QSAR modelling of pesticide modes of toxic action across trophic levels via multi‐label extreme gradient boosting

Abstract

BACKGROUND

In silico identifying modes of toxic action (MOAs) of toxic chemicals are of significance to gaining knowledge about the toxic severities and mechanisms of chemicals at the whole‐organism level. Currently, existing in silico models commonly focus on a diversity of chemical MOAs to one specific species, without considering their toxic effects to off‐target organisms.

RESULTS

In this work, we propose a quantitative structure–activity relationship (QSAR)‐based multi‐label extreme gradient boosting (XGBoost) model to consider the scenario that a pesticide exhibits multiple MOAs on different organisms across trophic levels. In the integral space of 16 taxa‐specific MOAs, pesticides are represented by the structural fingerprint MACCSKeys and Morgan, and each of the 16 taxa‐specific MoAs is treated as a class label. K‐fold cross‐validation (k = 10) shows that the proposed multi‐label XGBoost model achieves 0.81 micro recall, 0.72 macro recall, 0.55 perfect match ratio, and 69.5% Jaccard accuracy. Among the 16 taxa‐specific MOAs, 12 MOAs achieve satisfactory recall rates ranging from 0.643 (aquatic‐narcosis) to 0.986 (aquatic acetylcholinesterase inhibition). An external test shows that 84.62% of the herbicides, exhibiting plant photosynthesis inhibition, are correctly recognized. The holdout test shows that the proposed model, though possessing a much higher complexity of label space, outperforms or performs equivalently to existing multi‐class model (linear discriminant analysis).

CONCLUSION

Computational results show that narcosis majorly exhibits as an independent toxic effect, or an accessory/baseline toxic effect preferentially accompanied by reactive or other specific MOAs, and that the proposed multi‐label XGBoost model potentially benefits deriving baseline (narcosis) toxicity models for the studied organisms. © 2026 Society of Chemical Industry.

The model achieves the largest coverage of 16 taxa‐specific mode of action (MOA) classes. Multi‐label extreme gradient boosting assigns multiple MOAs to a pesticide simultaneously. Narcosis is unravelled as a baseline or broad‐spectrum MOA potentially accompanied by other specific MOAs. The model potentially facilitates deriving baseline (narcosis) toxicity models.

1. INTRODUCTION

Modes of toxic action (MOAs) have been well accepted as (eco)toxicological determinants to group hazardous chemicals into chemical classes for mechanistic insights. ¹ The pioneering work, ² referred to as the Verhaar scheme, proposes a fragment‐based expert system that broadly classifies organic chemicals into inert (also called baseline/minimum toxicity or nonpolar narcosis), less inert (polar narcosis), reactive, and specifically‐acting compounds. From mechanistic point of view, inert chemicals, whose toxicity primarily results from chemical hydrophobicity, elicit narcosis by non‐specifically acting at cell membrane; ³ less inert chemicals, slightly more toxic than the baseline toxicity, are typical in possessing acidity for their hydrogen‐bond donors. ⁴ Both reactive chemicals and specifically‐acting chemicals exhibit excess toxicity, the difference in which is that reactive chemicals non‐specifically react with biomolecules, while specifically‐acting chemicals selectively bind to certain receptor biomolecules (Hermens et al. ⁵ Ariens et al. ⁶ ). In the Verhaar scheme, ² , ⁴ the baseline toxicity (nonpolar narcosis) is modelled as the linear regression of logarithmic octanol–water partition coefficient (logK _ow) and the polar narcosis, reactive, and specific MOA classes are determined by the toxicity ratio (TR) between predicted baseline toxicity and experimental toxicity. Meanwhile, the Verhaar scheme provides a set of expert‐curated rules, which are based on the structural fragments of chemical molecules only, to predict MOA classes.

To date, in silico (eco)toxicological modelling of pesticides has been, or is being, devoted to quantitative toxicity regression or qualitative severity classification. These (eco)toxicological modelling strategies expand our vision for in silico MOA modelling. From a mechanistic point of view, the prevalent modelling strategies for effect concentrations of pesticides (e.g., LC₅₀ and EC₅₀) are based on quantitative structure–activity relationships (QSAR) ⁷ , ⁸ , ⁹ , ¹⁰ or quantitative activity–activity relationships (QAAR). ¹¹ , ¹² , ¹³ QSAR‐based modelling strategy commonly assumes that structurally similar chemicals exhibit similar biological activities or toxicities, whereas QAAR‐based modelling strategy assumes that (eco)toxicological knowledge is transferable across species directly from the relationships between their effect concentrations on the same group of pesticides. From modelling point of view, QSAR‐based models are further divided into local and global models, while QAAR‐based models are basically global models. Local models focus on one species, for example aquatic fishes (fathead minnow, daphnia magna, tetrahymena pyriformis and vibrio fischeri), ⁹ or on one chemical class of pesticides, for example endocrine disrupting chemicals. ⁷ Some work, though simultaneously modelling multiple species, for example the machine learning and deep learning models, ⁹ still belongs to local modelling strategy because the models are mutually independent, with one model developed for one species. Quite differently, global models model multiple species simultaneously into one integral model by means of common toxic patterns of structural fragments and physiochemical properties, ⁸ , ¹⁰ or population evolution and phylogeny. ¹¹ , ¹³ Comparatively, local models exhibit better performance on one specific species due to their lower model complexity, while global models facilitate transferring knowledge across aquatic species, animal groups and food commodities, ¹³ , ¹⁴ , ¹⁵ or even across tropic levels (e.g., algae, invertebrates and fish). ⁸ As a result, both the majority and the minor species are benefitted from the perspective of performance improvement. Currently, the major descriptor for baseline toxicity regression is the octanol–water partition coefficient (K _ow), which is shown in recent studies not to exhibit satisfactory performance, for instance the coefficient of determination (R ²) ranges from 0.353 to 0.551 in aquatic species. ¹⁶ K _ow, due to its limitation of descriptive ability, has been recommended to be replaced by the liposome–water partition coefficient, K _lipw, for the purpose of better discriminating polar narcosis from nonpolar narcosis. ¹⁷ , ¹⁸ , ¹⁹ For the insecticides and fungicides that act on off‐target aquatic species Daphnia magna, Wang et al. ²⁰ propose many more descriptors, such as heat of formation (Hf), polar surface area (PSA), polarity (S), molecular volume (V), hydrogen bonding basicity (B), and Cosmo volume (CV), to acquire sufficient descriptive ability. These toxicity regression models provide valuable cues to baseline toxicity modelling. Nowadays, the baseline toxicity models are mostly based on K _ow and/or K _lipw to infer the baseline (narcosis) toxicity of chemicals. The metric toxic ratios (TR), that is, the ratios of experimental toxicity values to the predicted baseline toxicity values, are used to infer polar narcosis, reactive, more specific MOAs ¹⁶ , ¹⁸ , ²⁰ , ²¹ , ²² and critical body residues (CBRs). ²³ The TR thresholds vary with species and heavily depend on experimental dose–response values. ²² Comparatively, fragment‐based QSAR models proposed in the Verhaar scheme ² , ⁴ only require structural information and do not depend on experimental dose–response values.

