A QSPR approach to the ecotoxicity of ionic liquids (Vibrio fischeri) using VolSurf principal properties

In silico structural descriptors for both cation and anion counterparts of ILs allow the prediction of Vibrio fischeri toxicity by means of a simple three parameter equation.

Abstract

Recently derived in silico structural descriptors for both IL cations and anions allowed the development of a QSPR model correlating ionic liquid structures to Vibrio fischeri toxicity using the partial least squares (PLS) approach. Interpretation of the PLS model confirmed the effect of IL cationic structural features such as the influence of cation side chain length, presence of heteroatoms, and non-aromaticity of the heterocyclic scaffold on toxicity. The PLS model also provided a quantitative evaluation of anion effects, previously not evidenced due to the structural similarity of the anions considered. A simple equation in which three descriptors (two for the cations and one for the anions) allow the prediction of Vibrio fischeri toxicity for over 8000 ILs is reported.

Introduction

Ionic liquids (ILs), low melting point salts composed of an organic cation and an inorganic or organic anion, are characterized by a low vapour pressure and represent very versatile process media that can serve as solvents,1–3 separating agents,4,5 catalysts,6,7 and electrolytes, and assume a variety of functionalities in different fields such as energy, materials and medicine.8 The low vapour pressure of ILs reduces the pollution as compared to common volatile organic solvents, but this property is not sufficient to define them as “green” solvents. In fact, release of ILs from industrial processes into aquatic environments may lead to water pollution due to their high solubility and high stability in water, which might render them persistent pollutants in wastewater. For this reason it is necessary to determine the environmental risk in aquatic ecosystems to comply with the terms of the European Union regulation for the Registration, Evaluation, Authorisation and Restriction of Chemicals (REACH) requiring a comprehensive knowledge of the properties and hazards of ILs.9

However, assessing ecotoxicity of ILs is not a simple issue10 because different biological assays can be selected as representative of IL hazards, such as cytotoxicity, toxicity towards invertebrates, vertebrates, fungi and bacteria, phytotoxicity, impact on enzymatic activity and protein stability.

Unfortunately, IL toxicity data are scattered in the literature and single toxicity tests are not available for many ILs. In order to simplify this complicated picture, a multivariate insight into IL toxicity11 allowed one to compact toxicity information by deriving 104 IL aquatic toxicity scores, as well as 87 bacterial and fungi toxicity scores. Such a limited number of ILs as compared to the enormous number of known and potentially possible ILs covers only a very narrow window of information regarding IL toxicity. Furthermore it is interesting to note that bacterial and fungi toxicity scores are not available for ILs for which aquatic toxicity data are reported and vice versa. This can be ascribed to different IL hydrophilic/hydrophobic properties preventing the determination of both kinds of toxicities. In fact highly polar ILs, being water soluble, when released in the environment determine aquatic toxicity, while more hydrophobic ILs can penetrate the bacterial membrane determining their bacterial toxicity.

Vibrio fischeri is a marine luminescent bacterium which emits light as a result of normal metabolic processes. A reduction in luminescent ability during exposure to contaminants or pollutants is taken as a measure of eco-toxicity. The assay with the bioluminescent bacterium Vibrio fischeri is an international standard ecotoxicological bioassay (DIN EN ISO11348)12 widely applied for toxicity determination13,14 and has been adopted in environmental toxicity studies,12,15 and for toxicity testing of chemicals.14,16 Such an extensive use can be ascribed to its low cost and simplicity and to the speed of this bioluminescence test as compared to other experimental procedures. Furthermore Vibrio fischeri, being a Gram negative bacterium used for aquatic toxicity evaluation, exhibits an optimum hydrophilic/lipophilic balance allowing measurements for a significantly high number of ILs.

The physicochemical properties as well as the toxicity of a single IL depend strongly on the molecular structure of its constituents, i.e. of the organic cation and of the inorganic or organic anion. However, concerning the risk assessment for humans and the environment this structural variability represents an enormous problem as it is impossible to measure experimentally physicochemical properties and to perform biological tests on the effects on human health and the environment for every single compound for such a huge number of ILs estimated to be in principle over a million. An interesting paper17 reports Vibrio fischeri standardized toxicity values for as many as 148 ILs including 64 cations and 30 anions and applies PLS-DA for discriminating ILs on the basis of their toxicity. In this approach the input variables for each IL were equal to 1 if the anion/cation was in the molecule, and 0 if not, and ILs’ expected toxicity was assessed with respect to toluene, an organic solvent traditionally used in industrial applications. However, this binary classification cannot be considered a proper Quantitative Structure Property Relationships (QSPR).

