Distribution of toxicity values across different species and modes of action of pesticides from PESTIMEP and PPDB databases

Pesticides classified by mode of action as a source of predictive models of toxicity.

Abstract

The continuous use of compounds contained in commodities such as processed food, medicines, and pesticides, demands safety measures, in particular, for those in direct contact with humans and the environment. Safety measures have evolved and regulations are now in place around the globe. In the case of pesticides, attempts have been made to use toxicological data to inform of potentially harmful compounds either across species, on different routes of exposure, or entirely new chemicals. The generation of models, based on statistical and molecular modeling studies, allows for such predictions. However, the use of these models is framed by the available data, the experimental errors, the complexity of the measurement, and the available computational algorithms, among other factors. In this work, we present the methodologies used for extrapolation across different species and routes of administration and show the appropriateness of developing predictive models of pesticides based on their type and mode of action. The analyses include comparisons based on structural characteristics and physicochemical properties. Whenever possible, the scope and limitations of the methodologies are discussed. We expect that this work will serve as a useful introductory guide of the tools employed in the toxicity assessment of agrochemical compounds.

Introduction

In 2006, the European Union created the regulation for the Registration, Evaluation, Authorization and Restriction of Chemical Substances (“REACH”) in order to protect the health of human beings and the environment.1 This is achieved with close surveillance of the production and use of chemical substances, ensuring that there are no carcinogenic, mutagenic, or teratogenic effects. Similarly, the potential harm on aquatic organisms, among other safety concerns is also carefully investigated. This regulation is intended for the control of chemical substances; but it encounters practical challenges. The constant evaluations to determine the sanitary safety of the substances increase the number of animal tests and as a result, there is a significant increase in the cost of the substances to be registered. These challenges call for alternative methods to reduce the costs of the experiments and the use of animals. In 1991, an effort to reduce animal testing led the European Commission to create the European Center for the Validation of Alternative Methods (ECVAM), currently under the name of the European Union Reference Laboratory for Alternatives to Animal Testing (EURL-ECVAM). Among the new alternatives to replace animal tests are the in silico methods, for the computational prediction of particular toxicity endpoints.2

Whether an endpoint is suitable for computational prediction depends on many aspects, including the availability of chemical and biological information, and preferably knowledge on the mode of action.3

The prediction of toxicological endpoints might be focused on different purposes; for example, the prediction of toxicity of new chemicals for the development of products for the agrochemical, pharmaceutical, food or cosmetic companies. Another scenario is the prediction of toxicological endpoints for regulatory purposes. It is important to note that for regulatory purposes, the predictions are only allowable for chemical impurities.

As the predicted variable is more complex, the accuracy of the models drops. For example, physicochemical properties can typically be predicted with 90% or higher accuracy, but endpoints involving biological data are less successful. Furthermore, models based on enzymatic evaluations are usually more accurate than those performed in living cells or entire animal models. Beyond models describing a single measurement, extrapolations across different experimental conditions, i.e. different species, routes of administration, etc., would be beneficial for the use of data already measured, to estimate more complex and costly experiments.

For side products (impurities) from agrochemical and other chemical industries, international regulatory agencies allow the prediction of toxicity as a substitution of experimental determinations, or at least, to decide whether a compound warrants experimental evaluation or not. Following the Organisation for Economic Co-operation and Development (OECD) principles, predictive models are generated using (Quantitative) Structure–Activity Relationship (Q)SAR studies.4 It is important to emphasize that QSAR is used in regulatory settings for the estimation of toxicity of impurities and the contribution of such impurities to toxicity should be minimal. Unfortunately, this is not the case for products manufactured under suboptimal conditions, which may contain, for example, dioxins.5

For a description of the OECD principles and their connection to the area of knowledge discovery on databases (KDD) and QSAR methods, the reader is refer to the literature.4 These methods consist of mathematical models that relate chemical structures with biological responses (i.e. toxicity). The correlations obtained in the (Q)SAR models are generally multilinear regression models6,7 but can also require more complex relations such as those obtained from artificial neural networks (ANN).

