the end point of interest. Methods and tools commonly used for this purpose include:

• Literature Review: Conducting a comprehensive review of the scientific literature to identify known structural features or substructures associated with the end point of interest. This may involve studying published SAR (structure–activity relationship) studies, experimental data, and mechanistic insights.

• Expert Knowledge: Consulting domain experts, such as medicinal chemists, toxicologists, or computational chemists, who possess expertise in the specific end point and can provide valuable insights into the relevant structural features.

• Data Mining and Cheminformatics Tools: Utilize data mining and cheminformatics tools to analyze large databases of chemical compounds and associated biological activity data. These tools can help identify statistically significant structural patterns or substructures correlated with the end point. Tools such as RDKit, ChemAxon, and KNIME offer functionalities for data mining, substructure searching, and statistical analysis of chemical and biological data.¹³⁴

• Pharmacophore Modeling: Construct pharmacophore models based on known active compounds or ligands for the end point of interest. Pharmacophore models identify the essential features (e.g., hydrogen bond donors/acceptors, aromatic rings) required for binding to the target. Tools like LigandScout, MOE, and Discovery Studio provide capabilities for pharmacophore generation and analysis.^121–123

• Fragment-Based Analysis: Decompose chemical compounds into fragments or chemical substructures and analyze their frequency and distribution among active and inactive compounds. Fragments that are overrepresented in active compounds may represent relevant structural features. Tools like Fragment Network and RDKit provide functionalities for fragment-based analysis and substructure mining.¹³³

• Machine Learning and Feature Importance Analysis: Employ machine learning algorithms, such as random forests or gradient boosting machines, to rank the importance of structural features in predicting the end point. Feature importance analysis techniques can identify the most influential descriptors or substructures. Machine learning libraries like scikit-learn (Sci-kit Learn Developers) in Python offer feature importance analysis methods for various algorithms.

By integrating these methods and tools, researchers can systematically identify and prioritize end point–specific relevant features for class-based studies, facilitating the development of predictive models and decision-making in chemical risk assessment and compound optimization.

Conclusions

This overview of class-based approaches outlined key concepts for supervised and unsupervised similarity methods and explored current applications of clustering and classification approaches. It addressed challenges and proposed solutions, along with advocating for future collaborations to promote better understanding, best practices, and increased use of CCAs in toxicity research and risk assessment. This serves two main purposes: educating users about the diverse CCAs and fostering enduring collaborations to tackle challenges associated with end point–specific similarity. The initial objective involved providing an in-depth overview of CCAs, emphasizing their relevance in scientific toxicity research and hazard identification—a pivotal step in risk assessment. Emphasis was placed on understanding the fundamental principles behind these techniques to ensure correct tool utilization and precise result interpretation.

The secondary objective was to emphasize the context-driven nature of similarity. Distinctions were drawn between end point–agnostic similarity, which is crucial for unsupervised analysis, and end point–specific similarity, which requires supervised selection of pertinent variables for accurate end point specificity. While combining supervised and unsupervised approaches is possible, it is crucial to exercise caution regarding the potential misuse of unsupervised analyses, especially when aiming for end point specificity. In the initial stages of selecting analogues for read-across, it is imperative to avoid using unsupervised methods based on general features denoting end point–agnostic similarity, such as starting with an arbitrarily selected fingerprint with a similarity metric. General similarity (visual or arbitrary)-based analogue selection does not inherently guarantee end point specificity, potentially leading to erroneous predictions.⁵ Analogues are also nontransferable across different studies because each end point has its own most influential features that define its underlying specific similarity.

This work inspired the creation of an automated workflow aimed at facilitating chemical grouping using supervised and unsupervised methods.¹⁰ This workflow was developed within the framework of the free and open-source KNIME environment.¹³⁴ It implements the guidance and key takeaways discussed herein, featuring an intuitive, documented, and guided user-friendly graphical interface aiming to democratize the use of CCAs for nonexperts and simplifying their application for the wider community.

For a broader application of CCAs, we believe it is important to establish a sustained international consortium of collaborators. The recommended consortium role would involve leveraging expertise, developing roadmaps and standards, establishing optimal communication and result-sharing mechanisms, publishing state-of-the-science papers, and investing in technological advancements. Through interdisciplinary collaboration and implementation of these recommendations, we can enhance the utilization of CCAs across diverse fields.

Conclusions and opinions are those of the individual authors and do not necessarily reflect the policies or views of EHP Publishing or the National Institute of Environmental Health Sciences.