The attempt to study and predict Vibrio fischeri and other biological target toxicities using QSPR approaches for ILs is not new. In the scientific literature several papers18 show the growing interest in this field, especially in QSAR studies for IL toxicity prediction recognized by the REACH regulation to limit time and costly consuming tests and reduce the number of in vitro and in vivo experiments.19 These papers usually adopt a somewhat “univariate” approach by considering the cation effect for a fixed anion and the anion effect for a fixed cation, assuming that toxicity does not vary with varying the anionic/cationic IL counterpart and that anion–cation interaction effects are absent. Other papers describe the IL structure by means of electrostatic and topological20,21 or heuristic descriptors.18 However, in our opinion correlation models should not only be reliable (i.e. have a good predictability), but also provide either the possibility to easily interpret the results and the correlation itself or to employ general, useful, reusable and not-heuristic “building blocks” as descriptors. In the present work we are adopting a QSPR approach able to include all the above aspects.

Experimental descriptor measurements are very expensive and time-consuming and extremely sensitive to IL purity, which is very difficult to achieve.22 To overcome this difficulty, we have derived in silico cation and anion physicochemical descriptors by means of the GRID approach in VolSurf+23,24 using information coded into 3D GRID Molecular Interaction Fields25–28 to derive in silico physicochemical and pharmacophoric molecular descriptors. This procedure has been applied to study structure–permeation relationships,23 to predict antitumour activities,29–34 and for modeling phospholipidosis induction.35,36

The validity of VolSurf+ descriptors for ILs was tested in a preliminary QSPR model with the aquatic toxicity scores as the responses, achieving a good correlation.37 It is evident that handling a high number of descriptors could be very difficult, especially for big IL datasets. Consequently the descriptors were compacted into the so-called IL cation and anion Principal Properties (or PPs: 5 PP+ and 4 PP– respectively)38 highly informative and easier to use. The above PPs allowed the development of Quantitative Structure Property Relationship (QSPR) models allowing assessment of IPC-81 rat cell line cytotoxicity and acetylcholinesterase inhibition38 for as many as 520 ILs.

Following these leads, the aim of the present work is to exploit the potentialities of IL PPs to develop a PLS model relating IL PPs to Vibrio fischeri toxicity and to predict unknown toxicity values for as many ILs as possible.

Results and discussion

Vibrio fischeri toxicity PLS model using IL cation and anion in silico PPs as descriptors

The chronic bioluminescence inhibition of Vibrio fischeri values reported here are log(EC₅₀) (μmol L^–1) for 65 ionic liquids including 35 different heterocyclic cations and 18 organic and inorganic anions (see Chart 1 below). These homogeneous toxicity values, spanning almost 5 log units, were taken from a literature database39 accessed in August 2014 and unfortunately no longer open access since September 2014. We have recently derived 9 PPs38 as physicochemical in silico descriptors for 218 cations (5 PPs+) and 38 anions (4 PPs–), and herein used to relate the ionic liquid structure to Vibrio fischeri toxicity by means of a PLS model.

Chart 1. Cation and anion structures used in this work as learning and test (L and T respectively in Table S3†) sets.

In order to verify the statistical significance of our results by means of a validation set, the dataset was randomly divided into a learning set (L) and a test set (T) containing 55 and 10 (15% of the total dataset) ILs respectively. 9 PPs were used to obtain a PLS model by using SIMCA40 from a data matrix containing 55 objects (ILs) and 9 descriptor variables (PPs). In such a model, two significant PLS components describe 76.9% of the total y variance (R²), with a predictive ability (Q²) of 0.681 (Table S1†). In the VIP (Variable Importance for Projection) plot, which shows the importance of each X-variable in explaining X-variation and correlation to Y (Fig. S1†), PP₁+ is the most important descriptor, followed by PP₂+, PP₅+ (with a high error which affects its significance)40 and PP₃–, while all other descriptors appear to be less important. Therefore, in order to limit the number of descriptors and simplify the model, a new PLS correlation model was built maintaining only three relevant X descriptors: PP₁+ and PP₂+ for cations and PP₃– for anions. The new simplified 55 × 3 matrix provided a two PLS component model explaining 78.9% of the total variance, and with a cumulative Q² of 0.770 (see Table S2† for model details) showing that the exclusion of low relevance six descriptors does not affect the “goodness” of the model. It is worth mentioning that the QSPR approach presented here considers simultaneous variations in both the cation (heterocyclic core, side chain length, presence of oxygen atoms in the side chain) and the anion structural features by means of 3 descriptors (PPs) whose physico-chemical interpretation has previously been commented on.38 In particular, PP₁+ embodies information related to cation solubility, size, flexibility and molecular weight, PP₂+ describes the interaction with water and the hydrophobic volume of the cation, while PP₃– is related to anionic size/shape and to the anion ability to form H-bonds as a donor or an acceptor.38