The use of (Q)SAR for the prediction of toxicity is a very active area in agrochemistry8,9 and risk assessment for emergency responses.10 All pesticide products contain active, inert ingredients and impurities. The active ingredients are the substances that perform the desired effect of the pesticide, and should be specified by name and quantity on the label. Inert ingredients might be used for example to increase the effectiveness of the active ingredients, making the pesticide easier to use or to increase the solubility of active ingredients, making up between 40 and 80% of the total weight/volume of the commercial formulations. Impurities are the byproducts of the chemical synthesis of the active ingredient, and usually account for 1–2% in products manufactured by agrochemical companies. When a pesticide becomes off-patent, different chemical synthetic reaction paths are used for the preparation of the active ingredient, and new impurities are produced. In such cases, the impurities should be chemically characterized and quantified and their toxicity and risk assessment should be investigated. In the absence of the experimental toxicity values for the impurities, toxicity predictions obtained from validated computational models11 are now accepted by regulatory agencies around the world, including but not limited to OECD country members. Due to the consequences observed from these studies, it is of utmost importance to develop and analyze the models with caution, knowledge and ethics. The general guidance of QSAR in regulatory settings as well as the scope and limitations of in silico models are described in the literature.4

The use of animal models relies on the hypothesis that the toxicity measured in the animal will inform of toxic compounds for higher order species. Uncertainty factors (UFs) have been used to account for different uncertainty sources such as animal-to-human extrapolation, intra-individual variability, extrapolation from shorter-term to longer-term exposures, etc.12 The idea behind UFs is to ensure that the toxicity of a compound is not underestimated. The decision of which factor to use (the number or factor needed) has been a topic of debate. UFs are based on conceptual information, for example, the 100 factor applied to the no-observed effect level (NOEL) in “animal-to-human extrapolation” is based on body-weight differences (a rat weighs about 100 times less than a 25 kg child) and on the assumption that the specific mode of action and metabolism is similar between humans and rats or mice. Also, the difference between short- and long-term exposures is accounted for by the acute/chronic ratio (A/C) of the same endpoint, which is specific for each compound. Needless to say that underestimation should be avoided at all cost, even at the cost of overestimation. In contrast to animal models, predictive models rely on the data from which they are built. In this context, one of the most comprehensive pesticide databases is the Pesticide Properties Database (PPDB). This database contains toxicological information, as well as uses status (by region or country), mode of action (MOA), type of crop, etc. PPDB has been developed and maintained for over 20 years by the Agriculture & Environment Research Unit (AERU) at the University of Hertfordshire. For each substance around 300 parameters are stored, including human health, environmental quality, and biodiversity risk assessments; a full description of the database is available elsewhere.13

Herein we collected a database of pesticides annotated with multiple toxicological endpoints, named PESTIMEP (PESTicide Multiple EndPoint), and explored the relations of experimental toxicity values measured at different species and at different routes of administration.2,14 In addition, we analyzed the pesticide profile by MOA of the PPDB database. Further information of PESTIMEP and PPDB databases is provided in the Materials and methods section of this work.

Materials and methods

Data preparation

Chemical structures were built using Marvin Sketch (v. 15.5.11, 2015) developed by ChemAxon (http://www.chemaxon.com).

PESTIMEP was collected from the CRC Handbook of Pesticides.15 PESTIMEP contains 158 pesticides and is annotated with 465 columns containing experimental data, including physicochemical properties and toxicity values. PESTIMEP is available upon request. In this work, we analyzed a subset of 146 pesticides containing at least 10 molecules evaluated in the same toxicity assay.

PPDB is available at https://sitem.herts.ac.uk/aeru/iupac/, and contains information of 2289 pesticides, from which we focused on 404 fungicides, 583 herbicides and 480 insecticides.

Molecules were prepared in MOE (v.16.08) by removing the metal ions, keeping only the largest fragments of the molecules, assigning protonation states, adding explicit hydrogen atoms, conserving chirality of the structures and rebuilding structures that were not correctly constructed. In all these steps the physiological pH of 7.4 was considered. Then, the structures were energy minimized using the MMFF94 force field.