Nowadays, in silico ecotoxicological studies on chemicals or pesticides have seemed to shift their focus from dose–response assays to QASR‐based insights of modes of toxic action (MOAs) ²⁴ , ²⁵ , ²⁶ , ²⁷ , ²⁸ or mechanisms of toxic action (MechoAs). ²⁹ , ³⁰ MOAs are defined as chemical‐induced functional changes at the cellular or whole organismal level, which are the toxic effects at the end of adverse outcome pathways (AOPs). Differently, MechoAs focus more on toxic mechanisms of chemicals, and are defined as the molecular initiating events (MIEs) activating AOPs. ³⁰ As a modification to the Verhaar scheme, ² , ⁴ the MOATox database ²⁴ further divides the specific MOAs into acetylcholinesterase (AChE) inhibition, electron transport inhibition, iono/osmoregulatory/circulatory impairment and neurotoxicity, so that the MOATox scheme expands to six broad MOA categories and 31 MOA subcategories. To date, the MOATox scheme ²⁴ has been intensively used to develop multi‐class machine learning models for MOA prediction, for example linear discriminant analysis (LDA), ²⁶ , ²⁸ random forest, ²⁶ Bayesian network, ³¹ and k‐nearest neighbors (k‐NN). ³² These models are essentially local models, adopting the strategy of one model for one species. In addition, these models commonly assume that a pesticide exhibits one toxic effect or one mode of toxic action, which does not hold true in many cases. First, a pesticide, possessing several distinct fragments, potentially acts on multiple specific receptors according to the Verhaar scheme, and thus it will manifest more than one mode of action. Second, as the studies ²¹ , ²² state, a pesticide often exhibits different MOAs to different off‐target aquatic species, and MOA is actually species‐dependent. More importantly, we are supposed to study pesticide modes of toxic action at the whole‐organism level ³⁰ so that we can gain insights into how a pesticide exhibits different MOAs to off‐target organisms. For instance, herbicides, originally designed to mortally target crop weeds, potentially exhibit inert or less inert toxicities to aquatic species. ²¹ To cover more species, we are required to adopt global strategies for pesticide MOAs modelling, where MOAs are labelled with species taxonomies. The taxa‐specific MOA database, ²⁷ hereinafter referred to as MOAtaxa, expands the species coverage to plant, fungi, insect, rodent, and acari, in which more specific MOAs are added as compared with the MOATox scheme, ²⁴ such as photosynthesis modulation, anticoagulation, biosynthesis inhibition, growth regulation, and neurotoxicity. In the MOAtaxa scheme, aquatic MOAs are not considered. Based on the MOAtaxa scheme, ²⁷ an LDA‐based global model was proposed to predict taxa‐specific MOAs. The studies (Kienzler and Barron et al. ³³ , ³⁴ ) comprehensively evaluate the consensuses among the Verhaar scheme, ² , ⁴ the ASTER scheme, ³⁵ and the MOATox scheme. ²⁴ All these models, based on the above‐mentioned schemes, are essentially multi‐class models. However, in silico MOAs modelling at the whole‐organism level actually belong to multi‐label learning scenarios from the perspective of machine learning, especially when the target and off‐target effects are both considered.

In this work, we propose a multi‐label extreme gradient boosting (XGBoost) model to predict pesticide MOAs on target and off‐target organisms across tropic levels. We aggregate the MOATox scheme ²⁴ and the MOAtaxa scheme ²⁷ to expand the organism space covering plant, fungi, insect, rodent, acari and aquatic species. As such, the multi‐label XGBoost model covers 16 taxa‐specific MOAs with chemical diversity significantly increased. The pesticides are represented using chemical structural fingerprints such as MACCSKeys and Morgan. The proposed multi‐label XGBoost model is evaluated using k‐fold cross‐validation (k = 10) and external test to yield its multi‐label learning performance metrics. The chemical substructures or fragments contributing most to the model performance are reported for mechanism understanding. Lastly, we take the aquatic species fathead minnow as case studies to unveil the relationships between toxicities and MOAs, and the similarities between MOAs.

2. MATERIALS AND METHODS

2.1. MOA data

2.1.1. Training data

The MOATox scheme ²⁴ focuses on aquatic species, and covers six broad MOA categories: narcosis, reactivity, electron transport inhibition, iono/osmoregulatory/circulatory impairment, AChE inhibition, and neurotoxicity. The six MoAs are further divided into 31 MOA subcategories. The MOAtaxa scheme ²⁷ covers five organisms belonging to different taxonomies, including plant, fungi, insect, rodent, and acari. Each organism is associated with one to three MOAs. For instance, plant is associated with the MOAs of biosynthesis inhibition, photosynthesis modulation, and growth regulation. Insect is associated with respiration modulation and neurotoxicity. The MOATox scheme ²⁴ curates MOAs of 1208 pesticides and the MOAtaxa scheme ²⁷ curates MOAs of 1309 pesticides. The two schemes are merged into an integral MOA space, so that each MOA is organism‐dependent, where the MOAs in the MOATox scheme ²⁴ are prefixed with ‘Aquatic’. As such, we totally obtain 16 taxa‐specific MOAs. For each pesticide, we resort to PubChem (https://pubchem.ncbi.nlm.nih.gov/) for SMILES retrieval. After removing the pesticides whose SMILES are not available or do not pass the RDKit parse, ³⁶ we obtain a total of 1855 pesticides. These pesticides are used as the training instances (see Supporting Information, File S1). The distribution and overlap of pesticides among the 16 taxa‐specific MOAs are shown in Fig. 1, where one dot denotes the number of pesticides possessed by the specified MOA and two connected dots denote the number of pesticides that belong to both the specified MOAs. There are 643 pesticides assigned two taxa‐specific MOAs, one taxonomy is aquatic and the other taxonomy is the organism from the MOAtaxa scheme. ²⁷ For instance, the pesticide dithicrofos (CAS number: 41219–31‐2) exhibits AChE inhibition on aquatic species and meanwhile exhibits neurotoxicity on insect. We can see that it is not plausible in the two schemes (MOATox and MOAtaxa) to treat MOA modelling as a multi‐class learning scenario. As shown in Fig. 1, a pesticide probably exhibits different MOAs on different organisms, for instance 240 pesticides exhibit neurotoxicity on insect and exhibit AChE inhibition on aquatic species. In addition, a pesticide also probably exhibits identical MOA(s) on different organisms, for instance 152 pesticides exhibit neurotoxicity on insect and aquatic species. This evidence shows that in silico MOA modelling at the whole‐organism level is a canonical multi‐label learning scenario. The pesticides are rather skewedly distributed among the 16 taxa‐specific MOAs, with the size ranging from 17 (Rodent‐Anticoagulation) to 464 (Aquatic‐Narcosis). If the minor MOAs, for example Rodent‐Anticoagulation, are independently modelled as local models, the performance would be rather worse due to limited training data. If aggregated, the minor MOAs are supposed to benefit from the knowledge that is transferred from the major MOAs. Based on the same reason, we adopt as class labels the 16 broad MOAs categories rather than the 31 MOAs subcategories, and merging the two schemes (MOATox and MOAtaxa) introduces the major aquatic species to the training data and MOA space to benefit the minor organisms.

Figure 1.

Pesticide distribution and overlap among the 16 taxa‐specific MOAs.

2.1.2. External test data

External tests, also called independent test data, are of importance if available to justify that the trained model generalizes well to unseen examples. In this work, we use the herbicide MOAs ³⁷ as external test data, where the MOA subcategories (amino acid synthesis, fatty acid biosynthesis, cell wall synthesis, and tetrahydrofolate synthesis) are merged into Plant‐Biosynthesis inhibitor, and the MOAs subcategories (electron transfer, heme synthesis for chlorophyll, and carotenoid synthesis of the herbicides) are merged into Plant‐Photosynthesis. After filtering out the herbicides already contained in the training data, we obtain a total of 47 herbicides used for external tests (see Supporting Information, File S1).

2.1.3. Case‐study data

We used the aquatic species fathead minnow (Pimephales promelas) as a case study, in which we attempted to reveal the relationships between toxicity values and modes of toxic action, and the similarities between MOAs. The ecotoxicological measurements on fathead minnow are taken from Sheffield and Judson. ¹⁰ After removing the pesticides already contained in the training data, we obtained 786 pesticides as case‐study data (see Supporting Information, File S1). Among the endpoints, we chose the LC₅₀ values (mg L⁻¹) of 96‐h exposure duration because this is often adopted as the reference duration of exposure for inferring pesticide baseline toxicities on aquatic species. When there were multiple LC₅₀ values from different endpoints or literature, we chose the minimum one to focus on the most toxic severity reported.