In the present case EC₅₀ represents the concentration of a compound at which 50% of its maximal effect is observed, therefore the higher this value, the less toxic the IL. The correlation between X descriptors and the Y response can be displayed by means of the loading correlation plot for both PLS model components (Fig. 1). PP₁+ shows a high positive influence on Vibrio fischeri toxicity, being very close to log(EC₅₀), whereas PP₃– has a negative contribution, being on the opposite quadrant. The projection of PP₂+ on the dotted line shows a slight positive contribution towards log(EC₅₀): this variable can be considered less important for the correlation with Y, but still useful in explaining X and modelling Y variance as Fig. 2 (the VIP plot) points out.

Fig. 1. Loadings correlation plot showing the descriptors (w*) and response (log EC₅₀) loadings superimposition; the correlation between each descriptor () and the dependent variable (◆) is elucidated by the PPs projection on the dotted line.

Fig. 2. VIP plot for the Vibrio fischeri ecotoxicity PLS model. Bars indicate the mean values and lines the standard deviation.

In agreement with previous studies,38 less toxic ILs have small cations bearing a non-aromatic cationic scaffold and short alkyl side chains highly soluble in water in the pH range 3–10 and, especially at physiological pH, show a great dispersion in water (PP₁+) and are able to interact with water and as H-bonding acceptors (PP₂+) (having oxygen atoms in the scaffold and/or in the side chain). These considerations are confirmed by the corresponding scores plot (Fig. S2†), which, when superimposed to the loadings plot in Fig. 1, evidences the IL cationic structures’ influence on toxicity. Similar features such as small dimensions and wide hydrophilic regions with H-bonding ability (PP₃–) are important for the anions. The VIP plot (Fig. 2) stresses the importance of PP₁+, which shows that for a good interaction between ionic liquids and the Vibrio fischeri bacterial membrane, the structure of the cations is more relevant as compared to that of the anionic counterpart. Evidently a very soluble and highly dispersed in physiological media IL is not able to interact with bacterial membranes and penetrate them and this results in a lower toxicity. This finding is consistent with the literature15,18 and our previous studies.11 Nevertheless, the anions cannot be considered as “silent audience” in the interaction. In fact, in Table S3,† ILs containing the same cation (1-butyl-3-methylimidazolium ILs 160–164, 166–168, 170) exhibit different log(EC₅₀) values, ranging from 1.83 for octylsulfate, to 2.47 for bis(trifluormethylsulfonyl)imide (also known as 1,1,1-trifluoro-N-[(trifluoromethyl)sulfonyl]methanesulfonamide) to 3.07 for bromide, and up to 3.60 for iodide and trifluoromethylsulfate. Such a variation is appreciated by our QSPR approach allowing one to evaluate anion effects, which cannot be evidenced when similar anions such as halides are considered.41,42

Fig. 3 (for IL labels, see Fig. S3†) reports the correlation plot between experimental and predicted log(EC₅₀) values for both learning (L) and test (T) set ILs. Only in a few cases the difference between experimental and predicted values is over one log unit, and could be considered less significant from a strictly statistical point of view. In particular, IL 231 (Table S3† and Fig. 3) with a cationic side chain including as many as 18 carbon atoms exhibits a high discrepancy between experimental and predicted toxicity values, being predicted as more toxic than it is: the longest cationic side chain ILs may form intermolecular aggregates43 which prevent them from interacting as surfactants with the biological target (i.e. resulting less toxic).

Fig. 3. Experimental vs. predicted log(EC₅₀) values plot for the Vibrio fischeri ecotoxicity PLS model. △, ILs in the learning set (L in Table S3†); [black circle], ILs in the test set (T in Table S3†).