Data analysis and visualization

Statistical descriptive analysis was performed with Statistica StatSoft TIBCO (v. 13.0). Data relationships based on pair-wise simple linear regressions (SLR), principal component analysis (PCA) and data visualization were performed using TIBCO SpotFire (v. 7.10.1) and Origin Pro (v. 9.1, 2016). PCA was based on the following normalized molecular descriptors: molecular weight, weinerPol, PEOE_VSA_FHYD, PEOE_VSA_HYD, log S, log P(o/w), vsa_acid, vsa_base, reactive, a_nCl, a_Np, a_Nh, a_No, a_acc, a_aro, a_count, a_don, a_nC, b_ar, b_rotN, lip_acc, lip_don, aromatic_atoms, aromatic_atoms2 and non-carbon_atoms previously generated in MOE (v.16.08). A brief description of the descriptors is provided in Table 1, and full description can be accessed in the MOE user guide.

Table 1. Physicochemical properties relevant in pesticides and their related molecular descriptors.

Physicochemical property	Related descriptor	Description (MOE)
Molecular weight	Molecular weight
Solubility	log S	Log solubility in water
log P	log P(o/w)	Log octanol/water partition coefficient
Polarity	WeinerPol	Weiner polarity number
Hydrophobicity	PEOE_VSA_HYD	Total hydrophobic vdw surface area
Acid	vsa_acid	VDW acidic surface area
Base	vsa_base	VDW basic surface area
Overall reactivity	Reactive	Reactivity
Organochloro compounds	a_nCl	Number of chlorine atoms
Organophosphorus	a_nP	Number of phosphorus atoms
Oxidant	a_nH	Number of hydrogen atoms
Reductor	a_nO	Number of oxygen atoms
HBD	a_don	Number of hydrogen bond donors
HBA	a_acc	Number of hydrogen bond acceptors
RB	b_rotN	Number of rotable bonds
Aromatic bonds	b_ar	Number of aromatic bonds
Aromatic atoms (%)	a_aro	Percentage of aromatic atoms
Non-carbon atoms	Non-carbon	Number of non-carbon atoms
Number of atoms	a_count	Number of atoms

Results and discussion

In the first section, we analyze the oral toxicity values across different species of the compounds contained in PESTIMEP, followed by a pairwise linear correlation analysis of the distributions of toxicity values for rat and mouse on different routes of administration; in the last part of this section, we present a visualization of the chemical space of these molecules. In the second section, we analyze the IUPAC database, classified by type of pesticide and by modes of action, also showing the corresponding chemical space analysis of this dataset, based on MOA and comparing to FDA approved drugs.

Toxicity data across different species

There are examples showing that toxicity across species does not follow linear correlations. Such is the case for birds, whose toxicity profile is not very relevant to humans; birds are very sensitive to most pesticides due to their poor detoxifying mechanisms,16 so they are good sentinel species to rely on but not to extrapolate from. This is consistent with the toxicities observed in PESTIMEP. Table 2 shows the species evaluated by 10 or more compounds as well as the minimum, maximum and median acute oral toxicity values of the compounds contained in PESTIMEP. The specie evaluated the most is rat, followed by the mouse model. Regardless of the number of molecules evaluated at each of the species, and without knowing if the same compounds were evaluated, birds are the most sensitive species of the set, by at least one order of magnitude, compared to larger animals, with acute oral toxicity values of 6.75, 8.1 and 57.5 for pigeon, wild bird and quail, respectively. The number of compounds evaluated on the entire PESTIMEP database, the list of acronyms and the distribution of toxicity values are shown in the ESI, Table S1 and Fig. S1 and S2,† respectively. The skewness of the data towards toxic compounds (rather than evenly distributed across the toxicity limits) hinders the development of robust predictive models. To further explore the PESTIMEP database, we performed pair-wise linear regressions across the data available, for a total of 288 linear regressions. A graphical representation is shown in Fig. S3.† However, since most of the compounds contained in this dataset correspond to highly toxic insecticides (which are among the most toxic pesticides for mammals), and roughly the most and least toxic compounds remain as such across the species, linear correlations are not informative. Notably, in some instances reported in the literature, a broad estimation within 5-fold to 10-fold is considered reasonably well.17 However, this cannot be generalized, for example neonicotinoids have a range of toxicities that span 4 or 5 orders of magnitude among aquatic organisms alone18 and the acute toxicities among this group of insecticides differ a lot between species: thiamethoxam and clothianidin are extremely toxic to bees but not so toxic to aquatic insects, whereas thiacloprid and acetamiprid are not very toxic to bees but extremely toxic to aquatic organisms.