2.2. Feature engineering and multi‐label XGBoost

2.2.1. Feature engineering

In the Verhaar scheme, ² a series of fragment‐based rules have been proposed as expertise to identify MOAs, for instance if a chemical possesses allylic/propargylic activation, benzylic activation, a three‐membered heterocyclic ring, activated carbon–carbon double/triple bonds, etc., the chemical will be identified as reactive compound. However, the scheme is criticized for its limited fragment diversity and rule space. ²⁶ As an alternative solution, the chemo‐physical and topological descriptors, such as E‐state values, E‐state counts, constitutional descriptors, topological descriptors, kappa, and hydrogen bond acceptor/donor counts, are used to depict pesticides. ²⁶ , ²⁷ In this work, we adopt chemical structural fingerprints such as MACCSKeys ³⁸ and Morgan ³⁹ as features according to the principle of QSARs. The MACCSKeys fingerprint defines the presence or absence of predefined functional groups, substructure motifs, or fragments. The Morgan fingerprint is a circular topological fingerprint representing the neighborhood atoms along a specified length of path. Both fingerprints are represented with 0/1 sparse bit‐vector. Here, the MACCSKeys fingerprint is 166 bits long and the Morgan fingerprint is set to be 2048 bits long (radius = 2). The two fingerprints are concatenated into an expanded feature space, facilitating overcoming the limited fragment space of the Verhaar scheme. ²

2.2.2. Multi‐label XGBoost

XGBoost ⁴⁰ is an ensemble model that iteratively constructs regression trees from the errors of the proceeding trees, aiming at steadily reducing the regression errors. The Python scikit‐learn package for XGBoost (version $\ge$ 1.6) supports multi‐label classification, where all instances have multiple non‐exclusive class labels. By default, XGBoost builds one meta binary model classifier for each class label, and each meta classifier identifies an instance as positive (1) or negative (0) using a decision probability ζ. In this work, we used a 16‐dimensional 0/1 vector to represent the 16 taxa‐specific MOA classes so that each element represents the presence or absence of the specified MOA a pesticide exhibits. Multiple 1‐elements in the label vector indicate that the pesticide exhibits multiple MOAs.

2.3. Multi‐label performance metrics

2.3.1. Binary metrics

Multi‐label XGBoost trains one meta binary classifier for one taxa‐specific MOA, and the meta classifier is evaluated using conventional binary learning metrics as defined in Eqn (1). ⁴¹

$$\left\{\begin{array}{c}P=\frac{\mathit{tp}}{\mathit{tp}+\mathit{fp}}\\ R=\frac{\mathit{tp}}{\mathit{tp}+\mathit{fn}}\\ F1=\frac{2\times P\times R}{P+R}\\ \text{accuracy}=\frac{\mathit{tp}+\mathit{tn}}{\mathit{tp}+\mathit{tn}+\mathit{fp}+\mathit{fn}}\\ \mathrm{MCC}=\frac{\mathit{tp}\times \mathit{tn}-\mathit{fp}\times \mathit{fn}}{\sqrt{\left(\mathit{tp}+\mathit{fp}\right)\left(\mathit{tp}+\mathit{fn}\right)\left(\mathit{tn}+\mathit{fp}\right)\left(\mathit{tn}+\mathit{fn}\right)}}\end{array}\right.$$ (1)

In Eqn (1), P denotes precision, R denotes recall, tp denotes true positive, fp denotes false positive, tn denotes true negative, fn denotes false negative, and MCC denotes Matthews correlation coefficient.

2.3.2. Multi‐label metrics

The meta classifiers are aggregated into a multi‐label XGBoost model, whose overall performance is estimated by averaging the instance‐level performance (denoted with the prefix micro) and the label‐level performance (denoted with the prefix macro). Formally, y denotes the set of test (sample, label) pairs and $\overline{y}$ denotes the set of predicted (sample, label) pairs. The micro metrics are defined in Eqn (2).

$$\left\{\begin{array}{c}\text{micro precision}=\frac{|y\cap \overline{y}|}{\left|\overline{y}\right|}\\ \text{micro recall}=\frac{|y\cap \overline{y}|}{\left|y\right|}\\ \text{micro}\mathrm{F}1-\text{score}=\frac{2\times \text{micro}\_P\times \text{micro}\_R}{\text{micro}\_P+\text{micro}\_R}\\ \end{array}\right.$$ (2)

Furthermore, $y_l$ denotes the subset of y with label l, ${\overline{y}}_l$ denotes the subset of $\overline{y}$ with label l, and L denotes the label set. The macro metrics are defined in Eqn (3). ⁴²

$$\left\{\begin{array}{c}P\left(y_l,{\overline{y}}_l\right)=\frac{\left|y_l\cap {\overline{y}}_l\right|}{\left|{\overline{y}}_l\right|}\\ R\left(y_l,{\overline{y}}_l\right)=\frac{\left|y_l\cap {\overline{y}}_l\right|}{\left|y_l\right|}\\ \mathrm{F}1\left(y_l,{\overline{y}}_l\right)=\frac{2\times P\left(y_l,{\overline{y}}_l\right)\times R\left(y_l,{\overline{y}}_l\right)}{P\left(y_l,{\overline{y}}_l\right)+R\left(y_l,{\overline{y}}_l\right)}\\ \text{macro precision}=\frac{1}{\mid L\mid }{\sum }_{l\in L}P\left(y_l,{\overline{y}}_l\right)\\ \text{macro recall}=\frac{1}{\mid L\mid }{\sum }_{l\in L}R\left(y_l,{\overline{y}}_l\right)\\ \text{macro}\mathrm{F}1-\text{score}=\frac{1}{\mid L\mid }{\sum }_{l\in L}\mathrm{F}1\left(y_l,{\overline{y}}_l\right)\\ \end{array}\right.$$ (3)

Similar to binary and multi‐class learning, we assume an instance i in the instance set I to be correctly predicted when the Jaccard index between its true label set $L_i$ and its predicted label set $L_i^{\prime }$ exceeds a predefined threshold $\delta$. According to the definition, the multi‐label accuracy termed Jaccard accuracy is defined in Eqn (4). ⁴³

$$\left\{\begin{array}{c}\text{Jaccard}\left(L_i^{\prime },L_i\right)=\frac{\mid L_i\cap L_i^{\prime }\mid }{\mid L_i\cup L_i^{\prime }\mid }\\ \text{Jaccar accuracy}=\frac{\mid \left\{i|\text{Jaccard}\left(L_i^{\prime },L_i\right)\ge \delta ,i\in I\right\}\mid }{\mid I\mid }\\ \end{array}\right.$$ (4)

Commonly, the predefined threshold $\delta$ assumes 0.5. If $\delta =1$, the Jaccrad accuracy turns to be perfect match ratio, that is, the percentage of pesticides whose taxa‐specific MOAs perfectly match their predicted labels.

2.3.3. Exceptional metrics

Besides the metrics of label match, we additionally provide some metrics to measure the level of label mismatch, such as null_rate, null_hit_rate, exception_rate, and max_exceptions, as defined in Eqn (5).

$$\left\{\begin{array}{c}\text{null}\_\text{rate}=\frac{|\left\{i|\left|L_i^{\prime }\right|=0,i\in I\right\}\mid }{\mid I\mid }\\ \text{null}\_\mathrm{hit}\_\text{rate}=\frac{\left|\left\{i|\left|L_i\cap L_i^{\prime }\right|=0\wedge \left|L_i^{\prime }\right|\ne 0,i\in I\right\}\right|}{\left|I\right|}\\ \text{exception}\_\text{rate}=\frac{\left|\left\{i|\left|L_i^{\prime }-L_i\right|\ne 0,i\in I\right\}\right|}{\left|I\right|}\\ \max \_\text{exceptions}=\underset{i}{\max }\left(\left|L_i^{\prime }-L_i\right|\right)\\ k\_\text{exceptions}\_\text{ratio}=\frac{\left|\left\{i\left|L_i^{\prime }-L_i\right|=k,i\in I\right\}\right|}{\left|I\right|},k>0\\ \end{array}\right.$$ (5)

The metric null_rate measures the percentage of pesticides whose predicted label set is empty, null_hit_rate measures the percentage of pesticides whose predicted labels do not hit the test label set, exception_rate measures the percentage of pesticides whose predicted labels are partly or completely exceptional to the test labels, max_exceptions denotes the maximum predicted labels exceptional to the test labels, and q_exceptions_ratio denotes the ratio of instances with exceptionally predicted labels, the number of which is equal to q. In addition, Hamming loss is also used to measure the rate of label set mismatch, as defined in Eqn (6). ⁴⁴

$$\text{Hamming}\_\text{loss}=\frac{1}{\mid L\mid }\mid p\Delta y\mid$$ (6)

where P denotes the predicted label set, y denotes the true label set, and $\Delta$ denotes XOR‐like operator.