The reliability of a PLS model may be evaluated by using different statistical tools. Internal cross validation gives a first important assessment about the goodness of fit and prediction by means of R² and Q² parameters and is supported by external validation by means of a test set. Furthermore, a permutation plot can give information about the statistical significance of a correlation model since it compares R² and Q² of the parental model with those of permutated models (arbitrarily set as 50) in which the responses are randomly exchanged. The resulting permutation plot for our correlation model reported in Fig. S4,† fulfills this validation criterion and excludes chance correlation.40 On the basis of the above considerations, predictions can be extended to a higher number of commonly used ILs (the same as the 520 arbitrarily chosen in ref. 38) also reported in Table S3† together with DModX (distance of a single object from the X-space)40 values. We are aware that the prediction affordability for such a high unprecedented number of ILs cannot be the same for each single IL. Some guidance on the “goodness” of the prediction can be evaluated by DModX values reported in Table S3† which give an estimate of the similarity of a new observation to the training set observations. Predictions for ILs exhibiting DModX values higher than 2.33 in Table S3† should be considered with great caution. Among the 520 ILs not belonging to either the learning or the test set, some compounds should be carefully considered in their prediction, due to their high DModX values (for instance, see ILs 85, 86, 117 bulky side chain imidazoliums; 255–264, 268–271, 328–330 long amido-side chain pyridiniums; 424–436 long bis ether imidazoliums). DModX values below 2.33 represent only one parameter indicating prediction reliability, which in specific cases of interest might be supported by experimental validation. Nevertheless, PLS predictions provide approximate toxicity estimates useful to orient the selection among several efficient ILs for specific applications. When considering the utility of the predictions, one should also take into account the generality of a data matrix containing five different cationic scaffolds and side chains and different anionic structures, as well as experimental errors in biological determinations and the effects of IL impurities influencing analytical determinations22 which can strongly affect the significance of any model. The present results confirm previous observations on structural IL features affecting toxicity using a multivariate approach11 in terms of the influence of cation side chain length, presence of heteroatoms, and non-aromaticity of the cation scaffold. Therefore, in our opinion this QSPR correlation spanning over 5 log units can be considered as satisfactory, since it tries to rationalize and quantitatively explain in a simple way the complex interactions between ionic liquids and living organisms.

The above considerations point out the prediction as well as the interpretation potentialities of the PLS approach, allowing one to evaluate the relevance of each single descriptor in determining the biological response.

Several studies adopting QSPR approaches to relate different descriptors to IL Vibrio fischeri toxicity have been published. Couling et al.20 carried out a QSPR study considering 25 ILs (including 10 pyridinium and 6 imidazolium cationic scaffolds) and using a multiple linear regression (MLR) approach. They reported an interpretation based on electrostatic and topological descriptors. However, descriptor values are not published and the results cannot be reproduced. Furthermore the model is not validated by an external validation set as recommended by the OECD principles for QSAR studies.44 Another paper45 reported a MLR correlation model in which IL structures for 9 cations and 17 anions were described by three main groups of descriptors: anions (A), cations (C) and substitutions (S) describing ecotoxicity as the summation of the contributions of each group. Descriptors are based on the Boolean approach: their numerical value is 1 if a specific group is present, and 0 if not. Each descriptor is multiplied for a parameter (a, c and s) indicating the descriptors’ contribution to the model. This approach is very simple although the specific information depending on the chemical features of cations and anions is limited due to the 0/1 variation in the descriptors. A four-parameter correlation was found42 for ionic liquids with halide (Cl^– or Br^–) anions. Descriptors were based on IL cation structures and the corresponding model was used to predict IL toxicity in water or in the gas phase. A heuristic procedure was used to select the most influent descriptors to obtain a reliable correlation model. Estimations of the descriptors’ collinearity and reproducibility of the results by using different QSPR approaches are not possible due to the lack of the descriptor numerical values. The paper conclusion asserted that the model could be applied to a broader range of ILs with different anions as, according to the literature,15 ILs with the same cation and different anions do not show any statistical difference in toxicity. The effect of anions is pointed out both by the results of our statistical approach and by experimental data reported in Table S3† showing two orders of magnitude difference in Vibrio fischeri toxicity for 1-butyl-3-methylimidazoliums. Das and Roy46 published an interesting paper using linear discriminant analysis with respect to toluene, and MLR and PLS approaches to derive QSPR models whose predictions are comparable with those reported in Fig. 3. Rigorous validation was performed either for the classification or the regression models by applying multiple strategies and in both the cases encouraging results were obtained for external and internal validation. The influence of IL features on toxicity is discussed, but unfortunately no descriptor values are reported and this represents a limitation to the QSPR model reproducibility. A very recent study47 considering toxicity towards Vibrio fischeri reported a five parameter MLR approach for a large dataset by using topological descriptors regarding the position and the features of non-hydrogen atoms used to mathematically derive cationic and anionic descriptors whose values are not reported.