Table 2. Statistics of acute oral toxicity values of different animal species evaluated by 10 or more compounds.

Species	Abbreviation	# Molecules Evaluated	Toxicity
			Min	Median	Max
Rat	rat	146	0.79	1010	10 001
Mouse	mus	101	0.30	635	15 001
Rabbit	rbt	46	10.00	546	7100
Guinea pig	gpg	40	2.30	502	300 000
Dog	dog	30	3.00	500	10 001
Quail	qal	29	1.00	57.5	16 001
Duck	dck	26	0.60	155	11 300
Chicken	ckn	25	8.00	373	7951
Wild bird	bwd	24	0.75	8.1	400
Mammalian	mam	14	10.00	435	12 600
Cat	cat	12	2.00	204	802
Pigeon	pgn	10	2.37	6.75	110

In an attempt to use toxicity data interspecies, in the 1980s, when the toxicity databases were scarce, uncertainty factors (UFs) were suggested and used for interspecies extrapolation of toxicity values.12 The general idea of UFs is to increase the toxicity, typically by 10-fold, when extrapolating the data. Other conversion factors, such as 5- or 3-fold, are also used, if there is information to support that decision. Imposing a high conversion factor ensures safety; however, a marked overestimation of toxicity might make this approach impractical. Since the late 1990s, UFs have been replaced with species sensitivity distributions (SSD). To date, SSDs have not been applied to mammal species because at most four or five species are used in toxicity testing (rat, mouse, rabbit and dog or cat) while the SSDs require at least six species to be considered valid.19 Moreover, SSD distributions show that the range of interspecific variation in toxicity data is compound specific, therefore, no direct generalizations can be made.

Toxicity data across different routes of administration

Toxicity data contained in PESTIMEP can also be explored for the same specie but evaluated under different routes of administration. Table 3 shows the number of compounds evaluated on rat and mouse on different routes of administration, and the table also summarizes the descriptive statistics of the LD₅₀ toxicity values. Not surprisingly, the intravenous and the topical skin administration were the most and least toxic respectively, for either animal model. In general, the median toxicity values show that rats were more sensitive than mice, except for the oral and subcutaneous administration. The toxicity values span a 100-fold difference in magnitude across different routes of administration. Looking at the distributions of the data (Fig. S1†), there is a clear predominance of toxic compounds in the dataset. This unbalanced distribution hinders the development of linear predictive models but roughly, compounds with high toxicity LD₅₀ values are highly toxic, regardless the route of administration.

Table 3. Statistics of toxicity values of different routes of administration evaluated on rats and mice using 10 or more compounds.

Species	Abbreviation	Administration	Abb	# mols	Toxicity
					Min	Median	Max
Rat	rat	Oral	orl	146	0.79	1010	10 001
		Topical skin	skn	77	2.40	2000	23 000
		Intraperitoneal	ipr	39	0.28	100	6810
		Subcutaneous	scu	23	0.28	395	15 001
		Intravenous	ivn	12	0.30	8.5	87
Mouse	mus	Oral	orl	101	0.30	635	15 001
		Topical skin	skn	13	8.00	2330	10 001
		Intraperitoneal	ipr	55	0.83	193	6811
		Subcutaneous	scu	22	0.25	285	23 800
		Intravenous	ivn	15	0.20	55	320

Physicochemical properties

Molecular descriptors are routinely used to characterize compound databases; the well-known Lipinski's rule of five (Ro5) has been applied and extended to agrochemical compounds.20 Moreover, quantitative estimation of pesticide-likeness has been proposed.21 Clarke and coworkers have previously described the physicochemical properties of pesticides and their relation to toxicity.22–24 After a description of the relevant properties of pesticides, we present the profile of the PESTIMEP dataset, in terms of descriptors related to toxicity.

Pesticides as dust particles can travel through air by the effect of wind, and can reach off-target areas. According to Hao,25 most pesticides have molecular weight (MW) less than 435, even smaller than the average drug MW. Depending on the hydrophobicity and volatility of the pesticide,26 it can pose a risk of accumulation and contamination of soil, water (rivers, lakes, streams) or air. In this regard, the octanol/water partition coefficient is a good indication of the bioaccumulation of pesticides.