3. RESULTS AND DISCUSSION

3.1. Multi‐label performance of k‐fold cross‐validation

3.1.1. Per class performance under decision threshold $\zeta =0.5$

The meta binary classifiers of the multi‐label XGBoost model yield the decisions in the form of probability, the threshold $\zeta$ of which is set to 0.5 in usual cases. The label is set to 1 when the predicted probability is above $\zeta$, denoting the presence of the corresponding taxa‐specific MOA; otherwise, the label is set to 0, denoting the absence of the taxa‐specific MOA. For XGBoost, we adopt grid search to empirically determine the hyperparameters such as n_estimators (number of trees, ranging from 300 to 3000 at interval = 50), max_depth (maximum tree depth, ranging from 5 to 30 at interval = 5), and learning_rate (weight shrinkage coefficient, ranging from 0.01 to 0.3 at interval = 0.05). The optimum hyperparamters are finally set to {n_estimators = 2000, max_depth = 25, learning_rate = 0.1}. The performance of k‐fold cross‐validation (k = 10) for each class label, under the probability threshold $\zeta =0.5$, is illustrated in Fig. 2, and the recall metrics for each taxa‐specific MOAS at the threshold ζ = 0.1, 0.3 and 0.5 are provided in Table 1. For all the taxa‐specific MOAs, the metric accuracy seems to be the most encouraging among the five metrics. However, accuracy will be severely biased towards the majority class when the training data are highly imbalanced. Especially when the negative class overwhelmingly predominates the positive class in terms of data size, the accuracy may be very encouraging, but the impact of the positive class is nearly negligible to the model training, therefore we have to comprehensively consider the other metrics. As the metric recall indicates, the multi‐label XGBoost model exhibits encouraging performance on 10 taxa‐specific MOAs, whose recall rates range from 0.6034 (Aquatic‐Narcosis) to 0.9684 (Aquatic‐AChE inhibition). There are two extreme cases, that is, Acari‐Neurotoxicity, and Insect‐Respiration Inhibitors/Regulators, where the multi‐label XGBoost model performs the worst, with the recall rates lower to 0.0857 and 0.0645. The discouraging performance mostly results from their limited training data, such as 35 and 31 training pesticides for the two worst cases as shown in Fig. 1. On the other four MOAs, the multi‐label XGBoost model achieves comparatively moderate performance, such as Fungi‐Reactivity (recall = 0.4054), Fungi‐Biosynthesis inhibitors (recall = 0.5871), Fungi‐Respiration Inhibitors/Regulators (recall = 0.5172), and Aquatic‐Reactivity (recall = 0.4091). These results are to some extent acceptable from the perspective of the imbalanced training data. As further seen from MCC point of view, the multi‐label XGBoost model exhibits less‐biased performance on most taxa‐specific MOAs except Acari‐Neurotoxicity, Insect‐Respiration Inhibitors/Regulators, and Aquatic‐Reactivity, whose MCC metrics are all below 0.5. Apart from data imbalance, the large number of class labels also renders in silico MOA modelling more challenging. Nevertheless, the multi‐label XGBoost model still outperforms the baseline linear discriminant analysis (LDA) ²⁷ in spite of its discouraging performance on several minor taxa‐specific MOAs.

Figure 2.

Performance of 10‐fold cross‐validation for each taxa‐specific MOA. The marked MOAs are assumed to achieve relatively lower performance (recall <0.6) and/or biased performance (MCC < 0.5).

Table 1.

Performance comparisons in terms of recall rate between multi‐label XGBoost and multi‐class LDA ²⁷

Taxa‐specific MOA class	Multi‐label XGBoost						Multi‐class LDA
	Cross‐validation (ζ)			Hold‐out (ζ)			Hold‐out
	0.5	0.3	0.1	0.5	0.3	0.1
Insect‐Neurotoxicity	0.8341	0.8756	0.9217	0.8925	0.9032	0.9247	0.98
Fungi‐Reactivity	0.4054	0.4595	0.5676	0.5714	0.5714	0.5714	0.33
Plant‐Growth Regulator	0.6379	0.6782	0.7644	0.6765	0.6765	0.7941	0.86
Fungi‐Biosynthesis inhibitors	0.5871	0.6323	0.7161	0.5161	0.5484	0.6774	0.59
Fungi‐Respiration Inhibitors/Regulators	0.5172	0.5862	0.7069	0.4545	0.5455	0.6364	0.40
Plant‐Biosynthesis inhibitors	0.7639	0.7897	0.8712	0.9130	0.9565	0.9565	0.90
Rodent‐Anticoagulant	0.8235	0.9412	0.9412	0.5000	0.5000	0.5000	1.00
Insect‐Respiration Inhibitors/Regulators	0.0645	0.0968	0.1935	0.0000	0.0000	0.0000	0.00
Plant‐Photosynthesis	0.7786	0.8550	0.8931	0.8462	1.0000	1.0000	1.00
Aquatic‐Neurotoxicity	0.8705	0.8808	0.9016
Aquatic‐AChE inhibition	0.9684	0.9719	0.9860
Aquatic‐Iono/Osmoregulatory/Circulatory impairment	0.7736	0.7925	0.8679
Aquatic‐Narcosis	0.6034	0.6573	0.7608
Aquatic‐Reactivity	0.4091	0.4727	0.5455
Acari‐Neurotoxicity	0.0857	0.1429	0.1714
Aquatic‐Electron transport inhibition	0.6818	0.6932	0.7614

3.1.2. Overall multi‐label performance with varying decision threshold $\zeta$

To comprehensively estimate the performance of the proposed multi‐label XGBoost model, we provide the label‐match metrics, such as macro‐ and micro‐ (precision, recall, and F1‐score) as defined in Eqns (2) and (3), and meanwhile provide the label‐mismatch metrics, such as null_rate, null_hit_rate, exception_rate, and max_exceptions, as defined in Eqn (5). We explicitly consider whether the proposed multi‐label XGBoost model is robust when the decision probability threshold $\zeta$ varies within a specific range. Too high a threshold $\zeta$ probably yields empty predicted label sets, while too low a threshold $\zeta$ will yield irrelevant labels. For this sake, we particularly investigated the impact of threshold $\zeta$ to achieve a proper trade‐off among the metrics. A rational guideline of choosing the threshold $\zeta$ is supposed to increase the recall rates of all the taxa‐specific MOAs, and meanwhile to accept a reasonable number of irrelevant labels, which may be not false positive but unreported. The overall performance with ζ ranging within [0.1, 0.5] at 0.1 intervals is shown in Fig. 3. With ζ decreasing from 0.5 to 0.1, the macro recalls increase from 0.6128 to 0.7231 (see Fig. 3(A)), the micro recalls increase from 0.7118 to 0.8147 (see Fig. 3(B)), and the Jaccard accuracies increase from 0.7348 to 0.8005 (see Fig. 3(C)). The macro and micro recalls both increase with the threshold ζ decreasing. Undoubtedly, a lower threshold ζ facilitates recognizing more taxa‐specific MOAs and reducing the risk of higher null rates. As shown in Fig. 3(D), the null rates decrease from 0.1342 to 0.0334 with ζ decreasing from 0.5 to 0.1. However, we cannot arbitrarily lower the threshold ζ, because more irrelevant labels (taxa‐specific MOAs) will be recommended by the model, and the metric precision will decrease accordingly. For instance, the macro precisions decrease from 0.8097 to 0.6681 with ζ decreasing from 0.5 to 0.1 (see Fig. 3(A)), and the micro precisions decrease from 0.8467 to 0.7158 (see Fig. 3(B)).

Figure 3.

Overall multi‐label learning performance with decision probability threshold ζ = 0.1, 0.2, 0.3, 0.4, and 0.5. (A) Averaging the binary performance at label levels yields the macro metrics, including macro precision, macro recall, and macro F1 score. (B) Averaging the binary performance at instance levels yields the micro metrics, including micro precision, micro recall, and micro F1 score. (C) Jaccard accuracy and perfect match ratio measure the overlap between predicted label set and true label set. (D) Metrics (null_rate, null_hit_rate and exception_rate) measure how predicted labels deviate from the test labels. (E) The ratio of instances, whose exceptionally predicted labels are total up to q (q_exceptions_ratio). (F) Hamming loss measuring the rate of label mismatch.