In our opinion the main advantage of the approach adopted in the present work over other papers adopting QSPR approaches is that the numerical values of the descriptors are available allowing the readers to reproduce the work or to apply alternative statistical models. However, commercial software such as VolSurf+ and PLS are not available to everybody. For the above reason we provide below eqn (1) in which the three descriptors for both cations and anions adopted to predict Vibrio fischeri toxicity values reported in Fig. 3 can be easily used to extend predictions to any combination of IL cations and anions for which PPs are available:38

1

Many potential users can benefit from the application of this simple equation allowing to predict Vibrio fischeri toxicity for thousands (in principle as many as 8284) of ILs, well above the 520 reported in Table S3.†

Materials and methods

The Vibrio fischeri toxicological properties, expressed as log EC₅₀ obtained after 30 min of exposure to the toxicants, were taken from a literature database39 in August 2014, and are reported in Table S3.†

Computational methods

In the present work, chemometric tools available in the SIMCA software package (SIMCA 13.0.3),40 namely Principal Partial Least Squares Projections to Latent Structures (PLS),48 were used. In particular, PLS allows to find relationships between the Volsurf-derived properties matrix (X matrix) and the response, in the present case the Vibrio fischeri toxicity, expressed as log EC₅₀, reported in Table S3.†

According to the procedure adopted by the SIMCA software, data were pre-processed by autoscaling all variables to unitary variance, i.e. by multiplying the variables by appropriate weights (the reciprocal of the variable standard deviation) to give them unit variance (i.e. the same importance) after subtracting the mean value. The PLS algorithm computes PLS components for each of the two matrices (X and Y matrices), searching simultaneously for a linear relationship between the X-scores and Y-scores of the PLS components trying to maximize the covariance between X and Y matrices.49 The main statistical parameters provided by the PLS method are R²X, R²Y (respectively sum of squares of all the X's and Y's explained by all extracted components) and Q², the fraction of the total variation of the Y's predicted by all PLS components, as estimated by cross validation (CV). Q² was computed as: 1 – PRESS/SS, where SS is the residual sum of squares and PRESS is the squared difference between observed and predicted values for the data kept out of the model fitting. The statistical results obtained by the PLS method are able to detect what variables in the X block are relevant to determine the dependent variables (Y block) by means of the VIP values. The VIP values reflect, in fact, the importance of terms in the model both with respect to Y, i.e. its correlation to all the responses, and with respect to X. SIMCA computes VIP values by summing over all model dimensions the contributions VIN (variable influence). For a given PLS dimension, a, (VIN)_ak² is equal to the squared PLS weight (w_ak)² of that term, multiplied by the % of residual sum of squares explained by that PLS dimension. The accumulated (over all PLS dimensions) value, VIP_k = Σ(VIN)_k², is then divided by the total percent of residual sum of squares explained by the PLS model and multiplied by the number of terms in the model.

Random selection of the learning and test sets is a good way of assessing the power of the model but since it is random the procedure should be repeated several times. The permutation plot40 is an internal validation tool comparing the goodness of fit (R² and Q²) of the original model with the goodness of fit of several models (50 in the present case) where the order of the Y-observations has been randomly permuted with the same X-matrix. The plot reported as Fig. S4 in the ESI† clearly indicates the validity of the PLS correlation model.

Conclusions

The availability of recently derived in silico structural descriptors for both IL cations and anions (5 and 4 PPs respectively)38 allowed the development of a QSPR model correlating ionic liquid structures to Vibrio fischeri toxicity by means of a simple three parameter equation whose application can be extended to a wider set of commonly used ILs. In silico predictions for an unprecedented high number of ILs provide a shortcut to the fulfilment of the REACH requirements by establishing priorities in selecting ILs for experimental hazard assessment and therefore allowing one to focus more expensive in vivo studies on more potentially toxic compounds thus reducing the number of tests on animals.