The chemical reactivity and soil absorption also affect the overall toxicity and accumulation of pesticides.27 For example, pesticides will accumulate longer if they are stable to redox reactions, hydrolysis, photolysis and biodegradation. The hydrophobicity and non-polarity of pesticides make them prone to accumulate in soils, instead than in waters. Thus, accumulation in soil is due to low degradability and hydrophobicity. However, unfortunately most of the contaminations of aquatic ecosystems is due to hydrophilic pesticides, particularly herbicides and systemic insecticides.28 Thus, water contamination is primarily due to highly soluble pesticides.

Classification of pesticides based on chemical core structures

Pesticides in general, beyond those contained in the PESTIMEP database, can be classified by different means, for example, by physicochemical properties, chemical structures, or mechanism of action, which are dependent on particular core stuctures.29 Representative core structures of pesticides are shown in Fig. 1.

Fig. 1. Representative core structures of pesticides. (A) Organochlorides (aldrin, chlorothalonil, and heptachlor epoxide). (B) Organophosphorates (mevinphos, TEPP, and phorate). (C) Carbamates (aldicarb, bendiocarb, and propoxur). (D) Pyrethroids (permethrin). (E) Sulfonylureas (bensulide and perfluidone).

Organochlorines are cyclic compounds that possess a high number of chlorine atoms in their structures. The presence of aromatic rings or aliphatic chains gives them low water solubility (0.001 to 10 ppm) and very high lipid solubility (log P of 4–7), in addition, they are very resistant to degradation. The recalcitrant nature of chlorinated pesticides is due to the strong electronegativity of the chlorine atoms, which makes it difficult for enzymatic processes to degrade the molecules. This applies to chlorinated insecticides and herbicides (i.e. atrazine) or fungicides (i.e. chlorothalonil). Their high lipophilicity promotes their accumulation in fatty tissues and they easily reach the nervous system where they produce their final effect, which is the hyperexcitability of the cells by alteration of cell membrane polarizability. Organochlorides may interact with molecular targets producing cell hyperexcitability by different pathways. DDT-type acts primarily on the peripheral nervous system inactivating sodium channels, and other chlorinated insecticides act in the central nervous system by their binding to GABA receptors in the synaptic membranes.30 Thus, the nervous system of the insects is a preferred target in the insecticide design.

Effective pesticides have been developed: pyrethroids, and pyrethrins, synthetic and natural derivatives, respectively. Although the same features of chlorinated molecules mentioned above apply to pyrethroids and pyrethrins, the vast majority of pyrethroids, bifenthrin being an exception, do not accumulate in the environment because they are readily degraded by bacteria, algae and many other organisms. Both groups, pyrethroids and pyrethrins, are derived from (+)-trans-chrysanthemic acid and share the cyclopropane core as the toxicophore group with sodium channels as their biological targets. The inclusion of electron donor or acceptor groups contributes to the efficacy of their interactions with voltage-gated sodium channels in the axonal membranes. Furthermore, pyrethrin and pyrethroid side-chain modifications with aromatic moieties, vinyl groups or H-bond acceptors increase the stability of the interaction of the insecticide with the sodium channel by intermolecular interactions. This interaction maintains the sodium channel open and avoids the repolarization of the cells, thereby paralyzing the insect until it is dead.31 The main difference between pyrethrin and pyrethroid derivatives is the presence of halogen atoms in their structure. In general, the inclusion of halogen atoms in the pyrethroid derivatives provides them with higher potency but also gives them a higher stability that allows their accumulation in the environment. Due to their easy biodegradability, pyrethrin-based insecticides are the preferred option, and they are considered as low-toxicity pesticides to humans.

In turn, most organophosphates and carbamates are prone to hydrolysis under alkaline conditions. In particular, the active moiety of organophosphates, the orthophosphoric acid, is an easy target for enzyme degradation. Water soluble pesticides will not accumulate in the soil and will undergo degradation via hydrolysis. Organophosphates and carbamates are insecticides that also produce their toxic effects disrupting the function of the nervous system by inhibiting the enzyme acetylcholine esterase, in an irreversible and reversible manner, respectively. The interaction of the insecticides with the enzyme is through a hydroxyl group of a serine amino acid of its active site. The oxygen atom of the serine group acts as a nucleophile attacking the electro-deficient atom of the insecticides. A similar effect occurs with the phosphorus atom of organophosphate compounds and the carbon atom in carbamates. This interaction becomes weaker or stronger depending on the structural variations in the organophosphate or carbamate derivative. Additionally, to increase the insecticide enzymatic inhibition potency, the enhancement of intermolecular interactions with the rest of the active site must be provided by side chains of the insecticide.32