Next, we further study how a lower threshold $\zeta$ affects the predictive robustness, from which to determine the acceptable threshold $\zeta$. As shown in Fig. 3(E), when the threshold ζ drops to 0.1, the number of irrelevant or exceptional taxa‐specific MOAs does not exceed 3, the 3_exceptions_ratio (maximum 3 irrelevant taxa‐specific MOAs) is lower to 0.43%, and the 2_exceptions_ratio is lower to 6.25%. These results show that the lower threshold ζ = 0.1 introduces 2 (3) irrelevant taxa‐specific MOAs to 6.25% (0.43%) of the pesticides, indicating that ζ = 0.1 takes a very low risk in recommending $\ge 2$ irrelevant taxa‐specific MOAs. As shown in Fig. 3(E), ζ = 0.1 at most yields one irrelevant taxa‐specific MOA to 29.76% of the pesticides. More importantly, lowering ζ to 0.1 only results in <%5 drop of perfect match ratio as compared to ζ = 0.2 (see Fig. 3(C)), and surprisingly, the Jaccard accuracy at ζ = 0.1 does not drop due to irrelevant labels, but reversely surges up to about 80%, indicating that the low threshold ζ = 0.1 yields more favorable performance gain from relevant labels than adverse impact from irrelevant labels. The results show that the threshold ζ = 0.1 is safe and acceptable in terms of recognizing more relevant labels and recommending a reasonable number of irrelevant labels. Meanwhile, the wide range of acceptable thresholds ζ indicates a large variance between the positive and the negative labels, and predictive robustness of the proposed multi‐label XGBoost model. If we need a higher confidence level of predictions, the threshold ζ = 0.3 is also worth considering because the number of irrelevant labels does not exceed 2 and the Hamming loss is very low (see Fig. 3(F)).

3.1.3. Applicability domain

QSAR modelling of chemical molecules requires that the test molecules are structurally similar to the training molecules. Beyond the similarity threshold or range, the predictions would become less reliable. To overcome the diversity of chemical molecules, applicability domain (AD) is usually required for a specific trained QSAR model. Here, we simultaneously estimate the applicability domain during the process of multi‐label performance estimation via k‐fold cross‐validation (k = 10). For each fold of training, we estimate the minimum MACCSKeys similarity and the Morgan similarity between the test set and the training set. For the k‐fold cross‐validation, the minimum MACCSKeys similarity ranges from 0.5 to 0.709, and the minimum Morgan similarity ranges from 0.177 to 0.345. In order to ensure, to the largest extent, that the proposed multi‐label XGBoost model to achieves the reported multi‐label performance, the minimum 0.5 MACCSKeys similarity and minimum 0.177 Morgan similarity are proposed as the applicability domain.

3.2. External test on herbicides

After deployment, the multi‐label XGBoost model, trained on the whole training data, has to be further evaluated using independent or external test data, which contains 47 herbicides exhibiting biosynthesis inhibition and photosynthesis inhibition in this work. The results show that 84.62% of the pesticides exhibiting photosynthesis inhibition are correctly recognized for all the ζ thresholds within the range [0.1, 0.5] at 0.1 intervals, suggesting that the predictions of photosynthesis inhibition are robust, so that steeply lowering the threshold ζ does not solicit bias towards the class label concerned. For biosynthesis inhibition, the recall rate remains 47.62% when ζ drops from 0.5 to 0.2. When the threshold ζ is lowered to 0.1, which is acceptable as revealed by k‐fold cross‐validation (k = 10), the recall rate steeply increases to 61.9%. The external test shows that the multi‐label XGBoost model generalizes well to herbicides. We explicitly study the off‐target effects of these herbicides on other organisms such as insect, fungi, and aquatic species. The predicted taxa‐specific MOAs for these herbicides at threshold ζ = 0.5ζ are provided in Supporting Information, File S1. Taking the herbicide phenisopham (CAS: 57375–63‐0) for example, the carbamate ester is predicted to exhibit multiple taxa‐specific MOAs including Plant‐Photosynthesis, Aquatic‐AChE inhibition, Aquatic‐Narcosis and Insect‐Neurotoxicity. Long‐term exposure to phenisopham has been confirmed by experimental evidence to cause toxicity to aquatic crustacean, rotifers, and algae. ⁴⁵ As compared to the targeted organism weeds, the herbicide potentially exhibits a lower or chronic toxicity to these off‐target species.

3.3. XGBoost feature importance

Feature importance analyses have been widely used in QSAR‐based toxicity regression models ⁴⁶ , ⁴⁷ in which molecular fragments are prioritized in terms of their contribution to pesticide toxicity. In tree models, such as decision tree, random forest, and XGBoost, the features of importance are often the critical nodes that diverge the instances into different paths. In the XGBoost model, the importance of a feature is weighted using the number of times that the feature is selected to split trees. The top 20 Morgan fragments contributing most to the model performance are shown in Fig. 4(A) and their abstract representations are provided in Fig. 4(B). Among the 2048‐bit Morgan fragments, bit 1750 is assigned the highest weight, indicating that the corresponding substructure plays a critical role in discriminating the 16 taxa‐specific MOAs. These important substructures or fragments are potentially supplements to the Verhaar scheme. ²

Figure 4.

XGBoost feature importance in discriminating taxa‐specific MOAs. (A) Weights of the top 20 Morgan fragments. (B) Two‐dimensional structures of Morgan fragments.

3.4. Comparisons with existing methods

Existing in silico baseline (narcosis) toxicity modelling ¹⁶ , ¹⁸ , ²⁰ , ²¹ , ²² , ²³ focuses on quantitative dose–response regression, only covering nonpolar narcosis, polar narcosis, reactivity, and specific MOAs. Existing in silico qualitative models, ²⁶ , ²⁷ , ²⁸ , ³¹ , ³² based on structural representation of pesticides only, cover more specific MOAs. In these models, the multi‐class LDA model ²⁷ most resembles the proposed multi‐label XGBoost model in terms of organism coverage across trophic levels and thus is used as baseline for performance comparison. The LDA model covers nine taxa‐specific MOAs without aquatic species and treats MOA modelling as classical multi‐class learning. Comparatively, the proposed multi‐label XGBoost model merges the MOATox scheme and the MOAtaxa scheme, greatly expanding the training data, and the coverage of MOAs and species.

The baseline multi‐class LDA only reports its hold‐out test performance without conducting k‐fold cross‐validation. For comparison, we evaluated the proposed multi‐label model on the same test set. However, the hold‐out training data used by the baseline multi‐class LDA does not contain multi‐label instances and thus is not applicable to the proposed multi‐label XGBoost. For this reason, we adopted the multi‐label training data merged from the MOATox scheme and the MOAtaxa scheme, except that the pesticides used as test data were removed from the training data. As shown in Table 1, the multi‐label XGBoost model on the whole outperforms the baseline multi‐class LDA model on most taxa‐specific MOAs, except several small‐sized taxa‐specific MOAs, such as Insect‐Neurotoxicity (93 pesticides), Plant‐Growth Regulator (35 pesticides), and Rodent‐Anticoagulant (2 pesticides). Notably, on the taxa‐specific MOAs where the baseline multi‐class LDA model shows poor recall rates, the proposed multi‐label model demonstrated encouraging performance, such as 0.5714 versus 0.33 recall on Fungi‐Reactivity, 0.7161 versus 0.59 recall on Fungi‐Biosynthesis inhibitors, 0.7069 versus 0.4 recall on Fungi‐Respiration Inhibitors/Regulators, and 0.1935 versus 0 recall on Insect‐Respiration Inhibitors/Regulators. On the other taxa‐specific MOAs, the two models demonstrate equivalent performance.