To broadly resemble physicochemical properties relevant to pesticides, Tice, Clarke, Hao, and Avram have proposed molecular descriptors, the ones we analyzed in this work are as follows: molecular weight, weinerPol, PEOE_VSA_FHYD, PEOE_VSA_HYD, log S, log P(o/w), vsa_acid, vsa_base, reactive, a_nCl, a_Np, a_Nh, a_No, a_acc, a_aro, a_count, a_don, a_nC, b_ar, b_rotN, lip_acc, lip_don, aromatic_atoms, aromatic_atoms2 and non-carbon_atoms. The intended property to be represented by each descriptor is summarized in Table 1, along with the definition of each descriptor, and the distributions are shown in Fig. S2,† along with the corresponding statistical descriptive analysis. Except for the descriptor reactive, which is categorical, the rest of the descriptors are continuous. To note, whereas some descriptors, such as structural atom counts might be related to acute toxicity (LD₅₀) others, such as solubility are meant to be related to environmental accumulation.

Comparison of pesticides and FDA approved drugs

To have a sense of the chemical profile of the compounds contained in this dataset, we analyzed the chemical space based on chemical scaffolds and physicochemical properties relevant in pesticides and compared it to the profile of drugs suggested previously. There are obvious differences between the toxicity of medical drugs and pesticides. In pesticides a key consideration is the targeted mode of action. In addition, the persistence and bioaccumulating properties need to be taken into account to assess the environmental impact.

Druglikeness of a compound is generally associated with characteristic physicochemical properties (Lipinski's Ro5) that account for desirable properties, such as bioavailability. According to Clarke24 pesticides may be associated with some other properties related to how agrochemicals behave primarily with their environment. Here we compared both sets of properties (druglike and pesticide-like properties) and visualize corresponding chemical space (Fig. 2). The chemical space of the PESTIMEP compared to FDA-approved drugs (downloaded from a ChEMBL server, ; https://www.ebi.ac.uk/chembl/) was obtained by PCA, a commonly used method for reducing the dimensionality of a data set.33 As shown in Fig. 2A–C, the representation of the chemical space of both databases is similar when only druglike properties are analyzed. This visualization shows that PESTIMEP covers a broad region of the space occupied by the FDA database, sharing their most densely populated area across the first two principal components. In both cases, as seen from the loading plots (Fig. 2D–F), the largest loading for the first principal component corresponds to log P, followed by HBA for FDA and molecular weight for PESTIMEP. In contrast, a different approach is obtained when pesticide-like properties are used to visualize the chemical space (Fig. 2G–L). At first glance, an opposite trend is observed for each database (Fig. 2G); while FDA-DB is mainly distributed along the first principal component, PESTIMEP does it for the second principal component. In order to extract more information about the descriptors that contribute the most to the variance observed, PCA plots (Fig. 2H and I) and their corresponding loading plots (Fig. 2K and L) for each database were graphed. The largest loadings for the FDA database in the first principal component are the number of non-carbon and aromatic atoms. In turn, for PESTIMEP, the main loadings accounting for most of the variances in the first principal component are hydrophobicity and polarizability, two highly correlated descriptors in this chemical space. The complete information of coefficients for each PCA is summarized in Tables S2–S5.†

Fig. 2. Chemical space representation of pesticides and approved drugs based on physicochemical properties relevant in drug development and pesticides. Chemical space using druglike properties: (A) FDA and pesticides, (B) FDA and (C) pesticides, and their corresponding loading plots (D), (E) and (F). Chemical space using pesticide-like properties: (G) FDA and pesticides, (H) FDA and (I) pesticides, and their corresponding loading plots (J), (K) and (L). Physicochemical properties: molecular weight (weight), octanol/water partition coefficient (log P), number of hydrogen bond donors (HBD), number of hydrogen bond acceptors (HBA), number of rotable bonds (RotB), polarizability (Pol), hydrophobicity (Hydroph), solubility (log S), acidity (Ac), basicity (Bas), reactivity (Rx), number of chlorine atoms (Cl), number of phosphorus atoms (P), oxygen atoms (nO), hydrogen atoms (nH), atoms in aromatic system (Arom) and atoms different from carbon (Non_C). The loading coefficients are summarized in ESI (Tables S3–S6†). The first three PC accounted for 60–93% of variance.

Analysis of the PPDB database by MOA

To explore a larger number of pesticide classes and modes of action, including herbicides, insecticides and fungicides, we analyzed the PPDB database. The distribution of the toxicity values by MOA, for herbicides, insecticides and fungicides, is shown in Fig. 3. Only those with more than ten compounds per MOA are included. Unlike the data contained in the PESTIMEP database, the toxicity values are more speared. Notably, the toxicity ranges differently by MOA. For example, while insecticides that act as inhibitors of chitin biosynthesis are the least toxic, those that inhibit AChE are on average the most toxic ones. Also, not surprisingly, on average insecticides are more toxic than herbicides and fungicides. These toxicity profiles provide guidance for the generation of toxicity predictive models. Those sets, by MOA, with a wider range of toxicity values and chemical diversity, will provide a more extended applicability domain. The generation of predicative models by MOA is in progress in our lab.

Fig. 3. Distributions of toxicity values by MOA for (A) insecticides, (B) herbicides, and (C) fungicides.

Fig. 4A shows a visualization of the chemical space based on physicochemical properties that have been identified as relevant for pesticides and drugs. This graph includes FDA approved drugs, pesticides contained in the T.E.S.T database and fungicides, herbicides and insecticides, from the PPDB database. Not surprisingly, the larger collections, FDA and T.E.S.T, occupy more spread areas. Based on this two-dimensional graph of physicochemical properties, these collections are not clearly distinguishable. Fig. 4B–D show the equivalent graph but for insecticides, fungicides and herbicides, respectively, color-coded by MOA. Interestingly, the classification of each pesticide type by MOA is distinguishable in this representation, supporting the idea of generating classification models across MOA.

Fig. 4. Chemical space representation of pesticides based on physicochemical properties relevant in pesticides. (A) FDA approved drugs, and pesticides classified by MOA and pesticides contained in T.E.S.T (B) insecticides, (C) fungicides, and (D) herbicides. Physicochemical properties: molecular weight (weight), octanol/water partition coefficient (log P), number of hydrogen bond donors (HBD), number of hydrogen bond acceptors (HBA), number of rotable bonds (RotB), polarizability (Pol), hydrophobicity (Hydroph), solubility (log S), acidity (Ac), basicity (Bas), reactivity (Rx), number of chlorine atoms (Cl), number of phosphorus atoms (P), oxygen atoms (nO), hydrogen atoms (nH), aromatic atoms (Arom) and atoms different from carbon (Non_C). The first three PC accounted for 60–93% of variance.

In this analysis, the calculated properties included atom counts, physicochemical properties resembling direct measurements based on internal QSAR models (log S, and log P) and descriptors related to the van der Waals surface area, such as hydrophobic, acidic, and basic. In addition, the reactivity descriptor is calculated, which is obtained from the consideration of reactive functional groups. The profiles obtained show that, rather than a differentiation across the type of pesticide, the properties behave differently by MOA, emphasizing the appropriateness of modeling and studying the pesticides considering the MOA classification.

Conclusions

The chemical and biological information contained in the PPDB and PESTIMEP databases is publicly accessible, and will continue to aid on the development of predictive models and decision making.

Building and curating databases are laborious and time-consuming required tasks. Collecting databases allow for the better use of experimental information already measured, the prediction of new entities, and to guide the design of new compounds or chemical libraries.

The information presented in this work shows the evolution models for the extrapolations of toxicity values across different species, from linear regression models to UFs and to SSDs. As more information is contained in the databases, and of good quality, better models and extrapolations will be possible. At present, however, extrapolations of toxicity values across different species are limited. Nonetheless, the PESTIMEP database allowed the identification of sensitive species and routes of administration, supporting what has been reported in the literature on individual cases. Particularly useful information obtained from the analysis of the PPDB database is the suitability and feasibility of analyzing the information based on MOA. Our current efforts are focused on this direction and the models will be published in due course.

Conflicts of interest

There are no conflicts of interest to declare.