Although molecular fragments have demonstrated advantages of capturing local chemical bonds and functional groups, such as where the carboxyl group is located, molecular descriptors (e.g., molecular volume, polarizability) can globally capture the long‐range effects spanning the molecular skeleton, such as directly quantifying the electron cloud distribution and spatial conformation of the entire molecule. Current molecular descriptors, such as Mordred, ⁴⁸ are commonly high‐dimensional and complicated to compute, so that some physiochemical properties are not computable, rendering many null values. Recently, a supervised dimensionality reduction technique termed ARKA was proposed to globally generalize commonly used molecular descriptors into two‐dimensional vectors, that is, (ARKA_1, ARKA_2). Such an extremely low‐dimensional representation of chemical molecules generalizes remarkably well to activity cliffs detection when applied to the nonlinear classifiers such as random forest and support vector machine. Interested readers are referred to Banarjee and Roy ⁴⁹ for algorithmic details. Here, we also compare the performance of different feature engineering strategies, especially molecular fingerprints versus the latest molecular descriptors ARKA. The 16 taxa‐specific MOAs can be viewed as 16 activity cliffs, using 1 to represent active (or exhibiting the MOA) and using 0 to represent inactive (or exhibiting no such MOA). For each taxa‐specific MOA, a two‐dimensional vector (ARKA_1, ARKA_2) is computed for a pesticide so that each pesticide is represented by a 32‐dimensional feature vector in the space of 16 taxa‐specific MOAs, which are further fed to the proposed multi‐label XGBoost to report the multi‐label performance metrics, as provided in the Supporting Information, Fig. S1. The extremely low‐dimensional representation ARKA still achieves encouraging 0.6573 macro precision and 0.7612 micro precision, but the recall rates are discouraging, for example micro recall lowered to 0.546, potentially due to its limited representative capacity of chemical molecules in a large label space. For ARKA, which has been proved effective in binary classification, it is worth further considering how to properly extend it to a large label space.

3.5. Case studies on aquatic species Pimephales promelas

The multi‐label XGBoost model, trained on the entire training data of 1855 pesticides and 16 taxa‐specific MOAs, is further used to predict the MOAs of the 786 case‐study pesticides, whose dose–response measurements on the aquatic species fathead minnow (Pimephales promelas) are available. The experimental LC₅₀ values (mg L⁻¹, 96 h) for these pesticides are taken from Sheffield and Judson. ¹⁰ Among the 786 pesticides, 226 pesticides are predicted to exhibit more than one taxa‐specific MOA (see Supporting Information, File S1).

3.5.1. Relationships between toxic severities and aquatic MOAs

As shown in Fig. 5, the pesticides predicted to exhibit the six aquatic MOAs are concatenatedly visualized into six groups in an integral Cartesian coordinates system, and within each group the pesticides are sorted ascendingly by their logarithm of LC₅₀ values. For credibility, only the pesticides predicted by the thresholdζ ≥ 0.9 were chosen for further analyses. The criteria for determining the toxic severities of pesticides on aquatic organisms are taken from Liu et al., ⁵⁰ that is, (i) exposure $\le$0.1 ppm (−1.0 log mg L⁻¹) is defined as very highly toxic; (ii) exposure between 0.1 and 1.0 ppm (−1.0–0 log mg L⁻¹) is defined as highly toxic; (iii) exposure between 1.0 and 10 ppm (0–1 log mg L⁻¹) is defined as moderately toxic; and (iv) exposure ≥10 ppm (≥1 log mg L⁻¹) is defined as slightly toxic or non‐toxic. As shown in Fig. 5, many more pesticides are predicted to exhibit narcosis toxicity. From the perspective of logLC₅₀ range, the MOA termed iono/osmoregulatory/circulatory impairment ranks the widest, with logLC₅₀ ranging from −3.39 to 4.22. Narcosis spans the second widest range of toxicities, with logLC₅₀ ranging from −2.70 to 4.0. According to the criterion of low toxicity ≥10 ppm (≥1 log mg L⁻¹), ≥50% of the pesticides that are predicted to narcosis exhibit very low toxicity, conforming to the baseline toxicity of narcosis. The quantitative LC₅₀ values (mg L⁻¹, 96 h) reversely validate the credibility of the predictions. Furthermore, when the pesticides are predicted to exhibit narcosis by the proposed multi‐label XGBoost model, their logK _ow or logK _lipw values could be used to extend the training data for in silico baseline toxicity modelling. The other pesticides, which are predicted to exhibit narcosis but are indicated by the LC₅₀ values to be moderately or highly toxic, potentially exhibit narcosis as accessory effect of some other specifically‐acting MOAs. In such cases, we resort to baseline toxicity model and toxic ratio (TR) to discriminate the narcosis from specifically acting MOAs in our future work (see Discussion). In a sense, MOAs like narcosis could be deemed as broad‐spectrum MOAs if they are found to span a wide range of toxicity values. Iono/osmoregulatory/circulatory impairment, defined as a specifically acting MOA, ³³ also spans a wide range of toxicity values, but it mostly does not exhibit baseline toxicity, and in some cases exhibits very highly toxicity (see Fig. 5), thus it is not proper to deem it as an accessory MOA like narcosis.

Figure 5.

Relationship between pesticide toxic severities and aquatic MOAs (ζ $\ge 0.9$).

The other specifically acting MOAs, such as neurotoxicity, AChE inhibition and electron transport inhibition, together with the reactivity MOA, span a much narrower range of LC₅₀ values, which could be comparatively deemed as specific‐spectrum MOAs. In particular, the MOAs neurotoxicity and electron transport inhibition exhibit a much higher risk of toxicity. These two MOAs' maximum logLC₅₀ values are as low as 2.0 and 2.26 and their minimum logLC₅₀ values are as low as −1.30 and −1.52 respectively. On average, the two aquatic MOAs (neurotoxicity, electron transport inhibition) exhibit a higher risk to aquatic species. Comparatively, the two MOAs AChE inhibition and reactivity can be induced by both low toxicity and highly toxic pesticides, but mostly are moderate or slightly toxic as indicated by the line logLC₅₀ = 0 (LC₅₀ = 1 mg L⁻¹) in Fig. 5. These results show that quantitative dose–response measurements and qualitative MOA are two facets that are mutually beneficial to enlarging the training data for quantitative baseline toxicity modelling. Importantly, the relationships between toxic severities and allow for the quantitative characterization toxicity profiles at the MOA level.

The toxicity profile of a MOA and the definition of broad‐/specific‐spectrum MOAs are dependent on the chosen pesticides, the availability of dose–response measurements, and the predicted MOAs. To ensure the quality of predictions, we further manually validated the credibility of the predicted aquatic MOAs via a QSAR strategy. For each predicted aquatic MOA, we chose one highly toxic and one less toxic pesticide for case studies. For each case‐study pesticide, we chose a control pesticide whose credible MOAs are available. From QSAR point of view, the chosen control pesticide should be structurally similar to the case‐study pesticide. Here, the structural similarities between pesticides are measured by the Tanimoto coefficient. The case‐study pesticides and chosen control pesticides are shown in Table 2. As can be seen from Table 2, the predicted MOAs of the case‐study pesticides are all consistent with the credible MOAs of the control pesticides, and for all the MOAs, the case‐study pesticides show high or moderately high structural similarities with the control pesticide. Comparatively, the case‐study pesticides with higher toxicity, as shown in the left‐hand column in Table 2, commonly exhibit higher structural similarities with their control pesticides than the case‐study pesticides with lower toxicity, shown in the right‐hand column in Table 2.

Table 2.

Case studies on aquatic species Pimephales promelas

MOAs	High risk (low logLC50)		Low risk (high logLC50)
	Structure (case study vs. control)	Toxicity and similarity	Structure (case study vs. control)	Toxicity and similarity
Aquatic‐Neurotoxicity		*−1.3 (0.5, 0.9)		*2.0 (0.21, 0.41)
Aquatic‐AChE inhibition		*−0.6 (0.69, 0.9)		*3.85 (0.44, 0.48)
Aquatic‐Narcosis		*−2.70 (0.56, 0.7)		*4.0 (0.34, 0.42)
Aquatic‐Reactivity		*−1.0 (0.38, 0.8)		*3.51 (0.25, 0.29)
Aquatic‐Electron transport inhibition		*−1.52 (0.1, 0.64)		*2.26 (1.0, 1.0)

For high risk or low risk, the pesticide on the left is used for case study and the pesticide on the right is used as control. The asterisked decimals denote the logLC₅₀ of the case‐study pesticides. The bracketed decimals denote the Tanimoto coefficient of Morgan and MACCSKeys fingerprints.

Taking aquatic neurotoxicity for the first example, the case‐study pesticide (CAS: 3547‐04‐4), chemically named 1‐chloro‐4‐[1‐(4‐chlorophenyl)ethyl]benzene and highly toxic with logLC₅₀ equal to −1.3, is a metabolite and environmental degradation product of the insecticide DDT. The pesticide is predicted by the proposed multi‐label XGBoost model to exhibit neurotoxicity. Its control pesticide (CAS: 72–54‐8), chemically named 1,1‐bis(4‐chlorophenyl)‐2,2‐dichloroethane, is also a metabolite of the organochlorine insecticide DDT. Between these two pesticides, the Morgan similarity is 0.5 and the MACCSKeys similarity is up to 0.9. The control pesticide (CAS: 72–54‐8) is known to exhibit neurotoxicity, consistent with the predicted MOA of the case‐study pesticide (CAS: 3547‐04‐4), which also complies with the QSAR assumption that structurally similar pesticides tend to exhibit similar bioactivities. The case‐study pesticide (CAS: 7785‐70‐8), chemically named (1R,5R)‐2,6,6‐trimethylbicyclo[3.1.1]hept‐2‐ene with low toxicity and logLC₅₀ equal to 2.0, acts as a plant or human metabolite. Its control pesticide (CAS: 5989‐27‐5), chemically named (4R)‐limonene, is an optically active form of limonene having (4R) configuration and is known to exhibit the mode of neurotoxic action. As compared with the case study (CAS: 3547‐04‐4), these two pesticides show a much lower structural similarity (Morgan similarity = 0.21, MACCSKeys similarity = 0.41). Take the aquatic AChE inhibition for the second example, the highly toxic case‐study pesticide (CAS: 311–45‐5, logLC₅₀ = −0.6), chemically named diethyl 4‐nitrophenyl phosphate, is highly similar to its control pesticide (CAS: 56–38‐2), named diethyl 4‐nitrophenyl phosphorothionate (CAS: 56–38‐2), with Morgan similarity equal to 0.69 and MACCSKeys similarity equal to 0.9. The low‐risk case‐study pesticide (CAS: 512–56‐1, logLC₅₀ = 3.85) named O,O,O‐trimethyl phosphate (CAS: 512–56‐1, logLC₅₀ = 3.85) is relatively less similar to its control pesticide (CAS: 10265–92‐6) named O,S‐dimethyl phosphoramidothioate, with Morgan similarity equal to 0.44 and MACCSKeys similarity equal to 0.48.

Non‐specific narcosis is commonly induced by low‐toxic pesticides, but in many cases, highly toxic pesticides also exhibit or are predicted to exhibit narcosis. For instance, the case‐study pesticide (CAS: 32598–13‐3) exhibits very high toxicity (logLC₅₀ = −2.70). In this case, the predicted narcosis is potentially induced as an accessory MOA by the pesticide. In this sense, the predicted narcosis, when not accompanied by other reactivity or specifically acting MOAs, needs to be cautiously considered. If the pesticides concerned exhibit high toxicity, the predicted narcosis is to a large extent accompanied by other MOAs. In view of the accessory or broad‐spectrum characteristic of narcosis, the predicted reactivity and specifically acting MOAs are potentially more credible. As shown in Table 2, to exhibit the same MOAs, a highly toxic case‐study pesticide tends to be more structurally similar to its control pesticide, whereas a less toxic case‐study pesticide tends to be less structurally similar to its control pesticide. As an exception, electron transport inhibition seems not to show the tendency. For instance, the case‐study pesticide (CAS: 7446‐08‐4, logLC₅₀ = −1.52) is highly toxic, but it reversely exhibits a lower structural similarity with its control pesticide (CAS: 7784‐46‐5) (Morgan similarity = 0.1, MACCSKeys similarity = 0.64). The case‐study pesticide (CAS: 1326‐82‐5, logLC₅₀ = 2.26) is less toxic, but it exhibits a high structural similarity with its control pesticide (CAS: 51–28‐5) (Morgan similarity = 1, MACCSKeys similarity = 1) and lower toxicity (see Table 2). The relationships between toxic severities and structural similarities, if not trivial, facilitate the study of pesticides sharing the same MOAs.

3.5.2. Similarities between aquatic MOAs

Similarities between MOAs are beneficial to unravelling the toxic effects of pesticide mixtures. Pesticides with similar MOAs are assumed to induce a synergistic mixture effect, while dissimilar MOAs are assumed to induce antagonist or additive mixture effects. ⁵¹ Unfortunately, the similarities between MOAs cannot be directly derived in most cases. In this work, we manage to estimate MOA similarities by the distribution of structural similarities between pesticides belonging to the two MOAs. All the pairwise similarities of pesticides are sorted ascendingly within each pair of MOAs (in total six aquatic MOAs) and are concatenated in an integral Cartesian coordinate system (see Fig. 6(A) for Morgan similarities and Fig. 6(B) for MACCSKeys similarities). For the six aquatic MOAs, there are 15 pairs of MOAs. Focusing on the peaks with Morgan similarity $\ge$0.4, we obtained the relatively more similar pairs of MOAs including Pair (3) (neurotoxicity, narcosis), Pair (7) (AChE inhibition, narcosis), Pair (10) (iono/osmoregulatory/circulatory impairment, narcosis), Pair (12) (iono/osmoregulatory/circulatory impairment, electron transport inhibition), Pair (13) (narcosis, reactivity), Pair (14) (electron transport inhibition, narcosis), and Pair (15) (reactivity, electron transport inhibition) (see Fig. 6(A)). Among these peak pairs, all the pairs are associated with narcosis except Pair (15), and narcosis is shown to be associated with all the other MOAs. The MACCSKeys similarities, when set to be above 0.6, also demonstrate the association of narcosis with all the other MOAs (see Fig. 6(B)). These results show that narcosis is potentially not only an independent baseline toxic effect induced by low‐toxic pesticides, but also is probably an accessory toxic effect accompanied by the other specific MOAs, which are often induced by highly ‐toxic pesticides.

Figure 6.

Similarities between predicted aquatic MOAs (ζ $\ge \mathbf{0.9}$). (A) Morgan similarities between predicted aquatic MOAs. (B) MACCSKeys similarities between predicted aquatic MOAs. I: Aquatic‐Neurotoxicity, II: Aquatic‐AChE inhibition, III: Aquatic‐Iono/Osmoregulatory/Circulatory impairment, IV: Aquatic‐Narcosis, V: Aquatic‐Reactivity, VI: Aquatic‐Electron transport inhibition.

The pesticide pairs exhibiting peak Morgan similarity are provided in Table 3. We can see that all the other aquatic MOAs tend to associate with narcosis through their low‐toxic pesticides, for example neurotoxicity (CAS: 1770‐80‐5, logLC₅₀ = 2.0), AChE inhibition (CAS: 54261–67‐5, logLC₅₀ = 1.46), iono/osmoregulatory/circulatory impairment (CAS: 54–21‐7, logLC₅₀ = 3.14), reactivity (CAS: 121–14‐2, logLC₅₀ = 1.39), and electron transport inhibition (CAS: 1326‐82‐5, logLC₅₀ = 2.26). These pesticides are structurally similar to the pesticides predicted to induce narcosis. These indirect associations potentially indicate that narcosis, when exhibiting as an accessory toxic effect, also can be induced by low toxicity pesticides, which potentially induce the major reactivity or specifically acting MOAs simultaneously. Combining the indication that highly toxic pesticides tend to induce more specific MOAs together with narcosis as an accessory toxic, as illustrated in Fig. 5, we could conclude that reactivity or specifically acting MOAs, induced either by highly toxic or low toxicity pesticides, potentially induce narcosis toxic effects simultaneously.

Table 3.

Similarities between predicted taxa‐specific MOAs

Aquatic‐Neurotoxicity vs. Aquatic‐Narcosis (0.39) [2.0, 0.65]	Aquatic‐AChE inhibition vs. Aquatic‐Narcosis (0.45) [1.46, −0.49]	Aquatic‐Iono/Osmoregulatory impairment vs. Aquatic‐Narcosis (0.56) [3.14, 2.0]
Aquatic‐Iono/Osmoregulatory impairment vs. Aquatic‐Electron transport inhibition (0.75) [−3.4, −1.39]	Aquatic‐Reactivity vs. Aquatic‐Narcosis (0.62) [1.39, 0.53]	Aquatic‐Electron transport inhibition vs. Aquatic‐Narcosis (0.44) [2.26, 1.94]
Aquatic‐Reactivity vs. Aquatic‐Electron transport inhibition (0.59) [1.39, 1.9]	Aquatic‐Narcosis vs. Insect‐Neurotoxicity (0.37) [1.34, −0.49]	Aquatic‐Narcosis vs. Plant‐Growth Regulator (0.28) [−0.8, 2.42]
Aquatic‐Reactivity vs. Insect‐Neurotoxicity (0.35) [0.82, −1.3]	Aquatic‐Neurotoxicity vs. Insect‐Neurotoxicity [−1.3]	Aquatic‐AChE inhibition vs. Insect‐Neurotoxicity [2.68]
Aquatic‐Narcosis vs. Plant‐Photosynthesis [2.0]

The bracketed decimals denote Morgan fingerprint similarity, and the square bracketed decimals denote logLC₅₀ values.

3.5.3. Similarities between aquatic MOAs and non‐aquatic MOAs

We further conducted analyses on MOA similarities at different trophic levels, especially the similarities between aquatic MOAs and non‐aquatic MOAs. As shown in Fig. 7, when the Morgan similarity threshold is set to 0.25, we obtained six relatively more similar pairs of MOAs, whose pesticides together with molecular structures and logLC₅₀ values are provided in Table 3. As shown in Fig. 7, among the MOA pairs, there are three pairs of MOAs that are significantly associated via common pesticides, that is, Pair (1) Aquatic‐Neurotoxicity versus Insect‐Neurotoxicity, Pair (2) Aquatic‐AChE inhibition versus Insect‐Neurotoxicity, and Pair (5) Aquatic‐Narcosis versus Plant‐Photosynthesis. As show in Table 3, the highly toxic pesticide (CAS: 3547‐04‐4, logLC₅₀ = −1.3), chemically named 1‐chloro‐4‐[1‐(4‐chlorophenyl)ethyl]benzene, is predicted to exhibit neurotoxicity on both insect and aquatic species. The pesticide (CAS: 1445‐75‐6, logLC₅₀ = 2.68), chemically named 2‐[methyl(propan‐2‐yloxy)phosphoryl]oxypropane, is predicted to exhibit AChE inhibition to aquatic and meanwhile to exhibit neurotoxicity to insect. The pesticide (CAS: 8027‐85‐8, logLC₅₀ = 2.0), chemically named 5‐bromo‐3‐butan‐2‐yl‐6‐methyl‐1H‐pyrimidine‐2,4‐dione, is predicted to induce aquatic narcosis and plant photosynthesis. The other three pairs of MOAs exhibiting high similarities, that is, Pairs (3), (4) and (6), are associated via two structurally similar pesticides. Among the three MOA pairs, two MOA pairs are associated with narcosis. As show in Table 3, Insect‐Neurotoxicity is associated with Aquatic‐Narcosis via its high‐toxic pesticide (CAS: 3309‐68‐0, logLC₅₀ = −0.49), and Plant‐Growth Regulator is associated with Aquatic‐Narcosis via its low‐toxic pesticide (CAS: 2008‐39‐1, logLC₅₀ = 2.42). Two associated MOAs are potentially bridged through the common chemical fragments of two pesticides, whose distinct fragments potentially drive the pesticides to exhibit different toxic effects. Molecular docking techniques promise to unravel the mechanism of how the common and distinct fragments of the two pesticides exhibit associated but distinct modes of toxic action.

Figure 7.

Morgan similarities between predicted aquatic MOAs and non‐aquatic MOAs (ζ $\ge 0.9$).

3.6. Discussion

QSAR‐based pesticide toxicity regression is an important research topic of ecotoxicological studies. Nowadays, we are more interested in how marine and terrestrial species respond to lethal doses of pesticides at the level of molecular events and pathways (mechanisms of toxic action, MechoAs), or at the level of toxic effects (modes of toxic action, MOAs). Starting from pioneering work, ² in silico modelling of pesticide MOAs evolves along two directions. One direction derives baseline (narcosis) toxicity regression models using experimental dose–response measurements, termed quantitative models, on which to further predict polar narcosis, reactivity, and more specific MOAs. These models are dependent on dose–response toxicity data and are hard to extend within specific MOAs. The other direction is to derive classification models based on chemical structures alone, termed qualitative models. These models are easy to build with less demanding data constraints, and are convenient to extend to model specific MOAs. However, these baseline toxicity models are primarily species‐specific and thus are not easy to extend to other species, especially across trophic levels.

Current in silico mechanistic modelling on pesticide toxicity focuses on MOAs, MechoAs, and AOPs. A pesticide potentially induces different toxicity responses from species, especially at different tropic levels, for instance the pesticide metribuzine (CAS: 21087‐64‐9) incurs impairment to plant photosynthesis and induces narcosis to aquatic species. Comprehensively investigating the toxic effects of pesticides on target and off‐target organisms presents a major challenge to in silico MOA modelling. From machine learning point of view, in silico modelling of this ecotoxicological problem is a typical multi‐label learning scenario. However, existing models commonly adopt multi‐class learning or independent binary learning to solve this problem.

In this work, we provide a solution to the challenging problem of MOA modelling to address three major concerns. First, a multi‐label XGBoost model is trained to assign potentially more than one taxa‐specific MOA to a pesticide; second, we attempt to unravel the implicative relationship between toxic severities and MOAs to facilitate baseline toxicity modelling; and, lastly, we attempt to unravel the similarities between the taxa‐specific MOAs via pesticide distribution. The results of k‐fold cross‐validation (k = 10) and the external test on herbicides show that the proposed multi‐label XGBoost model covering 16 taxa‐specific MOAs achieves encouraging multi‐label learning performance, outperforming or performing equivalently to existing in silico models. We further conducted several case studies on the aquatic species Pimephales promelas to validate the credibility of the predicted MOAs. Based on the distribution of logLC₅₀ values, the predicted MOAs could be divided into broad‐spectrum MOAs spanning a wide range of LC₅₀ values and specific‐ spectrum MOAs spanning a narrow range of LC₅₀ values. The similarities between MOAs show that Aquatic‐Narcosis is a potentially low‐toxic responsive symptom accompanied by the other specific MOAs. These findings are generally qualitative, and the quantitative relationship between toxicity and MOAs should result in more accurate baseline toxicity regression and successive quantitative modelling for specific MOAs.

4. CONCLUSION

In this work, the QSAR‐based multi‐label XGBoost model was proved to be effective in predicting taxa‐specific MOAs at different trophic levels via k‐fold cross‐validation, external tests, and hold‐out tests. The case studies on the aquatic species Pimephales promelas show that the multi‐label XGBoost potentially benefits deriving baseline (narcosis) toxicity models for all the species that the 16 taxa‐specific MOAs involve. Computational results show that narcosis could be an independent toxic effect or potentially an accessory toxic effect accompanied by reactive and other specific MOAs.

CONFLICT OF INTEREST

The authors declare that no conflict of interests.

AUTHOR CONTRIBUTIONS

MS conducted the study and wrote the paper.

Supporting information

Figure S1. The multi‐label performance metrics for ARKA + multi‐label XGBoost.

File S1. (1) The whole training data used for k‐fold cross validation (k = 10), (2) the external test set (herbicides) and their predicted MOAs (biosynthesis and photosynthesis inhibition), (3) the case‐study pesticides and their predicted MOAs on the aquatic species Pimephales promelas, (4) MACCSKeys and Morgan fingerprint representation plus label vector for the training data, (5) MACCSKeys and Morgan fingerprint representation plus label vector for the external test data, and (6) MACCSKeys and Morgan fingerprint representation plus label vector for the case‐study data (XLSX).

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.